Aligning Opinions: Cross-Lingual Opinion Mining with Dependencies

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 408–418, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Aligning Opinions: Cross-Lingual Opinion Mining with Dependencies Mariana S. C. Almeida∗† Cl´audia Pinto∗ Helena Figueira∗ Pedro Mendes∗ Andr´e F. T. Martins∗† ∗Priberam Labs, Alameda D. Afonso Henriques, 41, 2o, 1000-123 Lisboa, Portugal †Instituto de Telecomunicac¸˜oes, Instituto Superior T´ecnico, 1049-001 Lisboa, Portugal {mla,atm}@priberam.pt Abstract We propose a cross-lingual framework for fine-grained opinion mining using bitext projection. The only requirements are a running system in a source language and word-aligned parallel data. Our method projects opinion frames from the source to the target language, and then trains a sys- tem on the target language using the auto- matic annotations. Key to our approach is a novel dependency-based model for opin- ion mining, which we show, as a byprod- uct, to be on par with the current state of the art for English, while avoiding the need for integer programming or rerank- ing. In cross-lingual mode (English to Por- tuguese), our approach compares favor- ably to a supervised system (with scarce labeled data), and to a delexicalized model trained using universal tags and bilingual word embeddings. 1 Introduction The goal of opinion mining is to extract opinions and sentiments from text (Pang and Lee, 2008; Wilson, 2008; Liu, 2012). With the advent of so- cial media and the increasing amount of data avail- able on the Web, this has become a very active area of research, with applications in summariza- tion of customer reviews (Hu and Liu, 2004; Wu et al., 2011), tracking of newswire and blogs (Ku et al., 2006), question answering (Yu and Hatzivas- siloglou, 2003), and text-to-speech synthesis (Alm et al., 2005). While early work has focused on determining sentiment at document and sentence level (Pang et al., 2002; Turney, 2002; Balog et al., 2006), research has gradually progressed towards fine- grained opinion mining, where rather than deter- mining global sentiment, the goal is to parse text into opinion frames, identifying opinion expres- sions, agents, targets, and polarities (Ding et al., 2008), or addressing compositionality (Socher et al., 2013b). Since the release of the MPQA cor- pus1 (Wiebe et al., 2005; Wilson, 2008), a stan- dard corpus for fine-grained opinion mining of news documents, a long string of work has been produced (reviewed in §2). Despite the large vol- ume of prior work, opinion mining has by and large been limited to monolingual approaches in English.2 This is explained by the heavy effort of annotation necessary for current learning-based approaches to succeed, which delays the deploy- ment of opinion miners for new languages. We bridge the existing gap by proposing a cross-lingual approach to fine-grained opinion mining via bitext projection. This technique has been quite effective in several NLP tasks, such as part-of-speech (POS) tagging (T¨ackstr¨om et al., 2013), named entity recognition (Wang and Manning, 2014), syntactic parsing (Yarowsky and Ngai, 2001; Hwa et al., 2005), …

Content truncated to 500 words