CoMPM: Context Modeling with Speaker's Pre-trained Memory Tracking for Emotion Recognition in Conversation

TL;DR

This paper introduces CoMPM, a novel emotion recognition method in conversations that leverages pre-trained memory from language models, improving accuracy without relying on structured external data, and demonstrating language extension capabilities.

Contribution

We propose CoMPM, which integrates speaker's pre-trained memory with context modeling, achieving state-of-the-art results without using structured external knowledge.

Findings

Achieves top performance on multiple datasets.

Outperforms existing methods that use structured data.

Extends easily to non-English languages.

Abstract

As the use of interactive machines grow, the task of Emotion Recognition in Conversation (ERC) became more important. If the machine-generated sentences reflect emotion, more human-like sympathetic conversations are possible. Since emotion recognition in conversation is inaccurate if the previous utterances are not taken into account, many studies reflect the dialogue context to improve the performances. Many recent approaches show performance improvement by combining knowledge into modules learned from external structured data. However, structured data is difficult to access in non-English languages, making it difficult to extend to other languages. Therefore, we extract the pre-trained memory using the pre-trained language model as an extractor of external knowledge. We introduce CoMPM, which combines the speaker's pre-trained memory with the context model, and find that the pre-trained memory significantly improves the performance of the context model. CoMPM achieves the first or second performance on all data and is state-of-the-art among systems that do not leverage structured data. In addition, our method shows that it can be extended to other languages because structured knowledge is not required, unlike previous methods. Our code is available on github (https://github.com/rungjoo/CoMPM).