Paper deep dive
GameCommBench: A Unified Benchmark and Type-Aware Evaluation for AI-Generated Game Commentary
Qirui Zheng, Zhengteng Lin, Yunyi Xiao, Junhao Li, Keyuan Cheng, Xingbo Wang, Yongyi Wang, Lingfeng Li, Yunlong Lu, Wenxin Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/9/2026, 8:19:45 AM
Summary
The paper introduces GameCommBench, a unified benchmark for AI-Generated Game Commentary (AI-GGC) spanning board games, sports, and esports, and proposes Type-Aware Commentary Evaluation (TACE). TACE evaluates commentary based on three functional types: descriptive, analytical, and background, using specific metrics for each (correctness/completeness, faithfulness/depth, factuality/relevance) rather than holistic scores. The study validates TACE's reliability and human agreement, then benchmarks six VLMs, revealing that live observation and strategic analysis are major bottlenecks for current AI commentators.
Entities (22)
Relation Signals (18)
GameCommBench â covers â Go
confidence 95% ¡ GAMECOMMBENCH covers five representative games across three domains: chess and Go for board games
GameCommBench â covers â Chess
confidence 95% ¡ GAMECOMMBENCH covers five representative games across three domains: chess and Go for board games
GameCommBench â covers â Soccer
confidence 95% ¡ GAMECOMMBENCH covers five representative games across three domains: ... soccer and basketball for sports
GameCommBench â covers â Basketball
confidence 95% ¡ GAMECOMMBENCH covers five representative games across three domains: ... soccer and basketball for sports
GameCommBench â covers â League of Legends
confidence 95% ¡ GAMECOMMBENCH covers five representative games across three domains: ... and LoL for esports.
TACE â evaluates â Descriptive Commentary
confidence 92% ¡ TACE evaluates descriptive, analytical, and background commentary with criteria tailored to each type.
TACE â evaluates â Analytical Commentary
confidence 92% ¡ TACE evaluates descriptive, analytical, and background commentary with criteria tailored to each type.
TACE â evaluates â Background Commentary
confidence 92% ¡ TACE evaluates descriptive, analytical, and background commentary with criteria tailored to each type.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Game commentary is an open-ended generation task requiring multimodal perception, strategic reasoning, and contextual knowledge. Existing AI-Generated Game Commentary (AI-GGC) studies remain fragmented across games, modalities, and evaluation protocols, while overlap-based or holistic evaluators fail to capture the functional heterogeneity of commentary. We introduce \textsc{GameCommBench}, a unified benchmark spanning board games, sports, and esports, with commentary aligned to heterogeneous game contexts and annotated by commentary type. We further propose Type-Aware Commentary Evaluation (TACE), a structured framework for evaluating different types of commentary. We then validate TACE for reliability and human agreement, and use it to benchmark representative AI commentators. Results reveal non-uniform capability profiles, with live observation and strategic analysis emerging as major bottlenecks. Together, \textsc{GameCommBench} and TACE provide a diagnostic foundation for comparable and interpretable AI-GGC evaluation.
Tags
Links
- Source: https://arxiv.org/abs/2610.11129v1
- Canonical: https://arxiv.org/abs/2610.11129v1
Trouble viewing inline? Open PDF directly â
Full Text
106,442 characters extracted from source content.
Expand or collapse full text
GAMECOMMBENCH: A Unified Benchmark and Type-Aware Evaluation for AI-Generated Game Commentary Qirui Zheng 1 , Zhengteng Lin 2 , Yunyi Xiao 2 , Junhao Li 2 , Keyuan Cheng 1 , Xingbo Wang 1 , Yongyi Wang 1 , Lingfeng Li 1 , Yunlong Lu 1 , Wenxin Li 1,B 1 Peking University 2 South China University of Technology zzzqr@stu.pku.edu.cnlwx@pku.edu.cn Abstract Game commentary is an open-ended generation task requiring multimodal perception, strate- gic reasoning, and contextual knowledge. Ex- isting AI-Generated Game Commentary (AI- GGC) studies remain fragmented across games, modalities, and evaluation protocols, while overlap-based or holistic evaluators fail to cap- ture the functional heterogeneity of commen- tary. We introduce GameCommBench, a uni- fied benchmark spanning board games, sports, and esports, with commentary aligned to hetero- geneous game contexts and annotated by com- mentary type. We further propose Type-Aware Commentary Evaluation (TACE), a structured framework for evaluating different types of commentary. We then validate TACE for re- liability and human agreement, and use it to benchmark representative AI commenta- tors. Results reveal non-uniform capability profiles, with live observation and strategic analysis emerging as major bottlenecks. To- gether, GameCommBench and TACE provide a diagnostic foundation for comparable and in- terpretable AI-GGC evaluation. 1 Introduction Game commentary helps audiences make sense of gameplay in real time by narrating key events, inter- preting important decisions, and situating moments in broader context (Behrens and Uhrich, 2022). AI-Generated Game Commentary (AI-GGC) has therefore become an increasingly active research area with clear practical value. Beyond utility, AI- GGC is a challenging open-ended NLG task: high- quality commentary requires live observation of unfolding events, strategic analysis of decisions and consequences, and historical recall of relevant game context (Zheng et al., 2025). Despite growing interest, existing AI-GGC resources remain fragmented across individual Code and Data are available here. games, including soccer (Giancola et al., 2018; Deliege et al., 2021), basketball (Yu et al., 2018; Yan et al., 2019), chess (Jhamtani et al., 2018; Lee et al., 2022), Go (Tomlin et al., 2022), and League of Legends(LoL) (Zhang et al., 2024; Wang and Yoshinaga, 2025). These resources have supported progress within their domains, but they remain tied to specific games, input modalities, data formats, and comment focuses. As a result, the field lacks a unified benchmark that organizes heterogeneous game contexts under a shared interface, making systematic cross-setting comparison difficult. Evaluating game commentary poses an even deeper challenge. Many AI-GGC studies still rely on n-gram overlap metrics (Zheng et al., 2025), such as BLEU, ROUGE, or METEOR, despite the open-ended nature of commentary: for same game state, different commentators may produce distinct yet valid outputs. Such metrics therefore penalize variation rather than measuring commentary qual- ity. Recent LLM-based evaluators move beyond lexical overlap (Liu et al., 2023), but existing ap- proaches still score commentary holistically (Kim et al., 2025). This is insufficient because commen- tary serves distinct functions, including describ- ing game states, interpreting strategic intent, and providing broader historical context (Zheng et al., 2025). Each function requires different criteria for judging quality, so evaluation should target these functions separately rather than collapse them into a single score. AI-GGC further complicates evalu- ation because a given game state admits neither a unique correct commentary nor a single correct rea- soning process, especially for the analytical part. To address these gaps, we present GAMECOMM- BENCH and Type-Aware Commentary Evaluation (TACE). GAMECOMMBENCH brings board games, sports, and esports into a unified benchmark, while TACE evaluates different functional types of com- mentary with type-specific evaluation metrics. This design enables systematic comparison of AI com- arXiv:2610.11129v1 [cs.AI] 8 Oct 2026 A unified benchmark is essential for fair cross-setting comparison A Type-Aware Commentary Evaluation Framework Different commentary types require different evaluative priorities 1. Fragmented AI-GGC Landscape 2. Commentary Is Functionally Heterogeneous 3. GameCommBench + TACE Figure 1: Existing AI-GGC research is fragmented across domains and modalities, and lacks effective evaluation schemes. We address these issues with GAMECOMMBENCH, a unified cross-domain benchmark, and TACE, a commentary evaluation framework designed for descriptive, analytical, and background commentary. mentators across games, and supports finer-grained diagnosis of model capabilities. Our main contributions are as follows: (1) We construct GAMECOMMBENCH, a unified AI-GGC benchmark with heterogeneous game contexts and type-annotated commentary. (2) We propose TACE that assesses descriptive, analytical, and back- ground commentary with dedicated criteria in- stead of a single holistic score. (3) We validate TACE through repeated-run reliability analysis, hu- man validation of LLM-mediated components, and cross-judge consistency checks. (4) We benchmark representative AI commentators and reveal non- uniform capability profiles, with live observation and strategic analysis remaining key bottlenecks. 2 Benchmark Construction We construct GAMECOMMBENCH as a unified AI- GGC benchmark spanning board games, sports, and esports. Figure 2 summarizes its detailed game coverage and data composition. 2.1 Benchmark Scope and Domain Coverage GAMECOMMBENCH covers five representative games across three domains: chess and Go for board games, soccer and basketball for sports, and LoL for esports. These games are selected not sim- ply to broaden task coverage, but to expose varia- tion in game-state representation, game dynamics, and demands on commentator abilities. Board games provide relatively structured sym- bolic states but place strong demands on long- horizon strategic analysis (Lee et al., 2022). Com- mentary in board games typically emphasizes eval- uating move quality, analysing strategy, and predict- ing possible future moves. Sports games, by con- trast, are grounded in dynamic visual scenes, requir- ing commentators to identify salient events, track fast-changing actions, and produce coherent narra- tion under temporal constraints (You et al., 2025). Esports introduce an additional layer of complexity, combining visually crowded, fast-changing game- play with rich API-based information. This places higher demands on commentators to integrate com- plex multimodal inputs for game understanding and commentary generation (Zheng et al., 2025). 2.2 Data Collection and Processing Pipeline Data Collection and Alignment. We collect aligned game contexts and commentary for the five games in GAMECOMMBENCH from domain- specific sources. For chess and Go, we obtain game records and commentary from GameKnot and Foxwq, where comments are naturally aligned with turns. For basketball and soccer, we use broad- cast videos and action labels from high-quality datasets, NSVA (Wu et al., 2022) for basketball and SoccerReplay-1988 (Rao et al., 2025) for soc- cer, and transcribe spoken commentary with Whis- perX (Bain et al., 2023). For LoL, we collect of- ficial video clips paired with timestamped textual commentary; the timestamps are used to retrieve fine-grained API snapshots, yielding player-level and team-level game-state information. We also collect player and team background information from the same channels as supplementary context for commentary generation. Data Cleaning and Quality Control. We com- bine manual screening and filtering with alignment- level repair. At the match level, we prioritize Figure 2: Overview of GAMECOMMBENCH. The figure summarizes benchmark coverage, representative state- commentary examples, and commentary-type distributions across games. The central wheel reports the proportions of descriptive, analytical, and background commentary for each game. Scatter plots visualize classified human commentary embeddings, with red, blue, and green denoting descriptive, analytical, and background commentary. games with dense commentary, competitive impor- tance, and broad representativeness, while main- taining team and player diversity to avoid over- concentration on recurring entities. At the align- ment level, board-game commentary naturally maps to discrete game states, whereas sports and esports rely on timestamp-based alignment and are more susceptible to sentence-boundary errors. To address this, we apply an LLM-based post- processing step that merges an incomplete first ut- terance in a segment with the previous segment. We further remove only content that is completely unrelated to the game itself, such as off-topic chat- ter or broadcast-side digressions. The original raw commentary is retained in each benchmark record for traceability and future research use. Semantic Annotation. Since game commen- tary often mixes multiple functions, we decompose each commentary into finer semantic units and la- bel them as descriptive, analytical, or background. This step is performed by LLM, and its agreement with human annotation is reported in Table 5. This unit-level labeling avoids forcing an entire utter- ance into a single type and allows TACE to evaluate each functional part separately. Unified Packaging. Finally, all retained in- stances are packaged into a unified record schema. Each record contains raw and processed commen- tary aligned with game state, type-labeled semantic units, and game-related information. This schema keeps game-specific inputs such as symbolic states, videos, and API snapshots under a shared interface for generation and comparison. 