Paper deep dive
Generating Concept Lexicalizations via Dictionary-Based Cross-Lingual Sense Projection
David Basil, Chirooth Girigowda, Bradley Hauer, Sahir Momin, Ning Shi, Grzegorz Kondrak
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 4/18/2026, 1:34:15 AM
Summary
The paper introduces ExpandNet, a dictionary-guided, sentence-level projection method for automatically expanding WordNet-style lexical resources to new languages. By augmenting a base word aligner with a bilingual dictionary (DBAlign) and applying a filtering strategy, the method improves precision in sense generation while reducing reliance on large-scale corpus statistics.
Entities (5)
Relation Signals (3)
David Basil â affiliatedwith â University of Alberta
confidence 100% · David Basil... Department of Computing Science University of Alberta
ExpandNet â uses â DBAlign
confidence 100% · Our method, ExpandNet, introduces two dictionary-based enhancements... First, it employs a novel aligner, DBAlign
DBAlign â augments â Base Aligner
confidence 90% · DBAlign augments a pre-existing base aligner with evidence from a bilingual dictionary.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We study the task of automatically expanding WordNet-style lexical resources to new languages through sense generation. We generate senses by associating target-language lemmas with existing lexical concepts via semantic projection. Given a sense-tagged English corpus and its translation, our method projects English synsets onto aligned target-language tokens and assigns the corresponding lemmas to those synsets. To generate these alignments and ensure their quality, we augment a pre-trained base aligner with a bilingual dictionary, which is also used to filter out incorrect sense projections. We evaluate the method on multiple languages, comparing it to prior methods, as well as dictionary-based and large language model baselines. Results show that the proposed project-and-filter strategy improves precision while remaining interpretable and requiring few external resources. We plan to make our code, documentation, and generated sense inventories accessible.
Tags
Links
- Source: https://arxiv.org/abs/2604.14397v1
- Canonical: https://arxiv.org/abs/2604.14397v1
Trouble viewing inline? Open PDF directly â
Full Text
15,979 characters extracted from source content.
Expand or collapse full text
Generating Concept Lexicalizations via Dictionary-Based Cross-Lingual Sense Projection David Basil, Chirooth Girigowda, Bradley Hauer, Sahir Momin, Ning Shi, Grzegorz Kondrak Alberta Machine Intelligence Institute (Amii) Department of Computing Science University of Alberta, Edmonton, Canada Abstract We study the task of automatically expanding WordNet-style lexical resources to new languages through sense generation. We generate senses by associating target-language lemmas with existing lexical concepts via semantic projection. Given a sense-tagged English corpus and its translation, our method projects English synsets onto aligned target-language tokens and assigns the corresponding lemmas to those synsets. To generate these alignments and ensure their quality, we augment a pre-trained base aligner with a bilingual dictionary, which is also used to filter out incorrect sense projections. We evaluate the method on multiple languages, comparing it to prior methods, as well as dictionary-based and large language model baselines. Results show that the proposed project-and-filter strategy improves precision while remaining interpretable and requiring few external resources. We plan to make our code, documentation, and generated sense inventories accessible. 1. Introduction A wordnet is a semantic resource that organizes words into synonym sets, or synsets, providing a structured way to represent word meaning. Each synset represents a distinct lexical concept and contains the lemmas that express it. Wordnets are widely used throughout natural language processing. In word sense disambiguation (WSD), they serve as sense inventories, enabling the mapping of words in context to structured lexical knowledge [1]. They also support semantic parsing [2], machine translation evaluation [3], and the word- in-context task [4]. Despite their importance, automating wordnet construction remains challenging, particularly in the multilingual setting, where words from different languages must be organized into a shared semantic ontology. A central component of this process is sense generation. Because a word sense is defined as a pairing between a lemma and a synset, the act of generating a sense is equivalent to adding new lemmas to existing synsets. Prior work has leveraged parallel data to generate senses. Oliver and Climent [5â7] were the first to explicitly project English sense annotations onto lexical translations, providing a way to extend wordnets. More recently, Martelli, Procopio, Barba, and Navigli [8] combined WSD, translation, and alignment on an in-house English corpus to construct multilingual semantic dictionaries. While effective, these projection-based methods typically rely on corpus-level statistics, making them sensitive to corpus size and distribution, which can limit scalability. To address this limitation, we explore a finer-grained approach, namely sentence-level projection, to improve robustness across diverse data conditions. In this paper, we propose a sentence-level method for generating senses in a target language through dictionary-guided projection. Building on the semantic projection framework, we derive these new senses from translations of a sense-annotated source text. To ensure high-quality word alignment during this process, we combine a pretrained aligner with a bilingual dictionary. Crucially, this dictionary is also used to filter semantic projections, thereby reducing spurious alignments. To our knowledge, this represents the first application of dictionary-based filtering to sense generation. By operating entirely at the