Graph-based Topic Extraction Using Centroid Distance of Phrase Embeddings on Healthy Aging Open-ended Survey Questions

Abstract

Open-ended questions are a very important part of research surveys. However, they can pose a challenge when it comes to processing since manual processing requires a labour-intensive human effort. Automation of the task requires application of NLP methods since free text does not ensure standardized structure. To tackle this problem, we present a solution for topic discovery and analysis of open-ended survey items. We use graph-based representation of the text that adds structure and enables easier manipulation and keyphrase retrieval. Additionally, we use pre-trained fastText aligned word vectors to cluster similar phrases even if they are written in different languages. The goal is to produce topic word and phrase representatives that are easy to interpret by a domain expert. We compare the method with traditional LDA and two state-of-the-art algorithms: BTM and WNTM. The resulting keyphrases representing topics are more intuitive to the domain experts than the ones obtained by reference topic models in similar experimental settings.