Faculty Recruiting Support CICS

Neural Methods For Answer Passage Retrieval Over Sparse Collections

26 Oct
Monday, 10/26/2020 12:00pm to 2:00pm
Zoom Meeting
PhD Seminar
Speaker: Dan Cohen

Zoom meeting: https://umass-amherst.zoom.us/j/7950401127?pwd=UDVyMW8xaVVOTGlMeHBQYlVKOUdMdz09
Meeting ID: 795 040 1127
Passcode: 280608


Recent advances in machine learning have allowed information retrieval (IR) techniques to advance beyond the stage of handcrafting domain specific features. Specifically, deep neural models incorporate varying levels of features to learn whether a document answers the information need of a query. However, these neural models rely on a large number of parameters to successfully learn a relation between a query and a relevant document.

This reliance on a large number of parameters, combined with the current methods of optimization relying on small updates necessitates numerous samples to allow the neural model to converge on an effective relevance function. This presents a significant obstacle in the realm of IR as relevance judgements are often sparse or noisy and combined with a large class imbalance. This is especially true for short text retrieval where there is often only one relevant passage.

This problem is exacerbated when training these artificial neural networks, as excessive negative sampling can result in poor performance. Thus, we propose approaching this task through multiple avenues and examining their effectiveness on a non-factoid question answering (QA) task.

We first propose learning local embeddings specific to the relevance information of the collection to improve performance of an upstream neural model. In doing so, we find significantly improved results over standard pre-trained embeddings, despite only developing the embeddings on a small collection which would not be sufficient for a full language model. Leveraging this local representation, and inspired by recent work in machine translation, we introduce a hybrid embedding based model that incorporates both pre-trained embeddings while dynamically constructing local representations from character embeddings. The hybrid approach relies on pre-trained embeddings to achieve an effective retrieval model, and continually adjusts its character level abstraction to fit a local representation.

We next approach methods to adapt neural models to multiple IR collections, therefore reducing the collection specific training required and alleviating the need to retrain a neural model's parameters for a new subdomain of a collection. First, we propose an adversarial retrieval model which achieves state-of-the-art performance on out of subdomain queries while maintaining in-domain performance. Second, we establish an informed negative sampling approach using a reinforcement learning agent. The agent is trained to directly maximize the performance of a neural IR model using a predefined IR metric by choosing which ranking function from which to sample negative documents. This policy based sampling allows the neural model to be exposed to more of a collection and results in a more consistent neural retrieval model over multiple training instances.

Lastly, we move towards a universal retrieval function. We initially introduce a probe-based inspection of neural relevance models through the lens of standard natural language processing tasks and establish that while seemingly similar QA collections require the same basic abstract information, the final layers that determine relevance differ significantly. We then introduce Universal Retrieval Functions, a method to incorporate new collections using a library of previously trained linear relevance models and a common neural representation.

Advisor: Bruce Croft