FlexNeuART

FlexNeuART copied to clipboard

FlexNeuART (flex-noo-art)

Flexible classic and NeurAl Retrieval Toolkit, or shortly FlexNeuART (intended pronunciation flex-noo-art) is a substantially reworked knn4qa package. The overview can be found in our EMNLP OSS workshop paper: Flexible retrieval with NMSLIB and FlexNeuART, 2020. Leonid Boytsov, Eric Nyberg.

In Aug-Dec 2020, we used this framework to generate best traditional and/or neural runs in the MSMARCO Document ranking task. In fact, our best traditional (non-neural) run slightly outperformed a couple of neural submissions. Please, see our write-up for details: Boytsov, Leonid. "Traditional IR rivals neural models on the MS MARCO Document Ranking Leaderboard." arXiv preprint arXiv:2012.08020 (2020).

Regretffully, for adminstrative and licensing/patenting issues (there is a patent submitted), neural Model 1 code cannot be released. This model (together with its non-contextualized variant) is described and evaluated in our ECIR 2021 paper: Boytsov, Leonid, and Zico Kolter. "Exploring Classic and Neural Lexical Translation Models for Information Retrieval: Interpretability, Effectiveness, and Efficiency Benefits." ECIR 2021.

In terms of pure effectiveness on long documents, other models (CEDR & PARADE) seem to be perform equally well (or somewhat better). They are available in our codebase. We are not aware of the patents inhibiting the use of the traditional (non-neural) Model 1.

Objectives

Develop & maintain a (relatively) light-weight modular middleware useful primarily for:

• Research
• Education
• Evaluation & leaderboarding

Main features

• Dense, sparse, or dense-sparse retrieval using Lucene and NMSLIB.
• Multi-field multi-level forward indices (+parent-child field relations) that can store parsed and "raw" text input as well as sparse and dense vectors.
• Forward indices can be created in append-only mode, which requires much less RAM.
• Pluggable generic rankers (via a server)
• SOTA neural (CEDR, PARADE, BERT FirstP/MaxP/Sum) and non-neural models (multi-field BM25, IBM Model 1).
• Multi-GPU training and inference with out-of-the box support for ensembling
• Basic experimentation framework (+LETOR)
• Python API to use retrievers and rankers as well as to access indexed data.

Documentation

• Usage notebooks covering installation & most functionality (including experimentation and Python API demo)
• Legacy notebooks for MS MARCO and Yahoo Answers
• Former life (as a knn4qa package), including acknowledgements and publications

We support a number of neural BERT-based ranking models as well as strong traditional ranking models including IBM Model 1 (description of non-neural rankers to follow).

The framework supports data in generic JSONL format. We provide conversion (and in some cases download) scripts for the following collections:

• Cranfield (a small toy collection)
• MS MARCO v1 and v2 (documents and passages)
• Wikipedia DPR (Natural Questions, Trivia QA, SQuAD)
• Yahoo Answers
• Configurable dataset processing of standard datasets provided by ir-datasets.

Acknowledgements

For neural network training FlexNeuART incorporates a substantially re-worked variant of CEDR (MacAvaney et al' 2019).