How to Retrieve Examples in In-context Learning to Improve Conversational Emotion Recognition using Large Language Models?

Mengqi Wang; Shrikanth Narayanan; Tiantian Feng

REVIEW 5 major objections 9 minor 30 references

How to Retrieve Examples in In-context Learning to Improve Conversational Emotion Recognition using Large Language Models?

T0 review · 5 major / 9 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Retrieving one coherent, paraphrased example boosts LLM emotion recognition

desk verdict AER's reported gains are real but confounded: the five-round paraphrase vote, not retrieval, may be driving the improvement. read the letter →

arxiv 2506.20199 v2 pith:5J45ROEJ submitted 2025-06-25 cs.CL cs.AI

Mengqi Wang , Tiantian Feng , Shrikanth Narayanan This is my paper

classification cs.CLcs.AI

keywords ConversationalEmotionRecognitionIn-contextLearningExampleRetrievalAugmentedLargeLanguageModelsParaphraseAugmentationMacroF1ASRtranscripts

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

This paper is an empirical study of how to choose in-context examples for conversational emotion recognition. It proposes Augmented Example Retrieval: for each target utterance, find the most semantically similar reference utterance with an embedding model, generate four paraphrases of it with another LLM, and use one paraphrase per prompt across five voting rounds while randomly filling the other three emotion slots. On IEMOCAP, MELD, and EmoryNLP, this method raises macro F1 above both zero-shot prompting and random-example in-context learning, and the same pattern holds on noisy ASR transcripts. The paper also reports that enlarging the conversation window beyond a certain size does not help, and in two datasets hurts. A sympathetic reader would take away that example quality matters for subjective LLM tasks, not just the presence of examples.

What carries the argument

The machinery is the Augmented Example Retrieval (AER) pipeline. A sentence-embedding model maps every reference utterance and the target into a dense vector space, and cosine similarity selects the single most coherent reference utterance. That utterance is paraphrased four times by a separate instruction-tuned LLM, and each of the five prediction prompts contains one of the paraphrases labeled with its emotion, plus three randomly chosen examples covering the remaining three emotion classes. The five rounds are resolved by majority vote. The retrieval step is what distinguishes AER from random in-context learning, while the paraphrase-and-vote step turns one demonstration into a small ensemble.

What would settle it

Take the AER pipeline and replace the four paraphrases with the original retrieved example repeated four times, or with paraphrases whose label-preservation rate is measured. If macro F1 does not drop, the claimed benefit is not due to paraphrase diversity. Alternatively, have annotators label the four paraphrases of a sample of retrieved examples; if a substantial fraction disagree with the original emotion, the mechanism's premise fails.

Watch

Extended reading notes

Core claim

The central claim is that a prompt containing one retrieved coherent example, namely the reference utterance with the highest cosine similarity to the target, paraphrased into four variants, yields better four-class emotion predictions than zero-shot prompts or prompts with four randomly selected emotion-balanced examples. The evidence is macro F1 on IEMOCAP (0.575 vs 0.567 baseline), MELD (0.581 vs 0.576), and EmoryNLP (0.570 vs 0.547). The authors attribute the gain to retrieval relevance and paraphrase diversity: the model sees the most on-point demonstration in several surface forms, and the five-round majority vote stabilizes the prediction. They further claim the same AER procedure improves macro F1 on three ASR-transcribed versions of IEMOCAP, not only on human transcripts.

Load-bearing premise

The load-bearing premise is that the four paraphrases generated by the second LLM preserve the retrieved example's original emotion label and meaning; if a paraphrase shifts the emotion, the majority vote can be steered toward the wrong answer, and the paper does not validate this preservation.

Editorial extensions

If this is right

AER is a drop-in prompt construction method: any locally deployed LLM can use it without fine-tuning, which matters for privacy-sensitive emotion data.
Randomly selected, emotion-balanced in-context examples are not a reliable upgrade over zero-shot prompting; gains appear only for some datasets.
Conversation context is helpful only up to a saturation point, and on MELD and EmoryNLP the best setting uses no prior utterances at all.
AER also works on noisy ASR transcripts, improving macro F1 on all tested transcript types compared with the zero-shot baseline.

