Dafna Sheinwald

Dafna Sheinwald is a researcher whose work helped shape IBM Watson's question answering and later language-generation systems. Her early papers examined binary coding and compression with shared files, followed by work on fast enterprise search. She coauthored Watson's candidate-generation study and was named to the IBM team that received the 2013 AAAI Feigenbaum Prize. In 2021, she coauthored Project Debater, an autonomous system for competitive debate with people. Her later research covered active learning for language generation, neural table insights, Unitxt for generative-model data preparation, and an IBM patent granted in 2026.

Now at: IBM
Works: 39
Citations: 822
h-index: 11
Badges: Loyal

Publication numbers from OpenAlex, as of 19 October 2026.

Timeline

1992–2001

Data compression studies

In 1992, she studies prefix-free binary codes that preserve alphabet order and presents an order-preserving adaptation of Ziv-Lempel coding. In 2001, she and Michael Factor propose adding shared files to the Lempel-Ziv dictionary to improve compression.
2007

Just-in-time indexing

In 2007, she coauthors JiTI, a just-in-time indexing component that makes new content searchable as soon as it arrives, including for file-system and email search.
2012

Watson's candidate generation

In 2012, she coauthors a study of search and answer-candidate generation for IBM Watson's Jeopardy! questions. Its combined strategies place the correct answer in the candidate pool for 87.17% of questions in a blind test set.
2013

A Feigenbaum Prize

In 2013, AAAI awards the IBM Watson team the Feigenbaum Prize for real-time factual question answering; Sheinwald is one of the named team members.
2021

Project Debater

In 2021, she coauthors the description of Project Debater, an autonomous system that engages humans in competitive debate.
2023

Active learning and table insights

In 2023, she coauthors a systematic study of active learning for natural-language generation and nBIIG, a system that uses table analyses and RDF representations to generate textual insights with a neural model.
2024–2026

Unitxt and an IBM patent

In 2024, she coauthors Unitxt, a library for generative-language-model data preparation and evaluation, as an IBM-affiliated researcher. A 2025 patent application names her as an inventor on active learning for natural-language generation and is granted to IBM in March 2026.