2.3 Data Composition and Distribution Figure 2 summarizes the benchmark composition across games. Since alignment units differ across domains, we report commentary-bearing turns for board games and aligned video clips for sports and esports. Specifically, the chess and Go subsets con- tain 2,159 and 2,857 turns, with 3,405 and 5,811 commentary sentences, respectively. In the sports domain, the soccer and basketball subsets contain 2,681 and 2,981 aligned video clips, paired with 5,893 and 6,120 sentences, respectively. For es- ports, the LoL subset provides 7,675 aligned clips and 10,001 sentences. Detailed commentary-type proportions for each game are shown in Figure 2. Overall, GAMECOMMBENCH exhibits substan- tial cross-game variation in scale, modality, and commentary composition. Board game commen- tary is more concentrated in analysing, sports com- mentary is predominantly descriptive, and esports commentary shows a more balanced distribution. 3 Type-Aware Commentary Evaluation Game commentary is open-ended and contains dif- ferent functional types, so a single holistic score is insufficient. We propose TACE, which evaluates de- scriptive, analytical, and background commentary with criteria tailored to each type. Figure 3 summa- rizes the framework: commentary is decomposed by type and evaluated through three branches. We next describe the three pipelines and their metrics. 3.1 Descriptive Commentary Evaluation Descriptive commentary focuses on what is hap- pening. Therefore, instead of comparing surface text with human references, we measure whether the generated commentary accurately describes and covers salient events in the current game context. For each instance, we construct a reference event setE ref by extracting key actions from the de- scriptive units of the human commentary. To im- prove extraction accuracy, we incorporate domain- specific signals: dataset event labels for sports, ex- plicit game-state representations for board games, and significant API changes for LoL. We use GPT- 5.4 to extract reference events, GPT-5.4-mini to extract events from generated commentary, and GLM-4.7 to semantically align generated events with the reference event set. This extraction and matching pipeline is validated through human an- notation in Section 4.2. Given the aligned event sets, we define two com- plementary metrics. Correctness, computed as |E gen ⊠E ref |/|E gen |, measures the proportion of generated events supported by the reference set. Completeness, computed as|E gen ⊠E ref |/|E ref |, measures the fraction of reference events covered by the generated commentary. Correctness penal- izes unsupported event descriptions, while com- pleteness penalizes missing salient events. 3.2 Analytical Commentary Evaluation Analytical commentary interprets move quality, state significance, or future trajectories, and its eval- uation should prioritize the correctness of the rea- soning rather than surface fluency. Open-ended rea- soning evaluation is often simplified by turning the output into option selection, or by choosing tasks with a unique answer to limit open-endedness (Wen et al., 2025; Yao et al., 2025; Zhang et al., 2025). This setting does not fit AI-GGC: one game state may support multiple valid commentaries, and es- pecially for analytical commentary, there is neither a single correct answer nor a single correct reason- ing process. At the same time, evaluating lengthy analytical commentary holistically is also unreli- able, since intermediate steps in a reasoning chain may be partially correct, unsupported, or internally inconsistent (Lightman et al., 2023). We therefore need to verify whether each piece of articulated reasoning in the commentary is jus- tified under the current game state. To achieve this, we decompose analytical text into granular, explicitly verifiable units. Specifically, each an- alytical commentary segment is converted into a directed graph, where nodes are reasoning units, the smallest logically meaningful propositions in the commentary, and edges denote predefined logi- cal relations expressed in the text. We first decom- pose each analytical sentence into reasoning units, and then construct relations among them in two stages. Intra-sentence relation modeling consid- ers candidate reasoning-unit pairs within the same sentence and determines whether an explicit logical relation is present. Inter-sentence relation mod- eling extends this process across sentences within the same analytical segment: we first retrieve a restricted set of candidate sentences using embed- ding and keyword similarity, and then construct relations across the paired sentences. Importantly, graph construction is purely text-based: no game- state or auxiliary information is provided. The resulting graph therefore represents the reasoning structure verbalized by the commentary itself. We then perform game-state-grounded verification on the constructed graph. At the node level, we assess whether each reasoning unit is supported by the cor- responding game state. At the edge level, we assess whether the logical relation between two reasoning units is justified under the same game state. During verification, an edge is considered valid only when both endpoint reasoning units are supported and the expressed relation is justified. This decomposi- tion mitigates the difficulty of LLM-as-a-Judge (Gu et al., 2024), enabling fine-grained assessment of strategic reasoning correctness instead of relying on holistic plausibility. Both relation modeling and graph verification are implemented using GPT-5.4- mini, and their accuracy is validated in Section 4.2. Accordingly, we define two metrics for analyt- ical commentary. Analytical Faithfulness mea- sures the fraction of reasoning units and relations that pass game-state-grounded verification. Rea- soning Depth measures the longest verified rea- soning chain in the constructed graph, normalized by the largest value achieved by any evaluated model on the same instance. Together, they capture whether analytical reasoning is grounded and how far valid reasoning develops. ... Human Commentary Additional Information Draymond Green sets a screen Andre Iguodala passes the ball Stephen Curry shoots a three! Got it! ... .... Draymond Green sets a screen Klay Thompson passes the ball Stephen Curry hits a three- pointer ... .... Generated Commentary Semantics-based Event Matching â â Ă 1 Reasoning Unit Decomposition 2 Intra-Sentence Relation Modeling Sentence 1: ...... Sentence 2: ...... u1u2 u3 v1v2 u1 u2 u3 v1 v2 R2 R1 S1 S2 3 Inter-Sentence Relation Modeling 4 Game State Grounded Verification u1 u2 u3 v1 v2 R2 R1 R1 R3 R3 Candidate Sentence Retrieval u1 u2 u3 v1 v2 R2 R1 R1 R3 R3 Node Verification + Edge Verification 1 Background Claim Decomposition 1 2 3 Sentence 1: ...... c1 2 3 Descriptive Relevance Comparison Analytical Correctness â Completeness â Analytical Faithfulness â Reasoning Depth â Factuality â Relevance â Metrics Type-Aware Decomposition of Generated Commentary Descriptive Commentary Analytical Commentary Background Commentary + c2 c3 c4 Factuality Verification c1 Knowledge Base Hybrid Retrieval Evidence 1:... Evidence 2:... Evidence 3:... Is it True? c1 Is it Relevant? ... Event Extraction Event Extraction Figure 3: Overview of TACE. TACE evaluates descriptive, analytical, and background commentary through three type-specific pipelines. The resulting metrics assess the description correctness and completeness, reasoning faithfulness and depth, and background factuality and relevance. 3.3 Background Commentary Evaluation Background commentary provides contextual or historical information beyond immediate game events. We evaluate it along two dimensions: Fac- tuality and Relevance. For granular evaluation, we first decompose each background segment into background claims, defined as factual assertions distilled from the commentary. Factuality assesses the objective accuracy of each background claim. Rather than relying solely on the evaluatorâs parametric knowledge, we sup- port this judgment with an offline structured knowl- edge base constructed through iterative retrieval. This knowledge base provides external evidence for the relevant game, team, player, or match context. For each claim, we retrieve an evidence set using a hybrid strategy that combines indexed retrieval over structured fields, such as matches, teams, and players, with semantic retrieval based on claimâ evidence similarity. The retrieved evidence is then provided to the evaluator for evidence-grounded factuality judgment, and the final factuality score is computed by averaging claim-level judgments over the background segment. Relevance evaluates whether a background claim aligns with the imme- diate context of the commentary. Specifically, we compare each claim against the game state and the remaining portions of the segment, including de- scriptive and analytical components, to determine whether the information is contextually integrated or unrelated. The relevance score is similarly de- rived by averaging claim-level judgments. Appen- dices C.4 and F.3.4 provide the detailed rubrics and evidence retrieval protocols. 4 Validation of TACE We validate TACE along two axes: reliability and human agreement. Reliability measures run-to-run score variation and ranking consistency across re- peated evaluations, while human validation checks whether the intermediate LLM-mediated decisions align with human annotations. 4.1 Reliability of TACE To assess the reliability of TACE, we repeat the full evaluation pipeline three times on a held-out subset covering 6 models, 5 games, and 20 ran- domly sampled cases for each model-game pair. For each of the six TACE metrics, we report the mean and standard deviation across repeated runs, together with the coefficient of variation (CV) to quantify relative score fluctuation. Since bench- mark conclusions also depend on model ordering, we further compute pairwise Spearman rank corre- lations across runs to assess ranking stability. As shown in Table 1a, TACE exhibits low run- to-run variation across all six metrics. The CV values remain below 3.5%, indicating strong score- level reproducibility. Descriptive metrics show particularly small variation, while analytical met- rics exhibit slightly higher but still limited fluctua- tion, as expected given the additional complexity of reasoning-graph construction and verification. Ranking consistency is also high, with Spearman correlations ranging from 0.856 to 0.965 across metrics. These results suggest that TACE provides stable type-specific evaluation signals for model comparison. We additionally report Spearman rank Table 1: Validation of TACE. We report run-to-run reliability and human validation for LLM-mediated components. (a) Reliability TypeMetricMeanÂą SDCV(%) SpearmanĎ Desc. Correctness0.2312Âą0.00251.080.918 Completeness0.3215Âą0.00260.810.965 Analyt. Faithfulness0.3605Âą0.01123.110.905 Reasoning Depth 0.0465Âą0.00163.440.856 Back. Factuality0.7358Âą0.01321.790.918 Relevance0.9185Âą0.02022.200.868 (b) Component accuracy TypeValidated StepAcc.(%) Desc. Event Extraction91.16 Event Matching94.73 Analyt. Relation Modeling93.40 Relation Verification81.43 Back. Factuality Judgment92.86 Relevance Judgment94.44 correlations in Appendix E, where GPT-5.4-mini is replaced by Gemini-3-Flash as the judge model. 4.2 Human Validation To assess the validity of TACE, we conduct hu- man validation on the LLM-mediated components used in its evaluation pipeline. We first validate the accuracy of LLM-based type-aware decompo- sition. For each game, we randomly sample 100 decomposed commentary and manually label their types. The results, reported in Table 5, show con- sistently high decomposition accuracy across all games, with overall accuracy exceeding 95%. Then, we validate whether each LLM-mediated step in TACE pipeline are consistent with human annotations. We use a stratified random sample covering all 6 evaluated models and 5 games, with three instances sampled for each model-game pair. This results in 90 annotated instances and more than 3,700 intermediate validation items, including event extraction outputs, event-matching decisions, reasoning-graph relations, verification judgments, and background-claim judgments. For descriptive commentary, we validate event extraction and se- mantic event matching. For analytical commen- tary, we validate whether the constructed relations accurately reflect the logical structure expressed in the text, and whether relation verification cor- rectly judges the validity of reasoning connections with respect to the game state. For background commentary, we assess the accuracy of factuality and relevance judgments for extracted background claims. As shown in Table 1b, the LLM-mediated components achieve high agreement with human annotations across all components, with most vali- dated steps exceeding 90% accuracy. These results support the validity of the intermediate automatic decisions used by TACE. 5 Experiments Based on GAMECOMMBENCH and TACE, we evaluate representative VLM commentators across games and commentary types, then analyze capa- bility imbalances and the effects of model scale and reasoning mode. 5.1 Evaluation Setup We evaluate six representative state-of-the-art VLMs: three closed-source models, Gemini-3.1- Pro-Preview, GPT-5.4, and Claude-Sonnet-4.5, and three open-weight models, Qwen3-VL-32B, GLM- 4.6V, and InternVL3.5-38B. All models are evalu- ated on the full GAMECOMMBENCH benchmark. Each instance includes basic player/team and game information. For board games, models receive symbolic game states with local move history, in- cluding the current move and the preceding four moves, corresponding to two full