Interpretable Architectures and Algorithms for Natural Language Processing

• Published in: Doctoral Dissertations at the University of Agder
• Main Author: Yadav, Rohan Kumar
• Format: Dissertation
• Language: English
• Published: University of Agder 2022

This thesis has two parts: Firstly, we introduce the human level-interpretable models using Tsetlin Machine (TM) for NLP tasks. Secondly, we present an interpretable model using DNNs. The first part combines several architectures of various NLP tasks using TM along with its robustness. We use this model to propose logic-based text classification. We start with basic Word Sense Disambiguation (WSD), where we employ TM to design novel interpretation techniques using the frequency of words in the clause. We then tackle a new problem in NLP, i.e., aspect-based text classification using a novel feature engineering for TM. Since TM operates on Boolean features, it relies on Bag-of-Words (BOW), making it difficult to use pre-trained word embedding like Glove, word2vec, and fasttext. Hence, we designed a Glove embedded TM to significantly enhance the model’s performance. In addition to this, NLP models are sensitive to distribution bias because of spurious correlations. Hence we employ TM to design a robust text classification against spurious correlations. The second part of the thesis consists interpretable model using DNN where we design a simple solution for complex position dependent NLP task. Since TM’s interpretability comes with the cost of performance, we propose an DNN-based architecture using a masking scheme on LSTM/GRU based models that ease the interpretation for humans using the attention mechanism. At last, we take the advantages of both models and design an ensemble model by integrating TM’s interpretable information into DNN for better visualization of attention weights. Our proposed model can be efficiently integrated to have a fully explainable model for NLP that assists trustable AI. Overall, our model shows excellent results and interpretation in several open-sourced NLP datasets. Thus, we believe that by combining the novel interpretation of TM, the masking technique in the neural network, and the integrated ensemble model, we can build a simple yet effective platform for explainable NLP applications wherever necessary.