Decide, Don't Generate: Competitive Dimensional ABSA with Jev's Typed Decisions

• A title-filtered literature audit finds text-generative models in 38% of ABSA methods in 2021–2023 and about 65% in 2024–2025, prompting the question of whether competitive sentiment analysis requires generation.Source: §1 · p. 1
• Dimensional ABSA extends polarity analysis with continuous valence and arousal across given-aspect regression, triplet extraction, and categorized quadruplet prediction, with exact spans required for extraction credit.Source: §1 · p. 2
• The pipeline uses Jev’s rubric scores, label probabilities, and yes/no judgments with 488 CPU-fitted coefficients; it generates no text and updates no backbone weights.Source: §1 · p. 2, Figure 2 · p. 2
• 1.0645 RMSE is the lowest ten-corpus Task 1 aggregate of any participating system, while Tasks 2 and 3 score 52.09 and 44.06 cF1 above fine-tuned Llama-3.3-70B and GPT-OSS-120B.Source: §1 · p. 2

2.1 ABSA formulations

• Dimensional ABSA combines continuous valence–arousal ratings with exact tuple structure, and this paper revisits direct prediction by constructing tuples through candidate selection rather than text decoding.Source: §2.1 · p. 2

2.2 DimABSA systems and calibration

• Prior DimABSA systems use backbone adaptation, retrieval and ensembling, or repeated generation; this pipeline instead combines BM25 examples with Jev decisions and calibrates VA scores using benchmark labels.Source: §2.2 · p. 3

3.1 Tasks and decision interface

• Task 1 predicts valence-arousal ratings for a given aspect, while Tasks 2 and 3 predict aspect-opinion tuples and categorized quadruplets, respectively.Sentiment is represented as a two-dimensional vector on the benchmark’s 1–9 scale; explicit terms are input substrings, while permitted implicit aspects use NULL.Source: §3.1 · p. 3
• CHOICE supplies label probabilities, NOUL binary judgments, and SCORE rubric-level distributions with an expected zero-based index; questions target the aspect rather than the whole review.For valence and arousal, the method maps SCORE outputs to the benchmark scale.Source: §3.1 · p. 3

3.2 Task 1: given-aspect regression

• Task 1 uses two SCORE questions per aspect and nine fixed, training-derived demonstrations; corpus-specific regression calibrates raw valence and arousal outputs.The calibration standardizes features, includes extremity and interaction terms, and selects its ridge penalty with grouped cross-validation; development data select the calibration family.Source: §3.2 · p. 3

3.3 Task 2: dimensional triplet extraction

• Task 2 builds aspect and opinion candidates from token-level BIO probabilities, scores spans heuristically, applies boundary and relation checks, then uses a logistic reranker and greedy decoding to select pairs.Checks include exact-phrase and relation-convention judgments; reranker features combine model judgments, token support, boundary checks, and candidate statistics.Source: §3.3 · pp. 3–4

3.4 Task 3: category enrichment

• Task 3 scores categories for selected Task 2 pairs using model probabilities, add-one priors, and aspect and opinion lexical lookups, then attaches the top category without changing spans or valence-arousal values.The category feature weights are fitted on training gold pairs, with counts and retrieval excluding annotations sharing the sampled review’s normalized text.Source: §3.4 · p. 4

4.1 Setup

• Task 1 evaluates ten corpora in six languages, while Tasks 2 and 3 cover eight non-finance corpora; official scoring uses micro RMSE for Task 1 and macro cF1 for Tasks 2/3.The shared extraction test set contains 6,690 reviews; test labels are used only for evaluation.Source: §4.1 · pp. 4–5, Table 1 · p. 5

4.2 Competitive without generation or tuning

• 1.0645 micro RMSE is the lowest aggregate among 14 teams reporting all ten corpora, while Tasks 2 and 3 reach 52.09 and 44.06 cF1, above fine-tuned Llama-3.3-70B and GPT-OSS-120B baselines.The system trails the best published per-corpus score on eight of ten Task 1 corpora; extraction scores fall between sixth and seventh place for Task 2 and fifth and sixth for Task 3 among complete-coverage participants.Source: §4.2 · pp. 5–6, Table 1 · p. 5

