I am motivated by the challenge of building systems that can extract, understand, and reason
with information in natural language texts. My research interests are in two broad areas:
and Artificial Intelligence
. Here is a brief description
of my current research projects:
- Knowledge Extraction
My current knowledge extraction efforts focus on modeling how typical situations unfold. One studies scientific processes, and the other studies common-sense knowledge about events.
We are developing a question answering system that can recognize instances of processes.
We are designing a system that can automatically identify the key semantic roles and use them
in an alignment framework to answer questions. Project involves extending state-of-the-art
techniques in NLP and Machine learning. [K-CAP iKNOW 2015]
Common-sense knowledge about events
We are interested in modeling common-sense knowledge about events.
Our current work looks at generating effective representations for events through tensor-based compositions [AAAI 2018].
In my preivous work, I built a semantic resource called Rel-grams, a relational analogue to lexical n-grams.
Rel-grams capture co-occurrence between relations expressed in text and can be used for building open event schemas [EMNLP 2013, AKBC-WEKEX 2012].
- Reasoning with text-derived knowledge
As part of the AI2's
Project Aristo, I am involved in a couple of QA research threads.
The Allen Institute for Artificial Intelligence is generously funding this effort.
Probabilistic Reasoning for QA: I am interested in understanding how to
represent information in natural language texts in logical formalisms (e.g, FOL) and reason
with incomplete and possibly inconsistent knowledge. To this end, we investigate Markov Logic Networks as a probabilistic reasoning framework [StarAI 2015, EMNLP 2015].
Controlling aggregation in complex QA: One of the key challenges in reasoning with text-based knowledge is that reasoning can easily drift on to irrelevant concepts. We show how we can use question context to guide
reasoning in a page rank based formalism [ECIR 2018]. Our current work explores how this can be modeled in a neural framework.
- Privacy and Mobile:
We are interested in developing language based intelligence services that provide privacy protections to the users.
One of our current goals is to build a privacy dashboard that can inform users about possible inferences that can be made about them. Future work is going to look at also providing tools that limit these sensitive inferences.
These services ought to be able to run on the user's edge devices. Our PrIA work showed how to build news recommendation systems that run on user devices without revealing information about the user to the news aggregators [HotMobile 2017]. We are developing deep learning optimizations for NLP tools that allow for these models to run on limited resource devices. We showed how RNNs can be optimized [EMDL 2017] and have some results on porting a question answering system to phones.