Fast and Accurate Incomplete Utterance Rewriting

Omission and reference frequently occur in dialogues, which complicate the semantic structure of the dialogue and hinder the understanding of dialogue systems. Facing the challenge, researchers propose a new task, Incomplete Utterance Rewriting (IUR). IUR systems supplement information for utterances in a dialogue based on the previous context. By utilizing the substitutive nature of IUR, action-based models have recently led to much development on IUR compared to conventional sequence-to-sequence models. The current state-of-the-art method models IUR in a connected region locating scenario. As a result, the heavy burden for training and decoding hinders the efficiency significantly, especially when the dialogue context length increases. Thus, towards faster and more accurate IUR, this task should be modeled more naturally in a pair locating form where pairwise scorers locate pairs representing span and substitution relations. Following this idea, we propose a Cross Scorer Sharing (XSS) model to initially support pair locating. We experiment with XSS on both English and Chinese IUR datasets, and results have shown that our model leads to comparable or better performance than the previous state-of-the-art. For efficiency, our method is 132% faster when training and 85.1% faster when predicting.