Scholarly record

WORD EMBEDDING BASED PLACE THEME ANALYSIS IN USER-GENERATED DATA

Young-Min Lee

Abstract

We propose a theme analysis method that takes into consideration the perspectives of both a supplier (place) and a consumer (user) in unstructured place reviews written by users. To this end, ambiance on the place side and companion and time on the user side were considered as the theme elements. For the experiment, we used the review text data written in English provided by Yelp, and we selected coffee-related places (e.g., coffee shop, coffee house, etc.) from various place categories. From among them, we chose five places with the largest numbers of reviews. Word2vec, a word embedding technique, was used to analyze the detailed themes by the places. That is, after vectorizing each word using Word2vec, the correlation between theme words and theme representation words was calculated using the cosine similarity. As a result, we analyzed 17 theme representation words in the ambiance theme, eight in the companion theme, and three theme representation words in the time theme. Considering ambiance, companion, and time as the place theme, it is possible to search and recommend specific places such as ?quiet coffee shops?, ?restaurants to go to with children?, and ?good caf? in the morning?. In the future, applying this study to a recommendation system would provide users with various search results of places.

Publication details

Back to publication list