Reading between the lines

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

A reader wanting to isolate the cause of the gain could ablate retrieval, paraphrase count, and majority voting separately; the paper does not decompose which component carries the improvement, and this is an editorial extension.
Because IEMOCAP is evaluated against a reference database drawn from the other two corpora, AER's gain there doubles as evidence of cross-corpus transfer, a property the paper does not foreground.
The same retrieve-one-example, paraphrase, vote recipe could be applied to other subjective text tasks such as sentiment or sarcasm detection, since it only requires a labeled reference pool and a fixed label set.
The paper itself flags in Section 6 that only one 8-billion-parameter LLM and three datasets were used, and it does not validate that Mistral-7B's paraphrases preserve the original emotion label; this unmeasured premise is an editorial caution.

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.

Desk Editor's Note

read the letter

Short version: the paper's headline result is not demonstrated. AER changes two variables at once — retrieving a semantically similar example and majority voting over five paraphrases — while the zero-shot and random-ICL baselines get neither. So the +0.005 to +0.023 macro F1 gains in Table 3 could come from the ensemble alone, not from better retrieval. That is the load-bearing weakness.

What it does well: the AER pipeline is clearly described, the evaluation uses three standard datasets plus an ASR-robustness check, and the paper is honest about limitations. The context-size analysis (gains saturate on IEMOCAP and vanish on MELD/EmoryNLP) is a useful observation. The work is a legitimate extension of the TopK retrieval idea to conversational emotion recognition.

But the voting confound is real. Temperature is 0.0001, so repeated runs of the same prompt are not noise samples; the five rounds differ only by which paraphrase of the retrieved example is shown. That makes AER effectively a mini-ensemble over paraphrases, and no baseline gets a similar ensemble. Without a voting-matched control (e.g., random examples with five paraphrase votes), the retrieval contribution is unidentifiable. The paraphrase-label-preservation issue is secondary but also unvalidated: Mistral-7B-generated paraphrases are used without checking that the emotion label survives. There are also no error bars or significance tests, and the context size per dataset was selected on test data. Given gains are one to two F1 points, noise is a real concern. No code is released.

This is a reasonable workshop-level empirical note, not a broken piece of work. The central claim is plausible but not proven. A serious referee should ask for a voting-matched control and some measure of variance.

I would send this to peer review rather than desk reject, because the question is real and the method is concrete. If the authors can show the gain survives the voting-matched control, the paper has a useful contribution.

Referee Report

5 major / 9 minor

Summary. The manuscript investigates how to retrieve in-context examples to improve conversational emotion recognition (CER) with LLMs. The authors propose Augmented Example Retrieval (AER), which retrieves the most coherent example for each target utterance using SentenceTransformer embeddings, generates four paraphrases of that example with Mistral-7B, and runs five prediction rounds using each paraphrase, settling by majority vote. They compare AER with a zero-shot baseline and with random example ICL on IEMOCAP, MELD, and EmoryNLP, and report macro F1 gains of +0.008, +0.005, and +0.023 over the zero-shot baseline, respectively. They also report improvements on ASR-transcribed IEMOCAP data. The paper further analyzes the effect of conversational context size.

Significance. If the reported gains are reliable, the paper would provide a simple and practical recipe for improving CER by ICL, adding evidence that retrieval-based demonstration selection helps. The use of locally deployable open-source models and the positive ASR results are strengths. However, the evaluation is currently not rigorous enough to establish the central claim: the AER condition is confounded with a five-round majority-vote ensemble, the context size is chosen on the test data, the gains are small and without significance tests, and the paraphrase assumption is unvalidated. For these reasons, the significance of the contribution in its present form is limited.

major comments (5)

