Pith. sign in

REVIEW 3 major objections 6 minor 3 cited by

Large Language Models Enhanced by Plug and Play Syntactic Knowledge for Aspect-based Sentiment Analysis

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A memory-based plugin encoding constituent syntax, word dependencies, and CCG supertags can be trained separately and plugged into a frozen LLM to improve aspect-based sentiment analysis, the paper argues.

desk verdict A useful plug-and-play ABSA plugin that consistently helps, but the paper never shows syntax—rather than the plugin's own BERT encoder—is doing the work. read the letter →

arxiv 2506.12991 v1 pith:4Y6U5HOC submitted 2025-06-15 cs.CL

Yuanhe Tian , Xu Li , Wei Wang , Guoqing Jin , Pengsen Cheng , Yan Song This is my paper
classification cs.CL
keywords aspect-basedsentimentanalysislargelanguagemodelsplug-and-playpluginmemory-basedknowledgeinjectionconstituentsyntaxdependencyrelationsCCGsupertagsfrozenLLMadaptation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that a large language model can be improved for aspect-based sentiment analysis without being retrained, by attaching a separately trained memory-based plugin that supplies syntactic knowledge. The plugin encodes three kinds of syntax from off-the-shelf tools: constituency phrases, dependency relations linked to the aspect term, and CCG supertags, storing them as key-value memories that are read out by a BERT encoding of the sentence and aspect. The read-out vector is injected into the frozen LLM through a small hub module, either tuned alongside the plugin or used purely as a prompt hint. Across five benchmark datasets, every syntactic plugin improves over the vanilla LLM, and the three-plugin combination beats the compared baselines and previous systems, with the largest gain on MAMS (81.42 to 86.05 accuracy with LLaMA-2). If the claim holds, syntactic information can be treated as a swappable, low-cost component for adapting LLMs to fine-grained sentiment tasks.

What carries the argument

The load-bearing mechanism is the memory-based plugin: a trainable key-value store in which each key $k_m$ and value $v_m$ is an embedding of a syntactic construction, and the read-out is the attention-weighted sum of values given a BERT query over the sentence and aspect. Constituent syntax contributes keys for words inside the longest aspect-containing phrase and values pairing those words with the phrase label; dependency knowledge contributes ranked (word, relation) pairs from first- and second-order dependencies of the aspect; CCG contributes token-supertag pairs in a $\pm 3$-word window. The hub module, a multilayer perceptron, is what lets the plugin output enter the frozen LLM, and the two training strategies determine whether the plugin is tuned with the LLM frozen (Strategy 1) or trained alone and consulted through the prompt (Strategy 2). The memory size is set to five instances per plugin, chosen because larger memories add low-frequency, noisy context.

What would settle it

Run the exact same plugin and hub with the syntactic key-value entries replaced by randomly initialized or non-syntactic entries (for example, arbitrary word pairs with random relation labels), holding training and evaluation identical: if accuracy stays at the level of the syntactic plugins, the claim that syntax drives the improvement is falsified.

Watch

Extended reading notes

Core claim

The central claim is that plugging syntactic knowledge into an LLM through a memory-based plugin improves aspect-level sentiment prediction, and that this can be done with the LLM's parameters frozen. The plugin computes a query vector $h_{XA}$ by encoding the concatenation of sentence and aspect with BERT, scores memorized key vectors $k_m$ by dot-product softmax, and returns a weighted sum of value vectors $v_m$; the hub MLP projects this output into the LLM's embedding space, where it is prepended to the input embeddings (Strategy 1), or the plugin's own predicted label is inserted into the prompt (Strategy 2). Using dependency relations, constituency syntax, and CCG supertags as the key-value content, the full LLaMA-2 model with all three plugins achieves the best reported accuracy and F1 on LAP14, REST14, REST15, REST16, and MAMS, and the Qwen-2.5 plugin version also improves on SemEval-2015 joint ABSA. The authors conclude that the plugin coordinates multiple knowledge types, assigns higher memory weights to sentiment-relevant dependencies, and adapts LLMs to ABSA with less training time than full fine-tuning.