4.3 What each component contributes

• Task 1 calibration removes about half the raw-score error, while Task 2 loses 15.31 points without the learned reranker and Task 3 loses 6.03 cF1 without model category probabilities.For Task 2, restricting candidates to argmax BIO spans loses 6.09 points; for Task 3, model category decisions and the prior without lookups reach 44.08 against the full fusion’s 44.06.Source: §4.3 · p. 6, Table 3 · p. 6

4.4 Remaining extraction error is structural

• Gold valence-arousal values would add only 4.46 cF1 to Task 2; on average, 21.7% of gold pairs are never proposed and 26.8% are proposed but not selected.The 8.03-point Task 2-to-Task 3 loss is the smallest among systems matching or exceeding this system on both tasks; its 5.14-point Task 3 gap to PALI is attributed to Task 2 pairs.Source: Table 2 · p. 5, §4.4 · p. 7, Table 5 · p. 7

5 Conclusion

• Using Jev’s typed decisions, corpus statistics, and 488 CPU-fitted coefficients, the system achieves the lowest ten-corpus Task 1 aggregate among participating systems and beats fine-tuned Llama-3.3-70B and GPT-OSS-120B on extraction.It does so without text generation or backbone tuning.Source: §5 · p. 7
• The remaining extraction error lies in proposing and selecting span boundaries rather than sentiment values or categories.The learned combination of boundary evidence carries extraction.Source: §5 · p. 7
• Closing the extraction gap and comparing decision composition with generative systems under matched latency and cost conditions are proposed next steps.Source: §5 · p. 7

A Dataset Details

• Task 1 covers ten corpora in six languages and four domains; Tasks 2 and 3 cover the eight non-finance corpora.Source: A Dataset Details · p. 9
• Task 1 test data contain 9,658 reviews and 16,186 aspect annotations.Source: A Dataset Details · p. 9
• In the English laptop test set, Task 3 has 1,975 quadruplets and Task 2 has 1,974 triplets, so each task is evaluated against its own gold file.Task 3 reuses predicted Task 2 pairs, which does not require the gold inventories to agree.Source: A Dataset Details · p. 9

B Implementation and Reproducibility

• Task 2 rerankers use development labels, and their out-of-fold scores are not nested estimates of full feature and threshold selection.The final revision gained 1.50 cF1 out of fold and 0.41 on test while changing several components; folds also do not group parallel Russian/Tatar/Ukrainian translations.Source: B Implementation and Reproducibility · p. 9
• Task 1 calibration groups repeated texts and parallel translations, excludes fixed demonstrations and development/test overlaps, and replaces a calibration only after specified development and bootstrap checks.The procedure uses 256 training text groups per corpus, with ridge penalties selected from seven values.Source: B Implementation and Reproducibility · p. 9
• Overlap filtering excludes retrieved examples, but Task 2 lexicon and boundary counts still use the full training split.The audit found 23 train–test text overlaps in Japanese hotel and two normalized-text overlaps in the Task 2 subset.Source: B Implementation and Reproducibility · p. 9
• Task 2 uses corpus-specific retrieval and boundary settings, disables implicit aspects under stated rate conditions, and retrieves examples from the same corpus.Japanese hotel is the only corpus retaining the stated implicit-aspect policy.Source: B Implementation and Reproducibility · p. 9
• Task 3 category inventories vary by corpus, and fitting samples cover at least 1,000 annotated pairs per corpus with normalized-text review annotations left out of count features.The implemented inventories range from 12 categories in most restaurant corpora to 113–121 in laptop corpora.Source: B Implementation and Reproducibility · p. 9
• Tasks 2 and 3 share reviews, and Task 3 has one more English laptop test annotation; external comparisons use published rounded scores.Source: B Implementation and Reproducibility · p. 9, Table 6 · p. 10
• Each category variant refits its fusion weights, and the category glossary uses the eligible training pool even when sampled text is omitted from demonstrations.Thus “model + prior” is not the raw model argmax.Source: B Implementation and Reproducibility · p. 10

C Interpreting the Diagnostics

• 4.46 cF1 is the gain from scoring extracted Task 2 pairs with gold valence-arousal, leaving the extracted pairs unchanged.Source: C Interpreting the Diagnostics · p. 10
• The exact-VA diagnostic bounds gains from improving VA alone for the fixed pair set, but says nothing about candidate coverage, recall, or category selection.Source: C Interpreting the Diagnostics · p. 10
• Category accuracy conditions on matched pairs and ignores missing or spurious pairs; the Task 3-to-Task 2 cF1 ratio is close to, but not identical to, that accuracy.VA weights and gold tuple counts also enter the ratio.Source: C Interpreting the Diagnostics · p. 10