[Section 2.3, Table 3] The AER condition is confounded with a five-round majority vote: AER runs five prediction rounds, one per paraphrase of the retrieved example, and takes the majority vote, whereas the zero-shot and ICL-Random baselines are single-pass predictions. The reported macro F1 gains over baselines (+0.008 on IEMOCAP, +0.005 on MELD, +0.023 on EmoryNLP) therefore conflate the effect of retrieval with the effect of ensembling over paraphrases. The paper should include control experiments where the zero-shot and random-example baselines also use five paraphrases and majority voting, or otherwise isolate the retrieval contribution.
[Section 4.1, Figure 5, Table 3] The conversation context size that gives the best result is selected on the test set: the text states "we continue our experiments on the conversation context size with best results across three datasets." This is a test-set overfitting procedure that invalidates the reported comparison as an estimate of generalization. The context size should be selected on a validation split held out from the test data, and the final numbers reported on the test set only once.
[Section 4.3, Table 3] The reported improvements are very small in absolute terms (0.005–0.023 macro F1) and no significance tests, confidence intervals, or repeated runs are reported. Without any measure of variance, the claim that AER "consistently outperforms" the baselines is not statistically supported. At minimum, the authors should report results over multiple random seeds and a paired significance test across utterances.
[Section 2.3, Figure 3] AER depends on the assumption that the Mistral-7B-generated paraphrases preserve the original emotion label and semantic content, because the prompt presents these paraphrases as examples with the same label. The paper provides no validation of paraphrase quality or label preservation. A small human or automatic evaluation of paraphrase-label agreement, or an analysis of how label-preservation rate affects the final vote, is needed to determine whether the method's gains could be an artifact of label-shifted paraphrases misleading the model.
[Section 4.5] The paper states "We categorize other predicted emotions to neutral" when evaluating. Mapping all out-of-label predictions to the neutral class is an ad-hoc post-processing choice that can distort macro F1, especially on the imbalanced MELD and EmoryNLP datasets, and can artificially inflate neutral recall. The authors should either justify this mapping with reference to the task definition or report results without it in the appendix, and also report per-class F1 to check for systematic effects.

minor comments (9)

[Section 4.2] The sentence "there is about 2% improvement on the MELD dataset" appears to be a typo; Table 3 shows a slight decrease on MELD and the 2% improvement is on EmoryNLP (0.566 vs 0.547).
[Section 4.4, Table 4] It is not stated which dataset(s) the ASR evaluation uses; it appears to be IEMOCAP only, since this is the dataset with ASR transcripts. Please state this explicitly.
[Section 2.2] The description of random example selection is ambiguous: it says examples are selected from the augmented reference dataset, but it is unclear whether the random baseline also uses paraphrases and how they are chosen; clarify the exact procedure.
[Section 2.3] The claim that "We empirically investigate the retrieval from the in-domain and out-of-domain data sources" is not followed by any in-domain/out-of-domain comparison in the results; either add such experiments or rephrase.
[Section 3.1] "GenSEC Challenge" should likely be "GenSER Challenge" as in the abstract and introduction.
[Section 3.2] Typo: "inlcudes" should be "includes."
[Figures 2–3] The prompt templates displayed in Figures 2 and 3 are not fully legible; please include the exact prompt text in an appendix for reproducibility.
[Section 3.3] The paper does not mention the temperature used for Mistral-7B when generating paraphrases, which is relevant to the diversity and quality of the paraphrases; please report it.
[General] The paper does not report random seeds for the random example selection or the number of trials, which is important for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper reports a held-out empirical comparison with reference examples drawn from training splits, so no prediction reduces to its own input by construction.

full rationale

This paper is an empirical study: it compares zero-shot prompting, ICL with random examples, and ICL with augmented example retrieval on held-out test splits of IEMOCAP, MELD, and EmoryNLP. The retrieved examples come from training subsets of MELD and EmoryNLP, and the target utterances are from test splits, so the reference material is not the quantity being predicted. No parameter is fitted to the test labels and then renamed as a prediction; cosine similarity is computed with a fixed SentenceTransformer model, and the reported macro F1 values are direct evaluation outcomes. The paper does cite prior work by one of its authors, but that citation is contextual and not load-bearing for the central claim. The skeptic's concern about the five-round majority vote being confounded with retrieval quality is a legitimate experimental-design critique, but it is not circularity: an uncontrolled confound does not make the result equivalent to its inputs by definition. Accordingly, no circular step can be quoted, and the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on unverified domain assumptions rather than fitted parameters or invented entities. The free parameters are hand-chosen design choices (context size, number of examples, number of voting rounds), with the context size tuned per dataset on test performance.

