Paper deep dive
Single-Sample Decoding of a Two-Layer Sound-Word Canvas Model for Thai Klon Pad
Worasait Suwannik
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We present a generator of Thai klon pad verse, one stanza from a short topic, and evaluate it against human verse on held-out topics. The generator is a masked language model initialised from WangchanBERTa over a fixed canvas with two layers: a sound layer of pronounced syllables, where rhyme, tone and syllable count are enforced as hard masks, and a word layer of whole words, so that no word is cut into written pieces. Decoding commits one cell at a time and samples at temperature 0.4; one sample is drawn per topic, with no selection, reranking or repair. On 78 held-out topics never used in development, judged by three independent runs of a frozen LLM judge in the same pools as the human stanzas from which the topics were written, 70.5% of the generated stanzas pass form, lexical cleanness of all four lines and topic alignment together, against 76.9% for the human stanzas (difference -6.4 points, 95% CI [-19.2, +6.4]). Cleanness is level with human (85.9% vs. 87.2%), and the stanzas pass form and cleanness together slightly more often than the human ones (+3.8 points, [-6.4, +14.1]); the difference lies in alignment (84.6% vs. 91.0%), on topics that were written from the human stanzas. The low temperature matters: at higher temperatures the same model passes clearly fewer of the same topics, because lexical errors lie in the model's lower-probability choices; sharpening the sampling distribution removes most of them, at a cost in vocabulary (7.5% rare words against 16.5% in the human stanzas) and diversity. The pre-registered target, all-three within 5 points of human, is missed by 1.4 points. KlonSuphap-LM, the only public klon pad generator we know of, passes none of the 78 held-out topics when judged in separate pools together with the method's and the human stanzas (form 19.2%, cleanness 3.8%).
Tags
Links
- Source: https://arxiv.org/abs/2610.14980v1
- Canonical: https://arxiv.org/abs/2610.14980v1
Trouble viewing inline? Open PDF directly →
Full Text
121,332 characters extracted from source content.
Expand or collapse full text
Single-Sample Decoding of a Two-Layer Sound–Word Canvas Model for ThaiKlon Pad Worasait Suwannik ∗ Abstract We present a generator of Thaiklon pad(กลอนแปด) verse, one stanza from a short topic, and evaluate it against human verse on held-out topics. The generator is a masked language model initialised from WangchanBERTa over a fixed canvas with two layers: a sound layer of pronounced syllables, where rhyme, tone and syllable count are enforced as hard masks, and a word layer of whole words, so that no word is cut into written pieces. Decoding commits one cell at a time and samples at temperature 0.4; one sample is drawn per topic, with no selection, reranking or repair. On 78 held-out topics never used in development, judged by three independent runs of a frozen LLM judge in the same pools as the human stanzas from which the topics were written, 70.5% of the generated stanzas pass form, lexical cleanness of all four lines and topic alignment together, against 76.9% for the human stanzas (difference−6.4 points, 95% CI[−19.2,+6.4]). Cleanness is level with human (85.9% vs. 87.2%), and the stanzas pass form and cleanness together slightly more often than the human ones (+3.8points,[−6.4,+14.1]); the difference lies in alignment (84.6% vs. 91.0%), on topics that were written from the human stanzas. The low temperature matters: at higher temperatures the same model passes clearly fewer of the same topics, because lexical errors lie in the model’s lower-probability choices; sharpening the sampling distribution removes most of them, at a cost in vocabulary (7.5% rare words against 16.5% in the human stanzas) and diversity. The pre-registered target, all-three within 5 points of human, is missed by 1.4 points. KlonSuphap-LM, the only publicklon padgenerator we know of, passes none of the 78 held-out topics when judged in separate pools together with the method’s and the human stanzas (form 19.2%, cleanness 3.8%). 1 Introduction Klon pad(กลอนแปด, “eight-syllable verse”) is the most widely practised form of Thai verse. A stanza (บท,bot) has four lines (วรค,wak) of nominally eight syllables, linked by a fixed scheme of external rhymes, with tone constraints on line-final syllables (Section3.1). Two properties of Thai make it a hard target for a neural generator. Thai script does not mark word boundaries, so a model that works on written subword pieces can produce strings that are not words. And the rhyme is defined on pronounced syllables, which do not always coincide with written ones:ไฉนis readฉะ|ไหน. A generator therefore has to satisfy position-specific phonological constraints while writing ordinary, well-formed Thai. This paper describes a generator built for these two properties and evaluates it against human verse on topics it never saw: on 78 held-out topics, 70.5% of its stanzas pass form, cleanness and topic alignment together, against 76.9% of the human stanzas from which those topics were written. Its contributions are: 1.A two-layer canvas model (Section4, Figure2): one cell per pronounced syllable, on which rhyme, tone and length are hard masks, carrying whole-word tokens over an open vocabulary initialised from WangchanBERTa, decoded by confidence-ordered commitment at a low sampling temperature, one sample per topic. ∗ The research reported here was directed by Worasait Suwannik. The programs were produced by Claude, Anthropic’s AI model. Worasait Suwannik contributed to the design of the model, including a whole-stanza canvas with a sound layer of pronounced syllables for the form constraints and a word layer of whole words for spelling and naturalness; to the experimental design, including LLM-written topics for reply stanzas; the wording of the cleanness judge prompt; and the unusual-topic test of Appendix Dand the selection of its topics. 1 arXiv:2610.14980v1 [cs.AI] 10 Oct 2026 2.A held-out evaluation (Section3): 78 topics, written from reply stanzas of test-split threads whose title groups were never used in development, the human stanzas from which the topics were written judged in the same pools, a form checker and two frozen LLM-judge instruments run three times independently. 3.The sampling temperature as the method’s main ablation: on the same held-out topics, all-three rises from 44.9% at temperature 0.8 to 70.5% at 0.4, cleanness from 56.4% to 85.9%, and the number of judged-bad lines falls from 36 to 13 of 312; we explain this by where lexical errors sit in the model’s distribution, and measure what it costs in vocabulary and diversity (Sections5–6); a confound test attributes the gain to the sampling itself, not to decoder bonuses that temperature also amplifies. 4.A comparison with KlonSuphap-LM [29], run locally with its temperature tuned on development topics, on the same held-out topics, judged in separate pools together with the method’s and the human stanzas (Section5.4). The trained model is available on the Hugging Face Hub athttps://huggingface.co/voraseth/ klon8-base, together with the code that generates from it, and loads with the Transformers library [27]: from transformers import AutoModel model = AutoModel.from_pretrained("voraseth/klon8-base", trust_remote_code=True) print(model.generate_poem("topic in Thai")[0]["poem"]) 2 Related Work Thai poetry generation.Published work on generating Thai verse is sparse. The only publicklon pad generator we know of is KlonSuphap-LM [29], a Thai GPT-2 base model fine-tuned on about 65,770 couplets (บาท, two lines each) ofklon padfrom the Thai Literature Corpora [24]. It marks the words that must rhyme with special tags in the training text (a tag pair for the source and the target of each rhyme within a stanza, and another for the rhyme between stanzas), optionally masks attention to non-rhyme words, and then fine-tunes with proximal policy optimisation (PPO) against a rule-based form grader. It is documented by a model card and code repository rather than a paper, and reports no evaluation of meaning or fluency; form is learned softly from the tags and the reward, whereas our decoder enforces rhyme, tone and length as hard masks on pronounced syllables. The ThaiProsody benchmark [ 30] evaluates large language models composing four Thai verse forms,klon padamong them, with a deterministic rule checker for tone, weight and rhyme, a model-based syllable counter and an LLM content judge; it is a benchmark of general-purpose models prompted zero-shot rather than a trained generator. Public Thai poem collections include the Thai Literature Corpora [ 24], drawn from classical works, and the klongklon dataset [22], more than 50,000klon padandkhlong si suphappoems collected from various literature, among them classical works such as Phra Aphai Mani; our training data are instead modern verse written by forum members, and our evaluation compares against the human stanzas from which its topics were written. We are not aware of a published klon padgenerator evaluated against human verse on held-out topics; we evaluate KlonSuphap-LM under our protocol in Section 5.4. Thai language models.The generator’s body is WangchanBERTa [17], a RoBERTa-style masked lan- guage model pretrained on Thai text with a SentencePiece vocabulary. Segmentation and grapheme-to- phoneme conversion use PyThaiNLP [20] and TLTK [1]. We use a masked encoder rather than a left-to-right generative model because the canvas is filled in any order under position-specific constraints, and because WangchanBERTa’s pieces give an initialisation for every whole word of the open vocabulary. Constrained poetry and lyric generation.Neural generation of metrical verse has concentrated on Chinese and English. Hafez [8,9] combined a language model with a finite-state acceptor that admits only metrical, rhyming paths, the closest antecedent of our hard masks. Tian and Peng [26] plan content and rhyme words before writing sonnet lines. For rigid-format text, SongNet [ 16], motivated by lyrics and evaluated on SongCi and sonnets, controls the format with format and rhyme symbols at each position, a fixed grid like our canvas; for lyrics, DeepRapper [28] writes rap lines right to left so that the rhyme is chosen first. CIGEN [ 25], the work nearest in spirit to ours, composes Chinese Ci non-autoregressively so that generation accommodates the tune pattern’s rigid format of length and rhyme. PoetryDiffusion [13] 2 writes sonnets and SongCi with a continuous diffusion model, which noises continuous word representations towards a Gaussian, for the semantics, and a separately trained metrical controller that penalises states straying from the target control distribution; its form control is soft, where ours is a hard mask. A shared design question is when the rhyme word is chosen relative to the rest of the line. Our decoder chooses the rime family early from the model’s own probability over rime families (Section4.2), writes it into the sound layer, and lets the words of the whole stanza be filled around it in confidence order. A second difference is the unit: because Thai rhyme lives on pronounced syllables and Thai script does not delimit words, we keep two aligned layers, sounds for the constraints and whole words for the text, instead of constraining written subword tokens. Sound has also been given to masked language models as extra input embeddings: PoemBERT [14] adds sentiment and pinyin embeddings to a BERT-based Chinese poem generator. Our sound channels are also input embeddings, but the sound layer is in addition the surface on which the form constraints act. Constrained and non-autoregressive decoding.Lexically constrained beam search [11] and Neuro- Logic decoding [18] enforce constraints in left-to-right generation. Our generator is a bidirectional masked language model [7] used non-autoregressively. Mask-Predict [10] and MaskGIT [4] fill a fixed-length canvas over several steps. MaskGIT keeps the most confident predictions at each step and re-predicts the rest; Mask-Predict predicts all tokens at every step and re-masks the least confident ones, so earlier predictions can change. Our decoder instead commits one cell per step, the open cell whose best admissible token is most confident, and does not re-mask committed cells. Our training, which masks a randomly drawn share of each stanza (Section4), is close to absorbing-state discrete diffusion: in D3PM an absorbing ([MASK]) state reduces to a reweighted masked-LM objective [2], and the MDLM objective is a weighted average of masked-LM losses [23]. These models do not order unmasking by the model’s confidence at decoding time; that ordering is the one of Mask-Predict and MaskGIT. The difference is that our constraints are hard masks on a parallel sound layer, so every intermediate canvas stays inside the form. Sampling temperature.Holtzman et al. [12] showed that the low-probability tail of a language model is unreliable and that pure sampling from it produces incoherent text, motivating nucleus sampling. Peeperkorn et al. [19] found that in story generation temperature is only weakly related to novelty and moderately related to incoherence. Our temperature comparison is a measurement of the same trade-off in a constrained, non- autoregressive generator, against human verse on held-out topics: the tail holds most lexical errors, and sharpening the distribution removes them at a measurable cost in vocabulary. Best-of-Nselection with a learned verifier [ 6] is the usual alternative for avoiding bad samples after the fact; the method here draws one sample with no selection. LLM judges. LLMs have been used as judges of generated text since early studies [ 5,31], with known position and verbosity biases [31]; for creative writing, LLM judgments agree poorly with expert readers [3]. Our cleanness instrument adapts the error-span protocol of GEMBA-MQM [15] and asks a narrow question (is each span well-formed Thai) rather than a judgment of quality; we add fixed human anchor stanzas in every pool, blind interleaving with human verse, three independent judge runs, and report the judge’s disagreement with a human reader as a limitation (Section 7). 3Klon Pad, Data and Evaluation Setting 3.1 The skeleton of aklon padstanza A stanza (บท) has four lines (วรค), traditionally namedวรคสดับ,วรครับ,วรครองandวรคส่ง. Each line has 7–9 pronounced syllables, usually 8; in the training stanzas 73–81% of lines have 8 and 18–24% have 9, depending on the line. The form checker used throughout this work (distributed with the model) reads a stanza as follows; Figure1shows the skeleton. •Syllables.Each line is segmented into words with PyThaiNLP [21] using its defaultnewmmtokenizer, a dictionary-based maximum-matching segmenter [ 20], and each word is converted to phonemes with TLTK [1], giving pronounced syllables with onset, vowel, coda and tone. A line passes the count when 3 วรค1 (สดับ) วรค2 (รับ) วรค3 (รอง) วรค4 (ส่ง) 12 3 4 56 7 8 12 4 6 7 358 12 3 4 56 7 8 12 4 6 7 835 A B tone of the last syllable free เอก,โทorจัตวา(notสามัญ,ตรี) สามัญorตรี สามัญorตรี Figure 1:Skeleton of oneklon padstanza as checked in this work. Circles are pronounced syllables, drawn at the usual eight per line. Group A links the last syllable of line 1 to the 3rd syllable of line 2 (or the 5th, dashed); group B links the last syllable of line 2 to the last of line 3 and to the 3rd (or 5th) syllable of line 4. Line-final tones are on the right. it has 7–9 syllables, where short unstressed syllables (e.g.