Títol: | Automated Fact Checking |
Importa'l al teu calendari: |
---|---|---|
Tipus: | Xarrada d'investigació | |
Per: | James Thorne. University of Cambridge researching Natural Language Processing | |
Lloc: | Sala Claude Shannon (Escola Politècnica Superior IV) Soterrani | |
Dia/hora: | 12.00 08/04/2019 | |
Més informació: | https://jamesthorne.co.uk/ | |
Persona de contacte: | Saquete Boró, Estela (steladlsi.ua.es) | |
Resum: | With billions of individual pages on the web providing information on almost every conceivable topic, we should have the ability to collect facts that answer almost every conceivable question. However, only a small fraction of this information is contained in structured sources (Wikidata, Freebase, etc.) – we are therefore limited by our ability to transform free-form text to structured knowledge. There is, however, another problem that has become the focus of a lot of recent research and media coverage: false information coming from unreliable sources. In this presentation, I’m going to be discussing some of the unique challenges that are involved in Fact Checking: the task of predicting whether written information is true or false given evidence. I will first compare the task to fake news detection and highlight the need to have a rational explanation for the predictions made by the model and the trade-off between the types of evidence that can be used. I will use this as motivation for the Fact Extraction and VERification (FEVER) dataset and challenge which we released in 2018. The dataset comprises more than 185,000 human-generated claims extracted from Wikipedia pages. False claims were generated by mutating true claims in a variety of ways, some of which were meaning-altering. During the verification step, annotators were required to label a claim for its validity and also supply full-sentence textual evidence from (potentially multiple) Wikipedia articles for the label. I will explore some of the challenges for performing multi-sentence natural language inference on these these diverse texts for the FEVER challenge and explore two ways in which the models can be diagnosed: the first is through generating token-level explanations from the model without explicitly labelled training data and the second is with adversarial evaluation. Finally, I will conclude by discussing the future challenges in the task and the next iteration of the FEVER shared task. With FEVER , we aim to help create a new generation of transparent and interpretable knowledge extraction systems. |
[ Tancar ]