Abstract:
The paper presents a survey of methods solving the actual task of aspect-based sentiment analysis. Solutions for this task were proposed at multiple natural language processing conferences. Organizers of these conferences proposed evaluation platforms for methods for aspect-based sentiment analysis. This paper describes methods proposed by participants of two international evaluation platforms: SemEval-2015 focusing on English texts and SentiRuEval-2015 focusing on Russian texts.
SemEval-2015 organizers provided participants with the task 12.2 for English language and two domains: restaurants and laptops. The task was split to multiple subtasks two of which are described in this paper: opinion target expression (both explicit and implicit ones) extraction and sentiment polarity detection. Described methods for opinion target expression can be split to the following categories: sequence labeling, domain-specific terminology extraction and unsupervised learning. Methods for sentiment polarity detection varied from classification-based to unsupervised learning.
SentiRuEval-2015 organizers provided participants with the tasks A, B and C for Russian language and two domains: restaurants and automobiles. Task A was devoted to explicit aspect term extraction, task B – to explicit, implicit and factual aspect term extraction. Sentiment polarity detection was subject of the Task C. Described methods for aspect term extraction can be classified as following: sequence labeling, word2vec-based and neural network-based. Methods for sentiment polarity detection varied from word2vec-based to neural network-based.
Keywords:sentiment analysis, aspect extraction, text processing, machine learning.