Pith. sign in

REVIEW 4 major objections 5 minor 38 references

A Pluggable Multi-Task Learning Framework for Sentiment-Aware Financial Relation Extraction

T0 review · 4 major / 5 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 This paper claims that a pluggable auxiliary sentiment-perception task, built from a predicted sentiment token and the shortest dependency path between entity pairs, improves relation extraction performance on both financial and…

desk verdict A plausible plug-and-play sentiment+SDP module for financial RE with small but consistent F1 gains; needs a neutral-token control before the sentiment story holds. read the letter →

arxiv 2506.12452 v1 pith:KJKCHYVX submitted 2025-06-14 cs.CL cs.AI

Jinming Luo , Hailin Wang This is my paper · ORCID
classification cs.CLcs.AI
keywords relationextractionfinancialNLPsentimentanalysismulti-tasklearningshortestdependencypathattentionsupervisionimplicitlabelinformationbottleneck
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

Sentiment is usually ignored by modern relation-extraction models, and this paper tries to establish that it is a usable, and in finance a valuable, auxiliary signal. It proposes a pluggable module, SSDP-SEM, that prepends a predicted sentence-level sentiment token to each input, extracts the shortest dependency path between the two entities, and trains an auxiliary task whose only job is to steer attention toward those positions. With that module attached, all four reproduced attention-based baselines improve in F1 on the financial REFinD dataset (gains of +0.32 to +1.33 points) and on the general TACRED dataset (gains of +0.77 to +1.02 points). The intended upshot is that existing relation-extraction models can gain sentiment awareness without architectural redesign.

What carries the argument

The central object is the Implicit Sentiment Label (ISL), a binary position mask that marks the inserted sentiment token together with the shortest dependency path tokens between the entity pair. It is normalized to a distribution $q^{\mathrm{ISL}}$ that serves as the target for an auxiliary sentiment-perception task: the model's averaged attention weights are multiplied by the mask and penalized with a KL divergence against $q^{\mathrm{ISL}}$, so attention is trained to rest on sentiment- and syntax-bearing tokens. A second piece, the sentiment attention information bottleneck (SAIB), applies a soft attention over the final features with an entropy loss, forcing a sparse focus on a few key features. The combination is what carries the improvement: auxiliary supervision re-points where the model looks, and the bottleneck keeps that focus concentrated.

What would settle it

Shuffle the sentiment labels during training (swap positive and negative for half the corpus) while keeping everything else fixed: if the F1 gains over the baseline remain essentially unchanged, sentiment semantics are not the cause of the improvement, while a collapse would confirm they are. A complementary check is to insert a random filler token instead of the sentiment token and compare the same gains.

Watch

Extended reading notes

Core claim

The central claim is that adding sentiment and syntactic supervision as a secondary task improves relation extraction, especially when financial texts make sentiment relation-relevant. The module builds an Implicit Sentiment Label (ISL): it inserts a sentiment word (positive or negative, produced by an external sentiment model) at the start of the sentence and marks the positions of the shortest dependency path plus that sentiment token. A normalized version of this position vector supervises the model's attention weights through a KL-divergence loss, and a sentiment attention information bottleneck adds an entropy penalty that pushes the final-layer attention to concentrate on a small set of sentiment-laden features. The paper reports F1 gains on every reproduced baseline on both datasets, larger on REFinD, and ablations show that removing the sentiment component causes the largest drop. Its own error analysis notes that gains are concentrated on sentiment-sensitive entity pairs such as ORG:MONEY and that contradictory sentiment inside one sentence can still mislead the model.

Load-bearing premise

The load-bearing premise is that a single external whole-sentence sentiment label, prepended as one token, reliably captures the sentiment relevant to the target relation, and that steering attention toward that token and the shortest dependency path is a genuinely useful inductive bias rather than a surface cue.

