CQ-generation

Code for ICTIR 2021 paper "Towards Facet-Driven Generation of Clarifying Questions for Conversational Search".

Training

You can run training with:

python run.py --model_name 'gpt2' --use_faceted_data 1 --my_faceted_data 'data/ClariQ-FKw.tsv' 

For more control over hyperparameters please check out argparse arguments in run.py.

Inference

Given an initial query and facet keywords, the model will generate a clarifying question. Run inference with:

python run.py --model_name 'gpt2' --test_mode 1 --test_ckp 'gtp2_ckpt_epoch=6.ckpt' --use_faceted_data 1 --my_faceted_data 'data/ClariQ-FKw.tsv' 

Text generation is controlable with several parameters in run.py, including:

• temperature,
• top_k,
• top_p,
• min_output_len,
• max_output_len.

Pre-trained model

You can download fine-tuned GPT-2 model here.

Cite

If you found this paper useful please cite our ICTIR 2021 paper:

@inproceedings{sekulic2021towards,
author = {Sekuli\'c, Ivan and Aliannejadi, Mohammad and Crestani, Fabio},
title = {Towards Facet-Driven Generation of Clarifying Questions for Conversational Search,
year = {2021},
publisher = {Association for Computing Machinery},
booktitle = {Proceedings of the 2021 ACM SIGIR on International Conference on Theory of Information Retrieval},
location = {Virtual Event},
series = {ICTIR '21}
}