Edge-aware Graph Attention Networks: Joint Reasoning On Text and Knowledge Graphs for Biomedical Question Answering

Much of the real-world dataset, including textual data, can be represented using graph structures. The use of graphs to represent textual data has many advantages, mainly related to maintaining a more significant amount of information, such as the relationships between words and their types. In recent years, many neural network architectures have been proposed to deal with tasks on graphs. Many of them consider only node features, ignoring or not giving the proper relevance to relationships between them. However, in many node classification tasks, they play a fundamental role. This thesis aims to analyze the main GNNs, evaluate their advantages and disadvantages, propose an innovative solution considered as an extension of GAT, and apply them to a case study in the biomedical field. We propose the reference GNNs, implemented with methodologies later analyzed, and then applied to a question answering system in the biomedical field as a replacement for the pre-existing GNN. We attempt to obtain better results by using models that can accept as input both node and edge features. As shown later, our proposed models can beat the original solution and define the state-of-the-art for the task under analysis.

Vocabulary Transfer and Knowledge Distillation for Language Model Compression

Natural Language Processing (NLP) has seen tremendous improvements over the last few years. Transformer architectures achieved impressive results in almost any NLP task, such as Text Classification, Machine Translation, and Language Generation. As time went by, transformers continued to improve thanks to larger corpora and bigger networks, reaching hundreds of billions of parameters. Training and deploying such large models has become prohibitively expensive, such that only big high tech companies can afford to train those models. Therefore, a lot of research has been dedicated to reducing a model’s size. In this thesis, we investigate the effects of Vocabulary Transfer and Knowledge Distillation for compressing large Language Models. The goal is to combine these two methodologies to further compress models without significant loss of performance. In particular, we designed different combination strategies and conducted a series of experiments on different vertical domains (medical, legal, news) and downstream tasks (Text Classification and Named Entity Recognition). Four different methods involving Vocabulary Transfer (VIPI) with and without a Masked Language Modelling (MLM) step and with and without Knowledge Distillation are compared against a baseline that assigns random vectors to new elements of the vocabulary. Results indicate that VIPI effectively transfers information of the original vocabulary and that MLM is beneficial. It is also noted that both vocabulary transfer and knowledge distillation are orthogonal to one another and may be applied jointly. The application of knowledge distillation first before subsequently applying vocabulary transfer is recommended. Finally, model performance due to vocabulary transfer does not always show a consistent trend as the vocabulary size is reduced. Hence, the choice of vocabulary size should be empirically selected by evaluation on the downstream task similar to hyperparameter tuning.