free parameters (3)

conversation_context_size = IEMOCAP: 10, MELD: 0, EmoryNLP: 0
Chosen per dataset as the context size that maximized zero-shot macro F1 on the test sets (Section 4.1).
number_of_icl_examples = 4
Four example utterances with distinct emotions per prompt, fixed by design (Section 2.2).
prediction_rounds = 5
Five voting rounds with distinct paraphrase of the retrieved example; no ablation or variance reported (Section 2.3).

assumptions (4)

domain assumption Cosine similarity in SentenceTransformer embedding space is a valid proxy for selecting the most useful in-context example.
AER retrieves the reference utterance with highest cosine similarity to the target (Section 2.3). This is assumed, not validated by ablations.
domain assumption Paraphrases generated by Mistral-7B preserve the emotion label of the original example.
AER relies on four paraphrases for the retrieved example in five prediction rounds (Section 2.3, Figure 3).
ad hoc to paper LLM predictions outside the four target emotions can be mapped to neutral without distorting evaluation.
Section 4.5: 'We categorize other predicted emotions to neutral.' This mapping is introduced by the authors and affects macro F1.
domain assumption Gold emotion labels in the test datasets are accurate enough to measure small differences.
Section 6 acknowledges 'some utterances in the test datasets do not have accurate emotion labels,' undermining the precision claimed in Tables 3 and 4.

how reviews work

0 comments

Cite this review

Pith. "Pith review of How to Retrieve Examples in In-context Learning to Improve Conversational Emotion Recognition using Large Language Models?." pith.science (2026). https://pith.science/paper/5J45ROEJ

@misc{pith2026250620199,
  author       = {Pith},
  title        = {Pith review of: How to Retrieve Examples in In-context Learning to Improve Conversational Emotion Recognition using Large Language Models?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5J45ROEJ}},
  note         = {Machine review of arXiv:2506.20199}
}

read the original abstract

Large language models (LLMs) have enabled a wide variety of real-world applications in various domains. However, creating a high-performing application with high accuracy remains challenging, particularly for subjective tasks like emotion recognition. Inspired by the SLT 2024 GenSER Challenge, this study investigates approaches to improving conversational emotion recognition (CER) by LLMs. Specifically, we explore how to retrieve high-quality examples in in-context learning (ICL) to enhance CER. We propose various strategies based on random and augmented example retrieval and also analyze the impact of conversational context on CER accuracy. Experiments were conducted on the three datasets including IEMOCAP, MELD and EmoryNLP. The results show that augmented example retrieval consistently outperforms other techniques under investigation across all datasets, highlighting the importance of retrieving coherent targeted examples and enhancing them through paraphrasing.

Figures

Figures reproduced from arXiv: 2506.20199 by the authors.

**Figure 1.** Overview of our framework: Our goal is to use the LLMs to classify emotions from the utterances in the conversation. Our investigation focuses on evaluating whether the incorporation of reference utterances alongside the target utterances improves the emotion prediction by LLMs. prompt to improve the CER performance of GPT-3.5-turbo1 . They conducted two experiments: one without conversational context (zero-shot) … view at source ↗

**Figure 2.** Prompts for zero-shot with context and in-context learning. The purple color is the conversation context, the blue color is the target utterance, and the green color is the answers for reference utterances. significantly outperform standard approaches in ICL. We note that our study on conversational emotion recognition using ICL is inspired by the TopK approach proposed in [13]. In summary, the main contributions an… view at source ↗

**Figure 3.** Flow of our proposed ICL with Augmented Example Retrieval (AER). Given reference utterances, we prompt LLM to generate four paraphrases for each one and encode them with SentenceTransformer. Every time we have a target utterance during prediction, we encode it and find the reference utterance with the highest cosine similarity, while selecting another three examples with the remaining emotions arbitrarily. We design… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 5.** Figure 5: The fluctuation of macro F1 in different conversation context size for IEMOCAP, MELD, and EmoryNLP datasets. 4. Results 4.1. Baselines - Zero-shot without Examples The results reveal different prediction behaviors between IEMOCAP and the other two datasets from [PITH_…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 16 canonical work pages

[1]

How to Retrieve Examples in In-context Learning to Improve Conversational Emotion Recognition using Large Language Models?