sentence level, our method successfully eliminates the reliance on corpus-level statistics found in prior work. Our project-and-filter procedure rests on three assumptions about word senses and translation. First, we assume that each content word in context expresses a single lexical concept. This allows our system to eliminate competing synsets rather than retaining multiple arXiv:2604.14397v1 [cs.CL] 15 Apr 2026 Figure 1. An illustration of our project-and-filter algorithm. candidates for a single token. Second, word senses are represented in a multilingual wordnet as synset-lemma pairs, each corresponding to a single lexical concept. Synonymy therefore functions as a test of conceptual identity: lemmas that are synonymous realize the same sense. Third, when the alignment of words in a sentence and its translation is supported by lexical evidence (e.g., a bilingual dictionary or a clear cognate relationship), the alignment represents a literal translation, and the aligned words express the same concept. Thus, the target lemma can be assigned to the synset of the disambiguated source token. Figure 1 illustrates how dictionary filtering governs sense projection in our algorithm. The source sentence includes two sense-tagged words, with their corresponding translations identified through word alignment. The first aligned pair is confirmed by the bilingual dictionary, allowing the sense tag assigned to hate to be projected onto odio, creating the sense (odio,bn:00086717v). In contrast, since the second aligned pair is not confirmed by the dictionary, the translation of garden is considered non-literal, the alignment is discarded, and no sense is generated. 2. Methods In this section, we describe the two principal components of our sense generation method: the sense projection pipeline, and the dictionary-based alignment algorithm. 2.1. Sense Projection Our method builds upon the semantic projection framework, in which sense tags from a source-language corpus are transferred to a target language through word alignment. Because each sense tag corresponds to a lexical concept, projection associates target-language lemmas with existing sense tags, thereby generating new senses in the target language. This projection framework assumes a word-aligned parallel corpus with sense annotations on the source side. If a source wordxtagged with conceptsis aligned with a target wordy, then the lemma ofycan also be associated with concepts. We say that the semantic tagsis projected across the alignment link from x to y, which generates the sense(y, s). Our method,ExpandNet, introduces two dictionary-based enhancements to the projection framework. First, it employs a novel aligner,DBAlign, which incorporates evidence from a bilingual dictionary (Section 2.2). Second, it applies a filtering step that discards projections lacking dictionary support. The full procedure is summarized in Figure 3. The filtering step uses a bilingual dictionary to validate each aligned pair. LetDbe a bilingual dictionary consisting of translation pairs(x, y), wherexandyare words or phrases in the source and target languages that serve as valid translations of one another in at least one context. If a sense-tagged wordxis aligned with a target wordy, we project the sense Algorithm DBAlign(S, T, B, D) Aââ , |S|â length of S, |T|â length of T for passâ [Intersection, Dictionary, Base] do if pass = Intersection then C â Bâ© D else if pass = Dictionary then C â D else C â B while true do V â links in C not incident on tokens already linked in A if V =â then break (i, j)â DiagonalHeuristic(V,|S|,|T|) Aâ AâȘ(i, j) return A Figure 2. Our dictionary-based alignment procedure, DBAlign.Bdenotes links proposed by the base aligner,Ddenotes links between synonyms in the bilingual dictionary.S represents the source sentence, andTthe target sentence. Links are added in three passes, subject to a non-conflict constraint. When multiple candidate links are available, the diagonal heuristic selects the link closest to monotonic word order. only if(x, y)â D. Otherwise, the link is discarded and no sense is generated. However, to facilitate the generation of senses involving rare words and proper nouns, we always project a sense when x and y share the same orthographic form. We optionally apply a part-of-speech (POS) filter, which discards projections when the source and target words do not share the same basic POS (noun, verb, adjective or adverb). Words falling outside these categories are treated as function words and excluded from projection. We expect our method to be particularly useful in settings where comprehensive wordnets or sense inventories are not yet available, including many low-resource languages. While the semantic projection framework inherently assumes access to a bitext, we generate this automatically by translating a sense-annotated source corpus into the target language. Our pipeline relies on machine translation, a bilingual dictionary (for both alignment and filtering), and an optional POS tagger. Critically, these resources are generally more accessible than sense inventories in many languages. Algorithm ExpandNet(Bitext) for (sentence S, translation T)â Bitext do for (content word w)â S do sâ Synset(S, w) w y â DBAlign(S, T, w) if (w, w y )â Dict ⧠w y Ìžâ s then s.add(w y ) Figure 3. Our algorithmic wordnet expansion algorithm,ExpandNet, whereDictis a set of pairs(x, y) such thatyin the target language can express the same meaning asxin the source language in some context,DBAlign(S, T, w)returns the word in sentenceTaligned towas found by the alignment algorithm described in Section 2.2, andSynset(S, x): returns the ID of the synset corresponding to the concept expressed by x in S. 2.2. Dictionary-Based Alignment Dictionary-Based Align (DBAlign) augments a pre-existing base aligner with evidence from a bilingual dictionary. This integration improves the robustness of the base alignerâs output by prioritizing dictionary-supported