A Lebesgue Integral based Approximation for Language Modelling

点击翻译按钮获取中文摘要

Deep learning (DL) based Natural Language Processing (NLP) technologies have attracted significant interest in recent years. The current SOTA language models, a.k.a. transformer-based language models, typically assume that the representation of a given word can be captured by the interpolation of its related context in a convex hull. However, it has recently been shown that in high-dimensional spaces, the interpolation almost surely never occurs regardless of the underlying intrinsic dimension of the data manifold. The representations generated by such transformer-based language models will converge into a dense cone-like hyperspace which is often discontinuous with many nonadjacent clusters. To overcome the limitation of current methods in most DL-based NLP models, this project aims to deploy Lebesgue integral, which can be defined as an ensemble of integrals among partitions (i.e., discontinuous feature clusters), to approximate the posterior distributions of clusters given input word features in finite measurable sets by automatically identifying the boundary of such discontinuous set, which in turn could help to generate better interpretations and quantify the uncertainty. By our proposed Lebesgue integral based approximation, the input text will be characterised by two properties: an indicator vector encoding its membership in clusters (i.e., measurable sets), and another continuous feature representation for better capturing its semantic meaning for downstream tasks. This not only allows for a more faithful approximation of commonly observed countably discontinuities in distributions of input text in NLP, but also enables learning text representations that are better understood by humans.