Title:

Utilizing Passage Similarity Metric for Passage Ranking

Traditional text similarity metrics are not designed to capture fine-grained topical differences between paragraph-sized texts. We explore techniques to learn a subtopic similarity metric which can be used to cluster passages about same broad topic. We use pairwise supervision signals to generate training samples for our supervised approaches. We explore ways to utilize two state of the art language representation models, ELMo and BERT, and fine tune them for modelling a similarity metric for passage re-ranking. We find that an optimal combination of these supervised models along with unsupervised methods performs better than any of the individual methods.

Name

Comment*

Comments are viewable only by submitter

Captcha*