ShiyuNee / Facet-Generation

This is the source code for paper "A Comparative Study of Training Objectives for Clarification Facet Generation(SIGIR-AP 2023)".

Facet-Generation

This is the source code for paper A Comparative Study of Training Objectives for Clarification Facet Generation(SIGIR-AP 2023)

Requirements and Installation

• Python version >= 3.8
• PyTorch version >= 1.11.0
• Transformers version >= 4.27.2

Corpus

• MIMICS: microsoft/MIMICS: MIMICS: A Large-Scale Data Collection for Search Clarification (github.com)
  • You should download The Bing API's Search Results for MIMICS Queries

Models

We just provide model weights in the following websites(Seq-default contains both model weights and the tokenizer). The tokenizer for each model is the same with the tokenizer used in Bart-base(https://huggingface.co/facebook/bart-base)

• seq-default: https://huggingface.co/algoprog/mimics-bart-base
• seq-min-perm: https://huggingface.co/Shiyunee/seq-min-perm
• seq-avg-perm: https://huggingface.co/Shiyunee/seq-avg-perm
• set-pred: https://huggingface.co/Shiyunee/set-pred
• seq-set-pred: https://huggingface.co/Shiyunee/seq-set-pred

Usage

• data_process.py: prepare data for inference
• inference.py: generate facets for the given data
• evaluation.py: evaluate the results
• score.py: replace score.py in bert_score(package) with this file so we can load the model and tokenizer only once