Quick Notes - Day 16
All previous posts in this series here
Named Entity Recognition as Dependency Parsing
Authors: Juntao Yu, Bernd Bohnet, Massimo Poesio
Published at: ACL 2020 url
This paper takes ideas from dependency parsing model called a “biaffine model” and applies it for Named Entity Recognition (NER) to identify both flat as well as nested entities. They do this by reformulating the NER task as identifying the start and end of a span and assigning a category to this span. They then evaluated their approach with 8 datasets (3 of them have nested NERs) and showed that it did better on all. The code is opensourced.
I found the idea interesting, and found it cool that it did well with both flat and nested NER on various datasets. However, I felt the paper could have had at least a small section on discussing the results more e.g., where does it do well, where does it do poorly? is a particular kind of nested entity identified better? What could possible shortcomings of this approach be? How can they be overcome? - I find it extremely surprising that papers that claim sota don’t ask these questions nor talk about possible drawbacks of their approach. It is not that uncommon to see such papers though.
Deep Learning for Political Science
Authors: Kakia Chatsiou, Slava Jankin Mikhaylov
Published at: The SAGE Handbook of Research Methods in Political Science and International Relations, 2020 url
This is a book chapter on using deep learning methods for political science research. They start with an overview of what is machine learning, what are different forms of machine learning, what are different challenges in NLP, different kinds of data etc. Typical textbook stuff.
I am always curious about communicating the jargon of one discipline to another with the goal of using methods from one in another. Such writeups are also essential to drive inter-disciplinary research (such research is very important IMO). However, I felt this paper did not connect much to political science problems. For example, I got some idea of what are different use cases for machine learning/DL/NLP in pol. sci, but why specifically all the DL approaches? Why not more basic stuff? Is there evidence that DL approaches work better for political science research problems? - these sort of stuff is not discussed in the paper. It read more like a overview of different DL models, but not anything specific to PolSci. I was also not sure who the intended audience are, because I found this too technical for a typical pol.sci. student or researcher, and too basic for a DL/ML/NLP person interested in doing PolSci work.