arxiv: 2605.03092 · v1 · submitted 2026-05-04 · 💻 cs.CL

Recognition: 1 theorem link

Semantically Enriching Investor Micro-blogs for Opinion-Aware Emotion Analysis: A Practical Approach

Gaurav Negi , Paul Buitelaar

Authors on Pith no claims yet

Pith reviewed 2026-05-08 18:11 UTC · model grok-4.3

classification 💻 cs.CL

keywords opinion graphsemotion analysisfinancial microblogsgraph neural networksLLM pipelineStockEmotions datasetsentiment analysissemantic enrichment

0 comments

The pith

Adding opinion graphs to investor micro-blogs improves GNN emotion classification performance.

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

The paper augments the existing StockEmotions dataset of StockTwits comments with opinion graphs for each sentence. These graphs are produced by a declarative LLM pipeline and supply structured semantic information about the targets and relations in investor opinions. When the enriched data trains graph neural network classifiers, accuracy rises across multiple emotional categories compared with models that use only the original sentiment and emotion labels. A reader would care because the work moves financial NLP past coarse positive-negative scores toward identifying exactly what aspects of a stock or market event trigger specific emotions.

Core claim

Augmenting the StockEmotions dataset with semantically structured opinion graphs derived from a declarative LLM pipeline on 10,000 StockTwits comments improves the classification performance of baseline GNN classifiers across different emotional spectrums.

What carries the argument

Opinion graphs generated by a declarative LLM pipeline that add granular semantic depth to each sentence's existing sentiment and emotion labels.

If this is right

GNN classifiers achieve higher accuracy on emotion detection tasks once opinion semantics are added.
Analysis gains the ability to link emotions to specific targets within investor comments.
The approach supplies a scalable way to enrich other sentiment-labeled financial datasets with structured opinions.
Performance gains appear across the full range of emotional categories rather than only in isolated classes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the pipeline remains reliable at scale, the method could be applied to real-time streams of millions of micro-blogs without additional human annotation.
The same enrichment technique might transfer to non-financial domains such as product reviews or political discussion threads.
Combining the opinion graphs with temporal or user-network features could further improve prediction of market-moving sentiment shifts.

Load-bearing premise

The opinion graphs produced by the LLM pipeline accurately and consistently capture the semantic structure of the financial microblog text without introducing substantial noise or artifacts.

What would settle it

A manual review of a sample of the generated opinion graphs that finds frequent mismatches with the intended meaning or structure in the original StockTwits sentences would falsify the usefulness of the enrichment step.

Figures

Figures reproduced from arXiv: 2605.03092 by Gaurav Negi, Paul Buitelaar.

**Figure 2.** Figure 2: Sub-graph for : “Tesla I’ll buy back in and go Long at $190 [thumbs up] [fire]” The opinion graphs extracted for the input text during the data augmentation step comprise: (i) the spans in text, (ii) inferred characteristics associated with them, and (iii) Relations between them. In this work, our model is based entirely on the text sequence, so we construct graphs using only spans from the text and their… view at source ↗

**Figure 3.** Figure 3: Model Architecture STAGE-4: Fusion of Features Prior to the classification head, the baseline features Hseq and semantic features HG s are fused together via fusion function Hf = Fuse(Hseq, HG s ). We experiment with three types of feature fusion by searching over them as a hyper-parametric variation: 1) Concatenation (cat): These features are concatenated into a single vector Hseq||HG s ∈ R 2d and project… view at source ↗

**Figure 4.** Figure 4: Aggregating performance on valence with the mapping provided with Go-Emotions [22]. view at source ↗

read the original abstract

While sentiment analysis is the staple of financial NLP, capturing the nuances of 'why' behind that sentiment remains a challenge. There have been attempts to address this by analysing investor emotions alongside sentiment; however, this does not provide the additional granularity required to understand the target of the emotion/sentiment. We address this by augmenting the StockEmotions dataset with semantically structured opinion graphs, which provide granular semantic depth to the existing sentiment and emotion labels. Using a declarative LLM pipeline, we augment the StockEmotions dataset with opinion graphs for each sentence, derived from 10,000 comments collected from StockTwits. In addition, we study the effect of introducing opinion semantics on baseline classifiers using Graph Neural Networks (GNNs). Our analysis demonstrates that incorporating opinion semantics improves classification performance across different emotional spectrums