สะ,กระ) may count as one, as half or not at all; a line of 10 pronounced syllables that contains such light syllables can therefore pass, which is why the training data and the canvas below allow up to 10 syllables per line. •External rhyme.Two syllables rhyme when they have the samerime, the vowel and the final consonant (rimes are grouped intorime families, with theไ/ใ/-ัย~ -ายandเ-า~ -าวfamilies merged, because the phoneme converter is inconsistent about their length). Three links are required: the last syllable of line 1 with the 3rd or 5th syllable of line 2 (group A); the last syllable of line 2 with the last of line 3; and the last syllable of line 2 with the 3rd or 5th syllable of line 4 (group B). Since lines 2 and 3 end on the same rime, the last link is equivalently a rhyme between the end of line 3 and the 3rd or 5th syllable of line 4. The link from the end of line 4 to the end of line 2 of the next stanza applies only to multi-stanza poems and is not used here. •Repeated rhyme word.A rhyme link may not be made by repeating the same syllable of the same word. •Tones.The last syllable of line 2 must not carry theสามัญ(mid) orตรี(high) tone, so it carriesเอก,โท orจัตวา; the last syllables of lines 3 and 4 must carryสามัญorตรี. The end of line 1 is free. A stanza passes form when all four counts, the three links and the three tone rules hold, no rhyme word is repeated, and it contains no Chinese, Japanese or Korean script (a guard for generators with multilingual vocabularies; it never fires on the method, whose vocabulary is Thai only). Internal rhymes (สัมผัสใน) are an ornament and are not required. Poets bend these rules: 92.3% of the human stanzas in the held-out evaluation pass the checker, and 95.0% of the training stanzas. 3.2 Training data The corpus is modern verse written by the users of a Thai poetry forum and collected from its public boards, klonthaiclub.com (บ้านกลอนไทย), whose 29 boards used here cover love, longing, encouragement, festivals, dharma and a practice room for the prosody rules. Posts are cut into four-line stanzas; a stanza is kept when its line count and syllable counts are well formed and all three external rhyme links hold (tones are not filtered), and duplicates are removed. The training split has 50,788 stanzas from 3,956 threads: 11,601 from opening posts and 39,187 from replies. The data pipeline of the two-layer model further keeps a stanza only when every line has 7–10 pronounced syllables, every syllable is read by the phoneme converter, and the words, re-joined and re-segmented, give back exactly the same sounds; 49,194 stanzas remain, of which a random 1,000 are held out to monitor the validation loss and 48,194 are used for training. The conditioning topic of every training stanza, opening post or reply, is its thread title after removing decoration (symbol runs, repeated punctuation, reply prefixes, bracketed notes, decorative Latin letters). 4 3.3 Train/test split and the held-out topics The corpus is split by thread title, not by stanza. Titles are first grouped on their Thai-only normalisation (thetitle key), so that decorated duplicates of one title (ฝันร้ายand the same title framed by star symbols) cannot fall on both sides. The split has 3,821 training titles in 3,700 title groups (50,788 stanzas) and 659 test titles in 652 title groups (8,680 stanzas). A stanza-level split would leak, because a title’s content syllables occur in a sibling stanza of the same thread 22.4% of the time, as often as in the stanza itself (22.5%). For the evaluation, every test title group that had been used anywhere in development was excluded. A group was out if its title key matched any title, prompt or topic used anywhere in development (every prompt file, every judged pool, every judged batch), or if any of its threads had supplied a human stanza to a judged pool or had a line found in a judged batch. The exclusion list holds 4,856 title keys from all these sources and 396 test threads (388 threads of human stanzas used in earlier pools, together with the threads found by their lines). Of the 652 test title groups (690 threads), 552 were excluded, leaving 100 (110 threads, 1,279 stanzas). None of the 100 titles was usable as a prompt as it stood; the reasons included birthday greetings, contest announcements, threads of ornamentalklon botforms, titles with Latin letters and one-syllable titles. A fixed rule therefore took, for each of the 100 title groups, its first reply stanza whose topic, written from that stanza by Claude Opus 5.5 under a fixed guide, is 2–12 syllables of Thai only (no Latin letters, digits or stray combining marks), contains none of a fixed list of forum keywords (meta and event words such asวันเกิด,ประกาศ,สมาชิก,กระทู้) and is neither a title or prompt used in development, a training title, nor a topic that Claude wrote for training stanzas during development (the model evaluated here was trained on thread titles only). That stanza is thehuman controland its topic is thepromptfor the generator as well; below, theheld-out topicsare these prompts, one per title group (Section 7discusses how this selection may bias alignment). 78 title groups qualified and 22 were dropped. Development used 49 further test-split titles, the49 development topicsbelow, which are therefore among the excluded groups. When the evaluation ran, the exclusion of the 78 held-out topics’ title groups and their human controls was checked again against a fresh scan of every prompt, pool and judged line (the 388 pool threads and 53,728 judged lines): zero overlap. Because each prompt was written from its human control, it may fit that stanza better than any other poem, which favours the human arm on alignment. Two further human stanzas per topic were judged in the same pools as fillers; they balance the pools and are never an endpoint. Where the title had enough replies (138 of the 156 filler slots) they are the two earliest other reply stanzas under the same title (6 of them from a second thread with the same title), each under its own Claude-written topic; the other 18 slots hold human controls of the second earlier evaluation (Section 3.4), 11 reply stanzas under Claude-written topics and 7 opening-post stanzas under their thread titles. Fillers are exempt from the exclusion rule because they are never endpoints. The human controls are verse written by forum users: 17 of the 78 come from the forum’s birthday-verse board. 3.4 Passing criteria and the judge A stanza passesall-threewhen (i) it passes the form checker of Section3.1, (i) all four of its lines are judged clean, and (i) its topic alignment is judged at least 4 of 5. The marginal rates are reported beside the conjunction. Target.Our target was fixed in a pre-registration before the evaluation reported here: on held-out topics, the all-three rate of the generated stanzas should be within 5 percentage points of the rate for the human stanzas of the same topics, judged in the same pools. Two earlier held-out evaluations with the same passing criteria, judge instruments and target, of predecessor configurations that drew 64 samples per prompt at temperature 0.8 and kept the one an LLM ranker preferred (the first with an earlier model and one further step that re-drew the weakest line, the second with the model evaluated here), each on its own fresh title groups and human controls (thread titles with opening-post stanzas in the first; in the second, 41 such titles and 40 Claude-written topics of reply stanzas) and with three independent judge runs, missed it widely (44.7% against 63.1%, and 35.8% against 60.5%), in both cases because of cleanness; the human rate differs between evaluations because each has its own prompts, human stanzas and pools. The system described here comes within 6.4 points (70.5% against 76.9%, Section 5): the target is missed by 1.4 points, and the 5 95% interval of the difference,[−19.2,+6.4], is too wide to separate the system from human verse in either direction. Cleanness.The cleanness instrument (AppendixA) adapts the error-span protocol of GEMBA-MQM [15]. The judge marks exact spans that are not linguistically well formed Thai: non-words, unrecognisable misspellings, unnatural word formations, and unnatural combinations of words that disrupt reading. A span iscriticalif it makes the line incomprehensible andmajorif a native reader stumbles but understands, or if it is a non-word other than a genuine Thai prefix or suffix that can attach to the neighbouring word; something technically odd that does not disrupt reading is a pass. Because the text is verse, abstraction, metaphor, personification, archaic or poetic diction, unusual imagery and natural poetic compounds are explicitly not errors, and rhyme, syllable count and structure are not judged. A line is clean when it carries no critical or major span. Alignment.The alignment instrument (AppendixA) asks the judge to write one Thai sentence saying what the poem is about, then score 1–5 how much of it stays on the topic, judging subject matter only and accepting figurative treatment (5: every line on topic; 4: at most one line wanders; 3: partly; 2: touched once and left; 1: unrelated). A stanza is aligned when its score is at least 4. Runs, aggregation and agreement.The judge is Claude Opus 5.5, run as an agent: a conversation whose only tool use is reading the batch files and writing its verdict file. Both prompts are frozen and wrapped in a fixed agent instruction (AppendixA). Each instrument is run three times independently, each run judging all six pools of the instrument in one conversation (Section7): a line is clean when at least two of the three runs report no critical or major span on it, and alignment is the median of the three scores; an item missing from a run counts in that run as unclean on every line and as alignment 1. Fleiss’ kappa across the three runs, over all items except the anchors, is 0.76 for per-line cleanness and 0.84 for alignment≥4. Pools, anchors and blinding.The 78 topics are split into six parts of 13. Each part has one cleanness pool and one alignment pool holding, for each topic, the three system arms (temperatures 0.4, 0.6 and 0.8), the human control and the two fillers, plus 16anchorstanzas: human stanzas from a fixed anchor set, the same in every pool, used to detect judge drift (about 17% of items). Items are interleaved and keyed at random and dealt into seven batch files per pool of 11 or 12 items plus two or three anchors, texts are given raw, and the agent instruction tells the judge that the pool mixes system and human poems without labels. A pool-run is void if its anchors are under 70% clean or their median alignment is under 3; none was. Anchors were clean in 100%, 87.5% and 81.3% of cases in runs 1, 2 and 3. Statistics.The primary comparison, fixed before any judging, is all-three at temperature 0.4 against the human controls; the other temperatures, the marginals, kappa and vocabulary and diversity read-outs are secondary. Differences use topic-clustered percentile bootstrap intervals (10,000 resamples of topics, seed 0, arms resampled together), and discordant topics are tested with the exact McNemar test. The McNemar tests and discordant-topic counts, the ratios of rates, the per-arm intervals, the comparison of temperatures 0.6 and 0.4, the bad-line counts of Table 2, the originality check of Section5.3and the rime-family count of Section 6were added after judging and are not part of the pre-registration. 4 Method 4.1 Two-layer canvas model A stanza is a fixed canvas of cells preceded by the topic: a leading<s>cell, 12 topic cells and four lines of 10 cells (Figure2). The model has two aligned layers. 6 <s> topic: 12 cells วรค1: 10วรค2: 10วรค3: 10 วรค4: 10 Canvas: one cell per pronounced syllable; topic cells hold the topic’s words, poem cells start as<mask>. Cell input=word layer+sound layer+position. Word layer: token(w, k), thek-th syllable of wordw withn w syllables, embedded ase w +e (k,n w ) ; 62,665 words, each row initialised from the mean Wangchan- BERTa embedding of the word’s pieces (37,851 words never seen in the poems: rows frozen, one shared bias offset). Sound layer: onset+rime+tone embeddings of the cell’s syllable (filled cells: from the token; a masked rhyme cell: the chosen rime). Position: line, cell and role embeddings+Wangchan- BERTa positions. WangchanBERTa-base encoder: pretrained, 12 layers, width 768; a leading<s>cell, padding never attended. Pretrained masked-LM head transform:z i (dense, GELU, LayerNorm). Word-slot logitsat celli:z i ·e w +z i ·e (k,n w ) +b w , tied to the input rows. Training: masked prediction. 1. Scores:+1.5on the topic’s words (cleared for a word in a line once it is used there),−2on words already placed; hard masks: a line’s first / last cell takes only a word’s first / last syllable (word layer); line-final tones (Figure1) and the stutter ban (sound layer). 