SENPAI: Supporting Exploratory Text Analysis through Semantic & Syntactic Pattern Inspection

Abstract

Analyzing language for social computing tasks requires looking beyond individual words. For example, the word “please” generally signals politeness, but more so together with modal verbs (“could you please...”) than without (“please do this.”). Combining semantics and syntax into rich textual patterns is essential to capturing these nuances. What are the relevant patterns for a task, and how to find them? NLP practitioners choose patterns informed by theory, and find them through computational models. However, few tools allow identifying rich patterns without NLP expertise. We introduce SENPAI, a novel tool that discovers combined semantic and syntactic patterns. SENPAI fuses neural embeddings, dependency parsing, and graph mining to surface patterns directly from data. We apply SENPAI to measure credibility, politeness, and sentiment in text. Quantitatively, models powered by SENPAI perform similarly to theoretically-motivated ones. Qualitatively, SENPAI discovers patterns that are interpretable and meaningful. SENPAI enables building computational models without NLP expertise and discovering new linguistic constructs.