LexiCon: Lexically Constrained Review Generation via Robust Insertion

Abstract

Existing review generators struggle to generate specific information correctly (e.g., Caesar salad, Snapdragon CPU), which prevents generated reviews from being more informative. In this paper, we propose to introduce lexical constraints into review generation which can be any key phrases to be contained in reviews. Compared to soft constraints (e.g., aspects) used in previous work, lexical constraints easily incorporate specific information which can largely improve the diversity and informativeness of generated reviews. To this end, we present LexiCon, a novel insertion-based review generation framework that can generate personalized reviews containing lexical constraints. Specifically, the proposed method progressively inserts new tokens between existing tokens in a parallel manner until a sequence is completed. Experimental results show that LexiCon outperforms the strongest review generation model by 20% BLEU-2 (coherence) and 68% Distinct-2 (diversity) on average. Human evaluation also shows that LexiCon is more robust to various lexical constraints than the state-of-the-art lexically-constrained model for general purpose.

Tasks

Reproductions