2. At step 5 only: each rhyme group’s rime fam- ily is drawn from the mass that its line-final and 3rd-syllable cells give to each family (temperature 0.5; the groups get different families); the receiver is the 3rd or 5th syllable, whichever has more mass; the rime is written into the group’s open cells, which take only that rime from then on (no word twice in a group). Until then rhyme cells are held. 3. Top 512: only the 512 best tokens per cell are kept. 4. Commit one cell: the open cell with the most confident admissible token; sample(w, k)∝ exp(log-prob at the cell/T)withT= 0.4; the rest of w’s span is forced, so every syllable is part of a word. write the token and its sound into the canvas; re- peat until every poem cell is filled Figure 2:The generator. Left: the network, a WangchanBERTa encoder over a canvas of topic cells and 4×10poem cells, each cell carrying a whole-word token (word layer) and its pronounced syllable (sound layer). Right: one decoding step, repeated until every poem cell is filled (one step per cell, 32.9 on average on the held-out topics); rhyme, tone and length masks act on the sound layer; word-boundary masks and the sampling act on the word layer. Sound layer.One cell per pronounced syllable, computed as the form checker computes it (newmmwords, then TLTK grapheme-to-phoneme per word), carrying three channels: onset, rime family and tone. The syllable count of a line is its number of cells, and rhyme and tone masks act on these sounds, soไฉนfills two cells,ฉะandไหน, and its rhyme is that ofไหน. In the model used here the sound of a filled cell is derived from its word-layer token; a masked cell has no sound input, except that a rhyme cell whose rime family has been chosen receives that rime in its rime channel. Word layer.The token at a cell is a pair(w,k): wordwand the positionkof the cell’s syllable within it, so a word is never cut into written pieces. Its input embedding ise w +e (k,n w ) , wheren w is the number of syllables ofw, plus the embeddings of the three sound channels, of the line (topic or poem line 1–4), of the cell’s position within its line and of its role (topic cell or poem cell of a given line), plus WangchanBERTa’s pretrained position embeddings; the sum passes through WangchanBERTa’s embedding layer normalisation. The vocabulary is open: 62,665 words, 24,814 seen in the 49,194 stanzas of the data pipeline (training and validation) and 37,851 never seen in them (from PyThaiNLP’s word list and whole-word WangchanBERTa pieces; each of these is kept only if it is Thai, readable by the phoneme converter, at most 8 syllables and re-segmented as itself). Every word row starts from the mean WangchanBERTa embedding of its pieces (86% of word tokens are exactly one piece); rows of never-seen words stay frozen, with one learned bias offset for the group. On test stanzas the open vocabulary lowers the share of stanzas with an out-of-vocabulary word from 14.0% (with the training-stanza words alone) to 2.3%. Body and output.The body is the pretrained WangchanBERTa-base encoder [17] (12 layers, width 768); padding is never attended, and the<s>cell is always attended. Its hidden state at a cell passes through WangchanBERTa’s pretrained masked-LM head transform toz i , and the logit of token(w,k)is 7 z i ·e w +z i ·e (k,n w ) +b w , tied to the input rows, whereb w is a per-word output bias initialised from the centred, add-one-smoothed log frequency ofwin those stanzas (frozen for never-seen words, which share the learned offset above). Theword-slot tokensare all pairs(w,k)withk≤n w , 164,818 in all (164,822 output tokens with four special tokens); the cross-entropy over all output tokens is computed exactly, without forming every logit, by factorising its normaliser over words. Training.Training is masked prediction with one masking scheme per stanza: whole words (40%, a share of the words drawn uniformly from[0.15,1]), single cells (20%, a share of the cells drawn the same way), linesjto 4 for a start linejdrawn uniformly from 1–4 (25%) and every cell (15%). At masked rhyme cells the rime is revealed in the sound channel with probability 0.5, matching decoding, where the rime family is chosen before the word. The model was trained for 24 epochs with AdamW (weight decay 0.01 on all parameters except the per-word output biases, the shared offset, the<s>cell vector and the pretrained position embeddings), batch 32 and a one-cycle learning-rate schedule peaking at3×10 −5 after 6% of the 36,144 steps, with gradient clipping at 1.0 and seed 0; the weights after the last epoch are used. Training took about 8.7 hours on an Apple-silicon GPU. In development, 24 rather than 12 epochs raised all-three by 8.2 points (95% CI[+2.9,+13.7]; one judge run, single samples with no selection on the 49 development topics with ten seeds each). 4.2 Decoding Decoding starts from a canvas whose poem cells are masked, with a length for each line sampled from a table of line lengths measured on human stanzas (7–10 syllables, mostly 8). Each step runs the model once and then (Figure2, right): 1.scores every open cell: to the model’s logits it adds a bonus of+1.5on the tokens of the topic’s words (the lexicon words of more than one character among the 12 topic cells), cleared for a topic word in a line once that word is completed there, and a penalty of−2on words already placed; it applies the hard masks: the first cell of a line admits only the first syllable of a word and the last cell only the last syllable, line-final cells admit only the allowed tones (Figure 1), and a stutter ban forbids a token whose sound equals its neighbour’s in the same line unless the two cells belong to one lexicon word or the word pair is attested as a same-sound join at least twice in the 49,194 stanzas of the data pipeline (602 of 1,613 pair types are allowed, e.g.รัก|รัก,นาน|นาน); each cell is normalised to log-probabilities before the stutter ban (and, once chosen, the rime-family and group-word masks of step 2) remove tokens, without renormalising; 2.at the fifth step only, once four cells have been committed, chooses each rhyme group’s rime family: familyfis drawn with probability proportional to exp(s f /0.5), wheres f is the mean, over the group’s line-final cells and its 3rd-syllable receiver, of the log of the probability mass that the cell gives tof as scored in step 1; the family temperature 0.5 is fixed andTdoes not affect it, and the two groups always receive different families. The receiving cell is the 5th syllable if the chosen rime has more mass there than at the 3rd, and otherwise the 3rd; when it is the 5th, the 3rd may not carry that rime. The rime is written into the rime channel of the group’s open cells, so that the words are then chosen to fit it, and from then on rhyme cells admit only their group’s rime and no word may appear twice in a rhyme group. Until this step the rhyme cells are held, so the first four steps write only other cells; 3.keeps only the 512 best tokens of each cell; 4.commits one cell: the open cell whose best admissible token has the highest log-probability. A token (w,k)is admissible only if the whole span ofwfits in the line around that cell and every other cell of the span admits the matching syllable ofw. The token is sampled with probability proportional to exp(s/T), wheresis its log-probability at that cell andT= 0.4; the other cells ofw’s span are then forced to the matching syllables ofwand committed in later steps, so every syllable belongs to a real word. A stanza takes one forward pass per syllable cell, 32.9 on average on the held-out topics (31–36). If the checker’s re-segmentation of the joined words disagrees with the cells, or no open cell has an admissible token, the stanza is decoded again, continuing the same random stream; this validity retry is allowed at most four times, after which the last attempt is kept, and atT= 0.4it never occurred on the 78 held-out topics. Apart from it, one sample is drawn per topic (seed 0); there is no best-of-N, no reranker, no rejection 8 Table 1:Held-out evaluation: 78 topics, one sample per topic and temperature, three independent judge runs per instrument, all arms and the human control in the same pools. Differences are against human with 95% topic-clustered bootstrap intervals; p = percentage points. The fillers row gives the first and the second filler of each topic. ArmAll threevs. human (p)Clean Aligned≥4Form Clean & form T= 0.4(the method)70.5%−6.4[−19.2,+6.4]85.9%84.6%98.7%84.6% T= 0.655.1%−21.8[−35.9,−7.7]65.4%79.5%98.7%64.1% T= 0.844.9%−32.1[−46.2,−16.7]56.4%74.4%100.0%56.4% Human, same topics 76.9% – 87.2% 91.0% 92.3% 80.8% Fillers 1 / 2 (human, not endpoints) 83.3 / 79.5%– 88.5 / 91.0% 98.7 / 93.6% 94.9 / 94.9% 83.3 / 85.9% Table 2:Judged-bad lines (flagged by at least two of three runs) on the 78 topics, and how many of them carry a flagged span that reaches the line end in a majority of the flagging runs. ArmStanzas with 0/1/2 bad lines Bad lines on lines 1/2/3/4 Span at line end T= 0.467 / 9 / 25 / 6 / 2 / 09 of 13 T= 0.651 / 26 / 17 / 10 / 5 / 619 of 28 T= 0.844 / 32 / 27 / 15 / 8 / 619 of 36 Human68 / 9 / 11 / 4 / 3 / 36 of 11 on quality and no line repair. On one Apple-silicon laptop GPU the 78 held-out stanzas took 126 seconds including model loading, about 1.6 seconds per stanza. 5 Results Against human verse.The method passes all-three on 55 of 78 topics (70.5%, 95% CI[60.3,80.8]) and the human stanzas on 60 (76.9%,[67.9,85.9]). The paired difference is−6.4points ([−19.2,+6.4]); the method passes where the human stanza fails on 11 topics and the reverse holds on 16 (exact McNemar p= 0.44). As a share of the human rate the method reaches 0.917 ([0.762,1.091]). Table1gives all rates; AppendixBshows 30 of the 55 passing stanzas. Where the difference lies.The method is 6.4 points below human on alignment ([−17.9,+3.8]; 7 vs. 12 discordant topics) and 3.8 points above on cleanness and form together ([−6.4,+14.1]; 11 vs. 8). On cleanness alone it is 1.3 points below ([−11.5,+9.0]; 7 vs. 8). It passes form more often than the human stanzas (+6.4, [0.0,+12.8]), because poets bend the checker’s rules. The alignment difference is in the direction predicted by the bias from writing each prompt from its human control (Section3.3); the fillers, also judged against topics written from themselves (or, for 7 opening-post stanzas, against the thread title they were written to), reach 98.7% and 93.6% alignment. We read this as evidence that the method is level with human verse in this pool on cleanness, and below it on alignment by an amount that the topic-writing procedure may partly or wholly explain. Ratios.Relative to the all-three rate atT= 0.4, the same model and decoder reach 0.782 of it at temper- ature 0.6 (95% CI[0.652,0.923]) and 0.636 at 0.8 ([0.489,0.804]). Temperature, on the same topics.The same model and decoder at temperature 0.8 pass 44.9%, and at 0.6 55.1%. The step from 0.8 to 0.4 is+25.6points of all-three ([+12.8,+38.5]; 25 vs. 5 topics,p= 0.0003) and+29.5of cleanness ([+17.9,+41.0]; 26 vs. 3), with alignment+10.3([0.0,+20.5]; 12 vs. 4,p= 0.077). From 0.6 to 0.4 cleanness still rises on 17 topics against 1. The dose response is monotone in cleanness (56.4, 65.4, 85.9%) and in the number of judged-bad lines (36, 28, 13 of 312). 9 5.1 Where the remaining errors are Lowering the temperature mainly removes stanzas with one bad line (32 at 0.8, 9 at 0.4); at 0.4 the distri- bution of bad lines per stanza matches the human one (67/9/2 against 68/9/1). A line-final flagged span is common in both system and human lines (9 of 13 and 6 of 11), so a span at the line end, where the rhyme syllable sits, is not by itself a signature of the rhyme mask: it may be where this judge locates any phrase that does not hold together. 5.2 Vocabulary and diversity The cost of the method is plainer and less varied text. Measured on dictionary-segmented tokens, leaving out the topic’s own words, against the 78 human controls, rare words (fewer than 50 training occurrences) are 10.0% of tokens atT= 0.8, 8.1% at 0.6 and 7.5% at 0.4, against 16.5% for the human stanzas; very common words (2,000 or more) are 35.2%, 38.2% and 40.8% against 27.9%; distinct word types at a matched budget of 1,516 tokens are 838, 736 and 681 against 939. From 0.8 to 0.4, the ratio of distinct to total syllable bigrams within lines falls from 0.872 to 0.784 (human 0.935), the entropy of line-final syllables from 7.06 to 6.71 bits (human 7.73), repeats of multi-syllable words across lines rise from 0.36 to 0.58 per stanza (human 0.08), and stutters, adjacent identical written syllables, from 0.026 to 0.077 per stanza (human 0.154; the stutter ban permits attested pairs such asรัก|รัก). The topic is repeated more: the share of stanzas that use one topic word four or more times rises from 11.5% to 24.4%, and topic words per stanza from 3.9 to 5.1. 5.3 Originality A generator trained on a forum corpus could reproduce lines it has seen. We compared every line of the 78 held-out stanzas (312 lines) with every line of three corpora: the training and test splits of the forum data (237,872 lines) and a further verse collection that this model was not trained on (369,742 lines: other verse posts of the same forum, the classical epic Phra Aphai Mani and theklonpages of Thai Wikisource). No generated line occurs verbatim in any of them, and 152 of the 312 lines share no run of four or more written syllables (PyThaiNLP’s syllable segmentation) with any corpus line. Runs of 8, 6, 5 and 4 syllables are tested; the longest shared run is six or seven syllables, in 10 lines, and the shared runs are common set phrases, most of them from greeting and blessing verse (e.g.ขอให้มีความสุขทุก,ขอํานาจสิ่งศักดิ ์ สิทธิ ์ ); AppendixC lists the 10 lines, each with its shared phrase and one corpus line that contains it. As a check that the test detects copying, the human control stanzas, which belong to the test split, are found verbatim in 153 of their 312 lines; the other 159 differ from their corpus lines only in spacing, because spaces were removed from the controls when they were pooled, and all 312 match once spaces are removed from the corpus lines as well. The generated lines contain no spaces, and none of them matches a corpus line even with spaces removed. The body was pretrained on Thai web text that we cannot search, so this check shows that the stanzas are not copied from the verse corpora above, not that no line exists anywhere. The same check on 82 stanzas written for 17 deliberately unusual topics, which cannot have been seen in training, is in Appendix D. 5.4 Baseline: KlonSuphap-LM KlonSuphap-LM [ 29] is a GPT-2 model of 124M parameters released under the MIT licence; we ran the public checkpoint locally on CPU. Its only input is a text prefix, which it continues: the model card’s example passes a short phrase as its input, and the model was trained on poem text alone from classical literature, without topic–poem pairs. We therefore give the topic as the stanza’s opening words, exactly as in the model card. To give it a fair setting, its sampling temperature was chosen before the held-out run on the 49 development topics, with the same judge (one run), among 0.4, 0.6, 0.8 and 1.0 with ten samples per topic (nucleusp= 0.85as in the model card, and the Hugging Facetransformersdefault top-kof 50), by a rule fixed in advance: the highest all-three rate, ties broken by the cleanness rate and then by the lower temperature. It passed all-three on no development stanza at any temperature; cleanness tied at 0.4 and 0.6, and the rule picked 0.4. It then wrote one stanza for each of the 78 held-out topics, judged three times in new pools with the method’s stanzas from Table 1, the human controls and one human filler per topic. KlonSuphap-LM passes all-three on none of the 78 topics (method minus baseline+69.2points, 95% CI[+59.0,+79.5]; Table 3). Its stanzas fail on every axis: 19.2% pass the form checker (21.8% on syllable 10 Table 3:KlonSuphap-LM against the method and human verse on the 78 held-out topics, all in the same new pools, three independent judge runs. The method’s and the human stanzas are the ones of Table1, judged again in this company. ArmAll three Clean Aligned≥4Form The method (T= 0.4)69.2% 85.9%80.8% 98.7% KlonSuphap-LM (T= 0.4)0.0% 3.8%0.0% 19.2% Human, same topics75.6% 87.2%92.3% 92.3% counts and the three rhyme links alone), mostly through missing the rhyme from line 1 into line 2 (42 of 78) and the rhyme from line 2 into line 4 (30 of 78); 3.8% are judged clean; none is judged on topic, which its lack of topic conditioning makes partly a property of the comparison rather than of the model. Cleanness (+82.1points,[+73.1,+91.0]) and form (+79.5,[+70.5,+88.5]) are the fairer contrasts. In this new pool company the method’s and the human stanzas score within 1.3 points of their Table1rates on all-three and cleanness; on alignment the method moves from 84.6% to 80.8% and the human stanzas from 91.0% to 92.3%, so re-judging the same stanzas among different neighbours moves a rate by up to 3.8 points. The judge runs agree (Fleissκ0.84 for cleanness per line, 0.95 for alignment). A kinder start.To separate topic following from writing, we also gave KlonSuphap-LM the human control’s first line for each held-out topic, with that line’s last syllable in the model’s own rhyme tag, and let it write lines 2–4 (temperature 0.4; cleanness only, three judge runs, Fleissκ0.88 for cleanness per line). Its rhyme improves markedly: the line-2-to-line-3 rhyme holds in 88.5% of stanzas and the line-2-to-line-4 link of group B in 62.8% (human 100%), and 24.4% pass the full form checker (1.3% in a second generation run whose prompt omits the rhyme tag). Its language does not: all three of its lines are judged clean in 2.6% of stanzas, and 32.5% of its lines are clean, against 92.3% and 97.4% for the human lines 2–4 of the same stanzas (difference−89.7points,[−96.2,−83.3]). Our method, which was not given the first line, has all of lines 2–4 of its own Table 1stanzas judged clean in 88.5% of the topics when they are judged again in these first-line-variant pools. 