Load-bearing premise

The paper's central claim rests on the assumption that the accuracy gains come from the syntactic content stored in the memory, but no experiment varies that content, so the gains could in principle come from the plugin's BERT encoder and its trained sentiment classifier rather than from syntax.

Editorial extensions

If this is right

  • If the central claim is right, aspect-based sentiment analysis can be improved without touching the LLM's weights, cutting training time from over 12 hours to about 3 hours in the Qwen-2.5 experiments.
  • Adding a second or third syntactic plugin yields consistent further gains, so the knowledge sources appear complementary rather than redundant.
  • Strategy 2 shows that a plugin trained entirely without the LLM can still transfer its syntactic signal, making the approach usable when full fine-tuning is infeasible.
  • The same memory read-out works when the aspect term is not given in advance, as the SemEval-2015 joint ABSA result indicates, so the plugin need not be redesigned for aspect extraction.
  • Any knowledge expressible as key-value instances could, in principle, be swapped into the same plugin architecture, making the syntactic results a special case.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • What the paper does not ablate is the knowledge content of the memory: replacing the syntactic keys and values with random or non-syntactic tokens while keeping the BERT encoder and hub would tell whether the gain comes from syntax or from the plugin's own classifier.
  • Because the largest accuracy jump is on MAMS (+4.63) while LAP14 and REST14 gain only about 0.6 to 2.3 points, the plugin may matter most when the vanilla LLM is weakest; the benefit should be expected to vary with the base model and dataset difficulty.
  • In Strategy 2, the plugin's predicted label is part of the prompt, so part of the improvement may be a form of self-ensembling; comparing against a prompt containing a label from a non-syntactic classifier would separate the syntactic effect from the ensemble effect.
  • The architecture is stated to be knowledge-agnostic, so the same memory plugin could host semantic roles, sentiment lexicons, or topic memories; testing those would clarify whether syntax specifically, or structured knowledge generally, drives the improvement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a memory-based plugin that encodes three types of syntactic knowledge (constituent syntax, word dependencies, and CCG supertags) and integrates the encoded vector into a frozen LLM via a hub module. Two strategies are presented: joint tuning of the plugin and hub with the LLM frozen, and independent plugin training followed by prompt injection of the plugin's predicted label. Experiments on five ABSA benchmarks and one joint ABSA setting with Qwen-2.5 and LLaMA-2 report consistent improvements over vanilla LLMs and several prior works, with the full three-plugin model achieving the best results.

Significance. If the claimed improvements are attributable to syntactic knowledge, the plugin would be a practical, lightweight way to inject structured knowledge into LLMs for ABSA, and the code release would facilitate adoption. The evaluation covers multiple datasets, two LLMs, and two integration strategies, and the memory-size analysis is a useful diagnostic. However, the current evidence does not isolate the effect of syntax from the effect of the added fine-tuned encoder, and the lack of significance tests weakens the quantitative claims. The work's value depends on new ablations; as presented, the central attribution claim is not fully supported.