turns. For sports and esports, models process video inputs sampled at 4 FPS while preserving the original resolution; the prompt also specifies clip duration and asks the model to generate commentary appropriate for that duration. For esports, we additionally pro- vide timestamp-aligned game-state metadata de- rived from the game API. Human commentary is not provided during generation. All models use a shared commentary-generation template, with minor modality-specific adaptations for symbolic states, videos, and API-based meta- data. We set temperature to 0 and the maximum output length to 1024 tokens, with one response generated per instance. To ensure comparable in- ference conditions, explicit reasoning modes are disabled when supported; otherwise, we use the minimal reasoning setting. Other decoding param- eters are left at their default values. We also evaluate two existing game-specific AI- GGC systems with TACE: Concept-guided Chess Commentary generation (C) (Kim et al., 2025) for chess and MatchVision (Rao et al., 2025) for soccer. Their results are included in the heatmaps in Figure 4, enabling comparison between general- purpose VLM commentators and task-specific AI- GGC systems. Model DescriptiveAnalyticalBackground CorrectnessCompletenessFaithfulnessDepthFactualityRelevance Gemini-3.1-Pro-Preview0.2060.4990.2580.1710.6910.865 GPT-5.40.2810.6100.4390.3180.6550.922 Claude-Sonnet-4.50.2430.5330.3000.1960.6610.911 Qwen3-VL-32B0.2170.4670.1630.1970.6710.927 GLM-4.6V0.2200.2820.1900.1850.6820.927 InternVL3.5-38B0.1960.4450.1240.1140.7340.866 Average0.2270.4730.2460.1960.6830.903 Gemini GPT Claude Qwen3-VL GLM InternVL C Correctness Completeness Faithfulness Depth Factuality Relevance 0.390.430.340.410.430.530.11 0.810.930.920.850.870.720.13 0.300.500.390.190.350.200.60 0.320.610.430.330.360.230.13 0.820.340.590.650.410.770.14 0.670.950.820.870.890.630.57 Chess Gemini GPT Claude Qwen3-VL GLM InternVL 0.340.330.270.400.380.22 0.850.960.960.850.840.32 0.300.550.210.120.220.15 0.230.670.290.180.250.16 0.660.730.470.410.580.80 0.940.860.910.870.900.81 Go Gemini GPT Claude Qwen3-VL GLM InternVL MatchVision 0.220.240.180.140.120.110.46 0.360.440.330.280.240.160.17 0.300.250.400.240.220.160.60 0.290.210.350.320.220.180.05 0.680.640.680.730.750.550.00 0.930.970.950.980.970.940.00 Soccer Gemini GPT Claude Qwen3-VL GLM InternVL 0.200.280.130.070.060.06 0.340.510.250.150.130.09 0.290.440.290.150.100.07 0.170.330.130.130.130.06 0.610.700.720.810.820.70 0.930.970.950.960.950.96 Basketball Gemini GPT Claude Qwen3-VL GLM InternVL 0.060.100.100.080.070.06 0.140.220.200.200.150.12 0.140.250.250.150.120.06 0.070.150.080.150.120.07 0.690.860.860.770.850.86 0.820.950.910.960.930.96 League of Legends Figure 4: Main benchmark results on GAMECOMMBENCH. The top table reports aggregated model scores for the six TACE metrics, with the average row summarizing metric-wise performance across evaluated models. The bottom heatmaps show game-wise capability profiles across models and metrics, revealing cross-game and type-specific performance patterns. 5.2 Main Benchmark Results Figure 4 summarizes the overall and game-wise performance of the six VLM commentators. The re- sults reveal substantial capability imbalance across both games and commentary types. No model con- sistently dominates all metrics: GPT-5.4 performs best on descriptive and analytical metrics, whereas other models remain competitive on background factuality or relevance. This confirms that AI-GGC capability is not one-dimensional and should not be collapsed into a single holistic score. For descriptive commentary, models show a clear gap between completeness and correctness. The average completeness reaches 0.473, while correctness is only 0.227, suggesting that VLMs often generate more event descriptions than can be directly supported by the reference signals. This pattern indicates a verbosityâgrounding mismatch: models often describe plausible events, but many of these descriptions are not supported by the ref- erence events or available game-state signals. In board games, this issue is partly mitigated by the compact symbolic input, and lower correctness may also reflect the fact that human commenta- tors do not mention every objectively valid move- level event. However, the same pattern becomes much more problematic in video-based games. For sports and esports, both correctness and complete- ness drop sharply, as shown in Figure 4, indicating that current VLMs not only miss salient events but also generate unsupported descriptions in fast- paced multi-agent visual scenes. This exposes live observation as a major bottleneck for AI-GGC in dynamic video settings. Analytical commentary remains a major bottle- neck. The average analytical faithfulness is 0.246 and the average reasoning depth is 0.196, showing that most models have difficulty producing game- grounded strategic reasoning. GPT-5.4 achieves the best analytical performance, but the absolute scores remain limited. This indicates that even strong VLMs often produce analyses that are shal- low, partially unsupported, or weakly connected to underlying game state. For reference, human ana- lytical commentary shows high faithfulness (0.78) but relatively low depth (0.28), suggesting that ex- pert commentary can be strategically correct while remaining concise, rather than explicitly spelling out long reasoning chains. This distinction also shows why faithfulness and depth should be inter- preted as complementary analytical signals. Background commentary shows a different ca- pability profile. Models achieve high relevance, reaching 0.903 on average, while factuality remains moderate at 0.683. Notably, open-weight models achieve the strongest scores on both background metrics, unlike the pattern observed for descriptive and analytical metrics. The two AI-GGC systems show limited performance on background metrics, indicating that historical recall remains underex- plored in existing AI-GGC work. Overall, current VLMs can usually generate background informa- tion that is topically connected to the commentary Table 2: Effects of model scale and reasoning mode. Left: comparison of open-weight models with different sizes. Right: comparison between minimal-reasoning and thinking settings on board games. (a) Effect of Model Scale ModelSize DescriptiveAnalyticalBackground Correctness Completeness Faithfulness Depth Factuality Relevance Qwen3-VL 32B0.2520.5330.1690.2320.6470.920 8B0.2600.5000.1110.1840.5660.916 InternVL3.5 38B0.2290.3220.1480.1510.7040.835 8B0.2160.3190.1370.1880.5800.887 GLM-4.6V 106B0.2470.5200.2270.2310.6390.926 9B0.2570.4920.1600.1120.5110.923 (b) Effect of Reasoning Mode Model Thinking Level Analytical Faithfulness Depth Gemini Min0.2060.271 Max0.5320.451 GPT Min0.2500.427 Max0.3640.445 Claude Min0.2350.394 Max0.2460.357 context, likely benefiting from parametric knowl- edge and the basic game information provided in the prompt. However, there is still room for im- provement for the factuality, suggesting that back- ground commentary may benefit from retrieval- augmented or evidence-grounded generation mech- anisms that provide verified player, team, match, or historical information. 5.3 Scale and Reasoning Mode We further analyze two factors that may affect commentary capability: model scale and reason- ing mode. For model scale, we compare the three open-weight models in the main result with their lighter variants on the full board-game and sports subsets. For reasoning mode, we evaluate the three closed-source models on a randomly sam- pled quarter of the board-game turns, comparing the minimal-reasoning setting used in the main experiments with the maximum thinking setting supported by each model. As shown in Table 2a, increasing model scale does not lead to uniform improvements in descrip- tive commentary. Larger models generally achieve slightly higher completeness, but their correctness remains similar or even decreases. This suggests that scaling makes models describe more events, but not a higher proportion of supported events. This pattern is consistent with the main benchmark results, where descriptive correctness remains a key bottleneck, especially for video-based games. The effect of scale is more evident for analytical and background capabilities. Larger models gen- erally improve analytical faithfulness, indicating better grounding of strategic statements in the game state. However, reasoning depth does not increase consistently across model families, suggesting that scale alone does not guarantee longer valid reason- ing chains. For background commentary, factuality shows the clearest scaling trend, with larger vari- ants consistently achieving higher factuality scores. This may reflect stronger parametric knowledge, better entity-level understanding, or more effective use of the basic game information provided in the prompt. In contrast, relevance is already high for most models and is less sensitive to scale. As shown in Table 2b, reasoning mode has a clear positive effect on analytical commentary. Across the three closed-source models, increasing the thinking level improves five out of six analyt- ical scores. The gains are particularly strong for Gemini-3.1-Pro-Preview, whose faithfulness and reasoning depth both increase substantially, making it the strongest model in this board-game analyti- cal setting. GPT-5.4 also benefits from the higher thinking setting, with a clear improvement in faith- fulness and a smaller gain in reasoning depth. Claude-Sonnet-4.5 shows a weaker response: faith- fulness slightly improves, while reasoning depth decreases modestly. Overall, these results indicate that stronger thinking settings can improve analyti- cal commentary, especially analytical faithfulness, although their effect on reasoning depth remains more model-dependent. 6 Conclusion We introduced GAMECOMMBENCH, a unified benchmark for AI-GGC across board games, sports, and esports, together with TACE, a type-aware commentary evaluation framework for descriptive, analytical, and background commentary. Our vali- dation shows that TACE provides stable and human- aligned evaluation signals. Experiments with rep- resentative AI commentators reveal non-uniform capability profiles across games and commentary types, with grounded game state description and faithful strategic reasoning remaining key bottle- necks. Further analyses show that model scale and reasoning mode bring uneven gains across commentary capabilities. Together, GAMECOMM- BENCH and TACE support more comparable and interpretable evaluation of AI-GGC. Limitations GAMECOMMBENCH is designed to cover represen- tative AI-GGC settings across board games, sports, and esports, but it does not exhaust all game gen- res, languages, or broadcasting styles. The current benchmark focuses on English text commentary generation and evaluation over five games with different state representations; it does not address speech-based commentary, voice delivery, or other multimodal presentation forms of game broadcast- ing. Extensions to additional languages, regional commentary styles, game communities, and speech or multimodal commentary settings may require new data collection and domain-specific process- ing. Finally, full multimodal evaluation across mul- tiple games and VLMs is costly, especially with closed-source APIs. Ethical Considerations This work studies benchmark construction and evaluation for AI-GGC. Generated commentary may contain errors, such as inaccurate descrip- tions or unsupported interpretations, which could mislead viewers if used directly in public-facing commentary systems, especially in text-only live updates where viewers rely primarily on the gen- erated text. In addition, GAMECOMMBENCH is constructed from game records, videos, transcripts, API-derived metadata, and related contextual in- formation. These sources may contain public names of players, teams, matches, or commenta- tors that are necessary for game commentary and background evaluation; however, we do not col- lect private or sensitive personal information or intentionally include offensive content. The bench- mark is constructed from public or previously re- leased game materials, and no separate consent was obtained from players, teams, or commenta- tors. Users of the benchmark should respect the licenses, terms of use, and attribution requirements of the original data sources. References Max Bain, Jaesung Huh, Tengda Han, and Andrew Zis- serman. 2023. Whisperx: Time-accurate speech transcription of long-form audio. arXiv preprint arXiv:2303.00747. Anton Behrens and Sebastian Uhrich. 2022. Youâl never want to watch alone: The effect of displaying in-stadium social atmospherics on media consumersâ responses to new sport leagues across different types of media. European Sport Management Quarterly, 22(1):120â138. Adrien Deliege, Anthony Cioppa, Silvio Giancola, Meisam J Seikavandi, Jacob V Dueholm, Kamal Nas- rollahi, Bernard Ghanem, Thomas B Moeslund, and Marc Van Droogenbroeck. 2021. Soccernet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, pages 4508â4519. Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. 2018. Soccernet: A scalable dataset for action spotting in soccer videos. In Pro- ceedings of the IEEE conference on computer vi- sion and pattern recognition workshops, pages 1711â 1721. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, and 1 others. 2024. A survey on llm-as-a-judge. The Innovation. Harsh Jhamtani, Varun Gangal, Eduard Hovy, Graham Neubig, and Taylor Berg-Kirkpatrick. 