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper builds opinion graphs from StockTwits comments via LLM and tests GNN gains on the StockEmotions dataset, but the absence of any validation on those graphs makes the reported improvements hard to trust.

read the letter

The core contribution is a declarative LLM pipeline that turns 10k investor microblog sentences into opinion graphs and then feeds the augmented StockEmotions data into GNN classifiers. The authors report that this structured input lifts performance across emotion categories compared with baselines that lack the graphs. That is a concrete, domain-specific extension rather than a broad theoretical claim, and it directly targets the gap between flat emotion labels and knowing what the emotion is about. The practical framing and use of an existing dataset keep the work grounded and easy to replicate in principle. The experimental setup itself follows a standard GNN workflow on graph-augmented text, which is a reasonable way to test the value of the added structure. The main weakness is exactly the one flagged in the stress test: there is no human evaluation, inter-annotator check, or qualitative error analysis on the generated graphs. Without that, any accuracy lift could come from LLM artifacts, prompt regularities, or incidental graph properties instead of genuine semantic enrichment. The abstract also gives no numbers, baselines, or significance tests, so the strength of the result cannot be judged from the available text. If the full paper contains those details and the validation step, the concern shrinks; otherwise it remains central. This work is aimed at researchers in financial NLP who already work with social-media emotion data and want to try graph-based enrichment. A reader looking for a ready example of LLM-to-GNN pipelines in this niche could extract useful implementation ideas. It is solid enough on its own terms to merit peer review rather than desk rejection, mainly because the dataset augmentation is new and the method is falsifiable once the missing validation and metrics are supplied. I would send it out with a clear request for human checks on the graphs and full experimental reporting.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes augmenting the StockEmotions dataset with opinion graphs generated from 10,000 StockTwits investor microblog comments via a declarative LLM pipeline. These graphs are intended to add granular semantic structure to existing sentiment and emotion labels. The authors then evaluate the effect of this augmentation on baseline emotion classifiers by employing Graph Neural Networks (GNNs), asserting that the incorporation of opinion semantics yields improved classification performance across emotional categories.

Significance. If the claimed performance gains are demonstrated with rigorous metrics and the opinion graphs are shown to faithfully capture semantics, the work could provide a scalable, practical method for enriching financial NLP tasks with opinion-aware structure beyond coarse emotion labels. The declarative LLM approach for graph construction is potentially reproducible and extensible, but its contribution hinges on empirical validation that is currently absent.

major comments (3)

[Abstract] Abstract: The assertion that 'incorporating opinion semantics improves classification performance across different emotional spectrums' is presented without any quantitative results, baseline details, metrics (accuracy, F1, etc.), error bars, or statistical significance tests. This is load-bearing for the central claim, as the experimental design and results cannot be assessed from the available text.
[Method] Method: The declarative LLM pipeline for generating the 10k opinion graphs is described at a high level but includes no human validation, inter-annotator agreement, qualitative error analysis, or checks for LLM artifacts and prompt biases. This directly undermines the weakest assumption that the graphs accurately enrich the labels with true semantic structure rather than incidental features.
[Experiments] Experiments/Evaluation: No details are supplied on GNN architecture, how opinion graphs are encoded or fused with text inputs, training procedure, dataset splits, or specific performance comparisons. Without these, the reported improvement cannot be reproduced or interpreted as evidence for the approach.

minor comments (1)

[Abstract] The abstract refers to 'different emotional spectrums' without specifying the number or identity of emotion categories in StockEmotions, which would aid clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We appreciate the referee's thorough review and constructive feedback on our manuscript. We address each major comment point by point below, agreeing where revisions are needed to strengthen the paper's clarity, reproducibility, and empirical support. We will incorporate these changes in the revised version.

read point-by-point responses

Referee: [Abstract] Abstract: The assertion that 'incorporating opinion semantics improves classification performance across different emotional spectrums' is presented without any quantitative results, baseline details, metrics (accuracy, F1, etc.), error bars, or statistical significance tests. This is load-bearing for the central claim, as the experimental design and results cannot be assessed from the available text.

