- Diversity, Equity, and Inclusion
- Research News
- Department Life
- Oral History of Cornell CS
- CS 40th Anniversary Booklet
- ABC Book for Computer Science at Cornell by David Gries
- Department Timeline
- Job Postings
- Ithaca Info
- Internal info
- Graduation Information
- Cornell Tech Colloquium
- Student Colloquium
- Fall 2023 Colloquium
- Conway-Walker Lecture Series
- Salton Lecture Series
- Seminars / Lectures
- Big Red Hacks
- Cornell University High School Programming Contests 2023
- Game Design Initiative
- CSMore: The Rising Sophomore Summer Program in Computer Science
- Explore CS Research
- ACSU Research Night
- Cornell Junior Theorists' Workshop
On Entity Recognition, Entity Linking and Entity Salience Detection
Abstract: A core problem in natural language processing and information extraction is entity linking, which is the process of linking an entity mention in a document to its corresponding record in a knowledge base. Entity linking is relevant to a variety of downstream applications such as relation extraction, knowledge base population and question answering. Additionally, attributing importance to each entity in the document, namely salience detection has been useful in domains where understanding user generated content is important, such as finance, marketing or journalism.
In the first part of this talk, I will present some of our improvements in entity linking methodology. In particular, I will talk about a novel structured learning model for collectively linking entities in a document which is based on gradient-tree-boosting, and a novel inference method based on bidirectional beam search. With the presented approach, we show a 1% accuracy improvement over the state-of-the-art on AIDA-CoNLL.
In the second part I will talk about an important, but often neglected follow-up task to entity linking -- the task of establishing the salience of the entity to the article. Determining salience and being able to attribute specific sections of the articles to specific entities enables meaningful inference from these articles. I will use Twitter in my presentation, which presents unique challenges to deducing entity salience due to lack of context and brevity of text.
This is joint work with Yi Yang, Shefaet Rahman, Daniel Preotiuc-Pietro, Charley Chan, and Steven Lu.
Bio: Ozan Irsoy is an AI Research Scientist at Bloomberg, working with the Dialogue Understanding team. His research interests include NLP applications of neural networks, deep learning under latency and space constrained settings, model compression and model interpretation. Prior to his role at Bloomberg, he was a PhD student at Cornell University, advised by Claire Cardie. Before that, he was pursuing his BSc in Computer Engineering and BA in Math at Bogazici University.