[论文解读] EmotionX-IDEA: Emotion BERT -- an Affectional Model for Conversation

Yen-Hao Huang, Ssu-Rui Lee+4 more|arXiv (Cornell University)|Aug 17, 2019

Topic Modeling参考文献 10被引用 38

一句话总结

paper adapts BERT for dialogue emotion recognition by causal utterance modeling and dataset-specific pre-training, achieving micro F1 of 0.815 on Friends and 0.885 on EmotionPush.

ABSTRACT

In this paper, we investigate the emotion recognition ability of the pre-training language model, namely BERT. By the nature of the framework of BERT, a two-sentence structure, we adapt BERT to continues dialogue emotion prediction tasks, which rely heavily on the sentence-level context-aware understanding. The experiments show that by mapping the continues dialogue into a causal utterance pair, which is constructed by the utterance and the reply utterance, models can better capture the emotions of the reply utterance. The present method has achieved 0.815 and 0.885 micro F1 score in the testing dataset of Friends and EmotionPush, respectively.

研究动机与目标

• Motivate accurate emotion recognition in conversations by leveraging context from preceding utterances.
• Address domain differences between Friends (scripted dialogue) and EmotionPush (chat) data.
• Mitigate data scarcity and label imbalance in emotion annotation.
• Propose a BERT-based framework enhanced with causal utterance modeling and specialized pre-training.
• Evaluate on EmotionLines datasets to demonstrate improvements over baselines.

提出的方法

• Reframe dialogue into causal utterance pairs by concatenating each target utterance with its preceding utterance (x_t = concat(u_t, u_{t-1})).
• Develop two domain-tailored pre-trained models: FriendsBERT (scripted Friends data) and ChatBERT (Twitter/Chat-style data).
• Implement personality/tokenization schemes for Friends to encode character-specific cues; apply informal text processing for EmotionPush to handle slang, emojis, and entities.
• Fine-tune with a weighted NLL loss to address label imbalance and employ the [CLS] embedding for classification with a dense output layer.
• Train with standard BERT fine-tuning but adapt input structure to two-sentence format and causal utterance representations.
• Compare against traditional baselines (BOW, TFIDF) and TextCNN variants, reporting micro F1 as primary metric.

实验结果

研究问题

• RQ1 Can causal utterance modeling improve emotion detection in dialogue by leveraging preceding context?
• RQ2 Do domain-tailored pre-training strategies (FriendsBERT, ChatBERT) yield better performance on scripted versus chat-like dialogues?
• RQ3 Does weighted loss help address class imbalance in emotion labels and stabilize training?
• RQ4 How do single-sentence versus two-sentence input formats affect BERT-based dialogue emotion classification?
• RQ5 What are the comparative performances of the proposed models against traditional baselines on the EmotionLines dataset?

主要发现

• FriendsBERT and ChatBERT achieve superior micro F1 on testing data: 0.815 (Friends) and 0.885 (EmotionPush).
• Causal utterance modeling improves TextCNN by up to 6 percentage points in Anger and 2 in Joy, with overall F1 gains (~1-6 points depending on setup).
• Weighted loss in early training helps balance performance across majority and minority emotions, aiding minority classes (e.g., Anger, Sadness).
• Traditional baselines (BOW, TFIDF) achieve high overall F1 but poor performance on minority emotions.
• Two tailored pre-training regimes (FriendsBERT using Friends scripts; ChatBERT using Twitter data) help mitigate domain bias and data scarcity.
• The best-performing variants: FriendsBERT-base and FriendsBERT-large yield top validation metrics among Friends experiments.