Contextualized intent detection using generalized SemSpace and BLSTM

In this thesis, a novel method called Contextualized Deep SemSpace has been developed to address the problem of the intent detection problem in the field of natural language processing. In the proposed method, firstly, sense-based vectors are generated from WordNet data using the generalized SemSpace approach, words representing the context of each dataset are clustered in this vector space, and then the disambiguation process is carried out by selecting the closest sense candidate to the context cluster of the words whose sense was ambiguous. Finally, the sense-based contextualized SemSpace vectors we generated are trained with the Bidirectional Long Short-Term Memory (BLSTM) model. To measure the success of the resulting model, tests are conducted on six well-known intent detection benchmark datasets (ATIS, Snips, Facebook, AskUbuntu, WebApp, and Chatbot). According to the comparison results, the recommended method generates context-based word vectors similarly to large language models such as BERT, ELMo, and GPT which are available in the literature, beacuse it uses both sense-based vectors and a context-based word disambiguation method. It is also predicted that this approach can be used successfully in many problems in the field of natural language processing.