Sentence Vectors Debate
Matt expresses skepticism about the utility of pre-trained sentence vectors, while Sam shares his shift in perspective on their effectiveness for high-profile language understanding tasks. Although he acknowledges their analytical interest and potential for retrieval tasks, he argues that reducing sentence meaning to a single vector often falls short in applications like question answering and translation. The discussion highlights the advantages of models that provide a vector per token, suggesting a more scalable approach to capturing information content.In this clip
From this podcast
NLP Highlights
67 - GLUE: A Multi-Task Benchmark and Analysis Platform, with Sam Bowman
Related Questions