Events

Importance of Data and Controllability in Neural Language Generation

Natural language generation has become a popular playground for deep learning techniques. In this talk, I will demonstrate that creating high-quality training data and introducing controllability over different editing operations (such as paraphrasing, sentence splitting, etc.) can lead to significant performance improvements that overshadow gains from model variations. In particular, I will focus on the text simplification task that improves text accessibility, including: (1)  a monolingual word alignment model that can identify semantically related text spans between two sentences for analyzing human editing operations; (2) a controllable text generation approach that incorporates syntax through pairwise ranking and data argumentation; (3) a neural conditional random field (CRF) based semantic model to create parallel training data. I will also briefly discuss our other work on large-scale paraphrase acquisition from Twitter.