Michael Sullivan
About me
I am a fifth-year PhD student in linguistics at UB and a post-doc in the Computational Linguistics Group at UdS. My research interests currently lie in logical reasoning and tool use with LLMs. I have conducted research on shallow heuristics that NLI models leverage (instead of using actual reasoning ability), in particular with respect to negation. I am currently developing a language model that uses logical-form representations to generate sentence embeddings for inferencing tasks, in order to leverage those representations' sensitivity to negation (and other logical operators) and invariance with respect to syntactic paraphrase constructions (passivization etc.). Additionally, I am working on the automatic generation of tool use environments for training LLM agents with reinforcement learning.
My Erdős number is four.
Education
BA in Linguistics
With Research Distinction
Minors: Spanish, German
The Ohio State University (2016-2019)
MS in Computer Science and Engineering
Research/Honors Track
MS Project: Probing NLI Models with External Negation
University at Buffalo (2023-2024)
PhD in Linguistics (expected May 2025)
Semantics/Pragmatics Track
Qualifying Paper: Towards a Formal-Logical Distributional Semantics
University at Buffalo (2020-present)
Publications
Sullivan, Michael (2024). It is not True that Transformers are Inductive Learners: Probing NLI Models with External Negation. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 1924 - 1945.
Sullivan, Michael, Madani, Navid, Saha, Sougata, and Srihari, Rohini (2023). Positional Transformers for Claim Span Identification. In Forum for Information Retrieval Evaluation (Working Notes).
Saha, Sougata, Sullivan, Michael, and Srihari, Rohini (2023). Hate Speech Detection in Low Resource Indo-Aryan Languages. In Forum for Information Retrieval Evaluation (Working Notes).
Sullivan, Michael, Yasin, Mohammed N., and Jacobs, Cassandra L. (2023). University at Buffalo at SemEval-2023 Task 11: MASDA–Modelling Annotator Sensibilities through DisAggregation. In Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023).
Nominated for Best System Paper Award at SemEval 2023