Skip to content

Latest commit

 

History

History
51 lines (36 loc) · 2.17 KB

README.md

File metadata and controls

51 lines (36 loc) · 2.17 KB
Raw

Code for reproducing the SIGIR '22 paper

Zero-shot Query Contextualization for Conversational Search

Our approach (ZeCo2) contextualizes the last user question within the conversation history, but restrict the matching only between question and potential answer.

For more details, check our:

SIGIR'22 poster | paper

@inproceedings{krasakis-2022-zeroshot,
    author = {Krasakis, Antonios Minas and Yates, Andrew and Kanoulas, Evangelos},
    booktitle = {SIGIR 2022: 45th international ACM SIGIR Conference on Research and Development in Information Retrieval},
    month = {July},
    publisher = {ACM},
    title = {Zero-shot Query Contextualization for Conversational Search},
    year = {2022}}

Reproducing results:

  1. Install colbert
  2. Download colbert model checkpoint & update various paths (see the #TODO @ paths.py )
  3. Corpus indexing: create FAISS indexes using a ColBERT model (see [index your collection](README_ColBERT.md#ColBERT Indexing). Note that preprocessing scripts are available in /preprocessing. To make use our pipeline for retrieval and evaluation, you need to convert query&passage ids to integers and retain a mapping file (.intmapping) before indexing (see the preprocessing README and examples).
  4. Retrieve & rerank using the available pipeline:

python_pipeline.py --setting ZeCo2 --dataset cast19

The ColBERT checkpoint used (trained for 400K steps) is available here

Paper analysis section:

The two scripts used to reproduce the analysis section of the paper are:

token_embedding_change.py
embedding_closest_terms.py

You can already run the analysis since the final rankings are provided under data/rankings/

For ColBERT-related questions, instructions, etc. please refer to the original repository (forked from v0.2.0) or README_ColBERT.md, or feel free to raise an issue!