🧑‍🔬 Julian Eisenschlos

Julian Eisenschlos works in natural language processing, studying how pretrained language models can reason over structured inputs and stored facts. Their papers here belong to Language models & pretraining and Semantic parsing & meaning representation, with a link to knowledge graphs. A recurring question is what a model can draw from material other than plain running text, such as tables or facts that hold only during certain years. The usual approach is to adapt transformer encoders and their pretraining, then test them on benchmarks built for the purpose.

The most cited of these papers is TaPas: Weakly Supervised Table Parsing via Pre-training (2020). It answers questions about tables without first translating them into executable logical forms. Instead, it extends BERT to read a question together with a flattened table, adding embeddings that mark each token's row, column and numeric rank. The model then picks cells and an aggregation operator, and it learns from answers alone. The other paper, Time-Aware Language Models as Temporal Knowledge Bases (2022), examines what happens when a model trained on one snapshot of text meets facts that change over time. It introduces TempLAMA, a collection of fill-in-the-blank queries drawn from Wikidata, where the right answer varies with the year. The study finds that the model blends statements that were true in different years, loses track of years with little coverage, and makes poorly calibrated guesses about what comes later.

The 2022 paper marks a shift from question answering over tables toward the temporal side of factual knowledge in language models. Co-authors on these papers include Jonathan Herzig, Bhuwan Dhingra and William Cohen, along with Jacob Eisenstein and Daniel Gillick. Eisenschlos's bylines run from 2019 to 2025. The most recent ones name Google DeepMind in London, and earlier ones name Google in Mountain View and in Zurich. Some bylines also list the University of Cambridge and the Universidad Nacional de Córdoba.

Julian Eisenschlos has 2 papers in our library, in 2020–2022, most of them on Language models & pretraining, Knowledge graphs & ontologies, Semantic parsing & meaning representation; often with Bhuwan Dhingra, Daniel Gillick, Jacob Eisenstein.