Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AIMA at SemEval-2024 Task 10: History-Based Emotion Recognition in Hindi-English Code-Mixed Conversations

Mohammad Mahdi Abootorabi, Nona Ghazizadeh, Seyed Arshan Dalili, Alireza Ghahramani Kure, Mahshid Dehghani, Ehsaneddin Asgari

TL;DR

This work tackles emotion recognition in code-mixed Hindi-English conversations (MaSaC dataset) by building four base models around a RoBERTa-based encoder fine-tuned on GoEmotions and leveraging both prior and future conversational context. A two-step Hinglish-to-English translation pipeline (indic-trans to Hindi, then SeamlessM4T) enables processing of code-mixed input, while three architectures plus a data-augmented variant form four base models that are ensembled by majority voting. The Context-Aware GRU-Based model, which uses both preceding and following utterances and omits previous emotion to avoid error propagation, delivers the strongest single-model performance, with the ensemble achieving Weighted F1 of $0.4080$ on the test set, outperforming GPT-3.5 Turbo and LaBSE-based baselines. The study demonstrates the value of including historical and future context and through a two-step translation pipeline, provides a practical approach to code-mixed ERC with potential for extension to end-to-end and multimodal setups. Overall, this work advances code-mixed ERC by combining robust encoders, context modeling, and ensemble strategies, offering a foundation for more scalable emotion-aware dialogue systems in multilingual settings.

Abstract

In this study, we introduce a solution to the SemEval 2024 Task 10 on subtask 1, dedicated to Emotion Recognition in Conversation (ERC) in code-mixed Hindi-English conversations. ERC in code-mixed conversations presents unique challenges, as existing models are typically trained on monolingual datasets and may not perform well on code-mixed data. To address this, we propose a series of models that incorporate both the previous and future context of the current utterance, as well as the sequential information of the conversation. To facilitate the processing of code-mixed data, we developed a Hinglish-to-English translation pipeline to translate the code-mixed conversations into English. We designed four different base models, each utilizing powerful pre-trained encoders to extract features from the input but with varying architectures. By ensembling all of these models, we developed a final model that outperforms all other baselines.

AIMA at SemEval-2024 Task 10: History-Based Emotion Recognition in Hindi-English Code-Mixed Conversations

TL;DR

This work tackles emotion recognition in code-mixed Hindi-English conversations (MaSaC dataset) by building four base models around a RoBERTa-based encoder fine-tuned on GoEmotions and leveraging both prior and future conversational context. A two-step Hinglish-to-English translation pipeline (indic-trans to Hindi, then SeamlessM4T) enables processing of code-mixed input, while three architectures plus a data-augmented variant form four base models that are ensembled by majority voting. The Context-Aware GRU-Based model, which uses both preceding and following utterances and omits previous emotion to avoid error propagation, delivers the strongest single-model performance, with the ensemble achieving Weighted F1 of on the test set, outperforming GPT-3.5 Turbo and LaBSE-based baselines. The study demonstrates the value of including historical and future context and through a two-step translation pipeline, provides a practical approach to code-mixed ERC with potential for extension to end-to-end and multimodal setups. Overall, this work advances code-mixed ERC by combining robust encoders, context modeling, and ensemble strategies, offering a foundation for more scalable emotion-aware dialogue systems in multilingual settings.

Abstract

In this study, we introduce a solution to the SemEval 2024 Task 10 on subtask 1, dedicated to Emotion Recognition in Conversation (ERC) in code-mixed Hindi-English conversations. ERC in code-mixed conversations presents unique challenges, as existing models are typically trained on monolingual datasets and may not perform well on code-mixed data. To address this, we propose a series of models that incorporate both the previous and future context of the current utterance, as well as the sequential information of the conversation. To facilitate the processing of code-mixed data, we developed a Hinglish-to-English translation pipeline to translate the code-mixed conversations into English. We designed four different base models, each utilizing powerful pre-trained encoders to extract features from the input but with varying architectures. By ensembling all of these models, we developed a final model that outperforms all other baselines.
Paper Structure (12 sections, 1 figure, 1 table)

This paper contains 12 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. System Overview
  4. Preprocessing Data
  5. Model Architecture
  6. Simple History-Based Model
  7. Simple History-Based Model + Data Augmentation
  8. Full History-Based Model
  9. Context-Aware GRU-Based Model
  10. Experimental Setup
  11. Results
  12. Conclusion

Figures (1)

  • Figure 1: Three proposed base model architectures for predicting the emotion of the current sentence. (a): This model utilizes only the basic historical information from the conversation. (b): This model leverages information from all past sentences, in addition to the information used in the previous architecture. (c): This model employs GRU to leverage sequential information and incorporates future information to gain a more comprehensive understanding of the context of the current sentence.