Hybrid Emotion Detection with Word Embeddings in a Low Resourced Language: Turkish

Through natural language processing, subjective information can be obtained from written sources such as suggestions, reviews, and social media publications. Understanding and knowing the user experience or in other words the feelings/emotions of user on any type of product or situation directly aff...

Full description

Saved in:
Bibliographic Details
Published in:International journal of advanced computer science & applications 2024-01, Vol.15 (6)
Main Authors: Metin, Senem Kumova, Giraz, Hatice Ertugrul
Format: Article
Language:English
Subjects:
Computer science
Emotion recognition
Emotions
Experiments
Identification
Keywords
Language
Natural language processing
Product reviews
Psychologists
Representations
Semantic relations
Semantics
Similarity
User experience
User generated content
Words (language)
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Through natural language processing, subjective information can be obtained from written sources such as suggestions, reviews, and social media publications. Understanding and knowing the user experience or in other words the feelings/emotions of user on any type of product or situation directly affects the decisions to be taken on the regarding product or service. In this study, we focus on a hybrid approach of text-based emotion detection. We combined keyword and lexicon-based approaches by the use of word embeddings. In emotion detection, simply lexicon words/keywords and text units are compared in several different ways and the comparison results are used in emotion identification experiments. As this identification procedure is examined, it is explicit that the performance depends mainly on two actors: the lexicon/keyword list and the representation of text unit. We propose to employ word vectors/embeddings on both actors. Firstly, we propose a hybrid approach that uses word vector similarities in order to determine lexicon words, on contrary to traditional approaches that employs all arbitrary words in given text. By our approach, the overall effort in emotion identification is to be reduced by decreasing the number of arbitrary words that do not carry the emotive content. Moreover, the hybrid approach will decrease the need for crowdsourcing in lexicon word labelling. Secondly, we propose to build the representations of text units by measuring their word vector similarities to given lexicon. We built up two lexicons by our approach and presented three different comparison metrics based on embedding similarities. Emotion identification experiments are performed employing both unsupervised and supervised methods on Turkish text. The experimental results showed that employing the hybrid approach that involves word embeddings is promising on Turkish texts and also due to its flexible and language-independent structure it can be improved and used in studies on different languages.
ISSN:2158-107X
2156-5570
DOI:10.14569/IJACSA.2024.01506145