PaperlayerView exact revision on arXiv

Source-linked AI summary

Contrast and Generation Make BART a Good Dialogue Emotion Recognizer

Shimin Li, Hang Yan, Xipeng Qiu

arXiv:2112.11202v2cs.CL

TL;DR

Dialogue emotion recognition must model context and speaker dependencies while distinguishing semantically similar emotion categories. CoG-BART uses BART with supervised contrastive learning to distinguish similar emotions and auxiliary response generation to address contextual information. CoG-BART obtains state-of-the-art results on four ERC datasets; in MELD, it has an approximate absolute 1.24% improvement over the previous state-of-the-art BART-large.

  • Problem

    Dialogue emotion recognition must model context and speaker dependencies while distinguishing semantically similar emotion categories.

    Source: Introduction · p. 1

  • Method

    CoG-BART uses BART with supervised contrastive learning to distinguish similar emotions and auxiliary response generation to address contextual information.

    Source: Introduction · p. 2

  • Results

    CoG-BART obtains state-of-the-art results on four ERC datasets; in MELD, it has an approximate absolute 1.24% improvement over the previous state-of-the-art BART-large.

    Source: Introduction · p. 2, Main Results · p. 5

  • Takeaways & Limitations

    The authors report that supervised contrastive learning and auxiliary response generation improve ERC performance, and that the model does not depend on external resources such as graph-based methods.

    Source: Introduction · p. 2

    Representations can tend to be indistinguishable.

    Source: Introduction · p. 2

Abstract

from arXiv · show

In dialogue systems, utterances with similar semantics may have distinctive emotions under different contexts. Therefore, modeling long-range contextual emotional relationships with speaker dependency plays a crucial part in dialogue emotion recognition. Meanwhile, distinguishing the different emotion categories is non-trivial since they usually have semantically similar sentiments. To this end, we adopt supervised contrastive learning to make different emotions mutually exclusive to identify similar emotions better. Meanwhile, we utilize an auxiliary response generation task to enhance the model's ability of handling context information, thereby forcing the model to recognize emotions with similar semantics in diverse contexts. To achieve these objectives, we use the pre-trained encoder-decoder model BART as our backbone model since it is very suitable for both understanding and generation tasks. The experiments on four datasets demonstrate that our proposed model obtains significantly more favorable results than the state-of-the-art model in dialogue emotion recognition. The ablation study further demonstrates the effectiveness of supervised contrastive loss and generative loss.

Introduction

Emotion recognition in conversation assigns emotion labels to utterances in multi-speaker dialogue, where context and speaker relationships shape emotion. CoG-BART addresses contextual and speaker dependencies and similar emotion categories with a dialogue-level Transformer, supervised contrastive learning, and auxiliary response generation.

Introduction

  • ERC must account for context and speaker dependencies because the same expression can convey different emotions across contexts, and speakers can shift emotions in response to others.Semantically similar categories such as “frustrated” and “sad” are also difficult to distinguish.Source: Introduction · p. 1
  • Deeper graph networks can over-smooth representations, making similar sentiments indistinguishable.Source: Introduction · pp. 1–2
  • CoG-BART introduces a dialogue-level Transformer to model long-range context dependencies between utterances.The Transformer makes fewer prior structural assumptions than approaches that use pretrained features as graph-network nodes.Source: Introduction · p. 2
  • Supervised contrastive learning makes same-sentiment samples cohesive and different sentiments mutually exclusive, targeting difficulty with similar emotion categories.The method uses detached copies of batch hidden states as multiview representations because ERC categories can be highly unbalanced.Source: Introduction · p. 2
  • Auxiliary response generation is intended to improve context handling, while prepending speaker information models speaker–utterance dependency without additional parameters.Source: Introduction · p. 2
  • CoG-BART obtains state-of-the-art results on four ERC datasets, and ablations and case studies support the effectiveness of its contrastive and generative losses.Source: Introduction · p. 2

Related Work

Related work covers graph-based and pretrained-model approaches to dialogue emotion recognition, alongside contrastive-learning methods. The paper positions CoG-BART as an ERC-focused approach combining pretrained BART with supervised contrastive loss.

