Sentence-level heuristic tree search for long text generation

Abstract

AbstractIn this study, we primarily aim to address the exposure bias issue in long text generation intrinsic to statistical language models. We propose a sentence-level heuristic tree search algorithm, specially tailored for long text generation, to mitigate the problem by managing generated texts in a tree structure and curbing the compounding of biases. Our algorithm utilizes two pre-trained language models, an auto-regressive model for generating new sentences and an auto-encoder model for evaluating sentence quality. These models work in tandem to perform four critical operations: expanding the text tree with new sentences, evaluating the quality of the additions, sampling potential unfinished text fragments for further generation, and pruning leaf nodes deemed unpromising. This iterative process continues until a pre-defined number of [EOS] tokens are produced, at which point we select the highest-scoring completed text as our final output. Moreover, we pioneer two novel token-level decoding techniques—nucleus sampling with temperature and diverse beam search with sampling. These methods, integrated with our sentence-level search algorithm, aim to improve the consistency and diversity of text generation. Experimental results, both automated measures (including Jaccard similarity, Word2vec similarity, and unique word ratio) and human evaluations (assessing consistency, fluency, and rhetorical skills), conclusively demonstrate that our approach considerably enhances the quality of machine-generated long-form text. Through this research, we aim to inspire further innovations in sentence-level search-based text generation algorithms.