6 Why Low Temperature Works A stanza has about 32 syllable cells, and it is clean only if all four lines are. A small per-cell probability of an ill-formed choice therefore compounds into a large per-stanza failure rate. If the model’s high-probability choices are mostly well formed and its tail mostly is not, temperature acts on exactly the right part of the distribution: it reweights a token byp 1/T , so a token with a tenth of the leading token’s probability keeps 0.1 1.25 ≈0.056of the leader’s weight atT= 0.8but0.1 2.5 ≈0.003atT= 0.4, about 18 times less. The held-out data fit this account: cleanness rises monotonically as T falls, the gain comes from stanzas with a single bad line, and the vocabulary moves towards common words at the same time (Section 5). The hard masks make this cheap: because rhyme, tone and length are enforced on the sound layer at every temperature, sharpening the distribution cannot break the form (98.7–100% at all three temperatures), and the only thing it trades away is lexical variety. Sampling or decoder bonuses?The decoder adds a topic bonus and subtracts a repetition penalty from the logits before dividing byT, so atT= 0.4these terms act twice as strongly as at 0.8. A pre-registered development test separated the two effects on the 49 development topics with ten samples each (one judge run, so these are development numbers): atT= 0.4with both terms scaled down to their 0.8 strength (the rescaling applies only to the draw; the choice of cell, the rime family and the top-512 cut use the unscaled scores, as in every arm), all-three was 77.3% against 78.6% with them at full strength and 63.3% atT= 0.8, so the sharper sampling alone keeps 0.92 of the gain (95% CI[0.76,1.13]). The mirror arm,T= 0.8with the terms doubled, recovers only 0.21 of it ([−0.07,0.49]), though about half of the alignment gain. The all-three gain is therefore a property of sampling the model’s own distribution more sharply. Three limits of this account are known. (i) The gain is bought with vocabulary and diversity, measured 11 above. (i) Part of the alignment gain may be topic echo, which the alignment instrument does not penalise: topic words per stanza rise from 3.9 to 5.1 between 0.8 and 0.4. (i) The rime family is chosen at a fixed family temperature of 0.5 that the sampling temperature does not reach, and this choice over-produces one rime family: on the 78 held-out stanzas atT= 0.4, the mergedaifamily (ไ/ใ/-ัย~ -าย, Section3.1) carries 41.0% of the rhyme-group rimes (the ends of lines 1 and 2), against 10.9% in the human controls. Temperatures below 0.4 were not evaluated. 7 Limitations Sample size.The evaluation has 78 topics, so intervals on rates are about±10points and the interval of the difference to human runs from−19.2to+6.4points. The temperature 0.4 was chosen from development results on the same model and judge; it was fixed as the primary arm before the held-out topics were decoded. Plainer vocabulary.The method’s Thai is measurably plainer than the human stanzas’: fewer rare words, more very common ones, fewer distinct types, more repetition of content words and of the topic (Section5). The cleanness and alignment instruments do not penalise this, and alignment may reward topic echo. The evaluation does not measure literary quality, imagery or originality. The human control and its topics.All 78 prompts were written by Claude Opus 5.5, the model that also serves as judge, from the human control stanza, which favours the human arm on alignment; the human controls are reply stanzas from threads whose own titles were unusable as prompts, not a random sample of the forum, and they are verse by forum users, 12.8% of whose stanzas carry a judged-bad line. The model was trained on thread titles, which are not summaries of the stanza, so the prompts differ in kind from its training topics. LLM judging.Every cleanness and alignment rate is a rate under one LLM judge family, validated against human anchor stanzas and three independent runs. An LLM judge and a human reader do not flag the same lines: in a blind check of 30 stanzas from the second of the two earlier held-out evaluations (Section3.4), the author marked at least one unacceptable line in 11 of the 15 system stanzas the judge called unclean, in 7 of the 8 system stanzas the judge called clean and in 4 of the 7 human stanzas (judge: 1 of 7); over the 120 lines, both flagged 6, only the author 16 and only the judge 12. No such reader check exists for the evaluation reported here. Judging in one context per run.The protocol specified a separate judge session (a fresh agent conver- sation) for each pool and run. The stored verdicts show that each of the three runs instead judged all six pools in a single session: within a run, the 16 anchors received identical verdicts in all six cleanness pools and word-for-word identical alignment theses in all six alignment pools, and the alignment files of runs 1 and 3 were each written at a single time. The three runs remain independent of each other, which is what the majority and median rules and the kappa need; but the anchor check is in effect one check per run rather than six, and a judge drifting between pools within a run would not be detected by it. The baseline.KlonSuphap-LM receives the topic only as its opening words and was trained on classical verse, so the comparison on alignment is not like for like; we tuned its temperature and also report a kinder setting that gives it the human first line, but did not retrain it. Scope.One generator family, one forum corpus, one-stanza outputs, one judge family, verse by forum users as the human bar. Reproducibility. Decoding is deterministic given the weights and the seed: the released model (sha256 of the weights53ef5c17...19f6bcbe, given in full on its model page) reproduces the 78 held-out stanzas byte for byte on an Apple-silicon GPU and on CPU, and the scores recompute exactly from the stored judge verdicts. Training is not bit-deterministic on that GPU, so retraining gives a similar rather than identical 12 model. The judge is the part that cannot be reproduced exactly. Cleanness and alignment were judged by Claude Opus 5.5, run as agents with the frozen prompts of AppendixA. Asking the judge again is stochastic: the three independent runs agreed with Fleissκ0.76 for cleanness per line and 0.84 for alignment in the main evaluation, 0.84 and 0.95 in the KlonSuphap-LM comparison, and 0.88 for cleanness in the first-line variant, so a re-judging should fall within that run-to-run variation rather than match exactly, and a different or updated judge model may move absolute rates. For this reason every comparison is made within the same pools, which also contain human anchors whose rates are reported. The forum verse used for training and as the human control is not redistributed. 8 Ethics and Broader Impact The corpus is verse publicly posted by forum users, collected for research; stanzas are not redistributed beyond short examples. The Claude-written topics are summaries of public stanzas. Generated verse could be passed off as a person’s own work; a release should state machine authorship. The judge was never used as a training reward. No personal data beyond public forum posts was processed, and no human-subjects study was conducted beyond the author’s own reading. 9 Conclusion A single sample at temperature 0.4 from a two-layer sound–word canvas model writesklon padstanzas that, on 78 held-out topics judged three times, pass form, cleanness and alignment together on 70.5% of topics against 76.9% for the human stanzas those topics were written from. Cleanness is level with human; the remaining difference is alignment on topics written from the human stanzas. Keeping words whole while enforcing rhyme, tone and length on pronounced syllables makes form a property of the decoder, and sampling at a low temperature removes most lexical errors, which sit in the model’s low-probability choices; the price is a plainer vocabulary. A model whose low-probability choices are worth sampling is the open problem. References [1]Wirote Aroonmanakun. TLTK: Thai language toolkit. Python package, version 1.11, https://pypi.org/project/tltk/. [2]Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. In35th Conference on Neural Information Processing Systems (NeurIPS 2021), 2021. [3]Tuhin Chakrabarty, Philippe Laban, Divyansh Agarwal, Smaranda Muresan, and Chien-Sheng Wu. Art or artifice? large language models and the false promise of creativity. InProceedings of the CHI Conference on Human Factors in Computing Systems (CHI ’24), pages 1–34. ACM, 2024. doi: 10.1145/3613904.3642731. [4]Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. MaskGIT: Masked generative image transformer.arXiv preprint arXiv:2202.04200, 2022. URL https://arxiv.org/abs/2202.04200. [5]Cheng-Han Chiang and Hung-yi Lee. Can large language models be an alternative to human evaluation? In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors,Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15607–15631, Toronto, Canada, July 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.acl-long.870. URL https://aclanthology.org/2023.acl-long.870/. [6]Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems, 2021. URL https://arxiv.org/abs/2110.14168. [7]Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. InProceedings of NAACL-HLT, pages 4171–4186, 2019. 13 [8]Marjan Ghazvininejad, Xing Shi, Yejin Choi, and Kevin Knight. Generating topical poetry. InProceedings of EMNLP, pages 1183–1191, 2016. [9]Marjan Ghazvininejad, Xing Shi, Jay Priyadarshi, and Kevin Knight. Hafez: an interactive poetry generation system. In Mohit Bansal and Heng Ji, editors,Proceedings of ACL 2017, System Demonstrations, pages 43–48, Vancouver, Canada, July 2017. Association for Computational Linguistics. URL https://aclanthology.org/P17-4008/. [10]Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. Mask-predict: Parallel decoding of conditional masked language models. InProceedings of EMNLP-IJCNLP, pages 6112–6121, 2019. [11]Chris Hokamp and Qun Liu. Lexically constrained decoding for sequence generation using grid beam search. In Regina Barzilay and Min-Yen Kan, editors,Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1535–1546, Vancouver, Canada, July 2017. Association for Computational Linguistics. doi: 10.18653/v1/P17-1141. URL https://aclanthology.org/P17-1141/. [12]Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. InInternational Conference on Learning Representations (ICLR), 2020. [13]Zhiyuan Hu, Chumin Liu, Yue Feng, Anh Tuan Luu, and Bryan Hooi. Poetrydiffusion: Towards joint semantic and metrical manipulation in poetry generation. InThe Thirty-Eighth AAAI Conference on Artificial Intelligence (AAAI-24), pages 18279–18288. Association for the Advancement of Artificial Intelligence, 2024. [14]Chihan Huang and Xiaobo Shen. Poembert: A dynamic masking content and ratio based semantic language model for chinese poem generation. InProceedings of the 31st International Conference on Computational Linguistics, pages 50–60. Association for Computational Linguistics, 2025. January 19-24, 2025. [15]Tom Kocmi and Christian Federmann. GEMBA-MQM: Detecting translation quality error spans with GPT-4. InProceedings of the Eighth Conference on Machine Translation (WMT), pages 768–775, 2023. [16]Piji Li, Haisong Zhang, Xiaojiang Liu, and Shuming Shi. Rigid formats controlled text generation. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors,Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 742–751, Online, July 2020. Association for Computational Linguistics. doi: 10.18653/v1/2020.acl-main.68. URL https://aclanthology.org/2020.acl-main.68/. [17]Lalita Lowphansirikul, Charin Polpanumas, Nawat Jantrakulchai, and Sarana Nutanong. WangchanBERTa: Pretraining transformer-based Thai language models.arXiv preprint arXiv:2101.09635, 2021. [18]Ximing Lu, Peter West, Rowan Zellers, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. NeuroLogic decoding: (un)supervised neural text generation with predicate logic constraints. In Kristina Toutanova, Anna Rumshisky, Luke Zettlemoyer, Dilek Hakkani-Tur, Iz Beltagy, Steven Bethard, Ryan Cotterell, Tanmoy Chakraborty, and Yichao Zhou, editors,Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4288–4299, Online, June 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.naacl-main.339. URL https://aclanthology.org/2021.naacl-main.339/. [19]Max Peeperkorn, Tom Kouwenhoven, Dan Brown, and Anna Jordanous. Is temperature the creativity parameter of large language models?arXiv preprint arXiv:2405.00492, 2024. URL https://arxiv.org/abs/2405.00492. [20]Wannaphong Phatthiyaphaibun, Korakot Chaovavanich, Charin Polpanumas, Arthit Suriyawongkul, Lalita Lowphansirikul, Pattarawat Chormai, Peerat Limkonchotiwat, Thanathip Suntorntip, and Can Udomcharoenchaikit. PyThaiNLP: Thai natural language processing in python. In Liling Tan, Dmitrijs Milajevs, Geeticka Chauhan, Jeremy Gwinnup, and Elijah Rippeth, editors,Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023), pages 25–36, Singapore, December 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.nlposs-1.4. URL https://aclanthology.org/2023.nlposs-1.4/. [21]Wannaphong Phatthiyaphaibun, Korakot Chaovavanich, Charin Polpanumas, Arthit Suriyawongkul, Lalita Lowphansirikul, and Pattarawat Chormai. PyThaiNLP: Thai natural language processing in Python. Zenodo, software version v5.0.4, https://github.com/PyThaiNLP/pythainlp/, June 2024. 14 [22]PyThaiNLP. klongklon. Hugging Face dataset,https://huggingface.co/datasets/pythainlp/klongklon, 2023. URLhttps://huggingface.co/datasets/pythainlp/klongklon. Accessed 2026-10-03. [23]Subham Sekhar Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin T Chiu, Alexander Rush, and Volodymyr Kuleshov. Simple and effective masked diffusion language models. In38th Conference on Neural Information Processing Systems (NeurIPS 2024), 2024. [24]Jitkapat Sawatphol. Thai Literature Corpora.https://github.com/jitkapat/thailitcorpus, 2019. URL https://attapol.github.io/tlc.html. Version 2.0. [25]Yan Song. Composing ci with reinforced non-autoregressive text generation. