All-but-the-Top: Simple and Effective Postprocessing for Word Representations

TL;DR

This paper introduces a simple postprocessing method that improves the quality of word representations by removing the mean vector and dominant directions, leading to better performance across various NLP tasks.

Contribution

It proposes a straightforward yet effective postprocessing technique for word embeddings that enhances their utility across multiple linguistic tasks.

Findings

Processed embeddings outperform original ones on lexical tasks.

Improved results across sentence-level semantic tasks.

Method works across different languages and embedding algorithms.

Abstract

Real-valued word representations have transformed NLP applications; popular examples are word2vec and GloVe, recognized for their ability to capture linguistic regularities. In this paper, we demonstrate a {\em very simple}, and yet counter-intuitive, postprocessing technique -- eliminate the common mean vector and a few top dominating directions from the word vectors -- that renders off-the-shelf representations {\em even stronger}. The postprocessing is empirically validated on a variety of lexical-level intrinsic tasks (word similarity, concept categorization, word analogy) and sentence-level tasks (semantic textural similarity and { text classification}) on multiple datasets and with a variety of representation methods and hyperparameter choices in multiple languages; in each case, the processed representations are consistently better than the original ones.