Humpty Dumpty - Controlling Word Meanings via Corpus Poisoning

Explore a 17-minute IEEE conference talk that delves into the manipulation of word embeddings through corpus modifications. Learn how attackers can control the semantic meaning of words by altering their positions in embedding spaces. Discover the explicit expression used as a proxy for word distances and its causal relationship with embedding distances. Examine two adversarial objectives: making a word a top-ranked neighbor of another and moving words between semantic clusters. Understand the far-reaching implications of such attacks on various downstream tasks in transfer learning scenarios. Investigate how these manipulations affect query expansion in information retrieval systems, influence named entity recognition models, and impact machine translation. Finally, gain insights into how attackers can generate linguistically plausible corpus modifications to bypass language model-based defenses.