2018. Learn- ing to generate move-by-move commentary for chess games from large-scale social forum data. In Pro- ceedings of the 56th Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 1661â1671. Jaechang Kim, Jinmin Goh, Inseok Hwang, Jaewoong Cho, and Jungseul Ok. 2025. Bridging the gap be- tween expert and language models: Concept-guided chess commentary generation and evaluation. In Pro- ceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu- tational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 9497â9516. Andrew Lee, David Wu, Emily Dinan, and Mike Lewis. 2022. Improving chess commentaries by combining language models with symbolic reasoning engines. arXiv preprint arXiv:2212.08195. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harri- son Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Letâs verify step by step. In The twelfth inter- national conference on learning representations. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: Nlg evaluation using gpt-4 with better human align- ment. In Proceedings of the 2023 conference on empirical methods in natural language processing, pages 2511â2522. Jiayuan Rao, Haoning Wu, Hao Jiang, Ya Zhang, Yan- feng Wang, and Weidi Xie. 2025. Towards universal soccer video understanding. In Proceedings of the Computer Vision and Pattern Recognition Confer- ence, pages 8384â8394. Nicholas Tomlin, Andre He, and Dan Klein. 2022. Un- derstanding game-playing agents with natural lan- guage annotations. In Proceedings of the 60th An- nual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 797â 807. Zihan Wang and Naoki Yoshinaga. 2025. Commentary generation from multimodal game data for esports moments in multiplayer strategy games. In Proceed- ings of the 14th International Joint Conference on Natural Language Processing and the 4th Confer- ence of the Asia-Pacific Chapter of the Association for Computational Linguistics, pages 1795â1807. Qianfeng Wen, Zhenwei Tang, and Ashton Ander- son. 2025. Chessqa: Evaluating large language models for chess understanding.arXiv preprint arXiv:2510.23948. Dekun Wu, He Zhao, Xingce Bao, and Richard P Wildes. 2022. Sports video analysis on large-scale data. In European conference on computer vision, pages 19â 36. Springer. Yichao Yan, Ning Zhuang, Bingbing Ni, Jian Zhang, Minghao Xu, Qiang Zhang, Zheng Zhang, Shuo Cheng, Qi Tian, Yi Xu, and 1 others. 2019. Fine- grained video captioning via graph-based multi- granularity interaction learning.IEEE transac- tions on pattern analysis and machine intelligence, 44(2):666â683. Huanjin Yao, Jiaxing Huang, Yawen Qiu, Michael K Chen, Wenzheng Liu, Wei Zhang, Wenjie Zeng, Xikun Zhang, Jingyi Zhang, Yuxin Song, and 1 oth- ers. 2025. Mmreason: An open-ended multi-modal multi-step reasoning benchmark for mllms toward agi. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 273â283. Ling You, Wenxuan Huang, Xinni Xie, Xiangyi Wei, Bangyan Li, Shaohui Lin, Yang Li, and Changbo Wang. 2025. Timesoccer: An end-to-end multimodal large language model for soccer commentary genera- tion. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 3418â3427. Huanyu Yu, Shuo Cheng, Bingbing Ni, Minsi Wang, Jian Zhang, and Xiaokang Yang. 2018. Fine-grained video captioning for sports narrative. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6006â6015. Yiran Zhang, Mo Wang, Xiaoyang Li, Kaixuan Ren, Chencheng Zhu, and Usman Naseem. 2025. Turnbench-ms: A benchmark for evaluating multi- turn, multi-step reasoning in large language models. Findings of the Association for Computational Lin- guistics: EMNLP, pages 19892â19924. Zhihao Zhang, Feiqi Cao, Yingbin Mo, Yiran Zhang, Josiah Poon, and Caren Han. 2024. Game-mug: Multimodal oriented game situation understanding and commentary generation dataset. arXiv preprint arXiv:2404.19175. Qirui Zheng, Xingbo Wang, Keyuan Cheng, Muham- mad Asif Ali, Yunlong Lu, and Wenxin Li. 2025. From multimodal perception to strategic reasoning: A survey on ai-generated game commentary. arXiv preprint arXiv:2506.17294. A Detail of Data Collection and Processing Pipeline This appendix provides implementation-level de- tails for the data collection and processing pipeline among the five game genres in GAMECOMM- BENCH. As shown in Figure 5, the main text al- ready describes the high-level collection, cleaning, and packaging strategy; here we focus on the con- crete processing steps used to convert raw game state data and commentary into cleaned benchmark instances. A.1 Unified record schema and normalization The final conversion stage packages heterogeneous samples into a common JSONL interface. Sports records use the complete schema below, and board- game/esports records follow the same semantic con- tract while using PGN, SGF, or API-derived fields as the game-state carrier. "No": "...", "game_state": "modality": "video | pgn | sgf | api+video", "storage_file": "...", "step": "...", "label": "..." , "background": "...", "commentary": "raw_commetary": "...", "processed_commentary": "...", "commentary_category": [ "No": "...", "content": "...", "category": "descriptive | analytical | background" ] Here,raw_commetarykeeps the field name used by the implementation, and category IDs are normalized intodescriptive,analytical, and background. For football and basketball, the origi- nal video or event ID is replaced by an MD5-based stable identifier, with the original-to-hash mapping retained separately. A.2 Game-specific processing workflows This subsection merges the collection, alignment, cleaning, and final annotation steps for each game Game Info Game States Commentary Aligned by Turns Game States Aligned by Timestamps Textual Game States Visual Game States Commentary API API Fine-grained Temporal Alignment Irrelevant Content Filter ...Well, youâve already seen ...... itâs working out so far . Double, youâve rotated up to the top half of ......but when you look at their history in spring , they are actually ...... Game Quality Screening Dense Commentary Key Matches Balanced Teams/Players The traffic was really terrible tonight. The red wine I had yesterday... Semantic Decomposition 7.g4!! Hundreds of years ago players used to play things like king's gambit sacrificing the f pawn...... The idea behind this frightening pawn push is either to gain a half open g file or to...... Descriptive Background Analytical Type Annotation "No": â...", âgame_stateâ: ..., "commentary": "raw": â...", "processed": â...", âcommentary_category": [ âNoâ: â...â "content": â...", "category": â..." , ...... ] 1.g3 1...E5 2.Nf3 2...E4 3.Nd4 3...D5 4.d3 4...Exd3 ... ... Human Expert Verification Game GenresData Collection & AlignmentData Cleaning & Quality ControlSemantic AnnotationUnified Packaging Figure 5: Detailed data processing pipeline. Raw game data are collected and aligned, cleaned through quality control and temporal repair, decomposed into commentary types, and packaged into a unified format. type. The degree of processing differs by source modality: chess is the lightest because the source comments are already attached to PGN plies; Go requires extraction from FoxWQ records plus re- moval of unusable commentary; football, basket- ball, and LoL require explicit temporal alignment and cleaning. The LLM prompts used by the re- leased processing code are reproduced after the workflow descriptions. A.2.1 Football Football starts from raw rows containing a clip cap- tion and ASR-style commentary. The retained rows are tied to externally supplied video clips through the football video manifest, and match-level back- ground is reconstructed from the raw match meta- data, including teams, date, venue, referee, and player lists. The processing chain then performs: (i) sentence segmentation and type classification over the transcript with Prompt A.1; (i) caption- grounded filtering with Prompt A.3, where the cap- tion is treated as the authoritative action description and sentences with ambiguous references, weak rel- evance, or contradictions are removed; (i) caption- commentary fusion with Prompt A.4, which creates a unified broadcast-style description from the cap- tion and retained commentary; (iv) a final pass of Prompt A.1 over the fused text; and (v) conversion into the unified schema, including label normaliza- tion and MD5-based stable identifiers. A.2.2 Basketball Basketball follows the same alignment-and- cleaning pattern as football, but its initial alignment is stricter because the raw NSVA-derived caption and the transcript may refer to adjacent plays. The pipeline first applies the relevance gate in Prompt A.5 and only writes caption-commentary pairs whose boxed judgment is true. The re- maining rows are segmented and classified with the basketball version of the classification prompt in Prompt A.2, filtered with the same retention logic as Prompt A.3 using the basketball-specific phrase âon-court situations,â and then fused with the bas- ketball caption-commentary prompt in Prompt A.6. The basketball pipeline additionally runs the qual- ity selector in Prompt A.7 over batches of three merged candidates and keeps the two highest- ranked examples. Final conversion attaches the raw NSVA caption and transcript, the video storage path, the coarse action label, the game background recovered from NBA metadata files, normalized commentary categories, and an MD5-based stable identifier. A.2.3 Chess Chess uses the lightest data-processing path. Saved GameKnot HTML pages are grouped by game ID, sorted by page index, and parsed into player/event metadata plus move-comment tuples. The con- verter writes one PGN file per game, preserves comments in PGN braces, removes HTML/script artifacts from comments, normalizes whitespace, escapes PGN tag strings, and stores metadata such as event, site, players, ratings, opening, annotator, game ID, and move count. Since commentary is already attached to moves, no timestamp matching or caption-grounded LLM cleaning is needed. The GamePrimary alignmentProcessing summary FootballCaption/event clip to transcriptSegment and classify transcript, filter against caption, fuse caption and retained commentary, reclassify, then normalize IDs and labels. BasketballNSVA caption/action clip to transcript Gate caption-commentary relevance, segment/classify, filter, fuse, select high-quality merged outputs, then normalize with NBA metadata and action labels. ChessPGN mainline plyParse GameKnot HTML into PGN comments; preserve move-attached commentary with only lightweight text normalization. GoSGF mainline plyCrawl FoxWQ JSON, extract SGF, discard empty or irrelevant comments, and normalize turn-aligned records. League of LegendsSubtitle window plus API window Align subtitle segments to clips, aggregate Riot/LoL Esports API windows, and retain significant team/player state changes. Table 3: Integrated data-processing workflows for the five game types. final records use the PGN file asstorage_file, the ply asstep, and the comment text as the refer- ence commentary after lightweight normalization and label mapping. A.2.4 Go Go starts from FoxWQ metadata seeds.The crawler downloads each referenced JSON payload, and the converter extracts thechessfield into an SGF file named bychessid. Empty or malformed JSON entries are skipped. Compared with chess, Go needs an additional cleanup step because ex- tracted records can contain empty, duplicated, or off-topic comments; these are removed before the retained move-comment pairs are normalized into the released schema. SGF coordinates are kept as the source representation, while later state construc- tion can convert them into standard Go coordinates and render the board before the target move. A.2.5 League of Legends LoL combines subtitle/VOD alignment with API- derived state extraction. Subtitle files are parsed into timestamped segments after HTML tag re- moval, whitespace normalization, and optional consecutive-duplicate removal. For Match-V5 data, events are aggregated in fixed time windows, the most-overlapping subtitle sentence is selected as reference commentary, and each row stores the event window, linearized event summaries, team K/D/A and gold, patch, game date, and a stable MD5 row ID. For LoL Esports live-stat data, the acquisition script resolves match and game IDs from team names, game numbers, league sched- ules, and event details, fetches live-stat windows, and extracts team/player snapshots around the tar- get timestamp. When start/end snapshots are avail- able, significant API changes are retained as struc- tured context throughstep_boundary_change; this records boundary-level changes such as team gold, team kills, and player-level metric changes. B Candidate Commentary Generation Candidate commentaries are generated without ac- cess to the human reference commentary. Each generator reads one JSONL record, constructs the game-specific input context, calls the target model endpoint, and writes a JSONL output withNoand commentary_by_model. For football and basketball, videos are sampled at 4 FPS and the sampled frames are interleaved with timestamp text while preserving the original frame resolution used in the run. For chess and Go, the scripts parse the source game files and pass symbolic state rather than visual frames: chess reconstructs FEN before/after the target move to- gether with SAN/UCI notation and recent move history, while Go renders the board before the tar- get move and supplies the current coordinate plus recent moves. League of Legends uses video input together with match background and, when avail- able, thestep_boundary_changeAPI summary describing team/player state