Prajna Upadhyay will present her work on April 28th at 3pm.
It will be online at https://ecolepolytechnique.zoom.us/j/86323133834?pwd=QzFqdlpwalBwTUtRbzQxYWQwSXNLUT09
Title: Construction and Applications of a Technical Knowledge Base
Abstract:
Recent years have witnessed an exponential increase in the availability of technical material on the web, possibly due to the efforts to digitize education and open access to scientific publications. However, it is not necessary that all of this readily available material is comprehensible to most users. A user who is interested to work on a research problem or a student willing to take up a new course may come across multiple topics about which she has little / no idea. She may find a lot of learning material on that topic on the web, however, due to the lack of prerequisite knowledge, she will have to perform multiple searches before obtaining a basic understanding of that topic. To keep up with the latest research,she has to identify and understand different aspects of the topic to conduct a survey which can be overwhelming for her due to the large amount of technical material available on the web. It would be helpful if there was a system to recommend to her prerequisite concepts for basic understanding and research papers for advanced understanding of the topic.
To build such a system, we first have to store the knowledge from the technical domain in the form of entities and relations as a knowledge graph. This will help design applications to consume this information in a systematic way. So, I will first talk about the construction of TeKnowbase, which is a domain-specific knowledge base in Computer Science. I will then describe ASK (Aspect-based academic Search using domain-specific KBs), an application of TeKnowbase to improve the academic search experience of the user. More specifically, it assists the user in an aspect-based retrieval of research papers. The need for aspect-based retrieval arises when the user is also aware of an aspect along with the query to retrieve a research paper.While existing search engines can be used by expanding the query with terms describing that aspect, this approach does not guarantee good results because documents that contain the query and the aspect terms may not be relevant and a relevant document may contain terms apart from the query and the aspect, determined by the relationship between the two. To address this issue, ASK uses language models estimated for the query and the aspect using TeKnowbase to determine the relevance of a paper for a query and aspect. User evaluation of the papers retrieved by ASK shows that they were better than those retrieved using state-of-the-art pseudo-relevance feedback or diversification models.
Bio:
Prajna Upadhyay joined the CEDAR team as a research engineer in March 2021 after submitting her Ph.D. thesis from the Indian Institute of Technology, Delhi, India.
