Explaining the Stars: Weighted Multiple-Instance Learning for Aspect-Based Sentiment Analysis

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 455–466, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Explaining the Stars: Weighted Multiple-Instance Learning for Aspect-Based Sentiment Analysis Nikolaos Pappas EPFL and Idiap Research Institute Rue Marconi 19 CH-1920 Martigny, Switzerland nikolaos.pappas@idiap.ch Andrei Popescu-Belis Idiap Research Institute Rue Marconi 19 CH-1920 Martigny, Switzerland andrei.popescu-belis@idiap.ch Abstract This paper introduces a model of multiple- instance learning applied to the predic- tion of aspect ratings or judgments of specific properties of an item from user- contributed texts such as product reviews. Each variable-length text is represented by several independent feature vectors; one word vector per sentence or paragraph. For learning from texts with known as- pect ratings, the model performs multiple- instance regression (MIR) and assigns im- portance weights to each of the sentences or paragraphs of a text, uncovering their contribution to the aspect ratings. Next, the model is used to predict aspect ratings in previously unseen texts, demonstrating interpretability and explanatory power for its predictions. We evaluate the model on seven multi-aspect sentiment analysis data sets, improving over four MIR baselines and two strong bag-of-words linear mod- els, namely SVR and Lasso, by more than 10% relative in terms of MSE. 1 Introduction Sentiment analysis of texts provides a coarse- grained view of their overall attitude towards an item, either positive or negative. The recent abun- dance of user texts accompanied by real-valued la- bels e.g. on a 5-star scale has contributed to the de- velopment of automatic sentiment analysis of re- views of items such as movies, books, music or other products, with applications in social com- puting, user modeling, and recommender systems. The overall sentiment of a text towards an item often results from the ratings of several specific aspects of the item. For instance, the author of a review might have a rather positive sentiment about a movie, having particularly liked the plot and the music, but not too much the actors. De- termining the ratings of each aspect automatically is a challenging task, which may seem to require the engineering of a large number of features de- signed to capture each aspect. Our goal is to put forward a new feature-agnostic solution for ana- lyzing aspect-related ratings expressed in a text, thus aiming for a finer-grained, deeper analysis of text meaning than overall sentiment analysis. Current state-of-the-art approaches to sentiment analysis and aspect-based sentiment analysis, at- tempt to go beyond word-level features either by using higher-level linguistic features such as POS tagging, parsing, and knowledge infusion, or by learning features that capture syntactic and seman- tic dependencies between words. Once an appro- priate feature space is found, the ratings are typi- cally modeled using a linear model, such as Sup- port Vector Regression (SVR) with ℓ2 norm for regularization or Lasso Regression with ℓ1 norm. By treating a text globally, these models ignore the fact that the sentences of a text have diverse con- tributions to the overall sentiment or …

Content truncated to 500 words