Attention mechanism optimization for sub-symbolic-based and neural-symbolic-based natural language processing

• Main Author: Ni, Jinjie
• Other Authors: Erik Cambria
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/168430
• Institution: Nanyang Technological University

The capability for machines to transduce, understand, and reason with natural language lives at the heart of Artificial Intelligence not only because natural language is one of the main mediums for information delivery, residing in documents, daily chats, and databases of various languages, but also because it involves many key aspects of intelligence (e.g., logic, understanding, abstraction, etc.). Empowering the machine with more linguistic intelligence may benefit a wide range of real-world applications such as Machine Translation, Natural Language Understanding, Dialogue Systems, etc. At present, there are two popular streams of approaches for building intelligent Natural Language Processing (NLP) systems, i.e., sub-symbolic and neural-symbolic approaches. Sub-symbolic approaches learn implicit representations on the corpus that is unstructured, which is massive in amount but results in poor interpretability and reasoning ability of the learned models; neural-symbolic approaches integrate neural and symbolic architectures to incorporate structured symbolic data (e.g., semantic nets, knowledge graphs, etc.) as an external knowledge source, which makes the learned model more interpretable and logical, but the structured symbolic data is hard to be fully represented and it is comparatively scarce. As a result, both streams of approaches deserve studying, since they have their respective strengths and weaknesses, working complementarily in different tasks/scenarios. Meanwhile, attention-based models, such as Transformers, have achieved huge success in many NLP tasks such as Machine Translation, Language Modeling, Question Answering, etc. However, the attention itself has many issues, such as redundancy, quadratic complexity, weak inductive bias, etc. Besides, the previous applications of attention-based models in various NLP tasks are problematic, e.g., omitting the prior attention distribution, large computation complexity, weak long-term reasoning capability, etc. To this end, this thesis explores novel attention architectures for NLP tasks that are currently based mainly on sub-symbolic or neural-symbolic approaches to solve the existing issues and advance the state-of-the-art. In particular, for sub-symbolic-based tasks, we study Machine Translation, Language Modeling, Abstractive Summarization, and Spoken Language Understanding; for neural-symbolic-based tasks, we study Dialogue Commonsense Reasoning. The following lists the main contributions of this thesis: We study the redundancy and over-parameterization issues of Multi-Head Attention (MHA). We find that, in a certain range, higher compactness of attention heads (i.e., the intra-group heads become closer to each other and the inter-group ones become farther) improves the performance of MHA, which forces the MHA to focus on the most representative and distinctive features, providing guidance for future architectural designs. Accordingly, we propose a divide-and-conquer strategy that consists of Group-Constrained Training (GCT) and Voting to Stay (V2S). It mitigates the redundancy and over-parameterization issues of MHA. Our method uses fewer parameters and achieves better performance, outperforming the existing MHA redundancy/parameter reduction methods. We verify our methods on three well-established NLP tasks (i.e., Machine Translation, Language Modeling, and Abstractive Summarization). The superior results on datasets with multiple languages, domains, and data sizes demonstrate the effectiveness of our method. We ease the modality and granularity inconsistency problem when distilling knowledge from the teacher understanding model to the student ones, by refining the attention hidden states based on the attention map distribution. We propose to apply the Attention-based Significance Priors (ASP) to improve the semantic knowledge transfer from text to speech. We further propose the Anchor-based Adaptive Span Aggregation algorithm (AASA) that narrows the modal granularity gap of alignments. To the best of our knowledge, we are the first that evaluate multiple different alignment strategies beyond vanilla global and local alignments to study the feasibility of metric-based speech-text distillations. The results on three spoken language understanding benchmarks (i.e., Intent Detection, Slot Filling, and Emotion Recognition) verify our assumptions and claims. We improve the multi-source and long-term Dialogue Commonsense Reasoning (DCR) process, which is a new and difficult problem in NLP, by presenting a hierarchical attention-based decoding block. We propose the first Transformer-based KG walker that attentively reads multiscale inputs for graph decoding. Specifically, Multi-source Decoding Inputs (MDI) and Output-level Length Head (OLH) are presented to strengthen the controllability and multi-hop reasoning ability of the Hierarchical Attention-based Graph Decoder (HAGD). We further propose a two-hierarchy learning framework to train the proposed hierarchical attention-based KG walker, in order to learn both turn-level and global-level KG entities as conversation topics. This is the first attempt to learn models to make natural transitions towards the global topic in KG, where we present a distance embedding to incorporate distance information. Moreover, we propose MetaPath (MP) to concurrently exploit entity and relation information when reasoning, which is proved essential as the backbone method for KG path representation, providing a paradigm for KG reasoning. The results on the DCR dataset OpendialKG show that HiTKG achieves a significant improvement in the performance of turn-level reasoning compared with state-of-the-art baselines. Additionally, both automatic and human evaluation prove the effectiveness of the two-hierarchy learning framework for both short-term and long-term DCR.