IDIAP Speaker Series and Public Talks
4th March 2020, 11:00 AM:
Julien Perez: Iterative Reasoning Path Retrieval for Multi-Hop Question Answering
Abstract: Multi-hop Machine Reading necessitates retrieving multiple pieces of evidence over a possibly large collection of documents. One of the challenges comes from the fact that only a few lexical or semantic relationships are overlapping with the question. For this reason, classic information retrieval methods, which are mainly based on word matching techniques, even when distributional, have failed to achieve usable results in such a task. This talk introduces a series of novel neural iterative retrieval approaches that learn to find the sequence of necessary pieces of evidence, also called reasoning paths, to answer open-domain multi-hop questions. The approach trains a neural network based on a multi-headed transformer architecture that learns to retrieve evidence paragraphs conditioned on the previously retrieved documents. Any sequential search algorithm, like beam-search or MCTS, can then be coupled to our model to find the most probable sequence regarding a question. Finally, we use a neural approach as a stopping criteria for sequential retrieval. Experimental results show encouraging results in an open-domain QA dataset, HotpotQA. We will conclude the talk with current limitations and perspectives associated to this task.