Terminology Extraction Approaches for Product Aspect Detection in Customer Reviews

Proceedings of the Seventeenth Conference on Computational Natural Language Learning, pages 222–230, Sofia, Bulgaria, August 8-9 2013. c⃝2013 Association for Computational Linguistics Terminology Extraction Approaches for Product Aspect Detection in Customer Reviews J¨urgen Broß Institute of Computer Science Freie Universit¨at Berlin 14195 Berlin, Germany juergen.bross@fu-berlin.de Heiko Ehrig Neofonie GmbH Robert-Koch-Platz 4 10115 Berlin, Germany heiko.ehrig@neofonie.de Abstract In this paper, we address the problem of identifying relevant product aspects in a collection of online customer reviews. Be- ing able to detect such aspects represents an important subtask of aspect-based re- view mining systems, which aim at auto- matically generating structured summaries of customer opinions. We cast the task as a terminology extraction problem and ex- amine the utility of varying term acquisi- tion heuristics, filtering techniques, vari- ant aggregation methods, and relevance measures. We evaluate the different ap- proaches on two distinct datasets (hotel and camera reviews). For the best config- uration, we find significant improvements over a state-of-the-art baseline method. 1 Introduction Identifying significant terms in a text corpus con- stitutes a core task in natural language process- ing. Fields of application are for example glos- sary extraction (Kozakov et al., 2004) or ontology learning (Navigli and Velardi, 2004). In this work, we particularly focus on the application scenario of aspect-based customer review mining (Hu and Liu, 2004; Dave et al., 2003). It is best described as a sentiment analysis task, where the goal is to summarize the opinions expressed in customer reviews. Typically, the problem is decomposed into three subtasks: 1) identify mentions of rele- vant product aspects, 2) identify sentiment expres- sions and determine their polarity, and 3) aggre- gate the sentiments for each aspect. In this paper, we only consider the first subtask, i.e., finding rel- evant product aspects in reviews. More precisely, we define the problem setting as follows: Input is a homogeneous collection of customer reviews, i.e., all reviews refer to a sin- gle product type (e.g., digital cameras or hotels). The goal is to automatically derive a lexicon of the most relevant aspects related to the product type. For example, given a set of hotel reviews, we want to determine aspects such as “room size”, “front desk staff” “sleep quality”, and so on. In gen- eral, product aspects may occur as nominal (e.g., “image stabilization”), named (e.g., “SteadyShot feature”), pronominal (e.g., “it”), or implicit men- tions (e.g., “reduction of blurring from camera shake”). We explicitly restrict the task to finding nominal aspect mentions1. The contribution of this paper is to explicitly cast the problem setting as a terminology extrac- tion (TE) task and to examine the utility of meth- ods that have been proven beneficial in this con- text. Most related work does not consider this close relationship and rather presents ad-hoc ap- proaches. Our main contributions are as follows: – We experiment with varying term acquisition methods, propose a set of new term filtering ap- proaches, and consider variant aggregation tech- niques typically applied in TE systems. – We compare the utility of …

Content truncated to 500 words