alignments and explicitly handling multi-word expressions (MWEs). At a high level,DBAlignchooses a set of alignment links through a three-pass procedure. First, it accepts the intersection of links proposed by both the dictionary and the base aligner. We treat these alignments as high-confidence, as they are supported by both sources. Next, it incorporates any remaining dictionary-suggested links that do not conflict with previously accepted alignments. Finally, it adds remaining base-aligner links that do not conflict with the previously accepted set. This ordering prioritizes high-confidence alignments, before increasing coverage while avoiding conflicts. Throughout this process, two links are considered to conflict if they share a source or target token. During each pass, for every token on the source side, the target side is scanned, and candidates which satisfy the condition of that pass are found. For example, imagine we are on the first pass, and thus are searching for an alignment supported by both the base aligner and the dictionary. We wish to word-align the sentence âthat man is my husbandâ with âese hombre es mi esposoâ. The base aligner links man with both hombre and esposo, but man and hombre are translations according to the dictionary, while man and esposo are not. Then, man-hombre can be said to satisfy the pass condition and will thus be added to the final set of alignment links. In future steps, any alignments involving man or hombre will conflict with this link and thus be discarded. If two alignments for a word are both valid under the condition of the current pass, we employ the diagonal heuristic, which is adapted from the fast_align system [9]. The alignment links are selected so that the source and target words are as close as possible in terms of their relative positions in their respective sentences. Suppose we wish to word-align the sentences âhis front teeth were missingâ and âle faltaban los dientes frontalesâ and the base aligner links teeth to both dientes and frontales. While both of these alignment links meet the criterion for the current pass, it is the teeth-dientes link that will be chosen, since teeth is relatively closer to the location of dientes than frontales. To put it formally, theDBAlignalgorithm proceeds as follows: LetBbe the set of alignment links provided by the base aligner. That is, the set of all pairs of indices(x, y) such that the base aligner links the token at indexxin the source sentence to the token at indexyin the target sentence. LetDbe the set of all synonymous pairs suggested by the dictionary, the set of all pairs of indices(x, y) such that, according to the dictionary, the token at indexxin the source sentence is a translation of the token at indexyin the target sentence. This can hold true for MWEs, provided they are present in the dictionary. For example, according to our English-Spanish dictionary, derrochar is a translation of the English MWE fritter away. all MWEs are tokenized as a single unit. This means that DBAlignneeds only consider individual tokens. Then, the final set of alignment linksAis constructed from these sets using the three-pass algorithm described in detail in Figure 2. Acknowledgements We thank Sevryn Robinson and Junhyeon Cho for their assistance in refining our algorithm and applying it to Korean. This research was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) and the Alberta Machine Intelligence Institute (Amii). References [1]T. Pasini, A. Raganato, and R. Navigli. âXL-WSD: An extra-large and cross-lingual evaluation framework for word sense disambiguationâ. In: Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 35. 2021, p. 13648â13656. [2]R. Navigli, R. Blloshmi, and A. C. MartĂnez Lorenzo. âBabelNet Meaning Representation: A Fully Semantic Formalism to Overcome Language Barriersâ. In: Proceedings of the AAAI Conference on Artificial Intelligence 36.11 (June 2022), p. 12274â12279. [3] N. Campolungo, F. Martelli, F. Saina, and R. Navigli. âDiBiMT: A Novel Benchmark for Measuring Word Sense Disambiguation Biases in Machine Translationâ. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Ed. by S. Muresan, P. Nakov, and A. Villavicencio. Dublin, Ireland: Association for Computational Linguistics, May 2022, p. 4331â4352. [4]M. T. Pilehvar and J. Camacho-Collados. âWiC: the Word-in-Context Dataset for Evaluating Context-Sensitive Meaning Representationsâ. In: Proceedings of NAACL-HLT. 2019, p. 1267â 1273. [5]A. Oliver and S. Climent. âBuilding wordnets by machine translation of sense tagged corporaâ. In: Proceedings of the Global WordNet Conference 2012. Matsue, JapĂł 9-13 gener de 2012. 2012. [6]A. Oliver and S. Climent. âParallel corpora for wordnet construction: Machine translation vs. automatic sense taggingâ. In: International Conference on Intelligent Text Processing and Computational Linguistics. Springer. 2012, p. 110â121. [7]A. Oliver and S. Climent. âAutomatic creation of WordNets from parallel corporaâ. In: Proceed- ings of the Ninth International Conference on Language Resources and Evaluation (LRECâ14). Ed. by N. Calzolari, K. Choukri, T. Declerck, H. Loftsson, B. Maegaard, J. Mariani, A. Moreno, J. Odijk, and S. Piperidis. Reykjavik, Iceland: European Language Resources Association (ELRA), May 2014, p. 1112â1116. [8] F. Martelli, L. Procopio, E. Barba, and R. Navigli. âLexicoMatic: Automatic Creation of Multilingual Lexical-Semantic Dictionariesâ. In: Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). Ed. by J. C. Park, Y. Arase, B. Hu, W. Lu, D. Wijaya, A. Purwarianti, and A. A. Krisnadhi. Nusa Dua, Bali: Association for Computational Linguistics, Nov. 2023, p. 820â833. [9]C. Dyer, V. Chahuneau, and N. A. Smith. âA Simple, Fast, and Effective Reparameterization of IBM Model 2â. In: Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Ed. by L. Vanderwende, H. DaumĂ© I, and K. Kirchhoff. Atlanta, Georgia: Association for Computational Linguistics, June 2013, p. 644â648.