Editorial extensions

If this is right

  • If the central claim is correct, existing attention-based relation-extraction models can be upgraded with sentiment awareness by plugging in the module, without changing their core architecture.
  • The larger gains on REFinD than on TACRED would mean financial texts carry relation-relevant sentiment that standard RE training ignores.
  • The ablation results imply the sentiment token contributes more than entity-position or SDP supervision alone, since removing it costs the most in F1.
  • The paper's error analysis shows the benefit is uneven: sentiment-sensitive pairs like ORG:MONEY gain most, while ORG:DATA pairs show no improvement and mixed-sentiment sentences can produce false negatives.
  • The consistent but smaller gains on TACRED would indicate the sentiment-aware attention bias transfers partially beyond the financial domain.

Reading between the lines

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

  • A direct test the paper does not run is to replace the sentiment token with a random or shuffled label; if the gains survive that replacement, the mechanism is token insertion or extra supervision rather than sentiment semantics.
  • Because the sentiment label is a single sentence-level score from an external classifier, the method's ceiling is tied to that classifier's accuracy; using human-annotated or aspect-level sentiment could sharpen the signal.
  • The same supervision pattern could be tried in other sentiment-laden domains, such as medical or legal text, where the paper's design would need no changes beyond the relation label set.
  • The attention-supervision framing implies the module works by changing where models look; comparing shifted attention with human rationales would directly test that mechanism.
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

4 major / 5 minor

Summary. The paper proposes SSDP-SEM, a pluggable multi-task module for sentence-level relation extraction. The method prepends a sentiment token (obtained from an external sentiment model) to each input, extracts the Shortest Dependency Path (SDP) between the target entities, and constructs an Implicit Sentiment Label (ISL) that marks SDP tokens and the sentiment token. An auxiliary Sentiment Perception (ASP) task uses a KL-divergence loss to push the model's last-three-layer attention toward these positions, and a Sentiment Attention Information Bottleneck (SAIB) loss encourages sparse attention using a feature that includes the sentiment-token representation. Experiments on REFinD and TACRED report F1 gains for four reproduced attention-based baselines (Att-Bi-LSTM, R-BERT, RE-Improved, Casual), with ablations over entity-position, SDP-position, and sentiment-position components.

Significance. If the claimed effect is real, the paper offers a simple, model-agnostic recipe for injecting financial sentiment and dependency-path bias into attention-based RE models, and the reported gains are consistent in sign across all four reproduced models on both datasets. The module is clearly described, the loss terms are coherent in intent, and the use of two public benchmarks is appropriate. However, the central empirical claim is currently underdetermined: the design does not separate the effect of the sentiment token's semantic content from the effect of inserting an extra token, the ablation rows still inherit parts of the full pipeline, and the reported F1 differences (about +0.3 to +1.3 points) are not accompanied by error bars, significance tests, or hyperparameter details. The contribution is therefore plausible but not yet rigorously established.

