sweetpeach / inspired

Inspired

Dataset and code for the EMNLP 2020 paper "INSPIRED: Toward Sociable Recommendation Dialogue Systems".

Dataset

Annotated dialogs (train.tsv, dev.tsv, test.tsv) is available under data/dialog_data.

Survey result is under data/survey_data.

Movie database is in data/movie_database.tsv.

Code

We provide the code for our dialog model.

• data_preprocessing.py
• training.py
• interactive.py

Model Weights

Model weights can be downloaded here

Citation

@inproceedings{hayati-etal-2020-inspired,
    title = "INSPIRED: Toward Sociable Recommendation Dialog Systems",
    author = "Hayati, Shirley Anugrah  and Kang, Dongyeop  and Zhu, Qingxiaoyang  and Shi, Weiyan  and Yu, Zhou",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.654",
    pages = "8142--8152",
}

Contact

Shirley Anugrah Hayati - [email protected]

inspired's Issues

Unable to train the models

Hi,

I have been trying to reproduce the results reported in the paper, but so far after spending a lot of time, I am unable to execute and train the model. To be precise, I successfully executed the preprocessing of the dataset now struggling for model training.

Basically, the code often gets broken probably in the training phase due to different library versions.
Could you please give us the exact package version numbers especially for the transformers?

License for this dataset?

Hi,

Can we expect an license uploaded to this repo?

Best

How do you track history states and respond to a certain task?

Hi,
I just read your training's code and I wonder how your model responds to a given query since you only used LM loss with cross entropy for your separate model A/B. For example, A asks one question and this sentence is then trained for model A. B responses A's question but this response is only trained for model B. How could your model_B generates this respone like B's response while your model_B couldn't know A's question?

Question about pretrained model

Hi,

Thanks for your amazing work and code.

The vocab_size of pretrained is 50297. However, the vocab_size in tokenizer you released and paper is 50310. I guess you have uploaded the wrong pretrained model.

Could you please upload the right model again?

Missing Movie Db files

Dear @xiao00yang @sweetpeach I have successfully trained both language and recommendation models using 'training.py'. Now I am running the 'interactive.py' file and found that the file having movie data named 'filePath.tsv' is missing in the repository. Could you please provide that file or any related workaround?

Could you share the models of strategy and recommendation mentioned in the paper?

strategy_detector = torch.load("STRATEGY_CLS_MODEL_Path")
recommendation_detector = torch.load("RECOMMENDATION_CLS_MODEL_PATH")

Thanks.