Introduction The advancement in artificial intelligence based on deep learn- ing has catalyzed the development of progressively sophisti- cated large language models (LLMs) capable of understand- ing and interpreting human context [1]. This in turn has fa- cilitated a wide range of applications, including conversational agents, document processing, and ed...

work page Pith review arXiv 2024
[2]

First, we evaluate the zero-shot baseline by providing examples in the prompt

Methods Our experiment is designed as three steps. First, we evaluate the zero-shot baseline by providing examples in the prompt. Further, we study how the size of the conversation context im- pacts zero-shot emotion recognition. Next, we perform the ICL baseline experiment by including external reference datasets. Specifically, we construct the prompt by...

work page
[3]

SLT Baseline This work starts with the SLT 2024 GenSEC Challenge

Datasets and Models 3.1. SLT Baseline This work starts with the SLT 2024 GenSEC Challenge. The objective of the challenge is to scrutinize and enhance the accu- racy of emotion perception by large language models (LLMs) with ICL. Currently, it focuses on the prediction of four emo- tions [happy, sad, neutral, angry] . We evaluate the perfor- mance of the ...

work page 2024
[4]

Baselines - Zero-shot without Examples The results reveal different prediction behaviors between IEMOCAP and the other two datasets from Figure 5

Results 4.1. Baselines - Zero-shot without Examples The results reveal different prediction behaviors between IEMOCAP and the other two datasets from Figure 5. On IEMOCAP, we observe that an increase in the conversation con- text size tends to improve the macro F1 score. Specifically, there is a significant improvement of performance from 0.465 to 0.536 w...

work page
[5]

However, the LLM achieves the highest performance with no conversation context provided on the other two datasets, 0.576 in MELD and 0.547 in EmoryNLP

This increase in the performance saturated to around 0.567 once the conversation context size reaches 10. However, the LLM achieves the highest performance with no conversation context provided on the other two datasets, 0.576 in MELD and 0.547 in EmoryNLP. The macro F1 decreases to approximately 0.5 as the size of conversation context increases to 10. Ba...

work page
[6]

We specifically proposed an augmented example retrieval approach to prompt the LLMs with the most coherent example to the tar- get utterance

Conclusion In this work, we study how to retrieve examples in ICL to im- prove emotion recognition in conversations using LLMs. We specifically proposed an augmented example retrieval approach to prompt the LLMs with the most coherent example to the tar- get utterance. Our experiments show that ICL with randomly selected examples performs comparable to ba...

work page
[7]

1) Due to the restriction of GPU capacity, we chose to only experiment on Llama-3.1-8B- Instruct; more complex models should be investigated

Limitations and Future Work Despite the promise of the proposed AER method, there are lim- itations that need further refinement. 1) Due to the restriction of GPU capacity, we chose to only experiment on Llama-3.1-8B- Instruct; more complex models should be investigated. 2) The test datasets are limited and cannot assure the generalization of our method. ...

work page
[8]

Deep learning,

Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,”nature, vol. 521, no. 7553, pp. 436–444, 2015

2015

Show all 30 references

[9]

Empower typed descriptions by large language models for speech emotion recognition,

H. Wu, H.-C. Chou, K.-W. Chan, L. Goncalves, J. Du, J.-S. Jang, C.-C. Lee, and H.-y. Lee, “Empower typed descriptions by large language models for speech emotion recognition,” 09 2024

2024
[10]

Re- vise, reason, and recognize: Llm-based emotion recognition via emotion-specific prompts and asr error correction,

Y . Li, Y . Gong, C.-H. H. Yang, P. Bell, and C. Lai, “Re- vise, reason, and recognize: Llm-based emotion recognition via emotion-specific prompts and asr error correction,”arXiv preprint arXiv:2409.15551, 2024

2024 arXiv
[11]

Enhancing multimodal emo- tion recognition through asr error compensation and llm fine- tuning,

J. Kyung, S. Heo, and J.-H. Chang, “Enhancing multimodal emo- tion recognition through asr error compensation and llm fine- tuning,” in Proc. Interspeech 2024, 2024, pp. 4683–4687

2024
[12]

