Skip to content

Latest commit

 

History

History
38 lines (30 loc) · 2.83 KB

README.md

File metadata and controls

38 lines (30 loc) · 2.83 KB
Raw

Toward A Fine-Grained Analysis of Distribution Shifts in MSMARCO (MS-Shift) (arxiv)

Simon Lupart and Stéphane Clinchant

Abstract

Recent IR approaches based on Pretrained Language Models (PLM) have now largely outperformed their predecessors on a variety of IR tasks. However, what happens to learned word representations with distribution shifts remains unclear. Recently, the BEIR benchmark was introduced to assess the performance of neural rankers in zero-shot settings and revealed deficiencies for several models. In complement to BEIR, we propose to control \textit{explicitly} distribution shifts. We selected different query subsets leading to different distribution shifts: short versus long queries, wh-words types of queries and 5 topic-based clusters. Then, we benchmarked state of the art neural rankers such as dense Bi-Encoder, SPLADE and ColBERT under these different training and test conditions. Our study demonstrates that it is possible to design distribution shift experiments within the MSMARCO collection, and that the query subsets we selected constitute an additional benchmark to better study factors of generalization for various models.

Repository description

This repository contains all data files required to reproduced the experiments made in our work, splitted in the two TRAIN/EVAL folders. In addition, we also included an example notebook with cluster vizualisation.

TRAIN/EVAL 📂

  • TRAIN/queries_clustering.tsv is the repartitions of queries in each clusters.
Format is the following:
{query, topic_cluster, topic_train, whword_cluster, whword_train, length_cluster, length_train}

where topic_cluster is in the range [0-4]+[5] for all 'others' queries; whword_cluster in the range [0-2]+[3] {0:{what, definition}, 1:{how}, 2:{who, when, where, which}, 3:'others'}; length_cluster in the range [0-1] (0:short, 1:long). We also included the train columns to specify on which queries to train on.

  • EVAL contains the evaluation qrel.json and queries.tsv using the usual MS MARCO format.

Visualization of the clusters 🔭

  • In example/shift_visualization.ipynb we included scripts to display the PCA and T-sne of the clusters, together with the computation of similarities between clusters.

Cite

Please cite our work as:

@misc{https://doi.org/10.48550/arxiv.2205.02870,
  doi = {10.48550/ARXIV.2205.02870},
  url = {https://arxiv.org/abs/2205.02870},
  author = {Lupart, Simon and Clinchant, Stéphane},
  keywords = {Information Retrieval (cs.IR), FOS: Computer and information sciences, FOS: Computer and information sciences},
  title = {Toward A Fine-Grained Analysis of Distribution Shifts in MSMARCO},
  publisher = {arXiv},
  year = {2022},
  copyright = {Creative Commons Attribution Non Commercial Share Alike 4.0 International}
}