Attention describes cognitive processes that are important to many human phenomena including reading. The term is also used to describe the way in which transformer neural networks perform natural language processing. While attention appears to be very different under these two contexts, this paper presents an analysis of the correlations between transformer attention and overt human attention during reading tasks. An extensive analysis of human eye tracking datasets showed that the dwell times of human eye movements were strongly correlated with the attention patterns occurring in the early layers of pre-trained transformers such as BERT. Additionally, the strength of a correlation was not related to the number of parameters within a transformer. This suggests that something about the transformers’ architecture determined how closely the two measures were correlated.

Current graph-neural-network-based (GNN-based) approaches to multi-hop questions integrate clues from scattered paragraphs in an entity graph, achieving implicit reasoning by synchronous update of graph node representations using information from neighbours; this is poorly suited for explaining how clues are passed through the graph in hops. In this paper, we describe a structured Knowledge and contextual Information Fusion GNN (KIFGraph) whose explicit multi-hop graph reasoning mimics human step by step reasoning. Specifically, we first integrate clues at multiple levels of granularity (question, paragraph, sentence, entity) as nodes in the graph, connected by edges derived using structured semantic knowledge, then use a contextual encoder to obtain the initial node representations, followed by step-by-step two-stage graph reasoning that asynchronously updates node representations. Each node can be related to its neighbour nodes through fused structured knowledge and contextual information, reliably integrating their answer clues. Moreover, a masked attention mechanism (MAM) filters out noisy or redundant nodes and edges, to avoid ineffective clue propagation in graph reasoning. Experimental results show performance competitive with published models on the HotpotQA dataset.

Multi-hop question answering (QA) requires reasoning over multiple documents to answer a complex question and provide interpretable supporting evidence. However, providing supporting evidence is not enough to demonstrate that a model has performed the desired reasoning to reach the correct answer. Most existing multi-hop QA methods fail to answer a large fraction of sub-questions, even if their parent questions are answered correctly. In this paper, we propose the Prompt-based Conservation Learning (PCL) framework for multi-hop QA, which acquires new knowledge from multi-hop QA tasks while conserving old knowledge learned on single-hop QA tasks, mitigating forgetting. Specifically, we first train a model on existing single-hop QA tasks, and then freeze this model and expand it by allocating additional sub-networks for the multi-hop QA task. Moreover, to condition pre-trained language models to stimulate the kind of reasoning required for specific multi-hop questions, we learn soft prompts for the novel sub-networks to perform type-specific reasoning. Experimental results on the HotpotQA benchmark show that PCL is competitive for multi-hop QA and retains good performance on the corresponding single-hop sub-questions, demonstrating the efficacy of PCL in mitigating knowledge loss by forgetting.