Automatic Paraphrase Discovery based on Context and Keywords between NE Pairs

Automatic paraphrase discovery is an important but challenging task. We propose an unsupervised method to discover paraphrases from a large untagged corpus, without requiring any seed phrase or other cue. We focus on phrases which connect two Named Entities (NEs), and proceed in two stages. The first stage identifies a keyword in each phrase and joins phrases with the same keyword into sets. The second stage links sets which involve the same pairs of individual NEs. A total of 13,976 phrases were grouped. The accuracy of the sets in representing paraphrase ranged from 73% to 99%, depending on the NE categories and set sizes; the accuracy of the links for two evaluated domains was 73% and 86%.

Download PDF