Understanding, Answering and Asking Questions Applications in Education, Intelligent Personal Agents and Medical Diagnostics
Abstract
The amount of data that is being created and stored at every level keeps growing. There is great potential to extract key insights from these large datasets. And indeed, there has been significant progress in AI systems for extracting, storing and retrieving this information. As an example, PubMed houses more than 26 million citations for biomedical literature from MEDLINE, life science journals, and online books and provides users an opportunity to query and understand a very large number of electronic health information resources. These structured queries are used to derive key insights in the fields of medicine, nursing, dentistry, veterinary medicine, health care systems, and preclinical sciences.There are a number of challenges in building AI systems that can efficiently extract, represent and reason with such large information sources. If posed with some questions about a given piece of information in textual or visual form, modern AI systems falter at multiple levels. First, they may not be able to understand free form questions or natural language instructions beyond a prototypical typecast. Even if a good understanding of the question is achieved, there are challenges in answering complex questions - especially when the answer is not explicitly stated but must be inferred. Even if such question answering systems exist, they do not have the ability to explain and discuss those answers. Finally, unlike humans, our systems cannot ask rich and probing questions that can help them improve their understanding about the information content. We focus on three aspects of this problem - questionunderstanding, question answering and question asking.Understanding free-form natural language questions could be hard. We all encounter situations in our life where we cannot search for the correct information on Google. In fact, a majority of users of search engines do not phrase their queries correctly. They are often not familiar with the schema of knowledge bases and the web and ask ambiguous queries. Interpreting such free-format queries into a more structured representation is a key step in retrieval and helping address the user s information needs.Question answering (QA) is an age-old problem in AI and we have indeed made some progress in this area. Yet, it is well known that the state-of-the-art QAtechnology still lacks a deep understanding of the world. We draw from a number of suggestions from the QAcommunity and propose to use question answering on scientific literature, and in particular, standardized tests like GRE, SAT, etc. for driving our research. These include reading comprehensions, math and science question answering, visual question answering, etc. Good performance on many of such tests requires sophisticated understanding and can serve as an indicator of progress in AI.Finally, any QA system should have an ability of self-reflection, i.e., it should know the holes in its own knowledge. A way to tackle this is to allow the system to ask questions about its missing knowledge. We explore the task of question asking by allowing our system to communicate with humans in the form of a dialog, asking questions one by one to solicit knowledge. In this context, we are exploring the relationship between information seeking and well-known information theoretic measures with theoretical and empirical underpinnings such as information gain, lattice theory and the calculus of inquiry. This work will be useful in building intelligent personal agents such as Siri and Google Now, which can assist people with basic daily tasks such as making restaurant reservations or commuting via public transport. These agents would also be of use in the clinical setting as assistants to doctors that are able to interact with patients and solicit information such as pre-existing conditions, medical history and current symptoms from patients bymeans of a constrained natural language dialog.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2017
- Source ID
- N000141712463
Entities
People
- Eric P. Xing
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy