Intent Detection Using Contextualized Deep SemSpace

• Published in: Arabian journal for science and engineering (2011) Vol. 48; no. 2; pp. 2009 - 2020
• Main Authors: Orhan, Umut, Tosun, Elif Gulfidan, Ozkaya, Ozge
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.02.2023; Springer Nature B.V
• Subjects: Datasets; Deep learning; Engineering; Humanities and Social Sciences; multidisciplinary; Research Article-Computer Engineering and Computer Science; Science; Synset vectors; Bidirectional long short-term memory; Intent detection; Natural language understanding; WordNet; Generalized SemSpace
• ISSN: 2193-567X; 1319-8025; 2191-4281
• DOI: 10.1007/s13369-022-07016-9

In this study, a new approach called Contextualized Deep SemSpace is proposed for intent detection. First, the synset vectors are determined by training the generalized SemSpace method with the WordNet 3.1 data. Then, each word in an intent dataset is transformed into a synset vector by a contextualized approach, and finally, the synset vectors are trained with a deep learning model using BLSTM. Since the proposed approach adapts the contextualized semantic vectors to the dataset with a deep learning model, it treats like one of contextualized deep embeddings like BERT, ELMo, and GPT-3 methods. In order to measure the success of the proposed approach, some experiments have been carried out on six well-known intent detection benchmark datasets (ATIS, Snips, Facebook, Ask Ubuntu, WebApp, and Chatbot). Although the dependence of its vocabulary on WordNet causes a serious number of out of vocabulary problems, results showed that the proposed approach is the most successful intent classifier in the literature. According to these results, it can be said that deep learning-based contextualized synset vectors can be used successfully in many problems.