A novel evolutionary approach for learning syntactic features for cross domain opinion target extraction

Aspect-Based Sentiment Analysis (ABSA) is the problem of mining textual user reviews to gauge the orientation of a user towards the various characteristics of a product. Identifying words in the review towards which the user holds a sentiment, known as Aspect Target Extraction (ATE) or Opinion Target Extraction (OTE), is a crucial step in ABSA. Several start-of-the-art techniques, mostly employing Deep Learning methods, have been proposed to tackle this problem. However, when attempting ATE in a new domain, most of the supervised methods are crippled by a lack of labeled data within the domain and do not perform well with models built on other unrelated domains. Other unsupervised methods based on linguistic rules framed by experts have been proposed to mitigate the problem of the lack of data but this requires manual engineering of rules, which may not capture all aspect terms. Different from the previous approaches, we propose evolutionary methods to generate pattern graphs, which encode rules to identify cross-domain aspect terms. The proposed methods aim to mine linguistic patterns, which are quality indicators of the presence of aspect words. The automated pattern learning technique using the evolutionary approach enables automated learning of rules. This approach eliminates the need for human experts to coin the same, while at the same time enabling the production of complicated rules, which can identify cross-domain aspects effectively. Two slightly different approaches are proposed — the first based on identifying opinion targets by matching the learned pattern graphs against existing sentences and the second based on using the learned pattern graphs to construct features for building classification models for aspect identification. We conduct extensive experiments using 10 real-life datasets of varying sizes and different characteristics, and demonstrate the superiority of the proposed methods for cross-domain ATE. We find that the proposed approaches perform better as compared to the existing methods in 69 of the 90 cases, or in other words, 76% of the cases. •Novel evolutionary algorithm based techniques for cross-domain aspect term extraction.•Automatically evolve rules(encoded as pattern graph sets) for identifying opinion targets based on linguistic characteristics of text i.e. dependency relations between words and the word features.•Two techniques proposed — the first which identifies aspects based on the direct matching of the learned rule