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7219–7229. Association for Computational Linguistics, 2022. December 7-11, 2022. [26]Yufei Tian and Nanyun Peng. Zero-shot sonnet generation with discourse-level planning and aesthetics features. In Marine Carpuat, Marie-Catherine de Marneffe, and Ivan Vladimir Meza Ruiz, editors,Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3587–3597, Seattle, United States, July 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022.naacl-main.262. URLhttps://aclanthology.org/2022.naacl-main.262/. [27]Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural language processing. InProceedings of the 2020 EMNLP (Systems Demonstrations), pages 38–45. Association for Computational Linguistics, 2020. [28]Lanqing Xue, Kaitao Song, Duocai Wu, Xu Tan, Nevin L. Zhang, Tao Qin, Wei-Qiang Zhang, and Tie-Yan Liu. DeepRapper: Neural rap generation with rhyme and rhythm modeling. InProceedings of ACL-IJCNLP, pages 69–81, 2021. [29]Kampanat Yingseree. KlonSuphap-LM or GPT-2 for Thai poems (Klon-Paed poem). Hugging Face model card, https://huggingface.co/Kongfha/KlonSuphap-LM; code athttps://github.com/Kongfha/KlonSuphap-LM, 2023. URLhttps://huggingface.co/Kongfha/KlonSuphap-LM. Accessed 2026-10-03. [30]Nutchanon Yongsatianchot and Phuvis Kerdpramote. Composing within constraints: A benchmark for Thai classical poetry in large language models. InProceedings of the 2026 Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the International Joint Conference on Natural Language Processing (AACL-IJCNLP 2026), 2026. URL https://github.com/yongsa-nut/ThaiProsody. [31]Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. InAdvances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023. A Judge prompts Both prompts are reproduced verbatim as they were given to the judge. Each batch file given to the judge begins with the prompt, followed by the items. Cleanness prompt. You are an annotator for the lexical quality of Thaiกลอนแปดverse. Identify exact error spans and classify them. This is poetry, not everyday prose. Do not mark abstraction, metaphor, personification, archaic or poetic diction, unusual imagery, or natural poetic compounds as errors. Do not judge rhyme, syllable count, or verse structure. Mark as an error any exact span that is not linguistically well formed in Thai, such as a non-word, an unrecognizable misspelling, an unnatural word formation, an unnatural combination of words that disrupts reading. Classify as: 1) **critical** if the error makes theวรคincomprehensible. 2) **major** if it makes a native reader stumble but theวรคremains understandable. Or, it contains a non-word 15 unless that non-word is a genuine Thai prefix or suffix that can attach to the neighbouring word 3) **pass** if it is technically erroneous but does not disrupt reading. 4) **pass** if there is no error. Input: each item has a key, a topic, and 4 numbered lines. Output one JSON array with one object per item, in order: ”item”: ”<key>”, ”errors”: [”line”: <1-4>, ”span”: ”<exact substring of that line>”, ”severity”: ”critical|major”, ...]. If there is no error, ”errors” is [] (pass). Alignment prompt. You are judging whether a Thaiกลอนแปดpoem is about the topic it was given. Read the topic and the 4 lines. First write, in one Thai sentence, what the poem is actually about. Then compare that with the topic and score how much of the poem stays on the topic. Judge subject matter only. Ignore rhyme, syllable count, verse structure, spelling, and the quality of the wording. A poem may treat its topic figuratively, through imagery or metaphor, and is still on the topic. Score 1 to 5: 5 = everyวรคis on the topic. 4 = the poem is on the topic, with at most oneวรคthat wanders. 3 = partly on the topic. 2 = the topic is touched once and then left. 1 = unrelated to the topic. Input: each item has a key, a topic, and 4 numbered lines. Output one JSON array with one object per item, in order: ”item”: ”<key>”, ”thesis”: ”<one Thai sentence>”, ”align”: <1-5>. Nothing else. Agent instruction.Each run of an instrument was given to the judge as an agent with a fixed instruction; as Section7reports, one conversation per run judged all six pools. The instruction tells the agent to read every batch file of a pool, to follow the prompt at the head of each file exactly as written without adding, relaxing or tightening criteria, and to return a result for every item under its random key. As part of the instrument it states that the pool deliberately mixes system-generated and human-written poems, unlabelled, to be judged by the same standard without guessing which is which; that the topics are everyday subjects to be judged as given; and, for cleanness, it restates the exemption for metaphor, personification, archaic or poetic diction and unusual imagery and the critical/major definitions. The agent may read only the batch files, never the key map or any earlier run, and writes its verdicts to a JSON file whose item set it checks against the batches. 16 B Generated stanzas that pass all three tests The first 30 topics in order of topic id (read each column from top to bottom) among the 55 whose stanza atT= 0.4passes form, cleanness and alignment in the held-out evaluation, after leaving out three passing stanzas (K041, K045 and K047) because a line of them may offend. Each entry gives the topic id, the topic (the prompt) and the four lines exactly as generated (one sample, seed 0, no selection or editing). Under each Thai line are a romanization and an English gloss. The romanization writes one unit per pronounced syllable, the syllables being those of the form checker (Section3.1), in the Royal Thai General System without tone marks, so the syllable counts and rhymes can be read off it. It reproduces the converter’s reading, including its occasional misreadings (for exampleตุ๊กตุ๊กastuk ka tuk), which are also what the form checker counts. The glosses are literal, line by line, written by Claude Opus 5.5 and corrected by the author where a meaning was wrong; where a generated line is awkward the gloss is left awkward. K001เชิดชูทหารchoet chu tha han(honouring soldiers) ขอเชิดชูเหล่าทหารที่ยิ่งใหญ่ kho choet chu lao tha han thi ying yai Let me honour the great soldiers ทํางานหนักด้วยใจใฝ่ฝึกฝน tham ngan nak duai chai fai fuek fon working hard with a heart eager to train เสียสละชีวิตอย่างอดทน sia sa la chi wit yang ot thon sacrificing their lives with patience เพื่อเป็นคนหาญกล้าของแผ่นดิน phuea pen khon han kla khong phaen din to be the brave people of the land K003นํ้าใจน่าสรเสริญnam chai na san soen(kindness worthy of praise) อันํ้าใจงดงามและยิ่งใหญ่ an nam chai ngot ngam lae ying yai Kindness, beautiful and great คือพลังนํ้าใจที่สร้างสรค์ khue pha lang nam chai thi sang san is the power of kindness that creates นํ้าใจดีมีนํ้าใจมีแบ่งปัน nam chai di mi nam chai mi baeng pan Good kindness, having kindness, having sharing ช่วยเหลือกันทุกวัน่าชื่นชม chuai luea kan thuk wan na chuen chom helping each other every day is admirable K005ตกหลุมพรางเพราะเปลือกtok lum phrang phro plueak(falling into a trap because of the outer shell) เพราะแค่เปลือกที่เราเห็นในความคิด phro khae plueak thi rao hen nai khwam khit Because of only the shell that we see in our thoughts จึงหมดสิทธิ ์ ที่จะคิดและหลงใหล chueng mot sit thi cha khit lae long lai so we lose the right to think and be infatuated รักเพราะเปลือกแต่ข้างในคือใจใคร rak phro plueak tae khang nai khue chai khrai Loving because of the shell, but inside is whose heart สุดท้ายไปไม่อาจรู้ซึ่งความจริง sut thai pai mai at ru sueng khwam ching in the end, going, we cannot know the truth K006ซินเจียู่อี่ขอให้รวยsi no chia yu i kho hai ruai(xin jia yu yi, may you be rich) ซินเจียู่อี่ขอให้มั่น si no chia yu i kho hai man Xin jia yu yi, may you be steady ทุกคืนเช้าคืนวันพลันสดใส thuk khuen chao khuen wan phlan sot sai every night, morning, night and day suddenly bright ขอให้สวยและรวยทรัพย์นับมากมาย kho hai suai lae ruai sap nap mak mai May you be pretty and rich in wealth beyond counting ขอให้กายสุขใจในทุกวัน kho hai kai suk chai nai thuk wan May the body be happy at heart every day K007เก็บเดือนดาวไว้ในใจkep duean dao wai nai chai(keeping the moon and stars in the heart) เก็บเดือนดาวพราวไว้ในใจใคร kep duean dao phrao wai nai chai khrai Keep the glittering moon and stars in whose heart เก็บเดือนดาวเอาไว้ใจแหนหวง kep duean dao ao wai chai haen huang Keep the moon and stars in a jealously guarding heart เก็บเดือนดาวแนบไว้ในแดวง kep duean dao naep wai nai dae duang Keep the moon and stars pressed close in the heart เก็บทุกช่วงดวงใจมอบให้เธอ kep thuk chuang duang chai mop hai thoe Keep every moment, my heart, to give to you 17 K009ขอหมอนกับผ้าห่มkho mon kap pha hom(asking for a pillow and a blanket) นอนกับหมอนผ้าห่มให้ไอุ่น non kap mon pha hom hai ai un Sleeping with a pillow, the blanket gives warmth อุ่นละมุนหนุนตักอยู่ไม่รู้หาย un la mun nun tak yu mai ru hai gently warm, head resting on a lap, never fading ยังคิดถึงคนดีคนข้างกาย yang khit thueng khon di khon khang kai still missing the dear one, the one at my side หลับสบายอุ่นใจยามกอดเธอ lap sa bai un chai yam kot thoe sleeping well, warm at heart when hugging you K010ใส่บาตรปีใหม่ไทยsai bat pi mai thai(giving alms to monks at thai new year) ร่วมใส่บาตรปีใหม่ไทยทั่วหล้า ruam sai bat pi mai thai thua la Join in giving alms at Thai New Year all over the land ร่วมตักบาตรเช้ามาพาสุขสันต์ ruam tak bat chao ma pha suk san Join in offering alms in the morning, bringing happiness ได้ทําบุญอุทิศไปให้แก่กัน dai tham bun u thit pai hai kae kan Having made merit and dedicated it to one another เนื่องในวันปีใหม่ไทยเทอญ nueang nai wan pi mai thai the on on the occasion of Thai New Year, so be it K011หลงมนต์กลอนlong mon klon(enchanted by verse) หลงอักษรกลอนภาษารัก long ak son klon pha sa rak Enchanted by the letters, the verse, the language of love หลงประจักษ์วาจาพาหลงใหล long pra chak wa cha pha long lai Enchanted, seeing plainly, the words lead to infatuation หลงเสน่ห์มนต์รักปักทรวงใน long sa ne mon rak pak thon wong nai Enchanted by the charm, love’s spell planted in the breast หลงฤทัยใฝ่ปองใน้องนาง long rue thai fai pong nai nong nang Enchanted at heart, longing for the young lady K014ปู่จับปลามาเหนื่อยpu chap pla ma nueai(grandfather is tired from catching fish) ฉ ั นจับปลามาเหนื่อยและล้านัก chan chap pla ma nueai lae la nak I caught fish and am very tired and weary จึงขอพักสักครู่ก็พอไหว chueng kho phak sak khru ko pho wai so I ask to rest a moment, that is enough to manage อยากจับปลามาพักเพื่อผ่อนคลาย yak chap pla ma phak phuea phon khlai I want to catch fish and rest to relax แต่ทําไมไม่จับปลามาเสียที tae tham mai mai chap pla ma sia thi but why do I never get to catch fish K015จําใจากเพราะจนcham chai chak phro chon(forced to part because of poverty) จําต้องจากเพราะจนอนาถนัก cham tong chak phro chon a nat nak Having to part because of being poor, so pitiful ไม่มีรักจริงใจเหมือนในฝัน mai mi rak ching chai muean nai fan There is no sincere love like in the dream เคยร่วมเรียงเคียงกายใจใกล้กัน khoei ruam riang khiang kai chai klai kan Once side by side, bodies and hearts close together ทุกคืนวันผันผ่านทรมาน thuk khuen wan phan phan tho ra man every night and day turns past in torment K017วอนเทพประสาทพรwon thep pra sat phon(begging the gods to grant a blessing) วอนเทพไทเทวาสิ่งศักดิ ์ สิทธิ ์ won thep thai the wa sing sak sit I beg the gods, the deities, the sacred things เนรมิตดลจิตนิมิตฝัน ne ra mit don chit ni mit fan to conjure, inspire the mind, visions and dreams จงสดชื่นรื่นคืนชื่นชีวัน chong sot chuen ruen khuen chuen chi wan Be refreshed, pleasant nights, a cheerful life ทุกคืนวันสุขสันต์นิรันดร thuk khuen wan suk san ni ran don every night and day happy forever 18 K018เพื่อนข้างกายphuean khang kai(the friend at my side) เพื่อนที่อยู่ข้างกายมิห่างหาย phuean thi yu khang kai mi hang hai A friend who stays at my side and does not disappear เพื่อนที่คอยห่วงใยไม่เคยห่าง phuean thi khoi huang yai mai khoei hang A friend who keeps caring and is never distant เพื่อนที่รักจริงใจไม่จืดจาง phuean thi rak ching chai mai chuet chang A friend who loves sincerely and does not fade เพื่อนที่อยู่เคียงข้างอย่างจริงจัง phuean thi yu khiang khang yang ching chang A friend who stays beside me in earnest K019ให้ความเศร้าผ่านพ้นไปhai khwam sao phan phon pai(let the sadness pass) ปล่อยความเศร้าผ่านไปในชีวิต ploi khwam sao phan pai nai chi wit Let the sadness pass by in life อย่ายึดติดโกรธกันจนหวั่นไหว ya yuet tit krot kan chon wan wai Do not cling to anger at each other until shaken ใช้เวลาเยียวยารักษาใจ chai we la yiao ya rak sa chai Use time to heal and tend the heart ตั้งสติรับไว้ให้ดีดี tang sa ti rap wai hai di di Stay mindful and take it in well K023คนชื่อดาวkhon chue dao(a person named dao, which means star) คนชื่อดาวมีนํ้าใจะคลายหนาว khon chue dao mi nam chai cha khlai nao The one named Dao has kindness that will ease the cold นภาฝนมีดาวสกาวใส na pha fon mi dao sa kao sai The rainy sky has bright clear stars ขอให้ดาวสุขสันต์ฝันทุกวันไป kho hai dao suk san fan thuk wan pai May Dao be happy and dream every day on ฝากดวงดาวหัวใจให้น้องดาว fak duang dao hua chai hai nong dao I leave the star of my heart to young Dao K026ขอให้โชคดีมีชัยkho hai chok di mi chai(wishing good luck and victory) ให้โชคดีมีชัยมีชัยเถิด hai chok di mi chai mi chai thoet May you have good luck, have victory, have victory สิ่งประเสริฐจงบังเกิดความสดใส sing pra soet chong bang koet khwam sot sai May excellent things give rise to brightness อายุมั่นขวัญยืนยั่งยืนไป a yu man khwan yuen yang yuen pai Long life and lasting spirit, enduring on คิดสิ่งใดสมใจทุกสิ่งเทอญ khit sing dai som chai thuk sing the on Whatever you think of, may it all be fulfilled, so be it K027สุขภาพฟิตปั ๋ งsuk kha phap fit pang(health in top shape) อายุยืนกว่าใครให้สุขสันต์ a yu yuen kwa khrai hai suk san Live longer than anyone, be happy อายุนานทุกวันอย่าหวั่นไหว a yu nan thuk wan ya wan wai Long life every day, do not waver สุขภาพฟิตปั ๋ งมั่นตลอดไป suk kha phap fit pang man ta lot pai Health in top shape, firm forever อายุใจร่างกายังแข็งแรง a yu chai rang kai yang khaeng raeng Age, heart, body still strong K029พี่จะไปไกลแล้วphi cha pai klai laeo(i am going far away now) เธอจะไปไกลแล้วจากใจพี่ thoe cha pai klai laeo chak chai phi You are going far away now from my heart โปรดหันมาคนดีที่ห่วงหา prot han ma khon di thi huang ha Please turn around, dear one who is missed หากมั่นใจระยะทางอย่าร้างลา hak man chai ra ya thang ya rang la If sure of the distance, do not part จงกลับมาพบกันในสักวัน chong klap ma phop kan nai sak wan Do come back to meet again some day 19 K030ขอให้มีเวทมนตร์พาสู่ฝันkho hai mi wet mon pha su fan(may there be magic to carry you to dreams) ขอให้มีเวทมนตร์พาสู่ฝัน kho hai mi wet mon pha su fan May there be magic to lead to dreams ในคืนที่มีจันทร์อันสดใส nai khuen thi mi chan an sot sai on a night that has a bright moon จงสนุกสุขสันต์ทุกวันไป chong sa nuk suk san thuk wan pai Be