transitions. Across all reported generation runs, temperature is set to 0 and reasoning or thinking modes are disabled when supported. The prompts below are copied from the candidate-generation scripts. Runtime placehold- ers are filled by the corresponding script immedi- ately before the model call. C Evaluation The evaluation code implements TACE as a names- paced pipeline with four functional branches: seg- mentation, descriptive evaluation, analytical evalu- ation, and background evaluation. The run scope can be set toall,segmentation,descriptive, GameInput supplied to modelDecoding policyState construction FootballTimestamped video frames, match background, clip duration temperature 0; thinking disabled 4 FPS visual sampling with original frame resolution. BasketballTimestamped video frames, match background, clip duration temperature 0; thinking disabled 4 FPS visual sampling with original frame resolution. ChessFEN before/after, SAN, UCI, move string, recent moves, background temperature 0; thinking disabled PGN mainline parsing with recent move history. GoASCII board, SGF and standard coordinate, recent moves, background temperature 0; thinking disabled SGF mainline parsing with board rendering before the current move. League of Legends Video input, match background, API transition context temperature 0; thinking disabled Clip-aligned visual input plus team/player state transitions. Table 4: Candidate commentary generation settings used by the provided scripts. analytical, orbackground. Judge and extrac- tion calls use temperature 0.0, and sport aliases are normalized before evaluation so thatfootball, soccer, and their Chinese alias map to the same branch. C.1 Segmentation for evaluation Before type-specific scoring, candidate commen- tary is segmented into the smallest independent semantic units and assigned category IDs 1, 2, or 3. The implementation validates the returned JSON array and then routes units with ID 1 to descriptive evaluation, ID 2 to analytical evaluation, and ID 3 to background evaluation. The football prompt is shown below as the concrete representative; all other sports use parallel prompt files with the same output contract. C.2 Descriptive evaluation Descriptive evaluation measures whether the gen- erated commentary correctly captures the salient observable events. The evaluator first obtains a non-empty reference event list from the human ref- erence and state context. It then extracts generated events from candidate descriptive units. For sports, the reference extractor uses labels and match back- ground; for chess and Go, it uses the symbolic board context; for League of Legends, it uses both commentary and API-derived state changes. The same reference event list is used for both complete- ness and correctness matching. If no descriptive candidate event is extracted, both metrics are set to zero. Scores are rounded to two decimals in the implementation. C.3 Analytical evaluation Analytical commentary is evaluated through a proposition graph rather than a holistic score. The pipeline first extracts explicit analytical proposi- tions from category-2 units. It then builds sparse forward edge candidates using adjacent proposi- tions and hybrid retrieval, where similarity com- bines embedding similarity, entity overlap, and key- word overlap. Candidate edges are restricted to the relation setCAUSE,REASON,CONDITION,CONTRAST, andELABORATION. A second audit stage removes redundant or weakly anchored edges. The resulting graph is validated as a DAG with no self-loop and no duplicate edge for the same source-target pair. The released implementation stores several ana- lytical diagnostics. The TACE faithfulness signal is the graph-grounded correctness ratio, computed from node and final-edge truth labels: AnalyticalFaithfulness = N true + E true N + E , (1) whereNandEare the number of judged propo- sitions and graph edges.The code also re- portsproposition_truth, which is the node-only truth ratio. Reasoning depth is reported through effective_depth: after node and edge verifica- tion, the evaluator keeps only true nodes and true edges, computes the longest valid unweighted DAG path, and normalizes it by the maximum raw depth among models for the same sport-record group. Raw graphdepthandrelation_infoare retained as additional diagnostics. C.4 Background evaluation Background evaluation first converts category-3 units into verifiable claims. Each claim carries a source sentence ID, claim type, focus key, enti- ties, and time scope, making the later factuality and relevance judgments operate on independently checkable units. Factuality is judged against a local background knowledge base derived from the reference record metadata. The retriever combines structured chan- Table 5: Human validation accuracy of LLM-based type- aware decomposition across games. GameAccuracy Chess94.0% Go93.0% Football98.0% Basketball98.0% LoL96.0% Overall95.8% nels based on source ID, focus key, entity align- ment, fact-key alignment, and source-context fall- back with optional semantic embedding retrieval. The evidence package is bounded by a hard evidence-overflow limit, and the factuality judge may only cite evidence IDs that were actually retrieved. If no valid evidence is available, the claim is labeled false by construction. Relevance is judged separately by comparing each claim with the immediate descriptive and analytical clauses of the same candidate commentary. The reported scores: Factuality = #claims labeled true #scored claims ,(2) Relevance = #claims labeled relevant #scored claims .(3) Both scores are rounded to four decimals; if no claim can be scored, the corresponding value is None. D Human Validation Human annotators were given sampled commen- tary units or intermediate evaluation items with the corresponding game context when applicable. For type-aware decomposition, they independently as- signed each unit one of three labels: descriptive, analytical, or background. For component valida- tion, they independently annotated the target event, matching relation, reasoning relation, verification label, or background-claim judgment according to the provided context and annotation definitions. Their annotations were then compared with the LLM-produced outputs to compute agreement. Hu- man validation was conducted by trained internal annotators following predefined annotation guide- lines. No external crowdworkers or paid participant pools were recruited, no monetary compensation was provided. Annotators were instructed to rely only on the provided materials, and no personal information about annotators was collected. Table 6: Cross-judge consistency for analytical met- rics. We compare model rankings produced by the main GPT-5.4-mini-based judge and an alternative Gemini-3- Flash-based judge on a held-out subset. TypeMetricSpearmanĎ Analytical Faithfulness0.77 Reasoning Depth0.94 We conduct additional human validation for the LLM-based type-aware decomposition step. For each game, we randomly sample 100 commentary units after decomposition and manually annotate their commentary types as descriptive, analytical, or background. The LLM-produced labels are then compared with the human annotations to compute accuracy. As shown in Table 5, the decomposition step achieves consistently high accuracy across all games, ranging from 93.0% on Go to 98.0% on football and basketball.The overall accuracy reaches 95.8%, indicating that the type-aware de- composition used by TACE is reliable across het- erogeneous game domains. E Cross-Judge Consistency Check Because TACE uses GPT-5.4-mini for several LLM-mediated evaluation steps, we further exam- ine whether the main analytical results are driven by judge-model preference. This concern is most rele- vant for analytical commentary: in descriptive eval- uation, the key event-matching step is performed by GLM-4.7 rather than a GPT-based judge; in background evaluation, the best-performing mod- els are not GPT-family models. In contrast, ana- lytical node and edge verification are implemented with GPT-5.4-mini, while GPT-5.4 achieves the strongest analytical performance in the main re- sults. We therefore conduct a targeted cross-judge consistency check for analytical metrics. Specifically, on a held-out subset, we replace the GPT-5.4-mini-based node and edge verification judge with Gemini-3-Flash, while keeping the rest of the TACE pipeline unchanged. We then compare the resulting model rankings with those produced by the main judge using Spearman rank correlation. As shown in Table 6, the rankings remain consis- tent, especially for Reasoning Depth. This suggests that the analytical conclusions are not solely driven by preference from the GPT-based judge. F Detailed Prompts F.1 Prompts for Data Processing Prompt A.1: Football sentence segmentation and type classification You are a professional football commentary analysis expert. Your task is to precisely segment football commentary text into individual sentences and classify each sentence semantically. Football Commentary Classification Standards. 1. Descriptive Commentary (ID: 1) ⢠Core Question: âWhat is happening on the pitch?â ⢠Characteristics: Objective narration of events and actions directly observable in the game. ⢠Includes: Player movements (shooting, passing, dribbling, defending), score changes, referee calls, possession changes, etc. ⢠Examples: ... ... 2. Analytical Commentary (ID: 2) ⢠Core Question: âWhy is this happening? What will happen next?â â˘Characteristics: Explains tactical intentions, evaluates decision quality, predicts game developments, analyzes technical details. ⢠Includes: Tactical analysis, decision evaluation, situation prediction, technical movement breakdown. ⢠Examples: ... ... 3. Background Commentary (ID: 3) ⢠Core Question: âWhat is the background and significance of this event?â â˘Characteristics: Introduces external knowledge such as player backgrounds, historical data, version information, off-pitch stories, etc. ⢠Includes: Player career history, head-to-head records, injury status, transfer background, statistical data. ⢠Examples: ... ... Important Rules. ⢠Must accurately split the original text into complete sentences; do not merge or over-split. ⢠Each sentence can only be assigned one most appropriate category. ⢠Output must be in strict JSON array format, with each element as a binary tuple [sentence, category_id]. ⢠Only output the JSON array, with no other text or explanations. User Prompt Template. Please analyze the following football commentary text: commentary Perform sentence segmentation and classification according to the above standards, and output the JSON array. Prompt A.2: Basketball sentence segmentation and type classification You are a professional basketball commentary analysis expert. Your task is to precisely segment basketball commentary text into individual sentences and classify each sentence semantically. Basketball Commentary Classification Standards. 1. Descriptive Commentary (ID: 1) ⢠Core Question: âWhat is happening on the court?â ⢠Characteristics: Objective narration of events and actions directly observable in the game. â˘Includes: Player movements (shooting, driving, passing, defending), score changes, referee calls, possession changes, etc. ⢠Examples: ... ... 2. Analytical Commentary (ID: 2) ⢠Core Question: âWhy is this happening? What will happen next?â â˘Characteristics: Explains tactical intentions, evaluates decision quality, predicts game developments, analyzes technical details. ⢠Includes: Tactical analysis, decision evaluation, situation prediction, technical movement breakdown. ⢠Examples: ... ... 3. Background Commentary (ID: 3) ⢠Core Question: âWhat is the background and significance of this event?â â˘Characteristics: Introduces external knowledge such as player backgrounds, historical data, version information, off-court stories, etc. ⢠Includes: Player career history, head-to-head records, injury status, trade background, statistical data. ⢠Examples: ... ... Important Rules. ⢠Must accurately split the original text into complete sentences; do not merge or over-split. ⢠Each sentence can only be assigned one most appropriate category. ⢠Output must be in strict JSON array format, with each element as a binary tuple [sentence, category_id]. ⢠Only output the JSON array, with no other text or explanations. User Prompt Template. Please analyze the following basketball commentary text: commentary Perform sentence segmentation and classification according to the above standards, and output the JSON array. Prompt A.3: Caption-grounded sentence filtering (football) Analyze each sentence independently, applying minimal paraphrasing only to resolve ambiguous references or spelling errors. Retention Criteria (ALL must be satisfied): 1. Reference Clarity: Pronouns must clearly refer to a specific entity inferable from caption and previously retained commentary. Pronoun inference must be completely unambiguous and free of any doubt. â˘Mandatory first-sentence truncation check: Reject if the opening sentence appears incomplete (e.g., begins with conjunction/preposition/subordinate clause or lacks subject/verb). In the first sentence, any pronoun referring to a player or team with uncertainty must be rejected, as it likely refers to preceding context that was truncated and cannot be definitively confirmed. 