major comments (4)
  1. [III-B, Eq. (1), Table III] The experiments cannot currently attribute the F1 gains to sentiment semantics rather than to the insertion of an extra token. Section III-B inserts the sentiment token unconditionally as X = (Xsen : X), and Eq. (7) always feeds the sentiment-token feature r_sen into the SAIB attention module. Consequently, even the EP-only and EP+SP ablation rows in Table III contain the inserted sentiment token in the input and in the SAIB features, so the comparison between +IS and -IS changes the supervision target but not the input/token-insertion setup. A neutral control, e.g., inserting a fixed token such as "neutral" or a random nonce token instead of the sentiment label, is essential to support the abstract's claim that sentiment perception, rather than any additional token or regularization, drives the improvement.
  2. [Tables I, II, and III] The reported gains are small (REFinD F1 +0.32 to +1.33; TACRED F1 +0.77 to +1.02) and several precision and recall values decrease, yet no error bars, multiple seeds, significance tests, or hyperparameter settings are reported. Since run-to-run variance for attention-based RE models is typically on the order of one F1 point, the consistency claim in the abstract and Section IV-B is not statistically supported. The authors should report the number of seeds, standard deviations, and a paired significance test (e.g., bootstrap or paired t-test over test instances) for the four reproduced baselines.
  3. [Section III-C, Eqs. (5)-(6)] The ASP loss is not well-defined as written. Eq. (5) defines αISL = αavg ⊙ QISL, where αavg is an average of attention weights and QISL is a binary vector; αISL is therefore not normalized and is not a probability distribution over token positions. The KL divergence in Eq. (6) requires two distributions over the same support, and qISL is normalized by Eq. (2) but αISL is not. Unless the implementation renormalizes αISL before computing the KL term, the loss value is not a valid KL divergence and its gradient behavior is unclear. The authors should state the exact normalization used or replace the objective with a cross-entropy term over a softmax-normalized attention distribution.
  4. [Figure 3 vs. Eq. (1)] There is an internal inconsistency in the definition of the ISL signal. Equation (1) sets QISL_i = 1 if x_i belongs to X_sdp ∪ X_sen, and Section III-B inserts the sentiment token at the beginning of the input, so the first position should always be marked as 1. In all three examples in Figure 3, however, the displayed ISL vector starts with 0, and in Sentence 1 the ISL vector is identical to the SPL vector, with no position marked for the sentiment token. This discrepancy makes it unclear whether the implemented ISL actually includes the inserted sentiment token, and it directly affects the interpretation of the IS-ablation rows in Table III. The figure and equations should be reconciled.
minor comments (5)
  1. [Section III-E, Eq. (10)] Equation (10) writes P_i = σ(Σ_i αIB_i x_i), but the surrounding text says "r_i is the output feature of the baselines"; the symbol x_i is used both for input tokens and for features. Please use a consistent notation.
  2. [Section III-B, reference [24]] The sentiment model is called "senta_lstm" but the cited reference [24] is the PaddlePaddle platform paper. Please provide a direct citation or URL for the specific sentiment model and state its training data or checkpoint.
  3. [Tables I-III] The tables do not report the number of runs, standard deviations, or any significance indicators; the parenthesized deltas in Table III mix absolute and relative changes, which is confusing. Adding a clear notation for improvements and degradations would improve readability.
  4. [Table I, II, and IV] The reproduced model named "Casual" appears to correspond to the model in reference [33] whose title says "A Causal View of Entity Bias"; the table caption should use the correct name "Causal" unless the authors deliberately renamed it.
  5. [Figures 2 and 5] Figure 2 is difficult to read because of small fonts and overlapping labels, and Figure 5 reports only qualitative attention visualizations; the sparsity claim could be quantified, e.g., by reporting the effective number of attention heads or entropy before and after SAIB.

Circularity Check

1 steps flagged · score 3.0 of 10

The auxiliary ASP target is self-constructed from the inserted sentiment token and SDP tokens, but the reported RE F1 is measured on external benchmarks, so the central claim is not fully circular.

  1. fitted input called prediction [Section III-B 'ISL constructor' and Section III-C 'ASP with ISL Signal', Eqs. (1)-(6)]
    "We first use senta_lstm [24] to obtain the sentiment result Xsen = senta_lstm(X). And then insert this token into the original text as X = (Xsen : X). ... The ASP task trains the model to predict all positions of the ISL tokens. ... QISL_i = 1 if xi ∈ Xsdp ∪ Xsen otherwise 0 (1)"

    The target of the ASP task is the position distribution qISL, built by marking the positions of Xsen and Xsdp. Xsen is a token inserted by the method itself, and Xsdp is extracted from the same sentence via Spacy. Thus 'predicting all positions of the ISL tokens' is a self-supervised objective whose labels are a deterministic recoding of the method's own input transformation, not an independently measured sentiment signal. The inserted sentiment token is also present in every experimental configuration, including the +EP and +SP ablations in Table III, so the ASP loss partly teaches the model to attend to a token the authors placed in the input.