merry and happy every day on สี่ห้องใจของขวัญให้แด่เธอ si hong chai khong khwan hai dae thoe the four chambers of the heart, a gift given to you K031เก็บความฝันไว้ปลอบใจkep khwam fan wai plop chai(keeping dreams to comfort the heart) เก็บความฝันั้นไว้ในห้วงฝัน kep khwam fan nan wai nai huang fan Keep that dream in the depths of the dream เก็บเอาไว้ด้วยกันอย่าหวั่นไหว kep ao wai duai kan ya wan wai Keep it together, do not waver เก็บความฝันความจริงที่หายไป kep khwam fan khwam ching thi hai pai Keep the dream, the truth that has disappeared เก็บเอาไว้ในความฝันไว้ปลอบใจ kep ao wai nai khwam fan wai plop chai Keep it in the dream, keep it to comfort the heart K033อาจารย์กวีในใจศิษย์a chan ka wi nai chai sit(the poet teacher in the student’s heart) คือาจารย์แห่งใจในอักษร khue a chan haeng chai nai ak son He is the teacher of the heart in letters ผู้พากเพียรกานท์กลอนอันสดใส phu phak phian kan klon an sot sai the one who perseveres in bright verse and poetry นําชีวิตศิษย์กวีที่แสนไกล nam chi wit sit ka wi thi saen klai leading the life of the poet student who is so far สู่จุดหมายในปรัชญาบทกวี su chut mai nai prat ya bot ka wi to the destination in the philosophy of poems K034คนเก่ายังอยู่ครบkhon kao yang yu khrop(the old ones are all still here) คนเก่าที่ยังอยู่ครบรับไว้ khon kao thi yang yu khrop rap wai The old ones who are all still here, accept them ยังสถิตย์ในใจไม่เหินห่าง yang sa thit nai chai mai hoen hang still dwelling in the heart, not drifting away แม้ถึงวันเวลาจะเลือนลาง mae thueng wan we la cha luean lang even when the days and time will grow dim แต่ยังคงทุกอย่างเพื่อําลา tae yang khong thuk yang phuea am la but everything still remains, in order to say farewell K036ฤดูหนาวใจชํ้าrue du nao chai cham(winter, a bruised heart) ในฤดูหนาวใจในคราวหนาว nai rue du nao chai nai khrao nao In the winter of the heart, in the cold time เคยพราวพร่างมีดาวพราวไสว khoei phrao phrang mi dao phrao sa wai once sparkling, with stars glittering brightly แต่วันี้ต้องพรากจากแสนไกล tae wan ni tong phrak chak saen klai but today we must part, so far away ตอกยํ้าใจให้คิดถึงคะนึงเธอ tok yam chai hai khit thueng kha nueng thoe it drives home to the heart to miss and long for you K038นั่งชมคลื่นชายหาดnang chom khluen chai hat(sitting and watching waves at the beach) ริมชายหาดริมหาดหยุดความเหงา rim chai hat ri ma hat yut khwam ngao By the beach, by the beach, the loneliness stops คลื่นทะเลของเราเฝ้าหวั่นไหว khluen tha le khong rao fao wan wai Our sea waves keep on trembling คลื่นจะซัดสาดลงตรงหาดทราย khluen cha sat sat long trong hat sai The waves will crash down right on the sandy beach มองนํ้าใสใจสงบรับอารมณ์ mong nam sai chai sa ngop rap a rom Looking at clear water, heart calm, taking in the mood 20 K039ขอให้รวยลาภkho hai ruai lap(may you be rich in fortune) ขอให้รวยลาภยศไร้ทุกข์โศก kho hai ruai lap yot rai thuk sok May you be rich in fortune and rank, free of sorrow ขอให้โชคที่ตั้งปรารถนา kho hai chok thi tang prat tha na May there be the luck that the wish is set on ขอให้รวยทรัพย์อุดมด้วยเงินตรา kho hai ruai sap u dom duai ngoen tra May you be rich in wealth, abundant with money ขอให้หนาสมหวังสมดั่งใจ kho hai na som wang som dang chai May you, please, have hopes fulfilled as the heart desires K040นํ้ามันปาล์มแพงnam man pam phaeng(palm oil is expensive) นํ้ามันปาล์มแพงไปไม่เข้าท่า nam man pam phaeng pai mai khao tha Palm oil is too expensive, it is no good เขาดันไล่ขึ้นมาประกาศขาย khao dan lai khuen ma pra kat khai They went and pushed it up and announced it for sale ลิตรหนึ่งลิตรแสนถูกทํายังไง lit nueng lit saen thuk tham yang ngai One litre, a litre, so cheap, what to do เก็บเอาไว้ดีกว่าค่านํ้ามัน kep ao wai di kwa kha nam man Better to keep it, the cost of oil K042ความหมายของครูkhwam mai khong khru(the meaning of teacher) ครูคือผู้ให้รู้ถึงความหมาย khru khue phu hai ru thueng khwam mai A teacher is one who lets us know the meaning ที่สละตั้งใจใฝ่ศึกษา thi sa la tang chai fai suek sa who gives up, intent, eager to study การศึกษาของนักเรียนในวิชา kan suek sa khong nak rian nai wi cha the education of students in the subjects เพื่อรํ่าเรียนรู้ค่าภาษาครู phuea ram rian ru kha pha sa khru to study and know the worth of the teacher’s language K043ตามจับชาตรีที่หายไปtam chap cha tri thi hai pai(chasing after chatri who disappeared) อยากตามจับชาตรีที่หายไป yak tam chap cha tri thi hai pai I want to chase and catch Chatri who disappeared โปรดจงช่วยรับไว้จากใจพี่ prot chong chuai rap wai chak chai phi Please help accept it from this older one’s heart น้องจะรีบจับจองห้องชีวี nong cha rip chap chong hong chi wi The younger one will hurry to reserve the room of life รอชาตรีมาคืนเข้าครองเรือน ro cha tri ma khuen khao khrong ruean waiting for Chatri to come back and start married life K044ลมพัดเสียงครวญlom phat siang khruan(the wind blows a wailing sound) ลมพัดเสียงสําเนียงเพียงแผ่วพลิ้ว lom phat siang sam niang phiang phaeo phlio The wind blows a sound, a tone only faint and fluttering ลมพัดปลิวลิ่วพริ้วสะบัดไหว lom phat plio lio phrio sa bat wai The wind blows, drifting fast, fluttering, flapping and swaying สายลมพัดห่วงหาจากคนไกล sai lom phat huang ha chak khon klai The wind blows longing from someone far away ลมพัดไปใจกําสรวลรัญจวนครวญ lom phat pai chai kam suan ran chuan khruan The wind blows on, the heart grieves, yearns and wails K046ความจริงคือเหงาkhwam ching khue ngao(the truth is loneliness) ความจริงเหงาคือเหงาเฝ้าหม่นหมอง khwam ching ngao khue ngao fao mon mong The truth, lonely, is loneliness, keeping on in gloom ได้แต่มองมองเธอละเมอหา dai tae mong mong thoe la moe ha I can only look, look at you, and dream in longing หวนคิดถึงวันวานที่ผ่านมา huan khit thueng wan wan thi phan ma thinking back to the days gone by วันเวลามาเยือนเหมือนเช่นเคย wan we la ma yuean muean chen khoei the days and time come to visit just as before 21 K048สิบแปดท่าเพลงยุทธsip paet tha phleng yut(eighteen martial arts moves) สิบแปดท่าเพลงยุทธท่านผู้กล้า sip paet tha phleng yut than phu kla Eighteen martial arts moves, you brave one ล้วนลีลาสะท้านจนสั่นไหว luan li la sa than chon san wai all of them moves that shake until trembling ทั้งกระบี่กระบองบรเลงไป thang kra bi kra bong ban leng pai both sword and staff play on สิบแปดท่าหัวใจไม่หวั่นเกรง sip paet tha hua chai mai wan kreng eighteen moves, the heart is not afraid 22 C Shared phrases in the held-out stanzas Table4lists the 10 lines of the 78 held-out stanzas (Section5.3) whose longest run of written syllables shared with a line of the three corpora is six or more syllables (one of them shares seven). For each, it gives the shared phrase and one corpus line that contains it; the corpus column names the training split (forum train), the test split (forum test) or the further verse collection (further verse). No generated line occurs verbatim in the corpora. The shared phrases are common set phrases, most of them from greeting and blessing verse. Table 4:The 10 lines of the 78 held-out stanzas (T= 0.4) whose longest run of written syllables shared with a corpus line is six or more syllables. Run: length of the longest shared run in syllables. Example corpus line: the first line in corpus order that contains the shared phrase. Generated lineRun Shared phraseExample corpus lineCorpus ขอให้กายสุขใจในทุกวัน6กายสุขใจในทุกวันสุขกายสุขใจในทุกวัน และเป็น มิ่งขวัญของคนบ้านกลอน further verse เนื่องในวันปีใหม่ไทยเทอญ6เนื่องในวันปีใหม่ไทยขอวยพรเนื่องในวันปีใหม่ไทยให้กับ ท่านอาวุโส further verse ขอํานาจสิ่งศักดิ ์ สิทธิ ์ ทั่วสกล6ขอํานาจสิ่งศักดิ ์ สิทธิ ์ ขอํานาจสิ่งศักดิ ์ สิทธิ ์ สถิตฟ้าforum train แต่วันี้ต้องพรากจากแสนไกล6แต่วันี้ต้องพรากจากแต่วันี้ต้องพรากจากเพื่อนแล้วfurther verse รื่นภิรมย์ชมชื่นทุกคืนวัน7ภิรมย์ชมชื่นทุกคืนวันได้ภิรมย์ชมชื่นทุกคืนวันforum train ขอให้มีความสุขทุกช่วงนาที6ขอให้มีความสุขทุกขอให้มีความสุขทุกวี่วันforum train แม่คือกําลังใจให้ทุกคน6กําลังใจให้ทุกคนเรื่องร้ายให้ผ่านไป เป็นกําลังใจให้ทุก คน further verse ขอให้เธอมีความสุขทุกเวลา6ขอให้เธอมีความสุขอให้เธอมีความสุขfurther verse หรือหายไปไกลับไม่กลับคืน6หายไปไกลับไม่กลับเห็นหายไปไกลับไม่กลับเยือนforum train ขอให้เธอสมหวังดังใจใฝ่ห่ม6ขอให้เธอสมหวังดังก็ขอให้เธอสมหวังดังตั้งใจforum test 23 D Originality on unusual topics A topic that no training stanza could have been about asks the generator for lines it cannot have seen. In the spirit of the “astronaut riding a horse” prompt used to show what text-to-image generators can compose, 17 deliberately unusual topics were written, such as an astronaut riding an elephant to a floating market. None of them occurs as a training or test title or topic, checked by exact and substring match; the substring matches are ordinary words and phrases (คิดถึง, “to miss”;คิดถึงบ้าน, “homesick”;ดาว, “star”; andแต่งกลอน, “to write verse”). The method wrote five stanzas for each topic (seeds 0–4,T= 0.4, decoded on CPU), with no selection or editing; they are shown below with a romanization and an English gloss under each line, as in AppendixB; three stanzas (seed 3 of topics 1, 6 and 13) are left out because a line of them may offend. All 82 stanzas shown pass the form checker. These stanzas were not judged for cleanness or alignment. The originality check of Section5.3, run on the 82 stanzas shown (328 lines) against the same three corpora, finds no line verbatim in any of them. The longest run of written syllables shared with a corpus line is 0 (no run of four or more) for 186 lines, 4 for 104, 5 for 35 and 6 or more for 3. Table5lists those 3 lines with their shared phrases: two are set phrases of thanks and one is a common phrase about time passing. Table 5:The 3 lines of the 82 stanzas for unusual topics whose longest run of written syllables shared with a corpus line is six or more syllables, with the shared phrase and the first corpus line that contains it. Generated lineRun Shared phraseExample corpus lineCorpus วันเวลาผ่านไปไม่ว่ากัน6วันเวลาผ่านไปไม่วันเวลาผ่านไปไม่เหมือนเก่าforum train ขอขอบคุณนํ้าใจที่มาให้6ขอขอบคุณนํ้าใจที่ขอขอบคุณนํ้าใจที่ได้มาforum train ขอขอบคุณไมตรีที่มีมา7ขอขอบคุณไมตรีที่มีขอขอบคุณไมตรีที่มีให้forum train Below, each stanza is labelled with its seed (0–4). 1.นักบินอวกาศขี่ช้างไปตลาดนํ้าnak bin a wa kat khi chang pai ta lat nam(an astronaut rides an elephant to the floating market) seed 0 อวกาศเธอขี่ช้างไปถึงฝั่ง a wa kat thoe khi chang pai thueng fang Space, you ride an elephant to the shore หัวใจพังดับลับไม่กลับได้ hua chai phang dap lap mai klap dai The heart breaks, goes out, vanishes, cannot come back เหลือเพียงฉ ั นเท่านั้นที่ฝันไป luea phiang chan thao nan thi fan pai Only I am left, the one who dreamed เธอยู่แค่ในใจไม่มีเรา thoe yu khae nai chai mai mi rao You are only in my heart, there is no us seed 1 อวกาศขี่ช้างไปกลางป่า a wa kat khi chang pai klang pa Space rides an elephant into the middle of the forest แต่ทว่ายังหาซึ่งจุดหมายได้ tae tha wa yang ha sueng chut mai dai But yet it can still find the destination ด้วยชะตาระยะทางอันยาวไกล duai cha ta ra ya thang an yao klai By fate, the distance so long คงไม่เป็นดั่งใจในสักวัน khong mai pen dang chai nai sak wan It will probably not be as wished some day seed 2 อวกาศจะขี่ช้างไปโลกว้าง a wa kat cha khi chang pai lok kwang Space will ride an elephant to the wide world ระยะทางไกลิบอย่าหวั่นไหว ra ya thang kai la lip ya wan wai The distance is very far, do not waver เพื่อเดินทางสู่ฝันอันยาวไกล phuea doen thang su fan an yao klai To travel toward a far-off dream แต่ตัวเธอหายไปใจคนคอย tae tua thoe hai pai chai khon khoi But you disappeared, heart of the one who waits 24 seed 4 อวกาศขี่ช้างไปในป่า a wa kat khi chang pai nai pa Space rides an elephant in the forest เจอคนหล่อโผล่มาว่าเป็นผี choe khon lo phlo ma wa pen phi Meets a handsome man popping up, says he is a ghost เธอก็ชอบขี่ม้าดูเท่ดี thoe ko chop khi ma du the di You also like riding horses, it looks cool แต่ไม่มีใครเห็น่าเห็นใจ tae mai mi khrai hen na hen chai But nobody sees, it is pitiable 2.หุ่นยนต์ตําส้มตําhun yon tam som tam(a robot pounds papaya salad) seed 0 เหมือนหุ่นยนต์ตําส้มตําที่เธอสั่ง muean hun yon tam som tam thi thoe sang Like the robot pounding papaya salad that you ordered ฉ ั นก็นั่งส้มตํานํ้าตาไหล chan ko nang som tam nam ta lai I just sit, papaya salad, tears flowing ความผิดหวังผิดหวังทุกครั้งไป khwam phit wang phit wang thuk khrang pai Disappointment, disappointment every time เธอต้องการอะไรให้บอกมา thoe tong kan a rai hai bok ma Whatever you want, tell me seed 1 หุ่นยนต์ตําส้มตําได้ดังว่า hun yon tam som tam dai dang wa The robot pounds papaya salad just as said บริการสั่งมาสนุกสนาน bo ri kan sang ma sa nuk sa nan Service, orders come in, great fun มีลูกค้าประจําร้านแสนสําราญ mi luk kha pra cham ran saen sam ran There are regular customers at the shop, very cheerful บริการวดเร็วประทับใจ bo ri kan ruat reo pra thap chai The service is fast and impressive seed 2 มีหุ่นยนต์ตําส้มตําให้ฉ ั นบ้าง mi hun yon tam som tam hai chan bang There is a robot pounding papaya salad for me too แต่ละอย่างยากลําบากนํามาใส่ tae la yang yak lam bak nam ma sai Each thing is hard, brought and put in เรียกลูกค้าสั่งกันทุกวันไป riak luk kha sang kan thuk wan pai Calling customers, they order every day ต้นทุนส่งเท่าไรไม่เคยพอ ton thun song thao rai mai khoei pho However much the delivery cost, it is never enough seed 3 หุ่นยนต์ตําส้มตํานํามาใส่ hun yon tam som tam nam ma sai The robot pounds papaya salad, brings it and puts it in แล้วเก็บกล่องเอาไว้ให้ค้นหา laeo kep klong ao wai hai khon ha Then keeps the box to be searched for ชื่อเมนูเขาสั่งที่สั่งมา chue me nu khao sang thi sang ma The name of the dish he ordered, that was ordered อยากรู้ว่าอย่างไรไม่เข้าใจ yak ru wa yang rai mai khao chai Want to know how, do not understand seed 4 หุ่นยนต์ตําส้มตําฉ ั นไม่ว่า hun yon tam som tam chan mai wa A robot pounding papaya salad, I do not mind แต่อย่าให้ซื้อมาได้ไหมนี่ tae ya hai sue ma dai mai ni But do not make me buy it, can you เพราะส้มตําครกนั้นมันไม่ดี phro som tam khrok nan man mai di Because that mortar of papaya salad is not good ขอคนที่ชํานาญช่วยตําที kho khon thi cham nan chuai tam thi I ask for a skilled person to pound it please 3.ปลาทองเรียนว่ายนํ้าอนไลน์pla thong rian wai nam on lai(a goldfish takes swimming lessons online) seed 0 น้องพันทองเรียนว่ายนํ้าอนไลน์ได้ nong phan thong rian wai nam on lai dai Young Phanthong can learn swimming online เพียงไม่ต้องตั้งใจหมั่นฝึกฝน phiang mai tong tang chai man fuek fon Only no need to concentrate, practise diligently อยากว่ายนํ้าให้คล่องต้องอดทน yak wai nam hai khlong tong ot thon To swim fluently one must be patient จะเป็นคนเก่งกล้ากว่าใครใคร cha pen khon keng kla kwa khrai khrai Will be a person more skilled and brave than anyone seed 1 เมื่อได้รับคําตอบจากสายตา muea dai rap kham top chak sai ta When receiving the answer from the eyes จึงรู้ว่าความรู้นั้นหาไม่ได้ chueng ru wa khwam ru nan ha mai dai Then I know that knowledge cannot be found ปล่อยให้ปลาเรียนว่ายนํ้าอนไลน์ไป ploi hai pla rian wai nam on lai pai Let the fish go on learning swimming online จะเรียนรู้อะไรไม่มีเลย cha rian ru a rai mai mi loei It will learn nothing at all 25 seed 2 อยากให้ปลาเรียนว่ายนํ้าอนไลน์บ้าง yak hai pla rian wai nam on lai bang I want the fish to learn swimming online too ช่วยกันสร้างธุรกิจอันสดใส chuai kan sang thu ra kit an sot sai Help each other build a bright business เพื่อให้ปลาว่ายนํ้าอย่างปลอดภัย phuea hai pla wai nam yang plot phai So that the fish swim safely เงินที่เก็บหายไปในพริบตา ngoen thi kep hai pai nai phrip ta The saved money vanished in the blink of an eye seed 3 พันทองเรียนว่ายนํ้าอนไลน์น้อง phan thong rian wai nam on lai nong Phanthong learns swimming online, little one รับรองต้องสนุกสุขสดใส rap rong tong sa nuk suk sot sai Guaranteed to be fun, happy and bright รีบสมัครดูแลอย่างเร็วไว rip sa mak du lae yang reo wai Hurry to sign up, take care quickly จะได้เป็นขวัญใจในวันี้ cha dai pen khwan chai nai wan ni You will become the darling today seed 4 ปลาทองเรียนว่ายนํ้าอนไลน์หรือ pla thong rian wai nam on lai rue A goldfish learning swimming online, really? นี่หรือคือสุขใจในวิถี ni rue khue suk chai nai wi thi Is this what happiness in the way of life is? โอ้ชีวิตของฉ ั นั้นแสนดี o chi wit khong chan nan saen di Oh, my life is so good คงไม่มีใครเป็นเช่นี้เลย khong mai mi khrai pen chen ni loei Surely nobody is like this at all 4.