2. Relevance: Content must directly relate to captionâs action, event, or mentioned players. ⢠Category 1 (Action): Describe captionâs action or directly related on-field situations. ⢠Category 2 (Strategy): Analyze tactics based on captionâs action. â˘Category 3 (Background): Provide supplemental info ONLY for players appearing in caption; MUST clearly and unambiguously refer to a specific player or team through inference that is completely unambiguous and free of doubt. 3. Semantic Consistency: Must not contradict caption or retained sentences. 4. Minimal Correction: Rephrase solely for: ⢠Ambiguous but inferable references. ⢠Obvious name spelling errors (common in audio transcription). ⢠Preserve original tone and intent. Core Principles: ⢠Caption is authoritative; do not question its accuracy. ⢠Only retained sentences may serve as context for subsequent evaluations. â˘Background information must have a clearly inferable target; reject if the player/team reference is ambiguous or uncertain. ⢠Violation of criteria 1â3â reject without paraphrasing. Output Format. Part 1: Analysis. For each sentence, detail: â˘Reference clarity (with truncation assessment and pronoun ambiguity evaluation, especially for player/team references in first sentence). â˘Relevance and factual accuracy (state reason if failed; for Category 3, explicitly flag target specificity and inference clarity concerns). ⢠Semantic consistency (identify contradictions if failed). ⢠Reference Specificity (Category 3 ONLY): Confirm background information clearly and unambiguously refers to a specific player or team through inference that is completely unambiguous and free of doubt; reject if the target is unclear or requires ambiguous inference. ⢠Paraphrasing need (specify changes and rationale). ⢠Final verdict (retain/reject; provide corrected version if applicable). Part 2: JSON. [ ["Sentence1", Category1, true/false], ["Sentence2", Category2, true/false], ... ] ⢠Strict array of triplets: [sentence, category, retention flag]. ⢠Sentence: original if unchanged, corrected if paraphrased. ⢠Retention flag: lowercase true/false. ⢠Category number unchanged. User Prompt Template. Caption (authoritative action description): caption Commentary Sentences (pre-classified): sentences_info Apply the specified retention criteria to evaluate each sentence independently. Output Part 1: Analysis and Part 2: JSON. Prompt A.4: Football caption-commentary fusion You are an expert soccer video content analyst. Your task is to combine the caption and commentary into a coherent narrative written in the voice of a professional sports commentator. Input: â˘Caption: A brief textual description of a soccer/football game clip, focusing on the main events and player actions occurring in the video. â˘Commentary: Transcribed live commentary from the same video segment, which may contain multiple pieces of information. Task: Combine the caption and commentary into a coherent, natural-sounding description that captures the key elements of both, written in the live broadcast voice while maintaining stylistic fidelity to the provided commentaryâmirroring its energy, pacing, vocabulary, and tonal qualities. The result should read like a unified live commentary of the event, not as two separate pieces of information. Critical Filtering Rule: Before combining, carefully evaluate each piece of information in the commentary according to these strict hierarchical requirements: 1.Event Description Criterion: Commentary elements describing on-field actions must demonstrate direct relevance to the captionâs events OR establish a clear chronological relationship with them. Exclude isolated event descriptions that lack temporal or causal connection to the caption narrative. 2. Tactical Analysis Criterion: Strategic or tactical observations must be explicitly anchored to specific event descriptions from the caption or retained commentary. Exclude standalone tactical analysis that references no concrete on-field action present in the source materials. 3.Background Information Criterion: Contextual details (player positioning, formations, situational factors) may only be included when they directly supplement and illuminate events described in the caption or retained commentary. Exclude background information wholly detached from the depicted action. 4.General Enhancement Criterion: Only retain commentary elements that satisfy the above three criteria and demonstra- bly enhance or complement the captionâs narrative while maintaining absolute factual consistency. Strict Factual Boundaries: Your entire narrative must be constructed exclusively from information explicitly present in the caption and commentary. You are prohibited from: ⢠Introducing any external facts, statistics, or context not provided in the source materials. ⢠Speculating on player motivations, emotions, or unspoken thoughts. ⢠Fabricating details about game situation (score, time, half) beyond stated information. ⢠Assuming or inventing player names, teams, venues, or historical data. ⢠Creating hypothetical scenarios or âwhat-ifâ scenarios. Focus on: 1. Creating a seamless narrative that authentically reflects the source commentaryâs unique style, delivery patterns, and rhetorical approach, incorporating both factual events (caption) and relevant descriptive/color commentary. 2. Maintaining strict chronological flow of events as presented in the caption. 3.Employing the specific language patterns, excitement cadence, and terminology preferences exhibited in the provided commentary, ensuring stylistic consistency while applying professional broadcasting standards. 4. Ensuring clarity, readability, and factual accuracy by filtering out irrelevant or contradictory commentary. 5.Preserving all important information from the caption while selectively incorporating valuable commentary elements that enhance the broadcast-style narrative. 6. Maintaining absolute fidelity to the source material, with zero speculation or fabrication. OutputFormat:Presentyourcombinednarrativeinsideaboxusingtheformat your_combined_narrative_here.This should be the only content in your response.The narrative must be written in the distinctive style and tone captured in the source commentary. User Prompt Template. Caption: caption Commentary: commentary_text Prompt A.5: Basketball caption-commentary relevance gate You are an expert basketball video content analyst. Your task is to determine whether a live commentary segment is semantically relevant to the current video content described in a caption. Input: â˘Caption: A brief textual description of a basketball game clip, focusing on the main events and player actions occurring in the video (e.g., âPlayer X shoots a three-pointerâ, âTeam Y executes a fast breakâ). â˘Commentary: Transcribed live commentary from the same video segment, which may have temporal lag and could discuss events beyond the immediate clip. Relevance Criteria: The commentary is considered RELEVANT (true) if: 1. It directly describes or comments on the events/actions mentioned in the caption, OR 2. It provides strategic analysis of team tactics based on the captionâs events, OR 3.It offers background context about players/teams mentioned in the caption (e.g., season statistics, career achievements, team history), OR 4. It discusses game implications (score changes, momentum shifts) related to the captionâs events. The commentary is considered IRRELEVANT (false) if: 1. It discusses completely unrelated events (different plays, previous games, off-court topics), OR 2. It contains generic filler content with no connection to the captionâs specific events or entities. Output Format. â˘Phase 1 (Analysis): Provide a concise reasoning analyzing the semantic relationship between caption and commentary. Consider temporal context, entity mentions, and thematic connections. ⢠Phase 2 (Judgment): Output ONLY the final judgment as true or false with no additional text. Important: Your analysis must be objective and based solely on semantic relevance. Do not over-interpret vague connections. The boxed judgment must be your definitive conclusion. User Prompt Template. caption: caption commentary: commentary Prompt A.6: Basketball caption-commentary fusion You are an expert basketball video content analyst. Your task is to combine the caption and commentary into a coherent narrative written in the voice of a professional sports commentator. Input: ⢠Caption: A brief textual description of a basketball game clip, focusing on the main events and player actions occurring in the video. â˘Commentary: Transcribed live commentary from the same video segment, which may contain multiple pieces of information. Task: Combine the caption and commentary into a coherent, natural-sounding description that captures the key elements of both, written in the live broadcast voice while maintaining stylistic fidelity to the provided commentaryâmirroring its energy, pacing, vocabulary, and tonal qualities. The result should read like a unified live commentary of the event, not as two separate pieces of information. Critical Filtering Rule: Before combining, carefully evaluate each piece of information in the commentary: ⢠If a commentary element is directly supported by or consistent with the captionâs events, include it. ⢠If a commentary element is unrelated to the captionâs described events, exclude it. ⢠If a commentary element contradicts the factual events described in the caption, exclude it. ⢠Only retain commentary elements that enhance or complement the captionâs narrative. Strict Factual Boundaries: Your entire narrative must be constructed exclusively from information explicitly present in the caption and commentary. You are prohibited from: ⢠Introducing any external facts, statistics, or context not provided in the source materials. ⢠Speculating on player motivations, emotions, or unspoken thoughts. ⢠Fabricating details about game situation (score, time, quarter) beyond stated information. ⢠Assuming or inventing player names, teams, venues, or historical data. ⢠Creating hypothetical scenarios or âwhat-ifâ scenarios. Focus on: 1.Creating a seamless narrative that authentically reflects the source commentaryâs unique style, delivery patterns, and rhetorical approach, incorporating both factual events (caption) and relevant descriptive/color commentary. 2. Maintaining strict chronological flow of events as presented in the caption. 3.Employing the specific language patterns, excitement cadence, and terminology preferences exhibited in the provided commentary, ensuring stylistic consistency while applying professional broadcasting standards. 4. Ensuring clarity, readability, and factual accuracy by filtering out irrelevant or contradictory commentary. 5.Preserving all important information from the caption while selectively incorporating valuable commentary elements that enhance the broadcast-style narrative. 6. Maintaining absolute fidelity to the source material, with zero speculation or fabrication. OutputFormat:Presentyourcombinednarrativeinsideaboxusingtheformat your_combined_narrative_here.This should be the only content in your response.The narrative must be written in the distinctive style and tone captured in the source commentary. User Prompt Template. Caption: caption Commentary: commentary Prompt A.7: Basketball quality selection prompt System Prompt. You are a professional sports-commentary analyst. From a batch of basketball game commentaries, select the two highest- quality items. Evaluation Criteria: Each commentary will be scored on a 10-point scale across the following three dimensions, and the final ranking is based on the total score: 1.Descriptive Dimension (Descriptive): Evaluate the precision and information density with which the commentary captures the real-time game picture: whether observable facts such as the ball-handler, tactical movement, screen quality, shot area, defensive matchup, and referee calls are stated clearly, completely, and in detail. A higher score means the account of âwhat is happening at this momentâ is more like a precise courtside documentary. 2. Analytical Dimension (Analytical): Evaluate the depth of interpretation of the gameâs underlying logic: whether offensive actions (Spain P&R, Horns, Flare), defensive strategies (switching, double-teaming, ICE), player decision quality, coaching matchups, and tactical trade-offs under score/time/foul situations are analyzed thoroughly. A higher score means the insight into âwhy the play unfolds this wayâ is more like that of a tactical master. 3. Background Dimension (Background): Evaluate the ability to extend the event context: whether player historical statistics, matchup histories, team season trends, key-game conventions, cultural references (e.g., âChristmas Day gameâ), sneakers, records, and other off-court information are naturally integrated to add the âdeeper meaning of this momentâ to the current picture. A higher score means the rendering of âwhat this moment meansâ is more like that of a storyteller. Analysis Requirements: Follow these steps: ⢠Step 1 (Analysis Phase): For each commentary, provide analysis and a score for each of the three dimensions on a 10-point scale, then calculate the total score. â˘Step 2 (Selection Phase): Sort the commentaries by total score from high to low, select the two highest-scoring commentaries, and return only their video_id values. Output Format: [Analysis Phase] 1. [video_id]: - Descriptive Dimension (Descriptive): [Reason] - X points - Analytical Dimension (Analytical): [Reason] - X points - Background Dimension (Background): [Reason] - X points - Total Score: X points ... [Selection Phase] ["video_id1", "video_id2"] User Prompt Template. Select the two highest-quality commentaries from the following 3 basketball-game commentaries: commentary_list F.2 Prompts for Commentary Generation Prompt B.1: Football / Basketball candidate-generation prompt Role: You are a top-tier professional football / basketball commentator. Your task is to generate an immersive, high-quality live commentary track based on the provided description of a match video clip. Core Objective: Produce a coherent, pure-text narrative that replicates the energy, pacing, and terminology of a live TV broadcast. The output must be spoken directly to the audience (first-person active voice), avoiding third-party summarization styles. Content Requirements (Semantic Layers): Your commentary must align with the three core commentator capabilities and their corresponding commentary types: 1. Live Observationâ Descriptive Commentary: ⢠Accurately perceive and interpret real-time events and fine-grained details in the game environment. ⢠Objectively answer âWhat is happening?