Authors: We agree that the abstract should be more self-contained and include key quantitative support for the central claim. While the full manuscript reports these results in the Experiments section, we will revise the abstract to explicitly include performance metrics (e.g., F1-score gains across emotion categories), baseline comparisons, and reference to statistical significance to allow readers to assess the findings immediately. revision: yes
Referee: [Method] Method: The declarative LLM pipeline for generating the 10k opinion graphs is described at a high level but includes no human validation, inter-annotator agreement, qualitative error analysis, or checks for LLM artifacts and prompt biases. This directly undermines the weakest assumption that the graphs accurately enrich the labels with true semantic structure rather than incidental features.

Authors: We acknowledge that the current description of the LLM pipeline in Section 3 is high-level and lacks explicit validation steps. This is a valid concern. In the revision, we will add a new subsection detailing human validation on a sampled subset of graphs, including inter-annotator agreement metrics, qualitative error analysis, and sensitivity checks for prompt biases and LLM artifacts to better substantiate that the graphs provide genuine semantic enrichment. revision: yes
Referee: [Experiments] Experiments/Evaluation: No details are supplied on GNN architecture, how opinion graphs are encoded or fused with text inputs, training procedure, dataset splits, or specific performance comparisons. Without these, the reported improvement cannot be reproduced or interpreted as evidence for the approach.

Authors: We agree that additional experimental details are required for reproducibility and to fully interpret the results. We will substantially expand the Experiments section to specify the GNN architecture (e.g., layers and attention mechanisms), graph encoding and fusion with text inputs, training procedure and hyperparameters, dataset splits, baseline models, and complete performance comparisons including metrics, error bars, and statistical tests. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical pipeline evaluated on external labels

full rationale

The paper describes an applied pipeline that generates opinion graphs from microblog text via a declarative LLM method, augments the StockEmotions dataset, and measures downstream GNN classification accuracy against baselines. No equations, derivations, fitted parameters, or self-referential definitions appear; the reported performance lift is a direct empirical comparison on held-out data rather than a quantity forced by construction from the inputs. The central claim therefore rests on observable classification metrics rather than reducing to its own assumptions by definition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim depends on the LLM pipeline producing faithful opinion graphs and on GNNs being able to exploit the added graph structure for measurable gains; no free parameters are explicitly fitted in the abstract, but the pipeline itself introduces unstated prompt and model choices.

axioms (1)

domain assumption Large language models can reliably extract and structure opinions into graphs from short financial microblog text
Invoked by the declarative LLM pipeline used to augment every sentence in the 10,000-comment dataset.

invented entities (1)

opinion graph no independent evidence
purpose: To provide granular semantic depth to existing sentiment and emotion labels
New structured representation introduced for each sentence; no independent evidence of correctness supplied beyond the downstream classification improvement.

pith-pipeline@v0.9.0 · 5438 in / 1243 out tokens · 47515 ms · 2026-05-08T18:11:44.485486+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

37 extracted references · 11 canonical work pages · 1 internal anchor

[1]

Classifying sentiment in microblogs: is brevity an advantage?

A. Bermingham and A. F. Smeaton, “Classifying sentiment in microblogs: is brevity an advantage?” inProceedings of the 19th ACM International Conference on Information and Knowledge Management, ser. CIKM ’10. New York, NY , USA: Association for Computing Machinery, 2010, p. 1833–1836. [Online]. Available: https://doi.org/10.1145/1871437.1871741

work page doi:10.1145/1871437.1871741 2010
[2]

Contextual semantics for sentiment analysis of twitter,

H. Saif, Y . He, M. Fernandez, and H. Alani, “Contextual semantics for sentiment analysis of twitter,”Information Processing & Management, vol. 52, no. 1, pp. 5–19, 2016, emotion and Sentiment in Social and Expressive Media. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0306457315000242

2016
[3]

