CSDS

This is the official repo for paper CSDS: A Fine-grained Chinese Dataset for Customer Service Dialogue Summarization, accepted by EMNLP 2021 main conference.

Update

1. Evaluation

In the paper, we use the files2rouge package to run ROUGE scores by transferring all the Chinese characters into indexes. However, this process may cause effect to ROUGE-L score, since each summary will be recognized as one single sentence. Thus we modify the script by adding a special character after each sentence in the summary and set it to be the sentence splitting sign for files2rouge. We have already change the code in utils/cal_auto_metrics.py. And we present the new result as below.

2. Method

For CSDS, we found that the tri-gram blocking strategy does great harm, since many summaries actually have repeated tri-grams or four-grams, such as "用户说", 用户表示". Thus we modify the tri-gram blocking strategy in BERTAbs and TDS+SATM and report the new scores as below.

	ROUGE-1	ROUGE-2	ROUGE-L	BLEU	BS	MS
Longest	30.02/35.42/25.94	15.52/20.26/13.84	28.00/33.49/24.01	11.19/13.14/9.94	63.61/67.92/62.89	12.38/16.46/10.71
LexPageRank	36.32/35.15/30.81	19.43/19.29/16.56	34.67/33.82/29.37	13.48/14.14/12.65	66.60/67.23/65.27	15.01/13.94/12.26
SummaRuNNer	44.91/43.90/40.40	27.99/26.46/25.26	42.97/41.89/38.38	21.60/19.35/20.69	71.77/72.16/70.94	24.10/22.16/20.41
BERTExt	43.55/37.25/35.75	27.51/21.58/23.05	41.75/35.69/34.25	21.59/14.91/17.39	71.24/68.01/67.59	22.69/16.06/14.59
PGN	55.58/53.55/50.20	39.19/37.06/35.12	53.46/51.05/47.59	32.31/29.64/28.25	78.40/78.68/76.13	28.58/26.68/25.13
Fast-RL	57.95/57.33/53.07	41.39/40.43/37.59	55.99/55.17/50.76	33.04/33.39/30.44	79.57/80.29/77.72	29.78/28.55/27.18
Fast-RL*	57.70/58.40/52.83	41.24/41.68/37.38	55.76/56.11/50.54	32.94/33.53/30.11	79.76/81.06/77.52	30.12/29.95/26.89
BERTAbs	55.41/52.71/49.61	39.42/36.39/33.88	53.41/50.45/46.88	27.77/30.17/27.02	79.23/79.23/76.41	28.11/24.95/23.91
TDS+SATM	51.69/54.20/49.16	34.94/36.70/33.15	49.44/51.66/46.35	22.89/25.82/26.22	77.47/79.21/76.06	25.35/26.13/24.19
TDS+SATM*	53.14/53.82/47.37	35.98/36.64/31.55	50.68/51.56/44.65	26.47/25.47/22.72	77.81/79.29/75.52	26.11/26.12/23.09

Instructions

1. Introduction

We propose a new Chinese Customer Service Dialogue Summarization dataset (CSDS). It aims at summarizing a dialogue considering dialogue specific features. In CSDS, each dialogue has three different types of summaries:

• Overall summary: The summary condensing the main information of the whole dialogue.
• User summary: The summary focusing on the user's main viewpoints.
• Agent Summary: The summary focusing on the agent's responses.

Besides, each summary are split into several segments, where each segment represent a single topic with its topic label. (A few segments may not have topic labels.) An example annotation is given as below, and if you want to see the details of how data is represented in the json file, please check the introduction for CSDS.

2. Dataset Download

• Baidu NetDisk: https://pan.baidu.com/s/1KKKNuQO5af3JQuun1G3JDg 提取码：5dii
• Google Drive: https://drive.google.com/drive/folders/1IrpEUTR2ZanN0ZKmKcbaXjruTAD9JLkF?usp=sharing

3. Usage

Requirements:

• python == 3.7
• pytorch == 1.6
• files2rouge == 2.1.0
• jieba == 0.42.1
• numpy == 1.19.1
• tensorboard == 2.3.0
• tensorboardx == 2.1
• cytoolz == 0.11.0
• transformers == 3.3.1

Instruction for PGN

1. Go to the models/PGN/ directory.
2. Download the CSDS dataset, create a new folder named Data/ and put CSDS under the Data folder.
3. Download the tencent embedding and put it under the data_utils/embeddings folder.
4. Run the bash file run.sh to train and test.