major comments (3)
  1. [§2.1, §4.1 (Table 2)] The plugin in Eqs. (2)-(4) uses a fine-tuned BERT encoder to produce the query h^XA, and the keys/values are constructed from syntactic features. The reported gains over the vanilla LLM in Table 2 could therefore be caused by the added task-specific BERT representation rather than by the syntactic content of the memory. The paper does not include a control ablation that replaces the syntactic keys/values with random or non-syntactic embeddings while keeping the encoder, hub, and training procedure identical. The RAG baseline in Table 3 is not a valid control because it lacks the trained encoder and hub. Please add such an ablation to demonstrate that the syntactic knowledge, rather than the plugin's encoder or added capacity, is the active ingredient.
  2. [§2.3 (Strategy 2), Table 5] In Strategy 2, the plugin's predicted label y^P is inserted into the prompt template T', and the LLM is then asked to predict the sentiment. This creates a circularity because the LLM may simply copy the plugin's label, and Table 5 does not report the agreement rate between the plugin and the LLM. As a result, the Strategy 2 results do not show that the LLM benefits from the syntactic knowledge independently of the plugin's own classifier. Please report the overlap between y^P and the final prediction, or run a control condition with a random or majority-class placeholder label to quantify the LLM's reliance on the injected label.
  3. [§3.5, §4.1 (Table 2)] The results in Table 2 are averages over three random seeds, but no standard deviations or significance tests are reported. Several improvements are small (e.g., LAP14: 82.76 to 83.33 for Qwen-2.5; REST14: 85.71 to 87.98 for Qwen-2.5), so without variance estimates it is unclear whether these differences are statistically meaningful. Please report standard deviations and run a paired significance test (e.g., bootstrap) for the main comparisons against the vanilla LLM.
minor comments (6)
  1. [Abstract and §4.1] The abstract says 'outperforms strong baselines and previous approaches' and the conclusion says 'achieves state-of-the-art results on all datasets,' but in Table 3, Chai et al. (2023) reports a higher accuracy on REST16 (93.42 vs 93.32). Please qualify the state-of-the-art claim (e.g., with respect to F1) or adjust the wording.
  2. [§2.3, Eq. (6)] The text says 'we compute the sum of oP and the query vector hXA,' but Eq. (6) shows concatenation (⊕). Please make the operation consistent.
  3. [§3.2] The notation for the number of constituent keys/values (Mc) is not defined, whereas Md and Mt are defined in §3.5. Please define Mc or clarify the memory size for the constituent plugin.
  4. [§4.1, Table 2] The row for 'Qwen-2.5 (LoRA) +P(D)' appears to contain inconsistent values: REST15 accuracy 92.02 is much higher than all other models, and REST14 F1 74.35 is lower than the vanilla model's 76.90. Please check the numbers for alignment and typos.
  5. [§2.2] The description 'concatenate the embedding matrix E with the vector hP' is imprecise; since E is a matrix, the operation must be an appended token or a projection. Please clarify.
  6. [§3.5] The plugins are trained with BERT as the encoder; this means the plugin is not a purely symbolic memory but a learned encoder. The paper should acknowledge this in the description of what is 'plugged' into the LLM.

Circularity Check

1 steps flagged · score 6.0 of 10

Strategy 2's 'plug-and-play' result is partially circular: the plugin is a trained ABSA classifier whose predicted label is inserted verbatim into the LLM prompt, so the reported final prediction can reduce to the plugin's own fitted output, with no deviation or control analysis to isolate an independent LLM judgment.

  1. fitted input called prediction [Section 2.3 (Strategy 2), Eq. (6); Table 5]
    "To train the plugin on ABSA, we compute the sum of oP and the query vector hXA and pass the resulting vectors through a softmax classifier fP to predict the sentiment byP of the input ... In inference, when the plugin is working with the LLM, we firstly use the plugin to predict the sentiment label byP. Then we use the input sentence X, the aspect term A, and the plugin prediction byP to fill a prompt template T′ that has a particular slot to fill the plugin prediction. ... Afterwards, we feed the filled template into the LLM and follow the standard decoding process to get the prediction by."

    Eq. (6) defines the plugin as a full ABSA classifier trained with cross-entropy on gold labels. Strategy 2 then feeds that fitted classifier's test-time prediction byP directly into the LLM prompt, and final by is decoded from a template that already contains byP. If the chat-tuned LLM follows the prompt, by equals byP, so Table 5's Strategy 2 accuracy is the plugin's own fitted accuracy relabeled as an LLM-plus-syntax result. The paper reports no agreement statistics between by and byP and no control with the label slot removed or filled with a non-syntactic placeholder, so the LLM's and syntax's contributions cannot be separated from the plugin's trained sentiment classifier.

