Aspect-based sentiment analysis with word embeddings for Lithuanian

Opinion mining (sentiment analysis) problem is usually solved by applying a lexicon-based model (e.g., in www.semantika.lt project). Deep neural networks are gaining popularity in different text processing and classification tasks, sentiment analysis being one of them. The paper presents research on deep-learning algorithms for aspect-based sentiment analysis, aiming at sentiment definition for each identified aspect of a given entity (e.g., a person, a brand name, etc.). Most of the distributional word embedding models nowadays learn semantic representations of words ignoring their morphological structure. This appears to be a limitation, especially for languages with complex morphology, as their vocabularies are quite large, most words are infrequent, resulting in models being unable to learn acceptable semantic representations. The paper presents the research results on the application of a FastText n-gram (subword) based word embedding model, treating each word as a composition of character n-grams (subwords). Subword-level information is crucial for capturing word meaning and morphology, especially for out-of-vocabulary (OOV) entries and rare words. Comparison with the traditional word2vec (Skip-Gram) word embedding model showed, that the subword feature enhances learning for morphologically rich, heavily inflected languages, such as the Lithuanian language. It is shown that for Lithuanian texts FastText models with n-grams do significantly better on syntactic tasks (aspect-sentiment pair), because of the syntactic questions being related to morphology of the words. However, original word2vec models perform better on semantic tasks, since words in semantic analogues (distributional hypothesis) are unrelated to their n-grams, and the add-on information from irrelevant character n-grams can worsen the embeddings. [...]Informatikos fakultetasTaikomosios informatikos katedraVytauto Didžiojo universiteta