Instruction for Fast-RL/Fast-RL-mod

1. Go to the models/Fast-RL/ or models/Fast-RL-mod/ directory.
2. Download the CSDS dataset, create a new folder named dataset/ and put it under the dataset/ folder.
3. Copy the embedding file from models/PGN/data_utils/embeddings/dialogue_embed_word to models/Fast-RL/data_utils/embeddings/dialogue_embed_word or models/Fast-RL-mod/data_utils/embeddings/dialogue_embed_word.
4. Run the bash file run.sh to train and test.

Instruction for BERT (BERTExt/BERTAbs)

1. Go to the models/BERT/ directory.
2. Download the CSDS dataset, create a new folder named data/ and put CSDS under the data/ folder.
3. Download the pretrained BERT model, create a new folder named bert_base_chinese/ and put it into the bert_base_chinese/ folder.
4. Run the bash file run.sh to train and test.

Instruction for TDS-SATM/TDS-SATM-mod

1. Go to the models/TDS-SATM/ or models/TDS-SATM-mod/ directory.
2. Download the CSDS dataset, create a new folder named data/ and put CSDS under the data/ folder.
3. Download the pretrained BERT model, create a new folder named bert/chinese_bert/ and put it into the bert/chinese_bert/ folder.
4. Run the bash file run.sh to train and test.

Instruction for SummaRuNNer

1. Go to the models/SummaRuNNer directory.
2. Download the CSDS dataset, create a new folder named data/ and put CSDS under the dataset/ folder.
3. Copy the embedding file from models/PGN/data_utils/embeddings/dialogue_embed_word to models/SummaRuNNer/data/embeddings/dialogue_embed_word
4. Run the bash file run.sh to train and test.

Evaluation

1. We put the output of our trained models to the results/ folder, with the overall summary, user summary and agent summary separately. If you have trained your models, you could also put the outputs into the folder.

2. Run utils/cal_auto_metrics.py to evaluate through automatic metrics. Pay attention to change the file names if you want to test your own output.

3. Run utils/qa_num.py to evaluate the QA pair matching results (Precision, Recall, F1).

Acknowledgement

The reference code of the provided methods are:

• PGN
• FAST-RL
• BERTExt/BERTAbs
• TDS+SATM
• SummaRuNNer

We thanks for all these researchers who have made their codes publicly available.

Citation

If you want to cite our paper, please use this EMNLP proceeding version:

@inproceedings{lin-etal-2021-csds,
    title = "{CSDS}: A Fine-Grained {C}hinese Dataset for Customer Service Dialogue Summarization",
    author = "Lin, Haitao  and
      Ma, Liqun  and
      Zhu, Junnan  and
      Xiang, Lu  and
      Zhou, Yu  and
      Zhang, Jiajun  and
      Zong, Chengqing",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.365",
    pages = "4436--4451",
    abstract = "Dialogue summarization has drawn much attention recently. Especially in the customer service domain, agents could use dialogue summaries to help boost their works by quickly knowing customer{'}s issues and service progress. These applications require summaries to contain the perspective of a single speaker and have a clear topic flow structure, while neither are available in existing datasets. Therefore, in this paper, we introduce a novel Chinese dataset for Customer Service Dialogue Summarization (CSDS). CSDS improves the abstractive summaries in two aspects: (1) In addition to the overall summary for the whole dialogue, role-oriented summaries are also provided to acquire different speakers{'} viewpoints. (2) All the summaries sum up each topic separately, thus containing the topic-level structure of the dialogue. We define tasks in CSDS as generating the overall summary and different role-oriented summaries for a given dialogue. Next, we compare various summarization methods on CSDS, and experiment results show that existing methods are prone to generate redundant and incoherent summaries. Besides, the performance becomes much worse when analyzing the performance on role-oriented summaries and topic structures. We hope that this study could benchmark Chinese dialogue summarization and benefit further studies.",
}

If you have any issues, please contact with haitao.lin@nlpr.ia.ac.cn