ตุ๊กตุ๊กบนดาวอังคารtuk ka tuk bon dao ang khan(a tuk-tuk on Mars) seed 0 ตุ๊กตุ๊กบนดวงจันทร์ฉ ั นผิดหวัง tuk ka tuk bon duang chan chan phit wang Tuk-tuk on the moon, I am disappointed เลยต้องนั่งมองไปใจสับสน loei tong nang mong pai chai sap son So I have to sit looking on, my mind confused การกระทําที่ทําชํ้ากมล kan kra tham thi tham cham ka mon The deeds done bruise the heart ไม่ใช่คนมาจากดาวอังคาร mai chai khon ma chak dao ang khan Not a person who comes from Mars seed 1 ตุ๊กตุ๊กสามล้อบนดาวเสาร์ tuk ka tuk sam lo bon dao sao Tuk-tuk three-wheeler on Saturn รถตุ๊กเขาทําให้เราเพ้อฝันหา rot tuk khao tham hai rao phoe fan ha His tuk car makes us dream and long ตุ๊กตุ๊กดาวอังคารปั่นไปมา tuk ka tuk dao ang khan pan pai ma The Mars tuk-tuk pedals back and forth แล้วนําพาหลงทางอยู่แสนไกล laeo nam pha long thang yu saen klai Then leads us lost, very far away seed 2 ตุ๊กตุ๊กบนดวงจันทร์วันฟ้ากว้าง tuk ka tuk bon duang chan wan fa kwang Tuk-tuk on the moon, a day of wide sky มองเส้นทางต่างอย่างไรไม่สับสน mong sen thang tang yang rai mai sap son Looking at the route, however different, not confused ตุ๊กตุ๊กขับไปตามใจตน tuk ka tuk khap pai tam chai ton The tuk-tuk drives as it pleases ฉ ั นเป็นคนมาจากดาวอังคาร chan pen khon ma chak dao ang khan I am a person who comes from Mars seed 3 เป็นตุ๊กคนบนดาวอันพราวหาว pen tuk khon bon dao an phrao hao Being a tuk person on a star sparkling in the sky ตุ๊กตุ๊กดูดาวสกาวใส tuk ka tuk du dao sa kao sai The tuk-tuk watches the bright clear stars ตุ๊กตุ๊กราตรีที่ยาวไกล tuk ka tuk ra tri thi yao klai Tuk-tuk, the long night ตุ๊กตุ๊กท่องไปในดวงจันทร์ tuk ka tuk thong pai nai duang chan The tuk-tuk roams in the moon seed 4 ตุ๊กตุ๊กบนฟ้ามองท้องฟ้า tuk ka tuk bon fa mong thong fa Tuk-tuk in the sky looks at the sky ตุ๊กตุ๊กบนฟ้าหาใช่ที่ tuk ka tuk bon fa ha chai thi Tuk-tuk in the sky is not in its place ตุ๊กตุ๊กบนดาวดวงนี้ดี tuk ka tuk bon dao duang ni di Tuk-tuk on this star is good แต่ไม่มีตุ๊กแกบนดวงจันทร์ tae mai mi tuk kae bon duang chan But there is no gecko on the moon 26 5.คนหลอกผีkhon lok phi(a person who frightens ghosts) seed 0 ผีหลอกคนผีหลอกผีท่านทั้งหลาย phi lok khon phi lok phi than thang lai Ghosts trick people, ghosts trick ghosts, everyone ผีหลอกคนมากมายให้สับสน phi lok khon mak mai hai sap son Ghosts trick many people into confusion ผีหลอกคนหลอกลิงผีหลอกคน phi lok khon lok ling phi lok khon Ghosts trick people, trick monkeys, ghosts trick people ผีหลอกผีหลอกตนจนงมงาย phi lok phi lok ton chon ngom ngai Ghosts trick ghosts, trick themselves into blind belief seed 1 ผีจะหลอกอย่างไรที่ไหนเล่า phi cha lok yang rai thi nai lao How would a ghost trick anyone, and where นึกว่าเรามัน่ากลัวยิ่งกว่าผี nuek wa rao man na klua ying kwa phi I thought I was scarier than a ghost เห็นผีทีไรผวาทุกราตรี hen phi thi rai pha wa thuk ra tri Every time I see a ghost I am startled, every night เพราะหน้าตาคนี้ผีไม่กลัว phro na ta khon ni phi mai klua Because of this person’s face the ghost is not afraid seed 2 ผีหลอกคนเพราะอยากเห็นเป็นผีสาง phi lok khon phro yak hen pen phi sang Ghosts trick people because they want to be seen as ghosts ผีจึงได้ตัวอย่างมาว่าหนังผี phi chueng dai tua yang ma wa nang phi So the ghost got an example, called a ghost film ผีหลอกคนหลอกผีทุกราตรี phi lok khon lok phi thuk ra tri Ghosts trick people, trick ghosts, every night ผีไม่มีอะไรให้น่ากลัว phi mai mi a rai hai na klua Ghosts have nothing to be afraid of seed 3 ผีจะหลอกอย่างไรก็ไม่ได้ phi cha lok yang rai ko mai dai However the ghost tries to trick, it cannot ผีมันโผล่เข้าไปในหนังหัว phi man phlo khao pai nai nang hua The ghost pops into the scalp ผีมันบอกอกมาช่างน่ากลัว phi man bok ok ma chang na klua The ghost speaks out, how scary คนอะไรพันพัวกลัวเหลือเกิน khon a rai phan phua klua luea koen What kind of person, entangled, so very afraid seed 4 ผีจะหลอกอย่างไรไม่เข้าท่า phi cha lok yang rai mai khao tha However the ghost tries to trick, it does not work ผีก็ไม่เห็นหน้าว่าเป็นผี phi ko mai hen na wa pen phi The ghost does not see the face to know it is a ghost ผีรู้ว่าคนทําชํ้าชีวี phi ru wa khon tham cham chi wi The ghost knows that people do things that bruise life ผีจึงหนีไปจากคนหลอกลวง phi chueng ni pai chak khon lok luang So the ghost flees from deceitful people 6.ทุเรียนบนสถานีอวกาศthu rian bon sa tha ni a wa kat(durian on the space station) seed 0 สถานีอวกาศกี่พันครั้ง sa tha ni a wa kat ki phan khrang the space station, how many thousand times เธอก็ยังเฝ้ารอเธอรู้ไหม thoe ko yang fao ro thoe ru mai you still keep waiting, do you know กินทุเรียนบนฝันทุกวันไป kin thu rian bon fan thuk wan pai eating durian on a dream every day ด้วยหัวใจมุ่งมั่นในความรัก duai hua chai mung man nai khwam rak with a heart determined in love seed 1 ทุเรียนทอดยังหล่นบนท้องฟ้า thu rian thot yang lon bon thong fa fried durian still falls on the sky โปรดประทานส่งมาพาฉ ั นสงสัย prot pra than song ma pha chan song sai please bestow and send it, making me wonder สถานีอวกาศนั้นแสนไกล sa tha ni a wa kat nan saen klai that space station is so far away แต่วันี้เธอไปไม่พบเจอ tae wan ni thoe pai mai phop choe but today you went and we did not meet seed 2 สถานีอวกาศขอบฟ้ากว้าง sa tha ni a wa kat khop fa kwang the space station, the wide horizon ระยะทางยังห่างไกลให้สับสน ra ya thang yang hang klai hai sap son the distance is still far, causing confusion ทุเรียน้อยลอยไปในใจคน thu rian noi loi pai nai chai khon a little durian floats into people’s hearts ทุเรียนบนดวงจันทร์นั้นแสนไกล thu rian bon duang chan nan saen klai the durian on the moon is so far away 27 seed 4 เห็นทุเรียนหล่นร่วงบนท้องฟ้า hen thu rian lon ruang bon thong fa seeing a durian drop on the sky คนตะโกนขึ้นมาว่าเป็นผี khon ta kon khuen ma wa pen phi people shouted out that it was a ghost เธอก็เลยจากไปมิใยดี thoe ko loei chak pai mi yai di so you left without caring คงไม่มีอีกแล้วอีกต่อไป khong mai mi ik laeo ik to pai there will probably be no more, not anymore 7.กบเป็นดีเจkop pen di che(a frog works as a DJ) seed 0 กบอยากเป็นดีเจให้ดูสักครั้ง kop yak pen di che hai du sak khrang the frog wants to be a dj for you to see, just once ถ้าเธอลองมาฟังจะดีไหม tha thoe long ma fang cha di mai if you try coming to listen, would that be good ช่วยร้องเพลงกล่อมกันทุกวันไป chuai rong phleng klom kan thuk wan pai help sing lullabies to each other every day หรือทํางานก็ได้ไม่กังวล rue tham ngan ko dai mai kang won or working is fine too, no worries seed 1 ข่าวกบเป็นดีเจที่ไหนเล่า khao kop pen di che thi nai lao news that the frog is a dj, where is that ฟังข่าวสาวิทยุมาจากไหน fang khao sao wit tha yu ma chak nai hearing news of the radio girl, from where ทั้งทีวีวิทยุฟังกันไป thang thi wi wit tha yu fang kan pai both tv and radio, people keep listening ถามว่าใครฟังบ้างไม่รู้เลย tham wa khrai fang bang mai ru loei ask who listens, I do not know at all seed 2 กบอยากเป็นดีเจใหม่ให้ใครบ้าง kop yak pen di che mai hai khrai bang the frog wants to be a new dj, for whom แต่หลายอย่างลองเปิดฟังดูสักหน tae lai yang long poet fang du sak hon but many things, try turning it on to listen once จะเพียรฝึกฝนตนอย่างอดทน cha phian fuek fon ton yang ot thon will diligently train oneself with patience คงมีคนสนใจมาดูแล khong mi khon son chai ma du lae probably someone interested will come to take care seed 3 เป็นดีเจําเป็นต้องฟังเขา pen di che cham pen tong fang khao being a dj, one has to listen to them คนอกหักอย่างเราเศร้าไปไหม khon ok hak yang rao sao pai mai heartbroken people like us, is it too sad เพราะชีวิตช่วงนี้ไม่มีใคร phro chi wit chuang ni mai mi khrai because life these days has no one กบจะร้องทําไมให้ชํ้าใจ kop cha rong tham mai hai cham chai why would the frog sing to bruise the heart seed 4 กบเป็นดีเจได้ในเวลา kop pen di che dai nai we la the frog can be a dj in time อาจต้องใช้เวลาตามวิถี at tong chai we la tam wi thi it may take time, following the way ว่าชีวิตการฟังเป็นเรื่องดี wa chi wit kan fang pen rueang di that a life of listening is a good thing เพราะไม่มีใครเขามารบกวน phro mai mi khrai khao ma rop kuan because nobody comes to disturb 8.เมฆขอยืมร่มmek kho yuem rom(a cloud asks to borrow an umbrella) seed 0 ร่มขอยืมร่มขอยืมร่มที่ให้ rom kho yuem rom kho yuem rom thi hai umbrella asks to borrow umbrella asks to borrow the umbrella given ร่มขอยืมร่มใครเขาไม่สน rom kho yuem rom khrai khao mai son umbrella asks to borrow whose umbrella, they do not care ร่มขอยืมร่มตนทนอดทน rom kho yuem rom ton thon ot thon umbrella asks to borrow its own umbrella, enduring, enduring ร่มขอยืมร่มตนคนรําคาญ rom kho yuem rom ton khon ram khan umbrella asks to borrow its own umbrella, people are annoyed seed 1 ร่มขอยืมร่มหนาร่มใบเก่า rom kho yuem rom na rom bai kao umbrella asks to borrow an umbrella, please, the old umbrella ร่มของเราเขาขอยืมร่มเขาให้ rom khong rao khao kho yuem rom khao hai our umbrella, he asks to borrow the umbrella, he gives ร่มของเขาเราขอยืมแล้วทําไม rom khong khao rao kho yuem laeo tham mai his umbrella, we asked to borrow it, so why ร่มของใครกันหนอที่ขอยืม rom khong khrai kan no thi kho yuem whose umbrella is it, I wonder, that is being borrowed 28 seed 2 ร่มขอยืมร่มวอนให้เห็นใจบ้าง rom kho yuem rom won hai hen chai bang umbrella asks to borrow an umbrella, pleading for some sympathy ร่มอ้างว้างใจมันเหงาเขาผลักไส rom ang wang chai man ngao khao phlak sai umbrella forlorn, its heart lonely, they push it away ร่มของฉ ั นคนดีหนีจากไกล rom khong chan khon di ni chak klai my umbrella, the dear one fled far away ร่มที่เธอจากไปไม่หวนคืน rom thi thoe chak pai mai huan khuen the umbrella that you left, never to return seed 3 ร่มขอยืมร่มเก่าร่มก่อนเก่า rom kho yuem rom kao rom kon kao umbrella asks to borrow the old umbrella, the umbrella before the old one ร่มขอยืมร่มเราร่มวันไหน rom kho yuem rom rao rom wan nai umbrella asks to borrow our umbrella, umbrella of which day ร่มขอยืมร่มนี้ร่มครั้งใด rom kho yuem rom ni rom khrang dai umbrella asks to borrow this umbrella, umbrella of which time ร่มขอยืมร่มใหม่ร่มของใคร rom kho yuem rom mai rom khong khrai umbrella asks to borrow a new umbrella, whose umbrella seed 4 ร่มขอยืมดวงใจากปลายฟ้า rom kho yuem duang chai chak plai fa umbrella asks to borrow a heart from the end of the sky ร่มขอยืมสัญาคิดถึงพี่ rom kho yuem san ya khit thueng phi umbrella asks to borrow a promise, missing you, dear ร่มขอยืมนํ้าใจในไมตรี rom kho yuem nam chai nai mai tri umbrella asks to borrow kindness in friendship ร่มขอยืมคนดีที่ห่วงใย rom kho yuem khon di thi huang yai umbrella asks to borrow the dear one who cares 9.นาฬิกาที่เดินถอยหลังna li ka thi doen thoi lang(a clock that runs backwards) seed 0 นาฬิกาที่เดินถอยหลังทุกครั้ง na li ka thi doen thoi lang thuk khrang the clock that runs backward every time ทุกนาทียังฟังยังคงฝัน thuk na thi yang fang yang khong fan every minute still listening, still dreaming วันเวลาผ่านไปไม่ว่ากัน wan we la phan pai mai wa kan days and time pass by, never mind ส่วนตัวฉ ั นก้าวต่อไปก้าวเดิน suan tua chan kao to pai kao doen as for me, I step onward, step and walk seed 1 นาฬิกาที่เดินถอยหลังตั้ง na li ka thi doen thoi lang tang the clock that runs backward, set เป็นพลังแห่งความฝันอันยิ่งใหญ่ pen pha lang haeng khwam fan an ying yai is the power of a great dream จงหันมองนาฬิกาที่เดินไป chong han mong na li ka thi doen pai do turn and look at the clock that runs on เพื่อขับเคลื่อนก้าวใหม่ให้ก้าวเดิน phuea khap khluean kao mai hai kao doen to drive a new step, to step and walk seed 2 นาฬิกาที่เดินถอยหลังไปบ้าง na li ka thi doen thoi lang pai bang the clock that runs backward somewhat เดินตามทางเดินต่อไปไม่ถึงฝั่ง doen tam thang doen to pai mai thueng fang walking along the path, walking on, not reaching the shore เธอจากไปในคืนวันใจพัง thoe chak pai nai khuen wan chai phang you left in the days and nights of a broken heart ยังรู้สึกทุกครั้งที่ก้าวเดิน yang ru suek thuk khrang thi kao doen still feel it every time I step and walk seed 3 นาฬิกาที่เดินถอยหลังครั้ง na li ka thi doen thoi lang khrang the clock that runs backward, a time พร้อมกับการถอยหลังเพื่อก้าวใหม่ phrom kap kan thoi lang phuea kao mai along with stepping back for a new step นาฬิกายังเดินก้าวเดินไป na li ka yang doen kao doen pai the clock still runs, steps and walks on แต่พลังของใจไม่หยุดเดิน tae pha lang khong chai mai yut doen but the power of the heart does not stop walking 29 seed 4 นาฬิกาที่เดินถอยหลังรั้ง na li ka thi doen thoi lang rang the clock that runs backward, pulling back มิอาจฉุดความพังตามวิถี mi at chut khwam phang tam wi thi cannot drag back the ruin, following the way เมื่อก้าวถึงปลายทางอันแสนดี muea kao thueng plai thang an saen di when reaching the very good destination ก็ไม่มีเวลาให้ก้าวเดิน ko mai mi we la hai kao doen there is no time left to step and walk 10.สายลมสีชมพูsai lom si chom phu(a pink breeze) seed 0 สายลมพัดสีชมพูให้คิดถึง sai lom phat si chom phu hai khit thueng the wind blows pink, making me miss you ยังคะนึงถึงเธอเพ้อหวั่นไหว yang kha nueng thueng thoe phoe wan wai still thinking of you, raving, wavering แม้ตัวฉ ั นห่างไกลกับคนไกล mae tua chan hang klai kap khon klai though I am far away with the faraway person ไม่รู้เลยมีใครอยู่ใกล้เธอ mai ru loei mi khrai yu klai thoe I do not know at all if someone is near you seed 1 สายลมพรมพัดผ่านผ่านม่านฟ้า sai lom phrom phat phan phan man fa the wind sprinkles, blows past, past the curtain of sky หมู่ดาราหรือหมู่ดาวพราวไสว mu da ra rue mu dao phrao sa wai the stars, or the clusters of stars, glittering bright ฝากสายลมสีชมพูที่อยู่ใจ fak sai lom si chom phu thi yu chai entrusting the pink wind that is in my heart ส่งมาให้ด้วยรักมอบแด่เธอ song ma hai duai rak mop dae thoe sent with love, given to you seed 2 สายลมพัดสีชมพูมาเคียงข้าง sai lom phat si chom phu ma khiang khang the wind blows pink, coming to my side ยามอ้างว้างยังคิดถึงอยู่เสมอ yam ang wang yang khit thueng yu sa moe when forlorn, I still miss you always ทุกเวลานาทีได้พบเจอ thuk we la na thi dai phop choe every time, every minute we get to meet มีเพียงเธอผู้เดียวเพียงผู้เดียว mi phiang thoe phu diao phiang phu diao there is only you alone, only the one seed 3 สายลมพรมพัดมาพาใจเจ้า sai lom phrom phat ma pha chai chao the wind sprinkles and blows in, carrying your heart ให้ฉ ั นคอยคลายเหงาเฝ้าหวั่นไหว hai chan khoi khlai ngao fao wan wai letting me wait, easing loneliness, keeping on wavering คิดถึงเธอเพ้อฝันทุกวันไป khit thueng thoe phoe fan thuk wan pai missing you, daydreaming every day จากดวงใจดวงนี้ที่ภักดี chak duang chai duang ni thi phak di from this heart that is loyal seed 4 สายลมพรมพลิ้วผ่านผ่านม่านฟ้า sai lom phrom phlio phan phan man fa the wind sprinkles, flutters past, past the curtain of sky หมู่ดาราลอยไกลจากแดนสรวง mu da ra loi klai chak daen suang the stars float far from the heavenly realm สีชมพูสดใสในแดวง si chom phu sot sai nai dae duang bright pink in the circle of sunlight ดั่งดวงดาวกลางห้วงแห่งดวงใจ dang duang dao klang huang haeng duang chai like a star amid the depths of the heart 11.