full rationale

The paper's central claim is that the pluggable SSDP-SEM auxiliary task improves relation extraction, and this is tested by comparing F1 scores of reproduced baseline models with and without SSDP-SEM on the external REFinD and TACRED benchmarks. Those gains are not fitted to the test set and are not derived by construction from the training objective, so the main result has independent support. The main circularity-adjacent issue is the ASP auxiliary task itself: the ISL supervision signal is constructed from the very sentiment token the method inserts and from SDP tokens extracted from the same sentences, making the auxiliary 'prediction' self-referential rather than an external ground truth. This is a genuine limitation of the interpretive claim that the model learns sentiment semantics, and the absence of a neutral-token control further weakens the causal interpretation, but it is a confound and self-supervision issue rather than a reduction of the reported F1 result to the method's inputs. The paper also cites prior work by an author (Ref. [21]) on multi-task learning with SDP, but that citation is used only as related-work support and is not a load-bearing premise, uniqueness theorem, or fitted parameter. No other pattern of circularity is present.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim depends on an external sentiment classifier and a dependency parser as inputs, plus an unreported loss weight. No new physical or conceptual entities are introduced.

free parameters (1)
  • lambda_asp (ASP loss weight) = not reported
    Scales the auxiliary KL loss in Eq. 6; no value or tuning procedure is specified in Sections III-C or IV-A, so it is a fitted hyperparameter that affects all reported results.
assumptions (4)
  • domain assumption Senta LSTM sentiment labels are accurate enough to serve as ground truth for the auxiliary task.
    Invoked in Section III-B where X_sen = senta_lstm(X) is inserted; the paper provides no evaluation of sentiment label quality.
  • domain assumption Spacy's dependency parser yields correct shortest dependency paths for entity pairs.
    Section III-B extracts X_sdp using Spacy; parser errors would alter the ISL supervision signal and could bias the attention objective.
  • ad hoc to paper Attention weights from the final three layers are an appropriate place to inject token-position supervision.
    Section III-C aggregates the final three layers into one average attention vector; no ablation or justification for this layer choice is given.
  • standard math Softmax attention and KL divergence are differentiable and valid training objectives.
    Used in Eq. 3-6; treated as standard background, no proof needed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Pluggable Multi-Task Learning Framework for Sentiment-Aware Financial Relation Extraction." pith.science (2026). https://pith.science/paper/KJKCHYVX

@misc{pith2026250612452,
  author       = {Pith},
  title        = {Pith review of: A Pluggable Multi-Task Learning Framework for Sentiment-Aware Financial Relation Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KJKCHYVX}},
  note         = {Machine review of arXiv:2506.12452}
}
read the original abstract

Relation Extraction (RE) aims to extract semantic relationships in texts from given entity pairs, and has achieved significant improvements. However, in different domains, the RE task can be influenced by various factors. For example, in the financial domain, sentiment can affect RE results, yet this factor has been overlooked by modern RE models. To address this gap, this paper proposes a Sentiment-aware-SDP-Enhanced-Module (SSDP-SEM), a multi-task learning approach for enhancing financial RE. Specifically, SSDP-SEM integrates the RE models with a pluggable auxiliary sentiment perception (ASP) task, enabling the RE models to concurrently navigate their attention weights with the text's sentiment. We first generate detailed sentiment tokens through a sentiment model and insert these tokens into an instance. Then, the ASP task focuses on capturing nuanced sentiment information through predicting the sentiment token positions, combining both sentiment insights and the Shortest Dependency Path (SDP) of syntactic information. Moreover, this work employs a sentiment attention information bottleneck regularization method to regulate the reasoning process. Our experiment integrates this auxiliary task with several prevalent frameworks, and the results demonstrate that most previous models benefit from the auxiliary task, thereby achieving better results. These findings highlight the importance of effectively leveraging sentiment in the financial RE task.

Figures

Figures reproduced from arXiv: 2506.12452 by the authors.

