Kevin Stowe

Current Research Projects

I currently work as a Research Scientist for Pindrop where I focus on several areas relating to fraud detection, enrollment/authentication, responsible AI, and language modeling.

I work on automatic detection of machine-generated content, with the goal of detecting malicious chatbots and other entities. To this end we presented a paper relating evaluation and metrics at LREC 2026, and have a paper relating to bias in machine-generated text detection at ACL 2026.
I also work on systems for automatic enrollment/authentication of users. This involves using LLMs tuning, synthetic data creation, and encoder model training to automatically verify a user is who they say they are. Much of this work is currently under patent review; some details available in my blog post here
Other projects include robotic caller detection, improving intent detection modeling, and multi-modal foundation model training.

Previous Work

Creative Language Generation

I'm fascinated by creative language generation: exploring possibilities for generating all sorts of novel language (humor, metaphor, etc) using deep learning. Are end-to-end models sufficient, and if not, how can we improve them to handle this difficult area of language?

We built a system to generate conceptual metaphors based on BART, Frame Semantics, and other tools: see our ACL paper Metaphoric Generation with Conceptual Mappings
Relatedly, we built a dataset of idiomatic and metaphoric paraphrases to explore natural language inference capabilities on figurative language:IMPLI: Investigating NLI Models' Performance on Figurative Language

Crisis Informatics

Crisis informatics: how can we provide concrete benefits and relief to those impacted by natural disasters using NLP technology?
- I collaborated on a project attempting automatically adapt to rapidly changing language during crises, presented at the Findings of EMNLP 2022
- Our group explored social media collection and analysis for crises events, supporting citizen science efforts to annotate the never-ending stream of data that can contribute to our situational awareness, including the COVID-19 crisis. Our goal is to build a system for quick and efficient data collection and annotation to improve responses to natural hazards. Our collection uses the helpful Nasty Scraper for collection as well as the powerful INCEpTION tool for annotation. Presented at EACL 2023.
- Check out my and my students' work on a social good data collection pipeline, presented at the AAAI Fall Symposium on Social Good

Educational Content

I previously worked for Educational Testing Service where I worked on using LLMs (and smaller LMs) to generate educational content, focusing on control over relevant aspects of the content as well as fairness constraints specific to testing contexts.

PhD. Projects

Doctoral dissertation on syntactically motivated approaches to metaphors, with the CLEAR group under Profs. Martha Palmer and Jim Martin. I studied how syntax can influence metaphoric properties, and this can be leveraged to improve deep-learning metaphor detection.
Supported lexical resources at the University of Colorado, including VerbNet and SemLink.
Worked for the Project Epic lab with Profs. Leysia Palen and Kenneth M. Anderson. We collaborated with NCAR to analyze responses to hurricanes, and build improved computaitonal methods for data analysis and exploration in order to assist researchers understanding the effects of natural disasters.