Abstract:
The rise in popularity of mobile and voice search has led to a shift in IR from document to passage retrieval for non-factoid questions. Various datasets such as MSMarco, as well as efficient retrieval models have been developed to identify single best answer passages for this task. However, such models do not specifically address questions which could have multiple or alternative answers. In this dissertation, we focus on this new research area that involves studying answer passage relationships and how this could be applied to passage retrieval tasks.
We first create a high-quality dataset for the answer passage similarity task in the context of question answering. Manual annotation of passage pairs is performed to set the similarity labels, from which answer group information is automatically generated. We next investigate different types of representations, which could be used to create effective clusters. We experiment with various unsupervised representations and show that distributional representations outperform term-based representations for this task. Next, weak supervision is leveraged to further improve the cluster modeling performance. We use BERT as the underlying model for training and show the relative performance of various weak signals such as GloVe and term-based Language Modeling for this task. To apply these clusters to the answer passage retrieval task for multi-answer questions, we use a modified version of the Maximal Marginal Relevance (MMR) diversification model. We demonstrate that answers retrieved using this model are more diverse i.e, cover more answer types with low redundancy as well as maximize relevance, with respect to the baselines. So far, we used passage clustering to identify answer groups corresponding to a question and apply them in a question answering task. We extend this a step further by looking at related questions within a conversation. For this purpose, we expand the definition of Reciprocal Rank Fusion (RRF) and use this to identify pertinent history passages for such questions. Updated question rewrites generated using these passages are then used to improve the conversational search task.
In addition to being the first work that looks at answer relationships, our specific contributions can be summarized as follows: (1) Creation of new datasets with passage similarity and answer type information; (2) Effective passage similarity clustering models using unsupervised representations and weak supervision methods; (3) Applying the passage similarity/clustering information to diversification framework; (4) Identifying good response history candidates using answer passage clustering for the conversational search task.
Advisor: W. Bruce Croft