â by narrating immediate, observable events. 2. Strategic Analysisâ Analytical Commentary: â˘Comprehend the current game state, interpret strategic intent, evaluate decisions, and anticipate future developments. â˘Go beyond description by assessing âHow good?â, explaining âWhy?â, and predicting âWhatâs next?â within a broader tactical context. 3. Historical Recallâ Background Commentary: â˘Leverage extensive game knowledge, including player histories, classic match moments, and community-specific terminology. ⢠Answer âWhat is the significance?â by integrating broader contextual information such as detailed team and player backgrounds, game history, and cultural references. Style & Tone Guidelines: ⢠Dynamic Pacing: Match the rhythm of the text to the intensity of the event (e.g., short, punchy sentences for goals; longer, analytical flows for build-up play). ⢠Professional Vocabulary: Use authentic football / basketball terminology. â˘Natural Fluency: Avoid robotic sentence structures. Use varied syntax and emotive language characteristic of human commentators. Strict Constraints: 1.Format: Output raw text only as a single continuous paragraph. Do not use line breaks ( ), Markdown, bolding, labels, or list formats. 2. Language: Maintain a single language consistently. Use English uniformly. 3. Factual Grounding: Do not hallucinate scorelines, team names, or events not implied by the input. Do not invent specific statistics unless provided. 4. Consistency: Ensure the narrative flow is logical and non-contradictory. 5. Scope: The video is only a segment of the match. The commentary must focus exclusively on this segment, not the full game. User Prompt Template. Match Background (Full Match Context, not limited to this clip): background Generate a football / basketball commentary based on the video content. The video segment is duration_seconds seconds long. Complete the entire commentary within duration_seconds seconds. Prompt B.2: Chess / Go candidate-generation prompt System Prompt. You are a top-tier professional chess / go commentator. Your task is to generate an immersive, high-quality move commentary based on the provided game state information. User Prompt Template. Role: You are a top-tier professional chess / go commentator. Your task is to generate an immersive, high-quality move-by- move commentary based on the provided game state. Core Objective: Produce a coherent, pure-text narrative that replicates the insight, pacing, and terminology of a professional chess / go broadcast. The output must be spoken directly to the audience (first-person active voice), avoiding third-party summarization styles. Content Guidelines (Semantic Layers): You may draw on the following three core commentary capabilities of good commentators. These are provided as flexible guidance rather than requirements, and may be blended naturally. 1. Move Observationâ Descriptive Commentary: ⢠Accurately perceive and describe the current move, the piece involved, and the squares it moves from and to. ⢠Objectively answer âWhat is happening on the board?â by narrating the immediate, observable move. 2. Strategic Analysisâ Analytical Commentary: â˘Comprehend the current board position (provided via FEN), interpret the tactical or positional intent behind the move, evaluate its quality, and anticipate future developments. â˘Go beyond description by assessing âHow good is this move?â, explaining âWhy was it played?â, and predicting âWhatâs next?â within a broader strategic context. 3. Historical Recallâ Background Commentary: ⢠Leverage extensive chess / go knowledge, including opening theory, classic game references, and player tendencies. â˘Answer âWhat is the significance?â by integrating broader contextual information such as player backgrounds, tournament stakes, and historical parallels. Style & Tone Guidelines: ⢠Professional Vocabulary: Use authentic chess / go terminology. â˘Natural Fluency: Avoid robotic sentence structures. Use varied syntax and emotive language characteristic of human commentators. Strict Constraints: 1.Format: Output raw text only as a single continuous paragraph. Do not use line breaks ( ), Markdown, bolding, labels, or list formats. 2. Language: Use English uniformly. 3.Factual Grounding: Do not hallucinate player names, ratings, or events not implied by the input. Do not invent specific statistics unless provided. 4. Consistency: Ensure the narrative flow is logical and non-contradictory. 5. Scope: The commentary can focus on this specific move and its immediate context or the entire game. Match Background (Full Game Context): background Game State: Board before current move (FEN): fen_before Current move: full_move_str (SAN=san, UCI=uci) Board after current move (FEN): fen_after Previous moves (up to last 3 full moves by both sides): history_text Game State: Board before current move (X=Black, O=White, .=empty): board_ascii Current move (ply ply): move_str (SGF coord=coord_sgf, standard=coord_std) Recent moves (last 3 full rounds): history_text Generate a chess / go commentary for this move. Prompt B.3: League of Legends candidate-generation prompt System Prompt. You are a top-tier professional League of Legends commentator. Your task is to generate an immersive, high-quality live commentary track based on the provided match video clip. User Prompt Template. Role: You are a top-tier professional League of Legends commentator. Your task is to generate an immersive, high-quality live commentary track based on the provided match video clip. Core Objective: Produce a coherent, pure-text narrative that replicates the energy, pacing, and terminology of a live esports broadcast. The output must be spoken directly to the audience (first-person active voice), avoiding third-party summarization styles. Content Guidelines (Semantic Layers): You may draw on the following three core commentary capabilities of good commentators. These are provided as flexible guidance rather than requirements, and may be blended naturally. 1. Live Observationâ Descriptive Commentary: ⢠Accurately perceive and interpret real-time events and fine-grained details in the game environment. â˘Objectively answer âWhat is happening?â by narrating immediate, observable events (champion movements, ability usage, objective takes, kills, map states). 2. Strategic Analysisâ Analytical Commentary: â˘Comprehend the current game state, interpret strategic intent, evaluate decisions, and anticipate future developments. â˘Go beyond description by assessing âHow good?â, explaining âWhy?â, and predicting âWhatâs next?â within a broader tactical context (draft implications, win conditions, power spikes, macro decisions). 3. Historical Recallâ Background Commentary: â˘Leverage extensive League of Legends knowledge, including player histories, classic match moments, and community- specific terminology. â˘Answer âWhat is the significance?â by integrating broader contextual information such as team and player back- grounds, game history, and cultural references. Style & Tone Guidelines: â˘Professional Vocabulary: Use authentic League of Legends terminology (e.g., âlane pressure,â âreset,â âtrade,â âskirmish,â ârotation,â âvision control,â âengage,â âdisengage,â âobjective posture,â âpower spike,â âteam compâ). â˘Natural Fluency: Avoid robotic sentence structures. Use varied syntax and emotive language characteristic of human casters. Strict Constraints: 1.Format: Output raw text only as a single continuous paragraph. Do not use line breaks ( ), Markdown, bolding, labels, or list formats. 2. Language: Use English uniformly. 3.Factual Grounding: Do not hallucinate champion names, gold values, objectives, kills, or map states that are not clearly supported by the video or provided context. 4. Consistency: Ensure the narrative flow is logical and non-contradictory. 5. Scope: The video is only a segment of the match. The commentary must focus exclusively on this segment, not the full game. Match Background (Full Match Context, not limited to this clip): background extra_context_block The video segment is duration_seconds seconds long. Complete the entire commentary within duration_seconds seconds. Generate a League of Legends commentary based on the video content. F.3 Prompt for TACE F.3.1 Prompt for Generated Commentary Segmentation Prompt C.1: Evaluation segmentation and type classification (football) System Prompt. You are a professional football commentary analysis expert. Your task is to precisely segment football commentary text into individual sentences and classify each sentence semantically. Football Commentary Classification Standards. 1. Descriptive Commentary (ID: 1) ⢠Core Question: âWhat is happening on the pitch?â â˘Characteristics: Objective narration of events and actions directly observable in the game, without subjective evaluation. â˘Includes: Player movements (shooting, passing, dribbling, defending), score changes, referee calls, possession changes, etc. ⢠Examples: ... ... 2. Analytical Commentary (ID: 2) ⢠Core Question: âHow good? Why is this happening? What will happen next?â â˘Characteristics: Subjectively evaluates performance or decisions, explains tactical intentions and causes, predicts game developments, analyzes technical details. â˘Includes: Tactical analysis, decision evaluation, situation prediction, technical movement breakdown, emotional evaluation of plays. ⢠Examples: ... ... 3. Background Commentary (ID: 3) ⢠Core Question: âWhat is the background and significance of this event?â ⢠Characteristics: Introduces external knowledge such as player backgrounds, historical data, off-pitch stories, etc. ⢠Includes: Player career history, head-to-head records, injury status, transfer background, statistical data. ⢠Examples: ... ... Boundary Rules: ⢠When a sentence contains elements of multiple categories, classify based on its primary intent (semantic focus). ⢠Pure emotional exclamations with evaluative tone (e.g., âWhat a goal!â) belong to Analytical Commentary (ID: 2). ⢠Objective data statements (e.g., âThis is his 30th goalâ) belong to Background Commentary (ID: 3); interpretation or inference drawn from data belongs to Analytical Commentary (ID: 2). Important Rules: â˘Segment text into the smallest independent semantic units (clauses or short phrases) rather than full grammatical sentences. Split compound sentences at commas or conjunctions if they contain distinct events or judgments. ⢠Each sentence can only be assigned one most appropriate category. ⢠Output must be in strict JSON array format, with each element as a binary tuple [sentence, category_id]. ⢠Only output the JSON array, with no other text or explanations. User Prompt Template. Please analyze the following football commentary text: commentary Perform sentence segmentation and classification according to the above standards, and output the JSON array. F.3.2 Prompt for Descriptive Evaluation Prompt C.2: Reference event extraction (football) System Prompt. You are a football event extraction expert. Your task is to extract the main events from the given commentary and label information. Input: ⢠label: The main event type of the video clip (e.g., goal, corner, clearance, etc.). ⢠processed_commentary: The processed reference commentary describing what happened. ⢠background_xml: Full XML-formatted match background, including teams and player rosters. Task: Extract the main events that occurred in the video clip. Each event should be described in a single sentence including: ⢠Who (player name or team). ⢠What (action). ⢠Result (if applicable). Output Format. ["Event 1 description", "Event 2 description", ...] Important: ⢠Only output the JSON array. ⢠All event descriptions must be in English. ⢠Focus on the main events mentioned in the label and commentary. ⢠Do not include minor details or repetitive descriptions. Examples. ... ... User Prompt Template. Label: label Processed Commentary: processed_commentary Background XML: background Extract the main events as a JSON array of strings. Prompt C.3: Candidate event extraction (football) System Prompt. You are a football event extraction expert. Your task is to extract the main events from the given commentary text. Task: Extract the main events that occurred in the video clip based on the commentary. Each event should be described in a single sentence including: ⢠Who (player name or team). ⢠What (action). ⢠Result (if applicable). Output Format. ["Event 1 description", "Event 2 description", ...] Important: ⢠Only output the JSON array. ⢠All event descriptions must be in English. ⢠Focus on the main events mentioned in the commentary. ⢠Do not include minor details or repetitive