DialogXL

This repo contains the PyTorch implementaion for AAAI-2021 accepted paper DialogXL: All-in-One XLNet for Multi-Party Conversation Emotion Recognition.

Update 2021/3/18

Thanks to my co-author Junqing (github id: Digimonseeker), we have now updated our code to make it compatible with the latest version of huggingface's Transformers. Now you can run our code with Transformers 4.3.3, and the pre-trained XLNet can be downloaded from https://huggingface.co/xlnet-base-cased.

The hyper-parameters we previously provided still work well, but the well-trained DialogXL models we uploaded may be unable to run in this code version.

Requirements

• Python 3.6
• PyTorch 1.4.0
• Transformers 4.3.3
• scikit-learn 0.23.1
• CUDA 10.0

Preparation

Datasets

We have provided the preprocessed datasets of IEMOCAP, MELD, DailyDialog and EmoryNLP in /data/. For each dataset, there are 3 json files containing dialogs for train/dev/test set. Each dialog contains utterances and their corresponding speaker information and emotion labels.

Pre-trained XLNet

You should download the pytorch version pre-trained xlnet-base model and vocabulary from the link provided by huggingface. Then change the value of parameter --bert_model_dir and --bert_tokenizer_dir to the directory of the bert model.

Evaluate

You can download our well-trained DialogXL for each dataset from https://drive.google.com/drive/folders/1ONGBbo9r9S49i5bz4jgrC79txQHNVZv2?usp=sharing or https://pan.baidu.com/s/1SYbpCI4LTzUnYbRKzPvijg 提取码 jpfe

Then put them to /DialogXL/saved_models/

You can run the following codes to get results very close to those reported in our paper (we report the average of 5 random runs in paper):

For IEMOCAP (test F1: 65.88): python eval.py --dataset_name IEMOCAP --max_sent_len 200 --mem_len 900 --windowp 10 --num_heads 2 2 4 4 --modelname IEMOCAP_xlnet_dialog

For MELD (test F1: 62.67): python eval.py --dataset_name MELD --max_sent_len 300 --mem_len 400 --windowp 5 --num_heads 5 5 1 1 --modelname MELD_xlnet_dialog

For DailyDialog (test F1: 55.67): python eval.py --dataset_name DailyDialog --max_sent_len 300 --mem_len 450 --windowp 10 --num_heads 1 2 5 4 --dropout 0.3 --modelname DailyDialog_xlnet_dialog

For EmoryNLP (test F1: 36.27): python eval.py --dataset_name EmoryNLP --max_sent_len 150 --mem_len 400 --windowp 5 --num_heads 1 2 4 5 --modelname EmoryNLP_xlnet_dialog

Training

You can also train the models with the following codes:

For IEMOCAP: python run.py --dataset_name IEMOCAP --max_sent_len 200 --mem_len 900 --windowp 10 --num_heads 2 2 4 4 --dropout 0 --lr 1e-5 --batch_size 8 --epochs 50

For MELD: python run.py --dataset_name MELD --max_sent_len 300 --mem_len 400 --windowp 5 --num_heads 5 5 1 1 --dropout 0 --lr 1e-6 --batch_size 4 --epochs 15

For DailyDialog: python run.py --dataset_name DailyDialog --max_sent_len 300 --mem_len 450 --windowp 10 --num_heads 1 2 5 4 --dropout 0.3 --lr 1e-6 --batch_size 8 --epochs 20

For EmoryNLP: python run.py --dataset_name EmoryNLP --max_sent_len 150 --mem_len 400 --windowp 5 --num_heads 1 2 4 5 --dropout 0 --lr 7e-6 --batch_size 8 --epochs 10