Figure 1
Figure 1. Example sentence from the REFinD dev dataset. The entities “EN [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture of SSDP-SEM. This architecture presents a way of using [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Examples of ISL and their corresponding EPL and SPL. EPL focuses on entity token position, SPL combines SDP token positions with EPL, and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: F1 score distribution by entity pairs on REFinD, indicating the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Attention distribution visualization across baseline, ISL and their [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 31 canonical work pages

  1. [1]

    Toward expert-level medical question answering with large language models,

    K. Singhal, T. Tu, J. Gottweis, R. Sayres, E. Wulczyn, M. Amin, L. Hou, K. Clark, S. R. Pfohl, H. Cole-Lewis et al., “Toward expert-level medical question answering with large language models,” Nature Medicine, pp. 1–8, 2025

    2025

  2. [2]

    Knowledge enhanced graph inference network based entity-relation extraction and knowledge graph construction for industrial domain,

    Z. Han and J. Wang, “Knowledge enhanced graph inference network based entity-relation extraction and knowledge graph construction for industrial domain,” Frontiers of Engineering Management, vol. 11, no. 1, pp. 143–158, 2024

    work page 2024

  3. [3]

    Gap: A novel generative context-aware prompt-tuning method for relation extraction,

    Z. Chen, Z. Li, Y . Zeng, C. Zhang, and H. Ma, “Gap: A novel generative context-aware prompt-tuning method for relation extraction,” Expert Systems with Applications , vol. 248, p. 123478, 2024

    work page 2024

  4. [4]

    Causalbench: A comprehensive benchmark for evaluating causal reasoning capabilities of large language models,

    Z. Wang, “Causalbench: A comprehensive benchmark for evaluating causal reasoning capabilities of large language models,” in Proceed- ings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10), 2024, pp. 143–151

    work page 2024

  5. [5]

    Matching the Blanks: Distributional Similarity for Relation Learning

    L. B. Soares, N. FitzGerald, J. Ling, and T. Kwiatkowski, “Matching the blanks: Distributional similarity for relation learning,” arXiv preprint arXiv:1906.03158, 2019

    work page Pith review arXiv 1906

  6. [6]

    Kpi-edgar: A novel dataset and accompanying metric for relation extraction from financial documents,

    T. Deußer, S. M. Ali, L. Hillebrand, D. Nurchalifah, B. Jacob, C. Bauck- hage, and R. Sifa, “Kpi-edgar: A novel dataset and accompanying metric for relation extraction from financial documents,” in 2022 21st IEEE International Conference on Machine Learning and Applications (ICMLA). IEEE, 2022, pp. 1654–1659

    work page 2022

  7. [7]

    Fincon: A synthesized llm multi-agent system with conceptual verbal reinforcement for enhanced financial decision making,

    Y . Yu, Z. Yao, H. Li, Z. Deng, Y . Cao, Z. Chen, J. W. Suchow, R. Liu, Z. Cui, Z. Xu et al. , “Fincon: A synthesized llm multi-agent system with conceptual verbal reinforcement for enhanced financial decision making,” arXiv preprint arXiv:2407.06567 , 2024

    arXiv 2024

  8. [8]

    Refind: Relation extraction financial dataset,

    S. Kaur, C. Smiley, A. Gupta, J. Sain, D. Wang, S. Siddagangappa, T. Aguda, and S. Shah, “Refind: Relation extraction financial dataset,” in Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2023, pp. 3054– 3063

    work page 2023

Show all 38 references
  1. [9]

    Is positive sentiment in corporate annual reports informative? evidence from deep learning,

    M. Azimi and A. Agrawal, “Is positive sentiment in corporate annual reports informative? evidence from deep learning,” The Review of Asset Pricing Studies, vol. 11, no. 4, pp. 762–805, 2021

    2021

  2. [10]

    imetre: Incorporating markers of entity types for relation extraction,

    N. H. Vardhan and M. Chaudhary, “imetre: Incorporating markers of entity types for relation extraction,” arXiv preprint arXiv:2307.00132 , 2023

    2023 arXiv

  3. [11]

    Ahead of the text: Leveraging entity preposi- tion for financial relation extraction,

    S. Pasch and D. Petridis, “Ahead of the text: Leveraging entity preposi- tion for financial relation extraction,” arXiv preprint arXiv:2308.04534 , 2023

    2023 arXiv

  4. [12]

    Classifying rela- tions via long short term memory networks along shortest dependency paths,

    Y . Xu, L. Mou, G. Li, Y . Chen, H. Peng, and Z. Jin, “Classifying rela- tions via long short term memory networks along shortest dependency paths,” in Proceedings of the 2015 conference on empirical methods in natural language processing, 2015, pp. 1785–1794

