Systematic Generalization on gSCAN with Language Conditioned Embedding

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 27–36 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 27 Graph Attention Network with Memory Fusion for Aspect-level Sentiment Analysis Li Yuan†, Jin Wang†, 1, Liang-Chih Yu‡, *, 2 and Xuejie Zhang†, 3 †School of Information Science and Engineering, Yunnan University, Yunnan, P.R. China ‡Department of Information Management, Yuan Ze University, Taiwan *Innovation Center for Big Data and Digital Convergence Yuan Ze University, Taiwan Contact:{wangjin1, xjzhang3}@ynu.edu.cn, lcyu@saturn.yzu.edu.tw2 Abstract Aspect-level sentiment analysis(ASC) predicts each specific aspect term’s sentiment polar- ity in a given text or review. Recent stud- ies used attention-based methods that can ef- fectively improve the performance of aspect- level sentiment analysis. These methods ig- nored the syntactic relationship between the as- pect and its corresponding context words, lead- ing the model to focus on syntactically unre- lated words mistakenly. One proposed solu- tion, the graph convolutional network (GCN), cannot completely avoid the problem. While it does incorporate useful information about syn- tax, it assigns equal weight to all the edges between connected words. It may still incor- rectly associate unrelated words to the target aspect through the iterations of graph convo- lutional propagation. In this study, a graph attention network with memory fusion is pro- posed to extend GCN’s idea by assigning dif- ferent weights to edges. Syntactic constraints can be imposed to block the graph convolu- tional propagation of unrelated words. A con- volutional layer and a memory fusion were ap- plied to learn and exploit multiword relations and draw different weights of words to im- prove performance further. Experimental re- sults on five datasets show that the proposed method yields better performance than exist- ing methods. The code of this paper is avail- abled at https://github.com/YuanLi95/ GATT-For-Aspect. 1 Introduction Aspect-level sentiment classification is a fine- grained subtask in sentiment analysis (Wang et al., 2019; Peng et al., 2020). Given a sentence and an aspect that appears in the sentence, ASC aims to determine the sentiment polarity of that aspect (e.g., negative, neutral, or positive). For example, a review of a restaurant “The price is reasonable although the service is poor.” expresses a positive sentiment for the price aspect, but also conveys a negative sentiment for the service aspect, as shown in Figure 1. Such a technique is widely used to analyze online posts reviews, mainly from Ama- zon reviews or Twitter, to help raise the ability to understand consumer needs or experiences with a product, guiding a manufacturer towards product improvement. Aspect-level sentiment classifica- tion is much more complicated than sentence-level sentiment classification. ASC task is necessary to identify the parts of the sentence that describe the correspondence between multiple aspects. Tradi- tional methods mostly use shallow machine learn- ing models with hand-crafted features to build sen- timent classifiers for the ASC task (Jiang et al., 2011; Wagner et al., 2014).However, the process for manual feature engineering is …

Content truncated to 500 words