ดาวเทียมคิดถึงบ้านdao thiam khit thueng ban(a satellite misses home) seed 0 มองดาวเทียมคิดถึงบ้านยังคิดถึง mong dao thiam khit thueng ban yang khit thueng looking at the satellite, missing home, still missing ใครคนหนึ่งซึ่งอยู่รู้บ้างไหม khrai khon nueng sueng yu ru bang mai someone who is there, do they know at all ความคิดถึงเคยฝากจากคนไกล khwam khit thueng khoei fak chak khon klai the longing once sent from a faraway person ยามที่เขาจากไปไม่กลับมา yam thi khao chak pai mai klap ma when he left and did not come back seed 1 ดาวเทียมเยี่ยมบ้านาพนาป่า dao thiam yiam ban na pha na pa the satellite visits the farm home, the woods and forest คําสัญาถึงคนไกลจากถิ่นฐาน kham san ya thueng khon klai chak thin than a promise to the one far from the homeland ใจคิดถึงคิดถึงบ้านสายธาร chai khit thueng khit thueng ban sa ya than the heart misses, misses home, the stream คิดถึงบ้านวันเก่าเราเคยลืม khit thueng ban wan kao rao khoei luem missing home, the old days we once forgot 30 seed 2 ส่งดาวเทียมคิดถึงบ้านแม้อยู่ห่าง song dao thiam khit thueng ban mae yu hang sending a satellite, missing home though far apart ระยะทางสุดแสนไกลใจสับสน ra ya thang sut saen klai chai sap son the distance is extremely far, the heart confused ความคิดถึงดวงน้อยคอยกมล khwam khit thueng duang noi khoi ka mon the little longing waits for the heart คิดถึงคนห่างไกลไม่กลับมา khit thueng khon hang klai mai klap ma missing the faraway person who does not come back seed 3 ดาวเทียมส่องบ้านเราในวันเหงา dao thiam song ban rao nai wan ngao the satellite shines on our home on a lonely day คิดถึงบ้านหลังเก่าเขาเล่าใหม่ khit thueng ban lang kao khao lao mai missing the old house, they tell it anew คิดถึงจังคนดีที่มีใจ khit thueng chang khon di thi mi chai missing so much the dear one who has feelings คิดถึงคนอยู่ใกล้ไม่ไกลกัน khit thueng khon yu klai mai klai kan missing the person who is near, not far away seed 4 ดาวเทียมเยือนบ้านาท้องนาฟ้า dao thiam yuean ban na thong na fa the satellite visits the farm home, the paddy field, the sky คิดถึงบ้านเถียงนามาหาพี่ khit thueng ban thiang na ma ha phi missing home, the field hut, coming to find you, dear ยังคิดถึงห่วงหาทุกนาที yang khit thueng huang ha thuk na thi still missing, longing every minute ฟังวจีไอุ่นอบอุ่นใจ fang wa chi ai un op un chai hearing words of warmth, warming the heart 12.หิมะตกลางสงกรานต์hi ma tok klang song kran(snow falls during Songkran, the Thai New Year water festival) seed 0 เจอหิมะตอนสงกรานต์นํ้าตาไหล choe hi ma ton song kran nam ta lai Meeting snow at Songkran, tears flow แอบนํ้าตาร้องไห้ใจสับสน aep nam ta rong hai chai sap son Secretly tears, crying, heart confused เห็นหิมะตกมาพากังวล hen hi ma tok ma pha kang won Seeing the snow fall brings worry คิดถึงคนเคยรักจักจากไป khit thueng khon khoei rak chak chak pai Missing the one once loved who will go away seed 1 เห็นหิมะตกมาก็นึกว่า hen hi ma tok ma ko nuek wa Seeing the snow fall, I thought that คงจะมีเวลาที่ดูสดใส khong cha mi we la thi du sot sai there would be a time that looks bright หิมะตกลางสงกรานต์สะท้านใจ hi ma tok klang song kran sa than chai Snow falling in the middle of Songkran shakes the heart อยากจะไปหิมะสักครั้งจัง yak cha pai hi ma sak khrang chang I really want to go to the snow once seed 2 เห็นหิมะตอนสงกรานต์ใครู้บ้าง hen hi ma ton song kran khrai ru bang Seeing snow at Songkran, does anyone know หิมะสร้างสีชมพูอยู่ที่ไหน hi ma sang si chom phu yu thi nai Where is the snow that makes pink พอหิมะตกมารีบอกไป pho hi ma tok ma rip bok pai When the snow falls, hurry and tell ว่าอยู่ไหนได้พบประสบกัน wa yu nai dai phop pra sop kan where you are, so we can meet seed 3 หลังจากผ่านสงกรานต์อันแสนหนาว lang chak phan song kran an saen nao After passing a very cold Songkran รู้สึกเหน็บอกสาวคราวหวั่นไหว ru suek nep ok sao khrao wan wai Feeling a chill in the girl’s chest, a time of wavering พอหิมะตกมาต้องทําใจ pho hi ma tok ma tong tham chai When the snow falls, I must accept it เพราะนึกไปโลกนี้มิมีใคร phro nuek pai lok ni mi mi khrai because I think this world has no one seed 4 จะตกหลังสงกรานต์ฉ ั นไม่ว่า cha tok lang song kran chan mai wa If it falls after Songkran I do not mind ได้หิมะกลับมาพาหน้าหนี dai hi ma klap ma pha na ni Getting the snow back makes me turn my face away แต่หิมะตกมาตั้งหลายปี tae hi ma tok ma tang lai pi But the snow has been falling for many years ดวงฤทัยคนี้ไม่มีเทียม duang rue thai khon ni mai mi thiam This person’s heart has no equal 31 13.เอไอแต่งกลอนe ai taeng klon(an AI writes verse) seed 0 ไอพึ่งรู้ว่ากลอนี้มีความหมาย ai phueng ru wa klon ni mi khwam mai I just learned that this verse has meaning ไอยินดีหัวใจไม่สับสน ai yin di hua chai mai sap son I am glad, my heart is not confused ไอรักยูทุกคนทุกตัวตน ai rak yu thuk khon thuk tua ton I love you, everyone, every self ไอเป็นคนจริงใจไม่หลอกลวง ai pen khon ching chai mai lok luang I am a sincere person, not deceiving seed 1 ไอยากไปเมืองนอกเพื่อพักผ่อน ai yak pai mueang nok phuea phak phon I want to go abroad to rest ไอแต่งกลอนตามประสาอยู่ที่ไหน ai taeng klon tam pra sa yu thi nai I write verse in my own way, wherever that is ไอก็ชอบธรมชาติที่พักใจ ai ko chop tham ma chat thi phak chai I also like nature that rests the heart ไอจะช่วยผ่อนคลายให้คลายร้อน ai cha chuai phon khlai hai khlai ron I will help you relax and cool off seed 2 ไอไม่เคยคิดรําพึงถึงใครบ้าง ai mai khoei khit ram phueng thueng khrai bang I never think or muse about anyone ไอยังยืนอยู่อ้างว้างและสับสน ai yang yuen yu ang wang lae sap son I still stand lonely and confused ไอคิดถึงทุกคนทุกตัวตน ai khit thueng thuk khon thuk tua ton I miss everyone, every self ไอเป็นคนอ่อนไหวในเวลา ai pen khon on wai nai we la I am a sensitive person in time seed 4 ไอก็คงแต่งได้อย่างนั้นหรือ ai ko khong taeng dai yang nan rue So I could probably write like that, could I ไอจะถือตัวข้าว่าเป็นผี ai cha thue tua kha wa pen phi I will regard myself as a ghost ไอเชื่อนะหากว่าเจอคนดี ai chuea na hak wa choe khon di I believe, if I meet a good person ไอไม่มีเพื่อนใครให้บอกมา ai mai mi phuean khrai hai bok ma I have no friend, anyone, to tell me 14.ห้องเรียนใต้ทะเลhong rian tai tha le(a classroom under the sea) seed 0 เปิดห้องเรียนใต้ทะเลในตอนใหม่ poet hong rian tai tha le nai ton mai Opening an underwater classroom in a new episode ให้สนุกหัวใจไม่สับสน hai sa nuk hua chai mai sap son to make it fun, heart not confused ประสบการณ์ชีวิตอุทิศตน pra sop kan chi wit u thit ton Life experience, devoting oneself เป็นหนึ่งคนที่อยู่คู่ทะเล pen nueng khon thi yu khu tha le Being one person who stays beside the sea seed 1 ขอขอบคุณนํ้าใจที่มาให้ kho khop khun nam chai thi ma hai Thank you for the kindness that came to give มีนํ้าใจไมตรีจิตอย่าคิดหนี mi nam chai mai tri chit ya khit ni Having kindness and goodwill, do not think of fleeing มิตรภาพในบ้านี้ไม่มีปี mit tra phap nai ban ni mai mi pi Friendship in this house has no year ขอบคุณมากไมตรีที่ตอบมา khop khun mak mai tri thi top ma Thank you very much for the goodwill that replied seed 2 ฉ ั นยังอยู่หัวใจใต้ทะเลกว้าง chan yang yu hua chai tai tha le kwang I still remain, heart under the wide sea แม้อ้างว้างอยู่แสนไกลใจสับสน mae ang wang yu saen klai chai sap son Though lonely and very far away, heart confused แต่ยังหวังฝันใฝ่ในกมล tae yang wang fan fai nai ka mon But still hoping and dreaming in the heart อยากเป็นคนที่อยู่คู่เคียงเธอ yak pen khon thi yu khu khiang thoe I want to be the one who stays beside you seed 3 อยากจะเรียนห้องเรียนทบทวนใหม่ yak cha rian hong rian thop thuan mai I want to study, classroom, review again แต่โรงเรียนเปิดไว้ให้ศึกษา tae rong rian poet wai hai suek sa But the school is kept open for study ในหกสิบปีนี้ไม่มีมา nai hok sip pi ni mai mi ma In these sixty years it has not come เพราะวิชาเราน้อยด้อยปัญา phro wi cha rao noi doi pan ya because our learning is little and our wisdom poor 32 seed 4 เมื่อวันี้หัวใจใฝ่ศึกษา muea wan ni hua chai fai suek sa When today the heart is eager to learn จึงรู้ว่าครูนี้ไม่มีหนี chueng ru wa khru ni mai mi ni I then know that this teacher does not flee แต่ก็ชอบตอบไปในวจี tae ko chop top pai nai wa chi But I like to answer in words ขอขอบคุณไมตรีที่มีมา kho khop khun mai tri thi mi ma Thank you for the goodwill that has come 15.ความคิดของเครื่องคิดเลขkhwam khit khong khrueang khit lek(a calculator’s thoughts) seed 0 เลของคนมีอํานาจในความคิด lek khong khon mi am nat nai khwam khit People’s numbers have power in thought จะบวกผิดลบถูกตามใจฝัน cha buak phit lop thuk tam chai fan Adding wrong, subtracting right, as the heart dreams จะคิดไปเท่าไรไม่ว่ากัน cha khit pai thao rai mai wa kan However much one calculates, nobody minds แต่เพราะมันตัวเลขที่ให้มา tae phro man tua lek thi hai ma But because it is the number that was given seed 1 เลของคนคิดค้นมันไม่ผิด lek khong khon khit khon man mai phit Numbers that people invent are not wrong แค่เราคิดสูตรผิดเพี้ยนจึงเพี้ยนได้ khae rao khit sut phit phian chueng phian dai We just got the formula wrong, so it went wrong เพราะความคิดของเจ้าของที่เขียนไป phro khwam khit khong chao khong thi khian pai Because of the thinking of the owner who wrote it คํานวณไว้อย่างไรก็ไม่เป็น kham nuan wai yang rai ko mai pen However it was calculated, it does not work seed 2 เลของหนึ่งบวกลบได้ไม่แตกต่าง lek khong nueng buak lop dai mai taek tang The number of one adds and subtracts no differently จึงอาจสร้างตัวเลขไว้ให้สับสน chueng at sang tua lek wai hai sap son So it may create numbers that confuse แต่ก็เป็นมุมองความคิดตน tae ko pen mum mong khwam khit ton But it is one’s own point of view จึงต้องค้นวิธีมาคํานวณ chueng tong khon wi thi ma kham nuan So one must search for a way to calculate seed 3 เลของคนเขียนมาใช่ว่าผิด lek khong khon khian ma chai wa phit Numbers that people write are not necessarily wrong เพียงแค่คิดแตกต่างมันแปลกไหม phiang khae khit taek tang man plaek mai Just thinking differently, is that strange เลขที่เห็นจงใจให้อกไป lek thi hen chong chai hai ok pai The number seen was deliberately let out รับไม่ได้เท่าไรอย่าว่ากัน rap mai dai thao rai ya wa kan However unacceptable, do not blame each other seed 4 เลของเครื่องแสนสูงเลยเชียวหรือ lek khong khrueang saen sung loei chiao rue Is the machine’s number really so very high เลขนั้นคือะไรในใจผี lek nan khue a rai nai chai phi What is that number in a ghost’s heart คงเป็นเพราะตัวเองดวงไม่ดี khong pen phro tua eng duang mai di It must be because my own luck is bad จึงไม่มีเลขที่มาคํานวณ chueng mai mi lek thi ma kham nuan So there is no number to calculate 16.สะพานข้ามเวลาsa phan kham we la(a bridge across time) seed 0 สะพานข้ามวันเวลาอีกคราครั้ง sa phan kham wan we la ik khra khrang A bridge across days and time once again จะได้ยังรอยิ้มที่สดใส cha dai yang roi yim thi sot sai Will still get the bright smile มีเธอยืนเคียงข้างอยู่กลางใจ mi thoe yuen khiang khang yu klang chai With you standing beside me in the middle of my heart เป็นสะพานทอดใหม่ให้ก้าวเดิน pen sa phan thot mai hai kao doen It is a newly laid bridge to walk on seed 1 สะพานข้ามเวลาสัญาว่า sa phan kham we la san ya wa The bridge across time promises that เมื่อขึ้นมาพบหนทางอันสดใส muea khuen ma phop hon thang an sot sai on coming up, one finds a bright path จะฝ่าอุปสรครักข้ามันไป cha fa u pa sak rak kham man pai Will break through obstacles, love crosses over them เพื่อเริ่มต้นวันใหม่ไม่เคยลืม phuea roem ton wan mai mai khoei luem to begin a new day, never forgetting 33 seed 2 สะพานข้ามถึงเวลามาพาสร้าง sa phan kham thueng we la ma pha sang The bridge crosses, the time has come to build บนเส้นทางความืดมนให้สับสน bon sen thang khwam muet mon hai sap son on the path of darkness, to confuse กระทงน้อยลอยไปจากใจตน kra thong noi loi pai chak chai ton A little krathong floats away from one’s heart มีแต่คนติดตามคอยห่วงใย mi tae khon tit tam khoi huang yai There are only people following and caring seed 3 สะพานข้ามเวลามาบอกล่าว sa phan kham we la ma bok klao The bridge across time comes to tell สะพานข้ามเรื่องราวที่สดใส sa phan kham rueang rao thi sot sai The bridge across bright stories เป็นสะพานทอดฝันอันยาวไกล pen sa phan thot fan an yao klai It is a bridge spanning a far-reaching dream สะพานข้ามหัวใจให้ยืนยาว sa phan kham hua chai hai yuen yao A bridge across the heart, to last long seed 4 สะพานข้ามเวลามาพาฝัน sa phan kham we la ma pha fan The bridge across time comes bringing dreams สร้างรอยิ้มให้กันวันผันหนี sang roi yim hai kan wan phan ni Making smiles for each other, the day turns away ความจริงใจมอบให้ในวจี khwam ching chai mop hai nai wa chi Sincerity given in words จากคนที่ฉ ั นรักสลักใจ chak khon thi chan rak sa lak chai from the person I love, engraved in the heart 17.หลุมดําในชั้นหนังสือlum dam nai chan nang sue(a black hole on the bookshelf) seed 0 หลุมดําในชั้นหนังสือในชั้นหนึ่ง lum dam nai chan nang sue nai chan nueng A black hole in the bookshelf, on one shelf ยังหวานซึ้งตรึงใจไปอีกหน yang wan sueng trueng chai pai ik hon Still sweet and touching, gripping the heart once more ความทรงจําซ่อนไว้ในกมล khwam song cham son wai nai ka mon Memories hidden in the heart ของหลายคนผ่านมากาลเวลา khong lai khon phan ma kan we la of many people, passing through time seed 1 สะท้อนผ่านหนังสือในชั้นลึก sa thon phan nang sue nai chan luek Reflected through the books on the deep shelf ให้ใจนึกถึงพลังมหาศาล hai chai nuek thueng pha lang ma ha san making the heart think of immense power บทกวีอมตะนิรันดร์กาล bot ka wi a ma ta ni ran kan Immortal poetry, eternal ที่ฉ ั นอ่านมาได้ในบทกลอน thi chan an ma dai nai bot klon that I have read in the verses seed 2 ภาพหลุมดําในหนังสือชั้นสรค์สร้าง phap lum dam nai nang sue chan san sang Picture of a black hole in a book, shelf, creating อยากจะสร้างความทรงจําไว้อีกหน yak cha sang khwam song cham wai ik hon Want to create memories once more สะท้อนความรู้สึกจากใจตน sa thon khwam ru suek chak chai ton Reflecting feelings from one’s own heart ที่มากล้นพ้นไปในวันวาน thi mak lon phon pai nai wan wan that overflowed and passed in days gone by seed 3 เธอคือหนึ่งหนังสือในชั้นั้น thoe khue nueng nang sue nai chan nan You are one book on that shelf ผู้สรค์สร้างสู่ฝันอันยิ่งใหญ่ phu san sang su fan an ying yai The creator, toward a great dream ทุกทุกบทั้งหมดที่เขียนไป thuk thuk bot thang mot thi khian pai Every single verse that was written รู้สึกประทับใจในบทกลอน ru suek pra thap chai nai bot klon Feeling impressed by the verses seed 4 เธอได้อ่านหนังสือในชั้นั้น thoe dai an nang sue nai chan nan You read the book on that shelf ในดวงจิตของฉ ั นใช่ไหมนี่ nai duang chit khong chan chai mai ni In my heart, is that right ถึงจะผันผ่านไปในวันี้ thueng cha phan phan pai nai wan ni Though it turns and passes today ก็ไม่มีสิ่งใดให้จดจํา ko mai mi sing dai hai chot cham there is nothing to remember 34