Unsupervised Entity-Relation Analysis in IBM Watson

Text paraphrasing algorithms play a fundamental role in several NLP applications such as automated question answering (QA), summarization and machine translation. We propose a novel paraphrasing approach based on an entity-relation (ER) analysis of text. The algorithm uses a combination of deep linguistic analysis (part of speech, dependency parse information) and background resources (NGram, PRISMATIC KB, domain dictionaries) to detect and match entities and relations. We evaluate the ER approach in a QA setting by adding it to the suite of passage scoring algorithms in IBM Watson, a state-of-the-art question answering system. We show a statistically significant improvement in the ability of IBM Watson to identify justifying passages.