Beyond silent letters: Amplifying llms in emotion recognition with vocal nuances,

Z. Wu, Z. Gong, L. Ai, P. Shi, K. Donbekci, and J. Hirschberg, “Beyond silent letters: Amplifying llms in emotion recognition with vocal nuances,” arXiv preprint arXiv:2407.21315, 2024

2024 arXiv
[13]

Foundation model assisted automatic speech emotion recognition: Transcribing, annotating, and aug- menting,

T. Feng and S. Narayanan, “Foundation model assisted automatic speech emotion recognition: Transcribing, annotating, and aug- menting,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 12 116–12 120

2024
[14]

Large language model based generative error correction: A challenge and base- lines for speech recognition, speaker tagging, and emotion recog- nition,

C.-H. H. Yang, T. Park, Y . Gong, Y . Li, Z. Chen, Y .-T. Lin, C. Chen, Y . Hu, K. Dhawan, P. ˙Zelasko, C. Zhang, Y .-N. Chen, Y . Tsao, J. Balam, B. Ginsburg, S. M. Siniscalchi, E. S. Chng, P. Bell, C. Lai, S. Watanabe, and A. Stolcke, “Large language model based generative e...

2024
[15]

Iemocap: Interactive emotional dyadic motion capture database,

C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan, “Iemocap: Interactive emotional dyadic motion capture database,”Language resources and evaluation, vol. 42, pp. 335–359, 2008

2008
[16]

Meld: A multimodal multi-party dataset for emotion recognition in conversations,

S. Poria, D. Hazarika, N. Majumder, G. Naik, E. Cambria, and R. Mihalcea, “Meld: A multimodal multi-party dataset for emotion recognition in conversations,” 2019. [Online]. Available: https://arxiv.org/abs/1810.02508

2019 arXiv
[17]

Emotion detection on tv show tran- scripts with sequence-based convolutional neural networks,

S. M. Zahiri and J. D. Choi, “Emotion detection on tv show tran- scripts with sequence-based convolutional neural networks,” in Workshops at the thirty-second aaai conference on artificial in- telligence, 2018

2018
[18]

Large language model-based emotional speech annotation using context and acoustic feature for speech emotion recognition,

J. Santoso, K. Ishizuka, and T. Hashimoto, “Large language model-based emotional speech annotation using context and acoustic feature for speech emotion recognition,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, ...

2024
[19]

The llama 3 herd of models,

A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle et al., “The llama 3 herd of models,” 2024. [Online]. Available: https://arxiv.org/abs/2407.21783

2024 arXiv
[20]

Self-adaptive in- context learning: An information compression perspective for in-context example selection and ordering,

Z. Wu, Y . Wang, J. Ye, and L. Kong, “Self-adaptive in- context learning: An information compression perspective for in-context example selection and ordering,” arXiv preprint arXiv:2212.10375, 2022

2022 arXiv
[21]

Rethinking the role of demonstra- tions: What makes in-context learning work?

S. Min, X. Lyu, A. Holtzman, M. Artetxe, M. Lewis, H. Ha- jishirzi, and L. Zettlemoyer, “Rethinking the role of demonstra- tions: What makes in-context learning work?” in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 2022, pp. 11 048–11 064

2022
[22]

What makes good in-context examples for gpt- 3?

J. Liu, D. Shen, Y . Zhang, B. Dolan, L. Carin, and W. Chen, “What makes good in-context examples for gpt- 3?” arXiv preprint arXiv:2101.06804, 2021

2021 arXiv
[23]

Sentence-bert: Sentence embeddings using siamese bert-networks,

N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” 2019. [Online]. Available: https://arxiv.org/abs/1908.10084

2019 arXiv
[24]

Mistral 7b,

A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. de las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, L. R. Lavaud, M.-A. Lachaux, P. Stock, T. L. Scao, T. Lavril, T. Wang, T. Lacroix, and W. E. Sayed, “Mistral 7b,”
[26]

Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

2021
[27]

wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,” Advances in neural information processing systems , vol. 33, pp. 12 449–12 460, 2020

2020
[28]

Robust speech recognition via large-scale weak supervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” in International conference on machine learning . PMLR, 2023, pp. 28 492–28 518

