Reed Coke

I'm currently a second year EECS Ph.D. student and graduate student instructor at the University of Michigan. I'm very interested in artificial intelligence and, in particular, natural language processing. I'm working in the CLAIR group with Dr. Dragomir Radev to build a system that classifies a set typological rules in over 500 languages. The initial data are very sparse, so we are exploring the use of linguistic features to boost classification accuracy. The rules and languages are a subset of the World Atlas of Language Structures.

I received my Bachelors of Arts in Computer Science and Linguistics from Swarthmore College in 2013. While at the college I did research in multi-document summarization with Dr. Byron Gao at Texas State University and Dr. Sukanya Manna, who was then at Texas State University as well.

My research interests include multi-lingual language processing, information retrieval, graph-based natural language processing, multi-document summarization, applications of NLP to education and other social sciences, and extracting information from large, decentralized collections of text. I'm also quite interested in NLP methods that make use of linguistics to bolster statistical models.

Publications

Rahul Jha, Reed Coke, and Dragomir Radev. Surveyor: A system for generating coherent survey articles for scientific topics. In Proceedings of Twenty-Ninth AAAI Conference on Artificial Intelligence. 2015.

Manna, Sukanya, Byron J. Gao, and Reed Coke. A Subjective Logic Framework for Multi-Document Summarization. COLING (Posters). 2012.

Recent Projects

WALS - As described above, I have been working to classify some of the rules (listed as "features" on the website) for a variety of languages. As the current database is quite sparse, we are looking to use linguistic features derived from English-word-aligned Bibles in each of the target foreign languages to improve classification.

ISBN Recommender - I recently finished working on an interdisciplinary project focused on assisting students in comparative literature create network structures representing complex dimensions of translation theory. For my part, I created an easy-to-use recommender system that would suggest additional books for a student to add to their collections of works. As different collections might have overlapping works being analyzed along a different theme (say, representations of gender rather than foreignizing vs. domesticating translations), the recommendations need some amount of flexibility. To give this flexibility, my system allows the user to adjust the weighting of the top-N weighted search terms. This way, the user can inject some amount of targeted nuance into their queries.

Coherent Summarization - One of the common problems with multi-document summarization is that of coherence. While there are well-studied methods for selecting informative sentences, combining those sentences in a single, coherent document is a non-trivial task that has received far less attention. This project focused on creating summaries that were coherent at both a global (whole summary) and local (neighboring sentences) level, as well as informative.

Contact

reedcoke at umich dot edu
Office: 3961 BBB
2260 Hayward ST
48109, Ann Arbor, MI
Linked In