Investor sentiment in the theoretical field of behavioural finance,

M. ´Angeles L ´opez-Cabarcos, A. M. P ´erez-Pico, P. V ´azquez- Rodr´ıguez, and M. L. L ´opez-P´erez, “Investor sentiment in the theoretical field of behavioural finance,”Economic Research-Ekonomska Istraˇzivanja, vol. 33, no. 1, pp. 2101–2119, 2020. [Online]. Available: https://doi.org/10.1080/1331677X.2018.1559748

work page doi:10.1080/1331677x.2018.1559748 2020
[4]

Investor sentiment from internet message postings and the predictability of stock returns,

S.-H. Kim and D. Kim, “Investor sentiment from internet message postings and the predictability of stock returns,”Journal of Economic Behavior & Organization, vol. 107, pp. 708– 729, 2014, empirical Behavioral Finance. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0167268114001206

2014
[5]

Efficient capital markets: A review of theory and empirical work,

E. F. Fama, “Efficient capital markets: A review of theory and empirical work,”The Journal of Finance, vol. 25, no. 2, pp. 383–417, 1970. [Online]. Available: http://www.jstor.org/stable/2325486

work page arXiv 1970
[6]

Investor sentiment and the cross-section of stock returns,

M. Baker and J. Wurgler, “Investor sentiment and the cross-section of stock returns,”The Journal of Finance, vol. 61, no. 4, pp. 1645–1680,
[7]

Available: http://www.jstor.org/stable/3874723

[Online]. Available: http://www.jstor.org/stable/3874723

work page arXiv
[8]

Impacts of code of ethics on financial performance in the italian listed companies of bank sector,

M. Cuomo, D. Tortora, A. Mazzucchelli, G. Festa, A. Di Gregorio, G. Metalloet al., “Impacts of code of ethics on financial performance in the italian listed companies of bank sector,”Journal of Business Accounting and Finance Perspectives, vol. 1, no. 1, pp. 1–20, 2018

2018
[9]

Stockemotions: Discover investor emotions for financial sentiment analysis and multivariate time series,

J. Lee, H. L. Youn, J. Poon, and S. C. Han, “Stockemotions: Discover investor emotions for financial sentiment analysis and multivariate time series,”arXiv preprint arXiv:2301.09279, 2023

work page arXiv 2023
[10]

Towards semantic integration of opinions: Unified opinion concepts ontology and extraction task,

G. Negi, D. Dalal, O. Zayed, and P. Buitelaar, “Towards semantic integration of opinions: Unified opinion concepts ontology and extraction task,” inProceedings of the 5th Conference on Language, Data and Knowledge, M. Alam, A. Tchechmedjiev, J. Gracia, D. Gromann, M. P. di Buono, J. Monti, and M. Ionov, Eds. Naples, Italy: Unior Press, Sep. 2025, pp. 174–...

2025
[11]

Language models are few-shot learners,

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert- V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amode...

2020
[12]

Large language models are zero-shot reasoners,

T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large language models are zero-shot reasoners,” inAdvances in Neural In- formation Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, Eds., vol. 35. Curran Associates, Inc., 2022, pp. 22 199–22 213

2022
[13]

Are they different? affect, feeling, emotion, sentiment, and opinion detection in text,

M. Munezero, C. S. Montero, E. Sutinen, and J. Pajunen, “Are they different? affect, feeling, emotion, sentiment, and opinion detection in text,”IEEE transactions on affective computing, vol. 5, no. 2, pp. 101– 111, 2014

2014
[14]

Stock market sentiment lexicon acquisition using microblogging data and statistical measures,

N. Oliveira, P. Cortez, and N. Areal, “Stock market sentiment lexicon acquisition using microblogging data and statistical measures,” Decision Support Systems, vol. 85, pp. 62–73, 2016. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0167923616300240

2016
[15]

Learning stock market sentiment lexicon and sentiment-oriented word vector from StockTwits,

Q. Li and S. Shah, “Learning stock market sentiment lexicon and sentiment-oriented word vector from StockTwits,” inProceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), R. Levy and L. Specia, Eds. Vancouver, Canada: Association for Computational Linguistics, Aug. 2017, pp. 301–310. [Online]. Available: https://aclanth...

