The Effect of Downstream Classification Tasks for Evaluating Sentence Embeddings
This work addresses a methodological issue for researchers evaluating sentence embeddings, highlighting potential biases in common evaluation practices.
The paper investigates how using downstream classification tasks to evaluate sentence embeddings essentially treats them as a low-dimensional factorization of a sentence-task label matrix, showing that sentences with more labels across tasks have higher reconstruction loss, with the effect depending on the overall label distribution.
One popular method for quantitatively evaluating the utility of sentence embeddings involves using them in downstream language processing tasks that require sentence representations as input. One simple such task is classification, where the sentence representations are used to train and test models on several classification datasets. We argue that by evaluating sentence representations in such a manner, the goal of the representations becomes learning a low-dimensional factorization of a sentence-task label matrix. We show how characteristics of this matrix can affect the ability for a low-dimensional factorization to perform as sentence representations in a suite of classification tasks. Primarily, sentences that have more labels across all possible classification tasks have a higher reconstruction loss, however the general nature of this effect is ultimately dependent on the overall distribution of labels across all possible sentences.