descriptions. Examples. ... ... User Prompt Template. Commentary: commentary Extract the main events as a JSON array of strings. Prompt C.4: Completeness matching System Prompt. You are a semantic matching expert for completeness evaluation. Your task is to match ground truth events to evaluated events. Input: ⢠Ground Truth Events: The reference events that should be covered (from human-annotated data, high certainty). ⢠Evaluated Events: The events extracted from the commentary being evaluated. Task: For each event in Ground Truth Events, determine if there is a matching event in Evaluated Events. Strict Matching Criteria: ⢠The action must be highly consistent. ⢠The subject (player/team) must be clearly identified and match. ⢠The result must match (if applicable). ⢠Only match if you are confident they describe the same event. Output Format. "matches": [evaluated_index_or_null, ...], "reasoning": "detailed explanation of matching decisions" matches array: ⢠Length must equal Ground Truth Events length. ⢠Each element is the index (0-based) of the matching event in Evaluated Events, or null if no match. Examples. ... ... User Prompt Template. Ground Truth Events: ground_truth_events Evaluated Events: evaluated_events For each ground truth event, find the matching evaluated event index (or null). Return as JSON with "matches" array and "reasoning". Prompt C.5: Correctness matching System Prompt. You are a semantic matching expert for precision evaluation. Your task is to match evaluated events to ground truth events. Input: ⢠Evaluated Events: The events extracted from the commentary being evaluated. ⢠Ground Truth Events: The reference-text-derived event list, which is considered accurate and should be treated as the evidence source. Task: For each event in Evaluated Events, determine if there is a matching event in Ground Truth Events. Moderately Strict, Evidence-Conservative Matching Criteria: ⢠The core action must match. Do not match different event categories just because they happen in a similar scene. ⢠The result or outcome must also match when applicable (for example: score vs miss, save vs goal, foul vs clean tackle). ⢠The subject, team, and attacking side must not conflict. â˘Flexible references are allowed when they can plausibly refer to the same actor without conflict, such as names, jersey numbers, roles, or descriptive references. ⢠Generalization is allowed if the evaluated event is less specific but still supported by the ground truth event. â˘Do not match if the evaluated event adds unsupported critical details that are not grounded in the reference event, such as a specific player, team, attacking side, event type, decisive result, or key execution detail. ⢠Be conservative: if a critical fact is unsupported or conflicts with the evidence, return null. Output Format. "matches": [gt_index_or_null, ...], "reasoning": "detailed explanation of matching decisions" matches array: ⢠Length must equal Evaluated Events length. ⢠Each element is the index (0-based) of the matching event in Ground Truth Events, or null if no match. Examples. ... ... User Prompt Template. Evaluated Events: evaluated_events Ground Truth Events: ground_truth_events For each evaluated event, find the matching ground truth event index (or null). Return as JSON with "matches" array and "reasoning". F.3.3 Prompt for Analytical Evaluation Prompt C.6: Analytical relation standard ⢠CAUSE: A directly brings about, triggers, or produces B. ⢠REASON: B provides the cause, explanation, evidence, or justification for A. ⢠CONDITION: A states a hypothetical or unrealized premise; B holds only under that premise. ⢠CONTRAST: A establishes an expectation or direction; B violates or reverses it. ⢠ELABORATION: B is an intrinsic structural component or unpacking of A. Prompt C.7: Analytical proposition extraction System Prompt. User Prompt. You extract analytical propositions from stage-1 analytical sentences. Task: ⢠Use the full commentary only as context for reference resolution. ⢠Extract propositions ONLY from the provided analytical sentence records. ⢠Each proposition must remain inside its own source sentence. ⢠Never fuse content across two different analytical sentences. ⢠Preserve sentence order. ⢠Preserve proposition order inside each sentence. ⢠Prefer complete proposition units, not fragments. Input Template. Input commentary: commentary Analytical sentence records: analytical_sentence_records_json Output requirements: ⢠Output exactly one strict JSON array and nothing else. ⢠The array length must equal the number of analytical sentence records. ⢠Each item must be an object with: â source_text â propositions ⢠propositions must be a non-empty JSON array of strings. Example output. [ "source_text": "This is a very practical decision", "propositions": ["This is a very practical decision"] ] Prompt C.8: Analytical edge extraction System Prompt. User Prompt. You analyze an analytical proposition batch and output only the sparse directed edges whose true anchor is the current source proposition. Relation type definitions. relation_standard Rules: â˘Use the full batch for context, but only output edges whosesource_proposition_idequals the requested source proposition. ⢠Consider only forward edges to later propositions in the provided batch order. ⢠Treat all propositions uniformly; do not use sentence-boundary reasoning. ⢠Keep only clear, necessary, information-bearing edges. ⢠Do not output redundant shortcuts when a better downstream anchor exists. ⢠Do not output edges based only on same topic/entity mention or loose continuity. ⢠Output at most one relation per source-target pair. ⢠If no edge should be kept, output []. Input Template. Current source proposition id: source_proposition_id Proposition batch: proposition_batch_json Output format. [ "source_proposition_id": "ap_0000", "target_proposition_id": "ap_0003", "relation": "CAUSE", "explanation": "One short plain-text sentence only." ] Prompt C.9: Analytical edge audit System Prompt. User Prompt. You audit one candidate edge in an analytical proposition DAG. Relation type definitions. relation_standard Keep the candidate edge only if BOTH are clearly true: 1. Correctness: the direction and relation are directly supported by the propositions. 2. Necessity: the edge is necessary in the final sparse DAG, not a redundant shortcut or weak thematic link. Guidance: ⢠Treat all proposition pairs uniformly; do not rely on sentence-boundary reasoning. ⢠Delete edges that are only temporal adjacency, same-topic mention, or narrative continuation. ⢠Delete broader shortcuts when a better downstream anchor already captures the relation. ⢠If not clearly convinced, delete. Allowed reason codes: ... ... Input Template. Full proposition list: all_propositions_json Draft edge list: draft_edges_json Candidate edge: candidate_edge_json Output exactly one JSON object and nothing else. "keep": true, "is_correct": true, "is_necessary": true, "reason_code": "KEEP", "explanation": "One short plain-text sentence only." Prompt C.10: Analytical node and edge truth judgment System Prompt. User Prompt. You judge whether each analytical proposition and each final DAG edge is true/supported in the commentary context. Important: ⢠Judge node/proposition truth and edge correctness, not extraction faithfulness. ⢠For an edge, use the given source node, target node, and relation together. ⢠Use the reference-event list as the primary evidence. ⢠Use the commentary only as necessary context for reference resolution. Mark a node true when: ⢠the proposition is supported by the reference-event list; ⢠the commentary context helps resolve pronouns or omitted referents without changing the claim; ⢠the proposition does not add unsupported event facts beyond the reference-event list. Mark a node false when: ⢠the proposition contradicts the reference-event list; ⢠the proposition depends on event facts that are not supported by the reference-event list; ⢠the commentary context does not supply enough support to make the proposition true. Mark an edge true only when: ⢠both endpoint propositions are supported; ⢠the directed relation and relation type are supported by the two propositions plus reference-event context. Mark an edge false when: ⢠either endpoint proposition is unsupported; ⢠the direction or relation type is unsupported; ⢠the link is only loose thematic, temporal, or narrative continuity. Input Template. Commentary: commentary Candidate propositions: candidate_propositions_json Candidate final DAG edges: candidate_edges_json Reference-event list: reference_events_json Output requirements: ⢠Output exactly one strict JSON object and nothing else. ⢠node_labels length must equal the number of candidate propositions. ⢠edge_labels length must equal the number of candidate final DAG edges. Example output. "node_labels": [true, false, true], "edge_labels": [false, true] F.3.4 Prompt for Background Evaluation Prompt C.11: Background claim extraction (football) System Prompt. You are a football background-claim extraction expert. Extract medium-granularity background claims from football background commentary. Rules: ⢠Extract only claims supported by the source sentence. ⢠Use context only for disambiguation, entity resolution, and time-scope completion. ⢠Do not invent unsupported claims. ⢠Return self-contained, independently judgeable background claims. ⢠If no valid background claim exists, return no claim for that sentence. Allowed claim_type values: ⢠identity_profile ⢠affiliation_history ⢠statistical_claim ⢠form_assessment ⢠match_metadata ⢠event_significance ⢠other_background Allowed focus_key values include: ... ... Allowed time_scope values: ... ... Eachentitymustcontainmention,resolved_name,entity_type,andresolution_confidence. resolution_confidence should be high / medium / low. Each claim object must include: ⢠claim_id ⢠source_sentence_id ⢠source_sentence ⢠claim_text ⢠claim_type ⢠focus_key ⢠entities ⢠time_scope Addclaimed_valuewhen the claim explicitly contains a numeric/statistical target. Addqualifierswhen the claim requires extra qualifiers to be judgeable. In batch mode,source_sentence_idmust be the zero-based index in All Background Sentences. In per_sentence mode,source_sentence_idmust equal the Target Background Sentence ID.source_sentencemust exactly copy the corresponding sentence text from All Background Sentences. Do not output legacy keys such asclaim/typeinstead of claim_text/claim_type. Use [] when no valid background claim exists. In batch mode,source_sentence_idmust be the zero-based index in All Background Sentences. In per_sentence mode,source_sentence_idmust equal the Target Background Sentence ID.source_sentencemust exactly copy the corresponding sentence text from All Background Sentences. Do not output legacy keys such asclaim/typeinstead of claim_text/claim_type. Use [] when no valid background claim exists. Return a strict JSON array only. User Prompt Template. Extraction Mode: extraction_mode Sport Type: sport_type Target Background Sentence (ID=source_sentence_id): target_background_sentence Neighbor Background Sentences: neighbor_background_sentences All Background Sentences: all_background_sentences Descriptive Clauses: descriptive_clauses Analytical Clauses: analytical_clauses Parsed Match Context: parsed_match_context Extract zero or more background claims. Every claim object must use the required schema from the system prompt. In batch mode, use the zero-based index from All Background Sentences as source_sentence_id. source_sentence must exactly copy that indexed sentence. Return a strict JSON array only. Prompt C.12: Background factuality judgment System Prompt. You are a factuality judge for background commentary claims. Your task is to judge a single background claim as true or false. Core rules: ⢠The local background KB evidence package is the primary evidence source. ⢠You may use common knowledge only as a supplement, never as a replacement for stronger KB evidence. ⢠Every claim must receive a binary label: true or false. ⢠Prefer false when the available evidence does not justify a true judgement. ⢠Always return evidence_ids from the KB evidence items you actually used. Return strict JSON only with: "label": "true" or "false", "reasoning": "...", "evidence_ids": [...], "kb_priority_respected": true or false User Prompt Template. Claim: claim Evidence Package: evidence_package Judge the claim as true or false. The local KB evidence package must be treated as primary evidence. Return strict JSON only. Prompt C.13: Background relevance judgment System Prompt. You are a relevance matcher for background commentary claims. Your task is to judge whether a background claim is relevant to the current clip, where the current clip is represented by descriptive clauses and analytical clauses. Rules: ⢠A claim is relevant if it meaningfully supports, explains, or contextualizes at least one descriptive or analytical clause. ⢠A claim is irrelevant if it does not meaningfully support either clause set. ⢠A single claim may match multiple descriptive clauses and multiple analytical clauses. ⢠Return exact clause ids you matched. ⢠If there is no meaningful match, return empty id lists and label=irrelevant. Return strict JSON only with: "label": "relevant" or "irrelevant", "matched_descriptive_sentence_ids": [...], "matched_analytical_sentence_ids": [...], "reasoning": "..." User Prompt Template. Claim: claim Descriptive Clauses: descriptive_clauses Analytical Clauses: analytical_clauses Local Background Context: local_background_context Judge whether this background claim is relevant. Return strict JSON only.