DialogueNLP

This is a PyTorch implementation of the Committed paper for ACL 2021.

1 Preparation for Dataset

In order to facilitate project management, the data sets used in this project are linked as Download with password: ev80.

After the dataset is downloaded, it needs to be decompressed into this project. The extracted file is named dataset, which contains two Chinese datasets mentioned in this paper: Double and Weibo. The link of English datasets is as follows: Cornell Movie-Dialogu Corpus

2 Models

The proposed model is consist of a Retrieval Model and a Generation model. In order to align utterance and retrieval pairs at word-vector level, a novel Selective-Attention Guided Alignment module is proposed.

3 Training

This project is supported by Pytorch and other standard libraries. If you want to train the whole model, train.ipynb is a jupyter file, which can directly run and view the intermediate results in the training process.

Performance

We evaluted the proposed model on BLEU, Rouge, Relevance and Diversity, which the former metrics are word-overlap metrics and the latter two are embedding based metrics. The result are listed as follow:

Cornell	Types		Models	BLEU	Rouge			Relevance			Diversity
	Rtrv	Gene			R-1	R-2	R-L	Average	Extrema	Greedy	Dist-1	Dist-2
		√	S2S+Attn	38.79	37.82	17.87	33.73	0.361	0.201	0.346	0.049	0.088
		√	CVAE	45.23	41.89	20.86	39.49	0.381	0.256	0.374	0.076	0.145
		√	UniLM	50.65	44.24	23.07	40.27	0.401	0.294	0.387	0.121	0.189
	√		Retrieval	36.91	30.81	13.87	27.33	0.296	0.152	0.311	0.103	0.249
	√		Retrieval+Rerank	38.67	34.57	18.23	32.44	0.338	0.193	0.321	0.129	0.212
	√	√	Eidt	49.11	45.81	21.99	43.01	0.393	0.307	0.391	0.112	0.207
	√	√	Ours	52.67	48.72	24.45	43.28	0.417	0.331	0.407	0.121	0.231
Douban	Types		Models	BLEU	Rouge			Relevance			Diversity
	Rtrv	Gene			R-1	R-2	R-L	Average	Extrema	Greedy	Dist-1	Dist-2
		√	S2S+Attn	35.36	33.74	18.16	30.62	0.341	0.182	0.368	0.061	0.081
		√	CVAE	43.65	42.32	21.83	38.78	0.358	0.189	0.373	0.076	0.201
		√	UniLM	49.31	48.76	31.89	47.09	0.383	0.274	0.389	0.202	0.364
	√		Retrieval	36.28	30.19	14.88	28.36	0.298	0.164	0.327	0.131	0.466
	√		Retrieval+Rerank	40.17	36.49	17.67	35.44	0.362	0.211	0.378	0.137	0.431
	√	√	Eidt	47.65	48.27	29.81	46.53	0.378	0.243	0.366	0.134	0.189
	√	√	Ours	55.1	51.79	32.07	51.35	0.394	0.364	0.391	0.188	0.297
Weibo	Types		Models	BLEU	Rouge			Relevance			Diversity
	Rtrv	Gene			R-1	R-2	R-L	Average	Extrema	Greedy	Dist-1	Dist-2
		√	S2S+Attn	37.21	36.77	20.14	35.07	0.341	0.194	0.366	0.026	0.084
		√	CVAE	44.74	44.15	23.12	41.39	0.358	0.195	0.378	0.086	0.142
		√	UniLM	50.06	50.33	32.19	49.81	0.388	0.237	0.387	0.142	0.342
	√		Retrieval	35.79	32.41	15.21	28.03	0.304	0.162	0.315	0.111	0.472
	√		Retrieval+Rerank	38.92	38.29	18.17	35.14	0.354	0.201	0.378	0.167	0.494
	√	√	Eidt	50.64	50.82	26.71	48.38	0.396	0.234	0.387	0.152	0.158
	√	√	Ours	56.3	52.43	29.07	50.41	0.412	0.367	0.411	0.203	0.314