full rationale

Aside from Strategy 2, the paper's derivation chain is not circular under the rubric. The memory plugin (Eqs. 2-4) is a standard key-value attention mechanism over syntactic entries; the hub (Eq. 5) is an MLP; and the Table 2 results are external benchmark numbers, not consequences of the model's own assumptions. Self-citations appear (Tian et al. 2020 as the CCG toolkit; Tian et al. 2021 as motivation for syntax) but they are not load-bearing proof: the syntactic knowledge is still empirically injected and evaluated against external baselines. The largest weakness in the main experiments is an attribution confound, not circularity: the plugin adds a fine-tuned BERT encoder and MLP alongside the syntactic keys and values, and no ablation with random or non-syntactic keys/values isolates syntax as the active ingredient. That is a correctness risk. The circular component is specifically Strategy 2, where the plugin's trained sentiment prediction is written into the prompt, so the 'plug-and-play' result can collapse by construction into the plugin's own classifier. Because the paper's plug-and-play claim is explicitly supported by Table 5's Strategy 2, the overall score is 6 rather than lower; the main Table 2 strategy is not circular, so the paper is not scored higher.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper's contribution rests on several hand-chosen hyperparameters (memory size 5, phrase length <10, dependency order 1-2, CCG window ±3) and on the assumption that syntax parsers and the memory attention mechanism faithfully transmit useful information. No new physical or formal entities are introduced. The most important unmeasured assumption is that the plugin's BERT encoder is not the true driver of the gains.

free parameters (6)
  • Memory size (number of keys/values) = 5
    Set based on Figure 3 analysis on MAMS; performance peaks around 5 and plateaus. This is a hand-chosen hyperparameter.
  • Constituent phrase length threshold = <10 words
    Used to extract the longest phrase containing the aspect term (Section 3.2); no sensitivity analysis is provided.
  • Dependency order = first- and second-order
    Authors report that first-order only and third-order give worse performance (Section 3.3); this is a model selection choice.
  • CCG context window = ±3 words
    Hand-chosen window around the aspect term for CCG supertags (Section 3.4).
  • Number of selected dependency pairs = M_d = 5
    Same as memory size; only the most frequent pairs in training data are used (Section 3.3).
  • Training hyperparameters (learning rate, batch size, epochs) = not reported
    The central result depends on these, but they are not disclosed, affecting reproducibility.
assumptions (4)
  • domain assumption Off-the-shelf parsers (Stanza, NeST-CCG) provide accurate syntactic analyses for the input sentences.
    The plugin relies on these parses to construct keys and values (Sections 3.2-3.4). If the parses are noisy, the knowledge injected is degraded.
  • domain assumption The specific syntactic features (constituent phrases, word dependencies, CCG supertags) are useful for ABSA.
    Prior work supports this, but the paper does not prove that its particular encoding of these features is optimal.
  • ad hoc to paper The memory attention mechanism can effectively weigh the relevant syntactic knowledge for the LLM.
    This is the central design choice; no alternative architecture comparison is provided.
  • ad hoc to paper In Strategy 2, the LLM will use the plugin's predicted label as a helpful hint rather than blindly copying it.
    The paper does not analyze the agreement between plugin and LLM predictions, so this assumption is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models Enhanced by Plug and Play Syntactic Knowledge for Aspect-based Sentiment Analysis." pith.science (2026). https://pith.science/paper/4Y6U5HOC

@misc{pith2026250612991,
  author       = {Pith},
  title        = {Pith review of: Large Language Models Enhanced by Plug and Play Syntactic Knowledge for Aspect-based Sentiment Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4Y6U5HOC}},
  note         = {Machine review of arXiv:2506.12991}
}
read the original abstract

Aspect-based sentiment analysis (ABSA) generally requires a deep understanding of the contextual information, including the words associated with the aspect terms and their syntactic dependencies. Most existing studies employ advanced encoders (e.g., pre-trained models) to capture such context, especially large language models (LLMs). However, training these encoders is resource-intensive, and in many cases, the available data is insufficient for necessary fine-tuning. Therefore it is challenging for learning LLMs within such restricted environments and computation efficiency requirement. As a result, it motivates the exploration of plug-and-play methods that adapt LLMs to ABSA with minimal effort. In this paper, we propose an approach that integrates extendable components capable of incorporating various types of syntactic knowledge, such as constituent syntax, word dependencies, and combinatory categorial grammar (CCG). Specifically, we propose a memory module that records syntactic information and is incorporated into LLMs to instruct the prediction of sentiment polarities. Importantly, this encoder acts as a versatile, detachable plugin that is trained independently of the LLM. We conduct experiments on benchmark datasets, which show that our approach outperforms strong baselines and previous approaches, thus demonstrates its effectiveness.