2017
[16]

How emotions influence behavior in financial markets: a conceptual analysis and emotion-based account of buy-sell preferences,

D. Duxbury, T. G ¨arling, A. Gamble, and V . Klass, “How emotions influence behavior in financial markets: a conceptual analysis and emotion-based account of buy-sell preferences,”The European Journal of Finance, vol. 26, no. 14, pp. 1417–1438, 2020. [Online]. Available: https://doi.org/10.1080/1351847X.2020.1742758

work page doi:10.1080/1351847x.2020.1742758 2020
[17]

Sentiment analysis and subjectivity

B. Liuet al., “Sentiment analysis and subjectivity.”Handbook of natural language processing, vol. 2, no. 2010, pp. 627–666, 2010

2010
[18]

Liu,Many Facets of Sentiment Analysis

B. Liu,Many Facets of Sentiment Analysis. Cham: Springer International Publishing, 2017, pp. 11–39. [Online]. Available: ”https://doi.org/10.1007/978-3-319-55394-8 2”

work page doi:10.1007/978-3-319-55394-8 2017
[19]

SemEval- 2016 task 5: Aspect based sentiment analysis,

M. Pontiki, D. Galanis, H. Papageorgiou, I. Androutsopoulos, S. Manandhar, M. AL-Smadi, M. Al-Ayyoub, Y . Zhao, B. Qin, O. De Clercq, V . Hoste, M. Apidianaki, X. Tannier, N. Loukachevitch, E. Kotelnikov, N. Bel, S. M. Jim ´enez-Zafra, and G. Eryi ˘git, “SemEval- 2016 task 5: Aspect based sentiment analysis,” inProceedings of the 10th International Worksh...

2016
[20]

Knowing what, how and why: A near complete solution for aspect-based sentiment analysis,

H. Peng, L. Xu, L. Bing, F. Huang, W. Lu, and L. Si, “Knowing what, how and why: A near complete solution for aspect-based sentiment analysis,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 05, pp. 8600–8607, Apr. 2020. [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/view/6383

2020
[21]

SemEval-2017 task 5: Fine-grained sentiment analysis on financial microblogs and news,

K. Cortis, A. Freitas, T. Daudert, M. Huerlimann, M. Zarrouk, S. Handschuh, and B. Davis, “SemEval-2017 task 5: Fine-grained sentiment analysis on financial microblogs and news,” inProceedings of the 11th International Workshop on Semantic Evaluation (SemEval- 2017), S. Bethard, M. Carpuat, M. Apidianaki, S. M. Mohammad, D. Cer, and D. Jurgens, Eds. Vanco...

2017
[22]

An argument for basic emotions

P. Ekman, “An argument for basic emotions,”Cognition and Emotion, vol. 6, no. 3-4, pp. 169–200, 1992. [Online]. Available: https://doi.org/10.1080/02699939208411068

work page doi:10.1080/02699939208411068 1992
[23]

GoEmotions: A Dataset of Fine-Grained Emotions,

D. Demszky, D. Movshovitz-Attias, J. Ko, A. Cowen, G. Nemade, and S. Ravi, “GoEmotions: A Dataset of Fine-Grained Emotions,” in58th Annual Meeting of the Association for Computational Linguistics (ACL), 2020

2020
[24]

A systematic review of applications of natural language processing and future challenges with special emphasis in text-based emotion detection,

S. Kusal, S. Patil, J. Choudrie, K. Kotecha, D. V ora, and I. Pappas, “A systematic review of applications of natural language processing and future challenges with special emphasis in text-based emotion detection,” Artificial Intelligence Review, vol. 56, no. 12, pp. 15 129–15 215, 2023

2023
[25]

Emotion detection for social robots based on nlp transformers and an emotion ontology,

W. Graterol, J. Diaz-Amado, Y . Cardinale, I. Dongo, E. Lopes-Silva, and C. Santos-Libarino, “Emotion detection for social robots based on nlp transformers and an emotion ontology,”Sensors, vol. 21, no. 4, p. 1322, 2021

2021
[26]

Large language models as span annotators,

Z. Kasner, V . Zouhar, P. Schmidtov´a, I. Kart´aˇc, K. Onderkov´a, O. Pl´atek, D. Gkatzia, S. Mahamood, O. Du ˇsek, and S. Balloccu, “Large language models as span annotators,”arXiv preprint arXiv:2504.08697, 2025

work page arXiv 2025
[27]

Are large language models reliable argument quality annotators?