2023
[29]

Gemma 2: Improving open language models at a practical size,

G. Team, M. Riviere, S. Pathak, P. G. Sessa, C. Hardin, S. Bhu- patiraju, L. Hussenot, T. Mesnard, B. Shahriari, A. Ram ´e et al., “Gemma 2: Improving open language models at a practical size,” arXiv preprint arXiv:2408.00118, 2024

2024 arXiv
[30]

Msp-improv: An acted corpus of dyadic interactions to study emotion perception,

C. Busso, S. Parthasarathy, A. Burmania, M. AbdelWahab, N. Sadoughi, and E. M. Provost, “Msp-improv: An acted corpus of dyadic interactions to study emotion perception,” IEEE Trans- actions on Affective Computing, vol. 8, no. 1, pp. 67–80, 2016

2016
[2023]

Available: https://arxiv.org/abs/2310.06825

[Online]. Available: https://arxiv.org/abs/2310.06825

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

How to Retrieve Examples in In-context Learning to Improve Conversational Emotion Recognition using Large Language Models?

Introduction The advancement in artificial intelligence based on deep learn- ing has catalyzed the development of progressively sophisti- cated large language models (LLMs) capable of understand- ing and interpreting human context [1]. This in turn has fa- cilitated a wide range of applications, including conversational agents, document processing, and ed...

work page Pith review arXiv 2024

[2] [2]

First, we evaluate the zero-shot baseline by providing examples in the prompt

Methods Our experiment is designed as three steps. First, we evaluate the zero-shot baseline by providing examples in the prompt. Further, we study how the size of the conversation context im- pacts zero-shot emotion recognition. Next, we perform the ICL baseline experiment by including external reference datasets. Specifically, we construct the prompt by...

work page

[3] [3]

SLT Baseline This work starts with the SLT 2024 GenSEC Challenge

Datasets and Models 3.1. SLT Baseline This work starts with the SLT 2024 GenSEC Challenge. The objective of the challenge is to scrutinize and enhance the accu- racy of emotion perception by large language models (LLMs) with ICL. Currently, it focuses on the prediction of four emo- tions [happy, sad, neutral, angry] . We evaluate the perfor- mance of the ...

work page 2024

[4] [4]

Baselines - Zero-shot without Examples The results reveal different prediction behaviors between IEMOCAP and the other two datasets from Figure 5

Results 4.1. Baselines - Zero-shot without Examples The results reveal different prediction behaviors between IEMOCAP and the other two datasets from Figure 5. On IEMOCAP, we observe that an increase in the conversation con- text size tends to improve the macro F1 score. Specifically, there is a significant improvement of performance from 0.465 to 0.536 w...

work page

[5] [5]

However, the LLM achieves the highest performance with no conversation context provided on the other two datasets, 0.576 in MELD and 0.547 in EmoryNLP

This increase in the performance saturated to around 0.567 once the conversation context size reaches 10. However, the LLM achieves the highest performance with no conversation context provided on the other two datasets, 0.576 in MELD and 0.547 in EmoryNLP. The macro F1 decreases to approximately 0.5 as the size of conversation context increases to 10. Ba...

work page

[6] [6]

We specifically proposed an augmented example retrieval approach to prompt the LLMs with the most coherent example to the tar- get utterance

Conclusion In this work, we study how to retrieve examples in ICL to im- prove emotion recognition in conversations using LLMs. We specifically proposed an augmented example retrieval approach to prompt the LLMs with the most coherent example to the tar- get utterance. Our experiments show that ICL with randomly selected examples performs comparable to ba...

work page

[7] [7]

1) Due to the restriction of GPU capacity, we chose to only experiment on Llama-3.1-8B- Instruct; more complex models should be investigated

Limitations and Future Work Despite the promise of the proposed AER method, there are lim- itations that need further refinement. 1) Due to the restriction of GPU capacity, we chose to only experiment on Llama-3.1-8B- Instruct; more complex models should be investigated. 2) The test datasets are limited and cannot assure the generalization of our method. ...

work page

[8] [8]

Deep learning,

Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,”nature, vol. 521, no. 7553, pp. 436–444, 2015

2015