Figures

Figures reproduced from arXiv: 2506.12991 by the authors.

Figure 1
Figure 1. The overall architecture of the proposed approach of pluginning LLMs for ABSA. The left part illustrates [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. A part of the dependency tree for the sentence “The environment is fantastic although bar service is poor.”. The aspect term is highlighted by the green background color. The first- and second-order depen￾dencies are marked by the yellow and blue background colors, respectively. 3 Experiment Settings 3.1 Datasets Following previous studies, we run different mod￾els on five English benchmark datasets from differ￾ent … view at source ↗
Figure 4
Figure 4. An example sentence with the aspect term “knife” highlighted in boldface. The background colors illustrate the weights assigned for different dependen￾cies in the memory mechanism of the plugin, where deeper colors refer to higher weights. of knowledge. Meanwhile, our approaches require less training time compared with full tuning, which confirms that our approach is an effective solution for adapting LLMs to ABSA. … view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Text Reinforcement for Multimodal Time Series Forecasting

    cs.CL 2025-08 conditional novelty 6.0 of 10

    Reinforcement learning trains an LLM to generate improved text from time series, improving multimodal forecasting on Time-MMD.

  2. Balanced Training Data Augmentation for Aspect-Based Sentiment Analysis

    cs.CL 2025-07 conditional novelty 6.0 of 10

    DPO-optimized LLM data augmentation with label balancing improves ABSA accuracy and F1 on most English benchmarks, but the balancing benefit is inconsistent.

  3. The Grammar of Transformers: A Systematic Review of Interpretability Research on Syntactic Knowledge in Language Models

    cs.CL 2026-01 conditional novelty 4.0 of 10

    A systematic review of 337 articles shows Transformers handle formal syntax well but perform worse and more variably at the syntax-semantics interface, with the field over-reliant on English and BERT.

Reference graph

Works this paper leans on

61 extracted references · 36 canonical work pages · cited by 3 Pith papers

  1. [1]

    Gianni Brauwers and Flavius Frasincar. 2022. A survey on aspect-based sentiment classification. ACM Computing Surveys, 55(4):1--37

    work page 2022

  2. [2]

    Jiahao Cao, Rui Liu, Huailiang Peng, Lei Jiang, and Xu Bai. 2022. Aspect Is Not You Need: No-aspect Differential Sentiment Framework for Aspect-based Sentiment Analysis . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1599--1609, Seattle, United States

    2022

  3. [3]

    Heyan Chai, Ziyi Yao, Siyu Tang, Ye Wang, Liqiang Nie, Binxing Fang, and Qing Liao. 2023. Aspect-to-Scope Oriented Multi-view Contrastive Learning for Aspect-based Sentiment Analysis . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10902--10913, Singapore

    2023

  4. [4]

    Chenhua Chen, Zhiyang Teng, Zhongqing Wang, and Yue Zhang. 2022. Discrete Opinion Tree Induction for Aspect-based Sentiment Analysis . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2051--2064, Dublin, Ireland

    work page 2022

  5. [5]

    Romany Dekker, Danae Gielisse, Chaya Jaggan, Sander Meijers, and Flavius Frasincar. 2023. Knowledge injection for aspect-based sentiment classification. In International Conference on Database and Expert Systems Applications, pages 173--187. Springer

    work page 2023

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT : P re-training of D eep B idirectional T ransformers for L anguage U nderstanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171--4186

    work page 2019

  7. [7]

    Xuanwen Ding, Jie Zhou, Liang Dou, Qin Chen, Yuanbin Wu, Chengcai Chen, and Liang He. 2024. Boosting Large Language Models with Continual Learning for Aspect-based Sentiment Analysis . arXiv preprint arXiv:2405.05496

    arXiv 2024

  8. [8]

    Ruidan He, Wee Sun Lee, Hwee Tou Ng, and Daniel Dahlmeier. 2018. Effective attention modeling for aspect-level sentiment classification. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1121--1131

    work page 2018

Show all 61 references
  1. [9]

    Binxuan Huang and Kathleen M Carley. 2019. S yntax- A ware A spect L evel S entiment C lassification with G raph A ttention N etworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natura...

