Structured Aspect Extraction

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2321–2332, Osaka, Japan, December 11-17 2016. Structured Aspect Extraction Omer Gunes† † Department of Computer Science University of Oxford United Kingdom first.last@cs.ox.ac.uk Tim Furche†‡ ‡ Meltwater Shoreditch, London United Kingdom first.last@meltwater.com Giorgio Orsi∗‡ ∗School of Computer Science University of Birmingham United Kingdom g.orsi@cs.bham.ac.uk Abstract Aspect extraction identifies relevant features of an entity from a textual description and is typ- ically targeted to product reviews, and other types of short text, as an enabling task for, e.g., opinion mining and information retrieval. Current aspect extraction methods mostly focus on as- pect terms, often neglecting associated modifiers or embedding them in the aspect terms without proper distinction. Moreover, flat syntactic structures are often assumed, resulting in inaccurate extractions of complex aspects. This paper studies the problem of structured aspect extraction, a variant of traditional aspect extraction aiming at a fine-grained extraction of complex (i.e., hierarchical) aspects. We propose an unsupervised and scalable method for structured aspect ex- traction consisting of statistical noun phrase clustering, cPMI-based noun phrase segmentation, and hierarchical pattern induction. Our evaluation shows a substantial improvement over existing methods in terms of both quality and computational efficiency. 1 Introduction The abundance of web data has made information extraction (IE) of strategic importance for data-driven companies such as, e.g., retailers, because of its applications to information retrieval (Kannan et al., 2011), knowledge base construction (Shin et al., 2015), media intelligence (Zeng et al., 2010), and conversational agents (Cassell, 2000). Among all IE-related tasks, Aspect extraction (AE) (Zhang and Liu, 2014) is certainly one of the most challenging. AE aims at identifying features of an entity, e.g., a phone, from a free text description ans is commonly associated with (aspect-based) sentiment analysis as a means of extracting aspect terms and associated opinions in, e.g., product reviews or other forms of opinionated and evaluative text (Hu and Liu, 2004; Popescu and Etzioni, 2005). Similar methods have also been used for extracting attributes from, e.g., product listings and other forms of descriptive text (Ghani et al., 2006; Kannan et al., 2011). An obvious challenge in aspect extraction is dealing with noisy, unstructured text (NUT). NUT has consistently challenged traditional NLP tools, in particular POS taggers and dependency parsers, due to ungrammatical sentences and incorrect orthography. Example 1: Aspect extraction Victorian two bedroom mid terrace property located in Cambridge and comprising of living room with ORIGINAL!!! cupboards, and ORIGINAL!!! picture rail.Stairway off living room leads to two bedrooms. ↓ { bedroom mid terrace, picture rail.Stairway, cupboards, bedrooms, property} {Cambridge, ORIGINAL} Example 1 shows the output of IBM’s Alchemy Language service1 when asked to retrieve aspect terms and named entities. Alchemy Language extracts picture rail.Stairway as an aspect due to a missing space after the full stop, and ORIGINAL as a named entity, due to incorrect orthography. Another challenge is distinguishing aspect terms from their modifiers. The terms Victorian, mid terrace, and the (nested) expression two bedroom are qualifying the term property, …

Content truncated to 500 words