New Research Reveals the Key to Answering Ambiguous Questions Successfully
Published on Mon May 29 2023
Une tombe - A grave | Jeanne Menjoulet on flickrNew research has found that resolving ambiguous questions is crucial for successfully answering them. In a study focusing on questions about images, researchers created a dataset of ambiguous examples and annotated them to group answers by the underlying question they address. By rephrasing the questions for each group, they were able to reduce ambiguity. The findings reveal a linguistically-aligned ontology of reasons for ambiguity in visual questions.
To further investigate, the researchers developed an English question-generation model. Through automatic and human evaluations, they demonstrated that this model produces less ambiguous questions. Remarkably, the model was able to integrate answer group information without any direct supervision.
The study highlights the importance of understanding and addressing ambiguity in natural language. In particular, when it comes to visual question answering (VQA) models, reducing ambiguity can greatly enhance accuracy and effectiveness. These models, which seek to answer questions about images, have wide-ranging applications, including assistive technologies.
The researchers found that by rephrasing ambiguous questions to align with their respective answer groups, they were able to significantly decrease ambiguity. Interestingly, the automatic question-generation model showed promising results, producing questions with only a small drop in acceptability compared to human-generated questions. However, there is still room for improvement, as the model's questions were rated slightly less acceptable than those generated by humans.
Overall, this research provides valuable insights into the complexity of ambiguous questions and the potential for improving question clarity in VQA models. By reducing ambiguity, these models can enhance their ability to accurately answer questions about images, leading to advancements in various fields, such as assistive technology.