Search scope

• The audit searches selected ACL Anthology and AAAI proceedings using titles containing both “aspect” and “sentiment” or the standalone abbreviation “ABSA”; workshops and demonstrations are excluded.It additionally searches official NeurIPS, ICML, and ICLR records; no matching titles there does not mean those venues publish no ABSA research.Source: D Literature-Audit Protocol · p. 10

Candidate selection

• The search retained 277 papers from 293 venue-eligible candidates after excluding 16 dataset-only, diagnostic, or non-ABSA-prediction papers.Each included paper contributes once; comparison baselines and related-work mentions do not determine its category.Source: D Literature-Audit Protocol · pp. 10–11

Model-use codebook

• The codebook classifies model use across a proposed method’s stages: text-generative use counts as Generative, direct decisions without it as Discriminative, and latent or unresolved cases as Other.Masked-language-model encoders and task-specific pointer or transition decoders do not qualify as Generative on their own.Source: D Literature-Audit Protocol · p. 11

Unresolved records

• Seven of 30 papers in the AAAI expansion remain unresolved because full texts were unavailable for part of that expansion, so their eligibility is known but generative dependencies cannot be established.They remain in the denominator; the unresolved classification is retained as Other rather than assumed non-generative.Source: D Literature-Audit Protocol · p. 11

Aggregation and limitations

• 53 of 82 papers in 2024–2025 have identified text-generative use; assigning the four unresolved papers to Generative raises the share from 64.6% to 69.5%.The period totals pool papers, retain Other in the denominator, and omit years without matching papers.Source: D Literature-Audit Protocol · p. 11
• The audit cannot establish the field’s original method distribution from 2004–2013, which contains only four papers; unequal periods, venue coverage, title vocabulary, and task composition also limit interpretation.Coding used model assistance and targeted source checks without independent double annotation, so no interannotator agreement is available.Source: D Literature-Audit Protocol · p. 11

E.1 Shared valence–arousal rubrics

• For Tasks 1 and 2 pair scoring, the shared nine-level rubric asks about the sentiment or feeling directed at the target aspect, with arousal defined as calm-to-excited activation.When an aspect is NULL, the template describes it as implicit and never named in the text.Source: §E.1 · pp. 11–12, Table 8 · p. 12

E.2 Task 1: given-aspect regression

• For each review and aspect, Task 1 asks separate SCORE questions for valence and arousal, using nine fixed labeled demonstrations only to calibrate the 1–9 scale.The rubric uses nine ordered levels, while the benchmark supplies the scale and short dimension definitions.Source: §E.1 · p. 11, §E.2 · p. 12, Table 8 · p. 12

E.3 Task 2: dimensional triplet extraction

• Task 2 labels aspect and opinion tokens with B/I/O choices, then checks candidate spans and pairs against annotation boundary conventions and whether the opinion evaluates the aspect.Boundary guidance preserves compounds, negation, and degree modifiers while separating distinct coordinated targets or opinions.Source: §E.3 · pp. 12–13
• Task 2’s candidate-pair questions require complete annotated spans and a direct evaluative link; for implicit aspects, the question instead asks whether the opinion targets an implicit aspect.The same wording supports opinion extensions and retrieval views, while changing candidates and retrieved examples.Source: §E.3 · p. 13
• Task 2 also asks SCORE questions for valence and arousal on selected aspect–opinion pairs, then applies fitted affine calibration to the final system’s rescores.The pair-scoring questions use the shared nine-level criteria without the Task 1 demonstration clause or focus prefix.Source: §E.3 · pp. 13–14

E.4 Task 3: category enrichment

• Task 3 asks CHOICE to assign a category to each predicted pair using the corpus category inventory, retrieved annotated reviews, and a glossary of frequent training aspects.The fusion model combines CHOICE probabilities with training statistics; categories are offered only if observed in the eligible training pool, with attribute descriptions in Table 9.Source: §E.4 · p. 14
• Task 3 reuses Task 2’s valence–arousal predictions, and its category assignment does not change the predicted pair spans or those values.The category options encode an entity name followed by its attribute description.Source: §E.4 · p. 14, Table 9 · p. 14