MACHINE LEARNING
PROJECTS
PEOPLE
PUBLICATIONS
PATENTS
Our research spans many areas of text understanding and text generation, with a particular emphasis on factuality checking and factually guided text generation. NEC’s products help humans safely draw conclusions from large quantities of text that they don’t have the time to read. As a leader in the FEVER fact extraction and verification competitions, we have developed systems that achieved higher evidence precision and higher robustness to adversarial attack, and pioneered the ability to pursue missing evidence through multiple retrieval steps. With NEC X, we have commercialized this technology in a product that gives market researchers a deeper understanding of opinions expressed in customer reviews.
Ongoing research has made transformers easier to debug for classification tasks and to supplement with domain-specific knowledge. Current projects automatically summarize the most frequent opinions found in multiple documents, produce safer analysis and summarization of multi-modal data and test results found in clinical records, and make LLM generation less prone to hallucination. Our “Miru” tool supports decision making with visualization of the results of deep, user-guided document analysis.
Team Members: Christopher Malon, Martin Min
We study few-shot debugging of transformer based natural language understanding models, using recently popularized test suites to not just diagnose but correct a problem. Given a few debugging examples of a certain phenomenon, and a held-out test set of the same phenomenon, we aim to maximize accuracy
