A Joint Model for Ellipsis Identification and Recovery

An ellipsis is a gap in a sentence due to the pragmatics conventional use of grammar. Ellipsis is a ubiquitous phenomenon in daily conversation, especially in Chinese. A question-answering (QA) system can hardly automatically understand sentences with ellipsis. As a result, the QA system may produce wrong answer and thus cannot naturally interact with humans. Therefore, it is important to recover these ellipses in order to gain a better QA system. To automatically recover these ellipsis elements, we take the recovery system into two parts: zero anaphora identification and zero anaphora resolution. When connecting these two parts together, previous work always models the two steps separately, which suffers the error accumulation problem. In order to deal with this problem, we propose a joint model method that performs the zero anaphora identification and zero anaphora resolution simultaneously in a unified framework. Besides, we focus on Chinese dialogue text, which is collected from the interview of broadcast. The experimental results show that the proposed joint model method outperforms the state-of-the-art methods significantly.