Mimicking the Thinking Process for Emotion Recognition in Conversation with Prompts and Paraphrasing

Code for our paper, "Mimicking the Thinking Process for Emotion Recognition in Conversation with Prompts and Paraphrasing", IJCAI 2023, AI and Social Good track.

Dependencies

• torch==1.13.1+cu117
• transformers==4.17.0
• sklearn
• fitlog

Required data

Download required data for training of the second stage (including the base models we used, the raw texts of dialogues, the features of utterances, and the sample sets):

• MELD: MELD_data
• IEMOCAP: IEMOCAP_data
• DailyDialog: DailyDialog_data
• SentiWordNet: SentiWordNet_3.0.0

Training

Train the second stage model with the following commands:

• MEDL: bash train_prompt2.sh MELD 42 stage2 128 0.3 ./data/MELD/meld_stage1 3 bert_score
• IEMOCAP: bash train_prompt2.sh IEMOCAP 43 stage2 160 0.1 ./data/IEMOCAP/iemo_stage1 3 bert_score
• DailyDialog: bash train_prompt2.sh DailyDialog 2 stage2 128 0.2 ./data/DailyDialog/dd_stage1 5 bm25

The optim values of these hyper-parameters(alpha and k) can vary under different seeds or environments (similar to CoG-BART), and adjustments may be needed in different experimental settings.

Quick Start

We also provide checkpoints for reproduction:

• MELD: MELD_checkpoint 提取码：n5nn bash eval_prompt2.sh MELD 128 0.3 ./data/MELD/meld_stage2 3 bert_score
• IEMOCAP: IEMOCAP_checkpoint 提取码：zfta bash eval_prompt2.sh IEMOCAP 160 0.1 ./data/IEMOCAP/iemo_stage2 3 bert_score
• DailyDialog: DailyDialog_checkpoint 提取码：omu5 bash eval_prompt2.sh DailyDialog 128 0.2 ./data/DailyDialog/dd_stage2 5 bm25

Acknowledgment

Some code of this project are referenced from CoG-BART and DAG-ERC. We thank their open source materials for contribution to this task.