Why Did the Chicken Cross the Road? Rephrasing and Analyzing Ambiguous Questions in VQA

Anonymous

Abstract: Resolving ambiguities in questions is key to successfully answering them. Focusing on questions about images, we create a dataset of ambiguous examples; we annotate these examples, grouping the answers by the underlying question they address and rephrasing the question for each group to reduce ambiguity. An analysis of our data reveals a linguistically-aligned ontology of reasons why questions can be ambiguous. We then develop a model for question generation in English which we demonstrate via automatic and human evaluation produces less ambiguous questions. We further show that the question generation objective we use allows the model to integrate answer group information without any direct supervision. Finally, we employ our dataset and ontology for linguistic hypothesis testing, finding evidence against a previous syntactic account of ambiguity in ``why'' questions. We will release our code and data.

Paper Type: long