Implicit Syntactic Features for Target-dependent Sentiment Analysis

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 516–524, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Implicit Syntactic Features for Targeted Sentiment Analysis Yuze Gao†, Yue Zhang† and Tong Xiao‡ Singapore University of Technology and Design† Northeastern University‡ yuze gao,yue zhang@sutd.edu.sg xiaotong@mail.neu.edu.cn Abstract Target-dependent sentiment analysis in- vestigates the sentiment polarities on given target mentions from input texts. Dif- ferent from sentence-level sentiment, it offers more fine-grained knowledge on each entity mention. While early work leveraged syntactic information, recent research has used neural representation learning to induce features automatically, thereby avoiding error propagation of syn- tactic parsers, which are particularly se- vere on social media texts. We study a method to leverage syntac- tic information without explicitly build- ing parser outputs, by training an encoder- decoder structure parser model on stan- dard syntactic treebanks, and then lever- aging its hidden encoder layers when analysing tweets. Such hidden vectors do not contain explicit syntactic outputs, yet encode rich syntactic features. We use them to augment the inputs to a baseline state-of-the-art target-dependent sentiment classifier, observing signifi- cant improvements on various benchmark datasets. We obtain the best accuracies on two different test sets for targeted senti- ment. 1 Introduction Target-dependent sentiment analysis investigates the problem of assigning sentiment polarity labels to a set of given target mentions in input sentences. Some example are shown in Table 1. For instance, given a sentence “I like [Twitter] better than [Face- book]”, a target-specific sentiment model is ex- pected to assign a positive (+) sentiment label on “I like [Twitter]+ better than Facebook” “I like Twitter better than [Facebook]−” [lindsay lohan]0 goes on yet another emo rant on her twitter. Choose [NBI]+ for insulation, home energy au- dits, housing repairing, air sealing, windows and doors,furnaces, air conditioners and en energy efficient appliances. Table 1: Target-dependent sentiment analysis w1 w2 w3 w4 w5 w6 w7 w8 w9 w10 w11 w12 w13 w14 Left Context Target Right Context ⊕ R L - + 0 Figure 1: Sentence level context the target Twitter and a negative (−) sentiment la- bel on the target Facebook. The task has been addressed using neural net- work models, which learn target-specific represen- tations of the input sentence. These representa- tions are then used for predicting target-dependent sentiment polarities. In particular, Dong et al. (2014) derive the syntactic structure of input sen- tence using a dependency grammar, before trans- forming the tree structure to a target-centered form. A recursive neural network is used to trans- form the dependency syntax of a sentence into a target-specific vector for sentiment classification. More recently Vo and Zhang (2015) split the in- put sentence into three segments, with the target entity mention being in the center, and its left and right contexts surrounding it, as shown in Figure 1. 516 xb Hidden Vectors fb (Implicit Syntax) xt * xt Top sentiment Model Input [web; wet; posb] Bottom syntactic Model Figure 2: Model structure Rich word embedding features are extracted …

Content truncated to 500 words