    2015

  5. [13]

    Lexicalized dependency paths based super- vised learning for relation extraction

    H. Sun and R. Grishman, “Lexicalized dependency paths based super- vised learning for relation extraction.” Computer Systems Science & Engineering, vol. 43, no. 3, 2022

    2022

  6. [14]

    Sentfin 1.0: Entity-aware sentiment analysis for financial news,

    A. Sinha, S. Kedas, R. Kumar, and P. Malo, “Sentfin 1.0: Entity-aware sentiment analysis for financial news,” Journal of the Association for Information Science and Technology , vol. 73, no. 9, pp. 1314–1335, 2022

    2022

  7. [15]

    Multi-task learning for features extraction in financial annual reports,

    S. Montariol, M. Martinc, A. Pelicon, S. Pollak, B. Koloski, I. Lon ˇcarski, A. Valentin ˇciˇc, K. S. ˇSuˇstar, R. Ichev, and M. ˇZnidarˇsiˇc, “Multi-task learning for features extraction in financial annual reports,” in Joint European Conference on Machine Learning and Knowle...

    2022

  8. [16]

    Sentiment-aware multimodal pre-training for multimodal sentiment analysis,

    J. Ye, J. Zhou, J. Tian, R. Wang, J. Zhou, T. Gui, Q. Zhang, and X. Huang, “Sentiment-aware multimodal pre-training for multimodal sentiment analysis,” Knowledge-Based Systems , vol. 258, p. 110021, 2022

    2022

  9. [17]

    Emge: Entities and mentions gradual enhancement with semantics and connection modelling for document-level relation extraction,

    G. Chen, P. Chen, Q. Wang, H. Li, X. Zhou, X. Wang, A. Yu, and X. Deng, “Emge: Entities and mentions gradual enhancement with semantics and connection modelling for document-level relation extraction,” Knowledge-Based Systems, vol. 309, p. 112777, 2025

    2025

  10. [18]

    Enhancing language models for financial rela- tion extraction with named entities and part-of-speech,

    M. Li and K. H. Lim, “Enhancing language models for financial rela- tion extraction with named entities and part-of-speech,” arXiv preprint arXiv:2405.06665, 2024

    2024 arXiv

  11. [19]

    Relation extraction of medical concepts using categorization and sen- timent analysis,

    A. Mondal, E. Cambria, D. Das, A. Hussain, and S. Bandyopadhyay, “Relation extraction of medical concepts using categorization and sen- timent analysis,” Cognitive Computation, vol. 10, pp. 670–685, 2018

    2018

  12. [20]

    Sentiment classification with modified roberta and recurrent neural networks,

    R. Cheruku, K. Hussain, I. Kavati, A. M. Reddy, and K. S. Reddy, “Sentiment classification with modified roberta and recurrent neural networks,” Multimedia Tools and Applications , vol. 83, no. 10, pp. 29 399–29 417, 2024

    2024

  13. [21]

    Enhancing relation extraction using multi-task learning with sdp evidence,

    H. Wang, D. Zhang, G. Liu, L. Huang, and K. Qin, “Enhancing relation extraction using multi-task learning with sdp evidence,” Information Sciences, vol. 670, p. 120610, 2024

    2024

  14. [22]

    A simple approach to financial relation classification with pre-trained language models,

    L. Qiu, B. Peng, Y . Y . Hsu, and E. Chersoni, “A simple approach to financial relation classification with pre-trained language models,” in The 4th Workshop on Knowledge Discovery from Unstructured Data in Financial Services. Association for Computing Machinery, 2023

    2023

  15. [23]

    Nearest neighbor search over vectorized lexico-syntactic patterns for relation extraction from financial docu- ments,

    P. K. Rajpoot and A. Parikh, “Nearest neighbor search over vectorized lexico-syntactic patterns for relation extraction from financial docu- ments,” arXiv preprint arXiv:2310.17714 , 2023

    2023 arXiv

  16. [24]

