Minimalist Entity Disambiguation for Mid-Resource Languages
Benno Kruit · 2023
For text clustering, there is often a dilemma: one can either first embed each examples independently and then compute pair-wise similarities based on the embeddings, or use a crossattention model that takes a pair of examples as input and produces a similarity.The former is more scalable but the similarities often have lower quality, whereas the latter does not scale well but produces higher quality similarities.We address this dilemma by developing a clustering algorithm that leverages the best of both worlds: the scalability of former and the quality of the latter.We formulate the problem of text clustering with embeddingbased and cross-attention models as a novel version of the Budgeted Correlation Clustering problem (BCC) where along with a limited number of queries to an expensive oracle (a cross-attention model in our case), we have unlimited access to a cheaper but less accurate second oracle (embedding similarities in our case).We develop a theoretically motivated algorithm that leverages the cheap oracle to judiciously query the strong oracle while maintaining high clustering quality.We empirically demonstrate gains in query minimization and clustering metrics on a variety of datasets with diverse strong and cheap oracles.