Constituency Lattice Encoding for Aspect Term Extraction

Proceedings of the 28th International Conference on Computational Linguistics, pages 844–855 Barcelona, Spain (Online), December 8-13, 2020 844 Constituency Lattice Encoding for Aspect Term Extraction Yunyi Yang∗, Kun Li∗, Xiaojun Quan†, Weizhou Shen and Qinliang Su School of Data and Computer Science Sun Yat-sen University, Guangzhou, China {yangyy37, likun36}@mail2.sysu.edu.cn quanxj3@mail.sysu.edu.cn Abstract One of the remaining challenges for aspect term extraction in sentiment analysis resides in the extraction of phrase-level aspect terms, which is non-trivial to determine the boundaries of such terms. In this paper, we aim to address this issue by incorporating the span annotations of con- stituents of a sentence to leverage the syntactic information in neural network models. To this end, we first construct a constituency lattice structure based on the constituents of a constituency tree. Then, we present two approaches to encoding the constituency lattice using BiLSTM-CRF and BERT as the base models, respectively. We experimented on two benchmark datasets to eval- uate the two models, and the results confirm their superiority with respective 3.17 and 1.35 points gained in F1-Measure over the current state of the art. The improvements justify the effectiveness of the constituency lattice for aspect term extraction.1 1 Introduction In aspect-based sentiment analysis, aspect term extraction (ATE) serves as a fundamental task that aims to identify aspect terms from review texts (Hu and Liu, 2004; Popescu and Etzioni, 2005). An aspect term can be a word or a phrase that describes certain attribute of an entity (e.g., restaurant). For example, in the sentence “The food is very average, the Thai fusion stuff is a bit too sweet”, the terms “food” and “Thai fusion stuff” are the aspect terms to extract in this task. One of the challenges for ATE is to extract uncommon phrase-level aspect terms, which may contain complicated structures of words. One line of the research work on ATE leverages the syntactic structure of a sentence explicitly. Early research in this direction focuses on rule-based (Hu and Liu, 2004; Qiu et al., 2011) or feature engineering approaches (Jin and Ho, 2009; Li et al., 2010) that require manually designed rules based on part-of- speech tagging or dependency structure. They may also rely on features from predefined lexicons or syntactic analysis from annotated corpus. These approaches have unavoidable limitations in that they are labor-intensive and it takes extensive efforts to construct the rules and features. Besides, there are neural network approaches (Yin et al., 2016; Wang et al., 2016) that encode the dependency parse structure of a sentence in neural models. These work primarily focuses on syntactic dependency information of sentence, given that the dependency structure may indicate the position of a potential aspect term. Nevertheless, their tree-dependent nature of the encoding process makes the optimization inconvenient. Another line of work resorts solely to deep learning techniques such as LSTM (Liu et al., 2015), CNN (Xu et al., 2018), attention mechanism (Wang et al., 2017; Li et al., 2018) and Transformer (Devlin et al., 2019; Xu et al., 2019) to automate the …

Content truncated to 500 words