Passage Re-Ranking in Live QA NLP Pipelines with BERT
Disciplines
Artificial Intelligence and Robotics | Computer Sciences
Abstract (300 words maximum)
Passage ranking and document ranking are two common tasks in NLP. Many state of the art pipelines use BM25 to retrieve passages. The top results of this ranking are then re-ranked using a BERT transformer trained on the MS MARCO Passage data set. This system and variations have proved highly effective. In addition, questions and answers using BERT are also well explored topics. However, these systems are fundamentally limited by speed and resource consumption requirements. Given an arbitrary corpus and a collection of pre-trained models, we would like to prove that it is possible to create a live Question Answering machine without fine tuning for a particular topic. In particular, we employ a BERT re-ranker to find the first acceptable fit to pass to our QA transformer. This approach is fundamentally different from past research in that it is focused on first fit and not best fit. The goal of this research is to allow anyone to employ off the shelf components to create an effective, interactive question answering system.
Academic department under which the project should be listed
CCSE - Computer Science
Primary Investigator (PI) Name
Md Abdullah Al Hafiz Khan
pres1
Symposium-Presentation-LIVQA (1).pptx (318 kB)
pres2
Passage Re-Ranking in Live QA NLP Pipelines with BERT
Passage ranking and document ranking are two common tasks in NLP. Many state of the art pipelines use BM25 to retrieve passages. The top results of this ranking are then re-ranked using a BERT transformer trained on the MS MARCO Passage data set. This system and variations have proved highly effective. In addition, questions and answers using BERT are also well explored topics. However, these systems are fundamentally limited by speed and resource consumption requirements. Given an arbitrary corpus and a collection of pre-trained models, we would like to prove that it is possible to create a live Question Answering machine without fine tuning for a particular topic. In particular, we employ a BERT re-ranker to find the first acceptable fit to pass to our QA transformer. This approach is fundamentally different from past research in that it is focused on first fit and not best fit. The goal of this research is to allow anyone to employ off the shelf components to create an effective, interactive question answering system.