Sentence embeddings for automated factchecking
summary
Uses InferSent sentence embeddings trained on SNLI with GloVe inputs and a BiLSTM max‑pool encoder to spot verifiable claims in political subtitles. Annotated 25,000 sentences with five votes each via Prodigy, yielding about 1,000 positive examples and 4,500 negatives from ~12 hours of TV. Logistic regression on the embeddings outperforms naive averaged word vectors but confidence intervals stay wide because positives are few. Notes that embeddings and code are downloadable and transfer learning is the key lever.
AI-written summary · about
chapters (7)
- 0:00Full Fact’s mission
- 9:18Automating factchecking
- 16:49Defining political claims
- 23:19Infersent for sentence meaning
- 26:30Evaluating claim detection
- 30:48Volunteers and contributions
- 31:29Conclusion
related talks
01h 26mBagging to BERT- A tour of applied NLP
34mWord Embeddings for fun and profit in Gensim
31mAttacking Deep Learning-Based NLP Systems
33mBulk Labelling Techniques
45mDeep Learning for Natural Language Processing
39mSentiment Analysis, its techniques and applications
22mMaking Natural Language Processing Robust to Sociolinguistic Variation
29mHallucination by Design: How Embedding Models Misunderstand Language