Octa: Omissions and Conflicts in Target-Aspect Sentiment Analysis

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1651–1662 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1651 Octa: Omissions and Conflicts in Target-Aspect Sentiment Analysis Zhe Zhang ∗†, Chung-Wei Hang ∗†, and Munindar P. Singh‡ †IBM Corporation ‡Department of Computer Science, North Carolina State University {zhangzhe,hangc}@us.ibm.com, singh@ncsu.edu Abstract Sentiments in opinionated text are often deter- mined by both aspects and target words (or tar- gets). We observe that targets and aspects inter- relate in subtle ways, often yielding conflicting sentiments. Thus, a naive aggregation of sen- timents from aspects and targets treated sepa- rately, as in existing sentiment analysis mod- els, impairs performance. We propose Octa,1 an approach that jointly considers aspects and targets when inferring sentiments. To capture and quantify relation- ships between targets and context words, Octa uses a selective self-attention mechanism that handles implicit or missing targets. Specif- ically, Octa involves two layers of attention mechanisms for, respectively, selective atten- tion between targets and context words and attention over words based on aspects. On benchmark datasets, Octa outperforms leading models by a large margin, yielding (absolute) gains in accuracy of 1.6% to 4.3%. 1 Introduction People share their opinions about almost anything: tourist attractions, restaurants, car dealerships, and products. Such opinionated texts do not merely help people make decisions in their daily life, but also help businesses measure consumer satisfaction to improve their offerings. Sentiment analysis involves many aspects of Nat- ural Language Processing, e.g., negation handling (Zhu et al., 2014), entity recognition (Mitchell et al., 2013), topic modeling (Zhang and Singh, 2018, 2019). Importantly, opinionated texts often convey conflicting sentiments. Distinct sentiments may refer to distinct aspects of the domain in question— e.g., food quality of a restaurant or battery life of ∗Equal contribution. 1The data and source code of Octa can be found at https: //github.com/chungweihang/octa a smartphone. These predefined domain aspects may or may not appear in the texts. Aspect-Based Sentiment Analysis (ABSA) approaches (Wang et al., 2016; Xue and Li, 2018; Liang et al., 2019) predict sentiments from text about a given aspect. And, Target-Based Sentiment Analysis (TBSA) approaches (Chen et al., 2017; Fan et al., 2018; Li et al., 2018; Du et al., 2019; Zhang et al., 2019) predict sentiments of targets that appear in an opin- ionated text. Targets are usually entities in a review: e.g., a dish for a restaurant and a salesperson for a car dealership. We posit that aspects and targets provide subtle, sometimes contradictory, information about sen- timent and should therefore be modeled, not in isolation, but jointly. Considering them separately, as ABSA and TBSA approaches do, impairs perfor- mance. Take this review sentence from SemEval- 15 as an example: Conflicting Sentiments on Aspect We both had the filet, very good, didn’t much like the frites that came with. If we ask about aspect Food#Quality, by disre- garding targets during training, ABSA models fail to address the contradiction in sentiment about filet and frites, as do TBSA models, which focus on targets …

Content truncated to 500 words