    Paddlepaddle: An open-source deep learning platform from industrial practice,

    Y . Ma, D. Yu, T. Wu, and H. Wang, “Paddlepaddle: An open-source deep learning platform from industrial practice,” Frontiers of Data and Domputing, vol. 1, no. 1, pp. 105–115, 2019

    2019

  17. [25]

    Tacred revisited: A thor- ough evaluation of the tacred relation extraction task,

    C. Alt, A. Gabryszak, and L. Hennig, “Tacred revisited: A thor- ough evaluation of the tacred relation extraction task,” arXiv preprint arXiv:2004.14855, 2020

    2004 arXiv

  18. [26]

    Finbert: A pre-trained financial language representation model for financial text mining,

    Z. Liu, D. Huang, K. Huang, Z. Li, and J. Zhao, “Finbert: A pre-trained financial language representation model for financial text mining,” in Proceedings of the twenty-ninth international conference on interna- tional joint conferences on artificial intelligence , 2021, pp. 4513–4519

    2021

  19. [27]

    Spanbert: Improving pre-training by representing and predicting spans,

    M. Joshi, D. Chen, Y . Liu, D. S. Weld, L. Zettlemoyer, and O. Levy, “Spanbert: Improving pre-training by representing and predicting spans,” Transactions of the association for computational linguistics , vol. 8, pp. 64–77, 2020

    2020

  20. [28]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,

    V . Sanh, “Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,” arXiv preprint arXiv:1910.01108 , 2019

    1910 arXiv

  21. [29]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    2018 arXiv

  22. [30]

    Attention- based bidirectional long short-term memory networks for relation clas- sification,

    P. Zhou, W. Shi, J. Tian, Z. Qi, B. Li, H. Hao, and B. Xu, “Attention- based bidirectional long short-term memory networks for relation clas- sification,” in Proceedings of the 54th annual meeting of the association for computational linguistics (volume 2: Short papers) , 2016...

    2016

  23. [31]

    An improved baseline for sentence-level relation extraction,

    W. Zhou and M. Chen, “An improved baseline for sentence-level relation extraction,” arXiv preprint arXiv:2102.01373 , 2021

    2021 arXiv

  24. [32]

    Enriching pre-trained language model with entity information for relation classification,

    S. Wu and Y . He, “Enriching pre-trained language model with entity information for relation classification,” in Proceedings of the 28th ACM international conference on information and knowledge management , 2019, pp. 2361–2364

    2019

  25. [33]

    A causal view of en- tity bias in (large) language models,

    F. Wang, W. Mo, Y . Wang, W. Zhou, and M. Chen, “A causal view of en- tity bias in (large) language models,” arXiv preprint arXiv:2305.14695 , 2023

    2023 arXiv

  26. [34]

    Position- aware attention and supervised data improve slot filling,

    Y . Zhang, V . Zhong, D. Chen, G. Angeli, and C. D. Manning, “Position- aware attention and supervised data improve slot filling,” in Conference on empirical methods in natural language processing , 2017

    2017

  27. [35]

    Graph convolution over pruned dependency trees improves relation extraction,

    Y . Zhang, P. Qi, and C. D. Manning, “Graph convolution over pruned dependency trees improves relation extraction,” arXiv preprint arXiv:1809.10185, 2018

    2018 arXiv

  28. [36]

    Attention guided graph convolutional networks for relation extraction,

    Z. Guo, Y . Zhang, and W. Lu, “Attention guided graph convolutional networks for relation extraction,” arXiv preprint arXiv:1906.07510 , 2019

    1906 arXiv

  29. [37]

    A frustratingly easy approach for entity and relation extraction,

    Z. Zhong and D. Chen, “A frustratingly easy approach for entity and relation extraction,” arXiv preprint arXiv:2010.12812 , 2020

    2010 arXiv

  30. [38]

    On the use of silver standard data for zero-shot classification tasks in information extraction,

    J. Wang, T. Wang, and Z. Zeng, “On the use of silver standard data for zero-shot classification tasks in information extraction,” arXiv preprint arXiv:2402.18061, 2024

    2024 arXiv

Pith tools

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