    2019

  2. [10]

    Qingnan Jiang, Lei Chen, Ruifeng Xu, Xiang Ao, and Min Yang. 2019. A challenge dataset and effective models for aspect-based sentiment analysis. In EMNLP-IJCNLP-2019, pages 6280--6285

    2019

  3. [11]

    Haiyun Li, Qihuang Zhong, Ke Zhu, Juhua Liu, Bo Du, and Dacheng Tao. 2024. Iterative Data Augmentation with Large Language Models for Aspect-based Sentiment Analysis . arXiv preprint arXiv:2407.00341

    2024 arXiv

  4. [12]

    Shichen Li, Zhongqing Wang, Zheyu Zhao, Yue Zhang, and Peifeng Li. 2025. Exploring model editing for llm-based aspect-based sentiment classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 24467--24475

    2025

  5. [13]

    Xin Li, Lidong Bing, Wai Lam, and Bei Shi. 2018. Transformation networks for target-oriented sentiment classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 946--956

    2018

  6. [14]

    Bin Liang, Jiachen Du, Ruifeng Xu, Binyang Li, and Hejiao Huang. 2019. Context-aware E mbedding for T argeted A spect-based S entiment A nalysis. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4678--4683, Florence, Italy

    2019

  7. [15]

    Shuo Liang, Wei Wei, Xian-Ling Mao, Fei Wang, and Zhiyong He. 2022. B i S yn- GAT +: Bi-Syntax Aware Graph Attention Network for Aspect-based Sentiment Analysis . In Findings of the Association for Computational Linguistics: ACL 2022, pages 1835--1848, Dublin, Ireland

    2022

  8. [16]

    Yunlong Liang, Fandong Meng, Jinchao Zhang, Yufeng Chen, Jinan Xu, and Jie Zhou. 2021. An Iterative Multi-Knowledge Transfer Network for Aspect-Based Sentiment Analysis . In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1768--1780, Punta Cana, Do...

    2021

  9. [17]

    Meng Luo, Hao Fei, Bobo Li, Shengqiong Wu, Qian Liu, Soujanya Poria, Erik Cambria, Mong-Li Lee, and Wynne Hsu. 2024. PanoSent: A Panoptic Sextuple Extraction Benchmark for Multimodal Conversational Aspect-based Sentiment Analysis . arXiv preprint arXiv:2408.09481

    2024 arXiv

  10. [18]

    Yu, and Lijie Wen

    Fukun Ma, Xuming Hu, Aiwei Liu, Yawen Yang, Shuang Li, Philip S. Yu, and Lijie Wen. 2023. AMR -based Network for Aspect-based Sentiment Analysis . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 322--337...

    2023

  11. [19]

    Yukun Ma, Haiyun Peng, and Erik Cambria. 2018. T argeted aspect-based sentiment analysis via embedding commonsense knowledge into an attentive LSTM . In Thirty-Second AAAI Conference on Artificial Intelligence

    2018

  12. [20]

    Qianren Mao, Jianxin Li, Senzhang Wang, Yuanning Zhang, Hao Peng, Min He, and Lihong Wang. 2019. A spect-based sentiment classification with attentive neural turing machines. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 5139--5145

    2019

  13. [21]

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781

    2013 arXiv

  14. [22]

    Gaurav Negi, Rajdeep Sarkar, Omnia Zayed, and Paul Buitelaar. 2024. A Hybrid Approach To Aspect Based Sentiment Analysis Using Transfer Learning . arXiv preprint arXiv:2403.17254

    2024 arXiv

  15. [23]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training L anguage M odels to F ollow I nstructions with H uman F eedback. Advances in Neural Information Processi...

    2022

  16. [24]

    Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer

    Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistic...

    2018

  17. [25]

    Maria Pontiki, Dimitrios Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad Al-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, Bing Qin, Orph \'e e De Clercq, and 1 others. 2016. S emeval-2016 task 5: A spect based sentiment analysis. In International workshop...

