Rapid increase in conversational AI and user chat data lead to intensive development of dialogue management systems (DMS) for various industries. Yet, for low-resource languages, such as Azerbaijani, very little research has been conducted. The main purpose of this work is to experiment with various DMS pipeline set-ups to decide on the most appropriate natural language understanding and dialogue manager settings. In our project, we designed and evaluated different DMS pipelines with respect to the conversational text data obtained from one of the leading retail banks in Azerbaijan. In the work, the main two components of DMS—Natural language Understanding (NLU) and Dialogue Manager—have been investigated. In the first step of NLU, we utilized a language identification (LI) component for language detection. We investigated both built-in LI methods such as fastText and custom machine learning (ML) models trained on the domain-based dataset. The second step of the work was a comparison of the classic ML classifiers (logistic regression, neural networks, and SVM) and Dual Intent and Entity Transformer (DIET) architecture for user intention detection. In these experiments we used different combinations of feature extractors such as CountVectorizer, Term Frequency-Inverse Document Frequency (TF-IDF) Vectorizer, and word embeddings for both word and character n-gram based tokens. To extract important information from the text messages, Named Entity Extraction (NER) component was added to the pipeline. The best NER model was chosen among conditional random fields (CRF) tagger, deep neural networks (DNN), models and build in entity extraction component inside DIET architecture. Obtained entity tags fed to the Dialogue Management module as features. All NLU set-ups were followed by the Dialogue Management module that contains a Rule-based Policy to handle FAQs and chitchats as well as a Transformer Embedding Dialogue (TED) Policy to handle more complex and unexpected dialogue inputs. As a result, we suggest a DMS pipeline for a financial assistant, which is capable of identifying intentions, named entities, and a language of text followed by policies that allow generating a proper response (based on the designed dialogues) and suggesting the best next action.
Exploring and Improving Robustness of Multi Task Deep Neural Networks via Domain Agnostic Defenses