Understanding the Representational Power of Neural Retrieval Models Using NLP Tasks

The ease of constructing effective neural networks has resulted in a large number of varying architectures iteratively improving performance on a task. Due to the nature of these models being black boxes, standard weight inspection is difficult. We propose a probe based methodology to evaluate what information is important or extraneous at each level of a network. We input natural language processing datasets into a trained answer passage neural network. Each layer of the neural network is used as input into a unique classifier, or probe, to correctly label that input with respect to a natural language processing task, probing the internal representations for information. Using this approach, we analyze the information relevant to retrieving answer passages from the perspective of information needed for part of speech tagging, named entity retrieval, sentiment classification, and textual entailment. We show a significant information need difference between two seemingly similar question answering collections, and demonstrate that passage retrieval and textual entailment share a common information space, while POS and NER information is used only at a compositional level in the lower layers of an information retrieval model. Lastly, we demonstrate that incorporating this information into a multitask environment is correlated to the information retained by these models during the probe inspection phase.