    2016

  18. [26]

    Maria Pontiki, Dimitrios Galanis, Harris Papageorgiou, Suresh Manandhar, and Ion Androutsopoulos. 2015. S emeval-2015 task 12: A spect based sentiment analysis. In Proceedings of the 9th international workshop on semantic evaluation (SemEval 2015), pages 486--495

    2015

  19. [27]

    Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014. S em E val-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation ( S em E val 2014) , pages 27--35

    2014

  20. [28]

    Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. S tanza: A python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations,...

    2020

  21. [29]

    Han Qin, Guimin Chen, Yuanhe Tian, and Yan Song. 2021. Improving F ederated L earning for A spect-based S entiment A nalysis via T opic M emories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3942--3954, Online and Punta Cana...

    2021

  22. [30]

    Han Qin, Yuanhe Tian, Fei Xia, and Yan Song. 2022. Complementary L earning of A spect T erms for A spect-based S entiment A nalysis. In Proceedings of the 13th Language Resources and Evaluation Conference

    2022

  23. [31]

    Paul F Simmering and Paavo Huoviala. 2023. Large language models for aspect-based sentiment analysis . arXiv preprint arXiv:2310.18025

    2023 arXiv

  24. [32]

    Yan Song, Chia-Jung Lee, and Fei Xia. 2017. Learning W ord R epresentations with R egularization from P rior K nowledge. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 143--152

    2017

  25. [33]

    Yan Song and Shuming Shi. 2018. Complementary Learning of Word Embeddings. In IJCAI, pages 4368--4374

    2018

  26. [34]

    Yan Song, Shuming Shi, and Jing Li. 2018. Joint L earning E mbeddings for C hinese W ords and T heir C omponents via L adder S tructured N etworks. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4375--4381

    2018

  27. [35]

    Duyu Tang, Bing Qin, Xiaocheng Feng, and Ting Liu. 2016. E ffective LSTM s for T arget- D ependent S entiment C lassification. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers , pages 3298--3307

    2016

  28. [36]

    Hao Tang, Donghong Ji, Chenliang Li, and Qiji Zhou. 2020. Dependency G raph E nhanced D ual-transformer S tructure for A spect-based S entiment C lassification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6578--6588, Online

    2020

  29. [37]

    Siyu Tang, Heyan Chai, Ziyi Yao, Ye Ding, Cuiyun Gao, Binxing Fang, and Qing Liao. 2022. Affective Knowledge Enhanced Multiple-Graph Fusion Networks for Aspect-based Sentiment Analysis . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing,...

    2022

  30. [38]

    Hashimoto

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford Alpaca: An Instruction-following LLaMA model . GitHub repository

    2023

  31. [39]

    Yuanhe Tian, Guimin Chen, and Yan Song. 2021. Aspect-based sentiment analysis with type-aware graph convolutional networks and layer ensemble. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language T...

    2021

  32. [40]

    Yuanhe Tian, Pengsen Cheng, Guoqing Jin, Lei Zhang, and Yan Song. 2025. Representation decomposition for learning similarity and contrastness across modalities for affective computing. arXiv preprint arXiv:2506.07086

    2025 arXiv

  33. [41]

    Yuanhe Tian, Chang Liu, Yan Song, Fei Xia, and Yongdong Zhang. 2024. Aspect-based Sentiment Analysis with Context Denoising . In 2024 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Mexico City, Mexico

    2024

  34. [42]

    Yuanhe Tian, Yan Song, and Fei Xia. 2020. Supertagging C ombinatory C ategorial G rammar with A ttentive G raph C onvolutional N etworks. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6037--6044, Online

    2020

  35. [43]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. LLaMA 2: Open Foundation and Fine-tuned Chat Models . arXiv preprint arXiv:2307.09288

    2023 arXiv

  36. [44]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. A ttention is A ll you N eed. In Advances in N eural I nformation P rocessing S ystems 30 , pages 5998--6008

    2017

  37. [45]

    Olaf Wallaart and Flavius Frasincar. 2019. A hybrid approach for aspect-based sentiment analysis using a lexicalized domain ontology and attentional neural models. In The Semantic Web: 16th International Conference, ESWC 2019, Portoro z , Slovenia, June 2--6, 2019, Proceedings...

