Emotion-enriched word embeddings for Turkish

Abstract

Text, as one of the main communication methods, is frequently used as a data source for natural language processing (NLP) studies. Naturally, our thoughts, expressions, and actions are based on our feelings. Therefore, representing the language better to machines involves the problems of reflecting the actual meaning and also, the detecting emotion of the data source. While representing the textual data, word embeddings (e.g. Word2Vec and GloVe (Global Vectors for Word Representation)) which can extract semantic information are frequently used. However, these models may give unexpected results in sentiment and emotion detection studies because of the limitations of capturing emotive data. Because of occurring frequently in similar contexts, some words carrying opposite emotions may have similar vector representations. Nowadays, enriching the vectors by adding emotion or sentiment data is studied which aim to increase the success in emotion detection or classification tasks. The main purpose is to reorganize the vector space in a way that words having semantically and sentimentally similar in closer locations. In this study, three emotion enrichment models over two semantic embeddings (Word2Vec and GloVe) and a contextual embedding (BERT) (Bidirectional Encoder Representations from Transformers)) are applied to a Turkish dataset. Turkish is an agglutinative language. Thus, it is expected to produce different results in this problem, as it has a different structure from the languages that are frequently studied in this field. Besides, experiments on in-category/opposite-category cosine similarity based on eight emotion categories and classifi-cation with sequential minimal optimization, logistic regression and multi-layer perceptron are conducted. Ac-cording to experimental results emotionally enriched vector representations outperform the original models and give promising results.