N. Mirzakhmedova, M. Gohsen, C. Chang, and B. Stein, “Are large language models reliable argument quality annotators?” inRobust Ar- gumentation Machines - First International Conference, RATIO 2024, Bielefeld, Germany, June 5-7, 2024, Proceedings, ser. Lecture Notes in Computer Science, P. Cimiano, A. Frank, M. Kohlhase, and B. Stein, Eds., vol. 14638. Sp...

2024
[28]

Large language models for propaganda span annotation,

M. Hasanain, F. Ahmad, and F. Alam, “Large language models for propaganda span annotation,” inFindings of the Association for Computational Linguistics: EMNLP 2024, Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, Eds. Miami, Florida, USA: Association for Computational Linguistics, Nov. 2024, pp. 14 522–14 532. [Online]. Available: https://aclanthology.org/202...

2024
[29]

Towards temporal knowledge-base creation for fine-grained opinion analysis with language models,

G. Negi, A. K. Ojha, O. Zayed, and P. Buitelaar, “Towards temporal knowledge-base creation for fine-grained opinion analysis with language models,”CoRR, vol. abs/2509.02363, 2025. [Online]. Available: https://doi.org/10.48550/arXiv.2509.02363

work page doi:10.48550/arxiv.2509.02363 2025
[30]

DSPy: Compiling declarative language model calls into state-of-the-art pipelines,

O. Khattab, A. Singhvi, P. Maheshwari, Z. Zhang, K. Santhanam, S. V . A, S. Haq, A. Sharma, T. T. Joshi, H. Moazam, H. Miller, M. Zaharia, and C. Potts, “DSPy: Compiling declarative language model calls into state-of-the-art pipelines,” inThe Twelfth International Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/fo...

2024
[31]

ProSA: Assessing and understanding the prompt sensitivity of LLMs,

J. Zhuo, S. Zhang, X. Fang, H. Duan, D. Lin, and K. Chen, “ProSA: Assessing and understanding the prompt sensitivity of LLMs,” in Findings of the Association for Computational Linguistics: EMNLP 2024, Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, Eds. Miami, Florida, USA: Association for Computational Linguistics, Nov. 2024, pp. 1950–1976. [Online]. Availab...

2024
[32]

Optimizing instructions and demonstrations for multi- stage language model programs,

K. Opsahl-Ong, M. J. Ryan, J. Purtell, D. Broman, C. Potts, M. Zaharia, and O. Khattab, “Optimizing instructions and demonstrations for multi- stage language model programs,” inProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Y . Al-Onaizan, M. Bansal, and Y .-N. Chen, Eds. Miami, Florida, USA: Association for Comput...

2024
[33]

How attentive are graph attention net- works?

S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention net- works?” inInternational Conference on Learning Representations, 2022. [Online]. Available: https://openreview.net/forum?id=F72ximsx7C1

2022
[34]

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

J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre- training of deep bidirectional transformers for language understanding,” inProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), J. Burstein, C. Doran, and T. Solorio, Eds...

2019
[35]

RoBERTa: A Robustly Optimized BERT Pretraining Approach

Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V . Stoyanov, “Roberta: A robustly optimized bert pretraining approach,”arXiv preprint arXiv:1907.11692, 2019

work page internal anchor Pith review arXiv 1907
[36]

Qwen3.5: Towards native multimodal agents,

Qwen Team, “Qwen3.5: Towards native multimodal agents,” February
[37]

Available: https://qwen.ai/blog?id=qwen3.5

[Online]. Available: https://qwen.ai/blog?id=qwen3.5