    2019

  38. [46]

    Kai Wang, Weizhou Shen, Yunyi Yang, Xiaojun Quan, and Rui Wang. 2020. Relational G raph A ttention N etwork for A spect-based S entiment A nalysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3229--3238, Online

    2020

  39. [47]

    Qianlong Wang, Keyang Ding, Bin Liang, Min Yang, and Ruifeng Xu. 2023. Reducing Spurious Correlations in Aspect-based Sentiment Analysis with Explanation from Large Language Models . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2930--2941, Singapore

    2023

  40. [48]

    Yequan Wang, Minlie Huang, Li Zhao, and 1 others. 2016. A ttention-based LSTM for aspect-level sentiment classification. In Proceedings of the 2016 conference on empirical methods in natural language processing, pages 606--615

    2016

  41. [49]

    Hu Xu, Bing Liu, Lei Shu, and Philip Yu. 2019. BERT post-training for review reading comprehension and aspect-based sentiment analysis. In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technol...

    2019

  42. [50]

    Lu Xu, Lidong Bing, Wei Lu, and Fei Huang. 2020. Aspect sentiment classification with aspect-specific opinion spans. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3561--3567

    2020

  43. [51]

    Wei Xue and Tao Li. 2018. A spect B ased S entiment A nalysis with G ated C onvolutional N etworks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2514--2523

    2018

  44. [52]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  45. [53]

    Chen Zhang, Qiuchi Li, and Dawei Song. 2019. Aspect-based sentiment classification with aspect-specific graph convolutional networks. In EMNLP-IJCNLP, pages 4560--4570

    2019

  46. [54]

    Mao Zhang, Yongxin Zhu, Zhen Liu, Zhimin Bao, Yunfei Wu, Xing Sun, and Linli Xu. 2023 a . Span-level Aspect-based Sentiment Analysis via Table Filling . In ACL-2023, pages 9273--9284, Toronto, Canada

    2023

  47. [55]

    Mi Zhang and Tieyun Qian. 2020. Convolution over hierarchical syntactic and lexical graphs for aspect level sentiment analysis. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3540--3549

    2020

  48. [56]

    Yice Zhang, Yifan Yang, Bin Liang, Shiwei Chen, Bing Qin, and Ruifeng Xu. 2023 b . An Empirical Study of Sentiment-Enhanced Pre-Training for Aspect-Based Sentiment Analysis . In Findings of the Association for Computational Linguistics: ACL 2023, pages 9633--9651, Toronto, Canada

    2023

  49. [57]

    Zheng Zhang, Zili Zhou, and Yanna Wang. 2022. SSEGCN : Syntactic and Semantic Enhanced Graph Convolutional Network for Aspect-based Sentiment Analysis . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human ...

    2022

  50. [58]

    Anping Zhao and Yu Yu. 2021. Knowledge-enabled bert for aspect-based sentiment analysis. Knowledge-Based Systems, 227:107220

    2021

  51. [59]

    Guangmin Zheng, Jin Wang, Liang-Chih Yu, and Xuejie Zhang. 2024. Instruction Tuning with Retrieval-based Examples Ranking for Aspect-based Sentiment Analysis . arXiv preprint arXiv:2405.18035

    2024 arXiv

  52. [60]

    Changzhi Zhou, Dandan Song, Yuhang Tian, Zhijing Wu, Hao Wang, Xinyu Zhang, Jun Yang, Ziyi Yang, and Shuhao Zhang. 2024. A comprehensive evaluation of large language models on aspect-based sentiment analysis. arXiv preprint arXiv:2412.02279

    2024 arXiv

  53. [61]

    Senbin Zhu, Hanjie Zhao, Xingren Wang, Shanhong Liu, Yuxiang Jia, and Hongying Zan. 2024. ZZU-NLP at SIGHAN-2024 dimABSA Task: Aspect-Based Sentiment Analysis with Coarse-to-Fine In-context Learning . arXiv preprint arXiv:2407.15341

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.