Dialog Emotion Recognition

  • Graph-based ERC methods model utterance and speaker relations with task-specific structures, while pretrained approaches use feature extractors or fine-tune models to represent utterances and context.Examples include DialogGCN and RGAT for graph approaches, and HiTrans, COSMIC, and DialogXL for pretrained approaches.Source: Graph-based model · p. 2, Pre-train-based model · p. 2, Related Work · p. 2
  • Ide and Kawahara train BART for generation and classification but prioritize response generation, whereas CoG-BART focuses on ERC and adds supervised contrastive loss.Source: Pre-train-based model · p. 3

Contrastive Learning

Methodology

CoG-BART predicts each utterance’s emotion using context and speaker information. It combines BART-based dialogue modeling with supervised contrastive learning and auxiliary response generation, training these alongside classification loss.

Problem Definition

  • Dialogue emotion recognition predicts each utterance’s label from conversation context and speaker information.Source: Problem Definition · p. 3

Supervised Contrastive Learning for ERC

  • Speaker-prefixed utterances are encoded by BART, pooled, and passed through a dialogue-level Transformer to model long-range context dependencies.Source: Utterance Encoding · pp. 3–4, Dialogue Modeling · p. 4
  • Supervised contrastive learning groups same-label examples and separates categories; detached hidden-state copies provide multiview samples when a batch contains a category only once.Source: Dialogue Modeling · p. 4

Auxiliary Response Generation

  • The model generates the following utterance from the current one as an auxiliary task to encourage use of contextual information in emotion recognition.Source: Auxiliary Response Generation · p. 4

Model Training

  • Training combines classification, supervised contrastive, and response-generation losses as a weighted sum whose weights total one.Source: Model Training · pp. 4–5

Experimental Settings

Experiments use four dialogue benchmarks, with dataset-specific evaluation metrics. Training selects the best validation checkpoint and averages test scores across five random seeds.

Experimental Setup

  • The best validation-set checkpoint supplies test results, averaged across five random seeds; training uses AdamW with linear-scheduled warm-up.Source: Experimental Setup · p. 5

Datasets

Metrics

  • Weighted average F1 is used for MELD, EmoryNLP, and IEMOCAP; DailyDialog uses micro-F1 with the neutral label excluded.Source: Metrics · p. 5

Results and Analysis

Across four dataset evaluations, CoG-BART compares favorably with baselines, while analyses report dataset-dependent effects of contrastive loss and component removals.

Main Results

  • CoG-BART improves significantly on EmoryNLP and achieves results similar to graph-based models on IEMOCAP, but has lower micro-F1 than some graph models on DailyDialog.IEMOCAP contexts contain more utterances than the other three datasets; the authors note that pre-trained models are usually incapable of handling excessively long contexts.Source: Main Results · pp. 5–6
  • CoG-BART improves MELD by approximately 1.24% over BART-large and 0.53 points over COSMIC.The comparisons are reported against the previous state-of-the-art BART-large and graph-based COSMIC, respectively.Source: Main Results · p. 5, Table 2 · p. 6, Table 3 · p. 6

The Potency of Supervised Contrastive Learning

  • As SCL’s proportion increases, same-class samples become more cohesive while uniformity between samples is maintained; CoG-BART’s weighted-average F1 on IEMOCAP differs by 2.8 points between α=0.4 and α=0.8, and the best α varies by dataset.CoG-BART performs best at α=0.2 on MELD and α=0.4 on IEMOCAP.Source: Qualitative Analysis of SCL · p. 6, Quantitative Analysis of SCL · pp. 6–7

Effect of Response Generation

  • Case studies indicate response generation can guide emotion predictions toward contextual cues when the current utterance alone is misleading; changing β also slightly affects overall performance.The paper reports this effect in its Figure 4 cases and Table 4 analysis.Source: Table 4 · p. 6, Effect of Response Generation · p. 7, Figure 4 · p. 7

Ablation Analysis

  • Removing any of the three modules worsens MELD performance; on IEMOCAP, speaker removal has the largest effect, while removing SCL alone reduces performance by 1.95 points.The three modules are response generation, SCL loss, and speaker-label splicing; the authors also report that dialogue-level Transformer removal affects longer-context modeling.Source: Ablation Analysis · p. 7

Conclusion

CoG-BART combines supervised contrastive learning with auxiliary response generation for emotion recognition in conversation. Results across four datasets and ablations support the effectiveness of both components.

Conclusion

  • Across four datasets, comparisons with current state-of-the-art baseline methods demonstrate CoG-BART’s effectiveness for emotion recognition in conversation.Source: Conclusion · p. 7
  • Supervised contrastive learning distinguishes similar emotions by reducing intraclass distance and increasing inter-class variance.Source: Conclusion · p. 7
  • Auxiliary response generation helps the model use context to discern semantically similar utterances with different emotions across varying contexts.Source: Conclusion · p. 7

Related papers

Loading 2112.11202v2…