🧑‍🔬 Tomáš Kočiský

Tomáš Kočiský works on machine reading comprehension and question answering. Their papers here ask how a system can be trained and tested on whether it understands a text rather than guessing from surface cues. The method is to build large collections of documents paired with questions and answers. These are drawn from news articles in one case and from books and film scripts in the other.

The most cited of these papers is Teaching Machines to Read and Comprehend (2015). It addressed the shortage of large supervised reading datasets by producing one automatically from CNN and Daily Mail stories. Each article's highlight bullets were turned into fill-in-the-blank queries by removing one entity at a time. Entity mentions were grouped with a coreference system so that answers had to come from the document itself. The second, The NarrativeQA Reading Comprehension Challenge (2018), collected 1,567 long stories, half from Gutenberg books and half from movie scripts. Annotators who had read only a Wikipedia plot summary wrote the questions and answers in their own words, which keeps the wording of the questions apart from the wording of the full text.

Between 2015 and 2018 the work moved from short news articles with automatically generated queries to whole narratives with human-written questions. The 2018 paper asks for understanding across an entire book or script. The bylines of these papers give the University of Oxford (2013–2018), Google DeepMind in London (2015–2019) and later Google in Mountain View (2019–2021). Co-authors on the papers here include Karl Moritz Hermann, Edward Grefenstette, Phil Blunsom and Chris Dyer.

Tomáš Kočiský has 2 papers in our library, in 2015–2018, most of them on Question answering; often with Edward Grefenstette, Karl Moritz Hermann, Phil Blunsom.