自然语言处理算法的范式转换:统计真理与人类偏好

Statistical Truth and Human Preferences:On the Paradigm-shift in NLP Algorithms

基于数据的深度学习方法在自然语言处理领域方兴未艾,但种种问题也开始显现。语言的经典的组合原则被打破,让机器“理解”语言变得不再重要,符号之间的黏性完全由样本统计决定,违背了人类语言理解和习得的主观经验。本文针对这一系列问题,详尽阐述了经典数字计算机的运算和决策过程,并对其处理部分模糊性问题时采取的策略进行简要介绍,指出即使是在统计学习过程中,人类偏好也无时无刻不在扮演着关键作用,规律性不可能无条件地从海量数据中自动浮现。逻辑学家和数据科学家有责任揭示这些主观偏好,这不仅是算法可解释性的外在需要,也是提高运算效率的正途。

Data based statistical learning as a method keep on gaining its momentum in NLP related fields.Yet by posing against our conventional understanding of language,the practice has greatly weakened the consistency of linguistics by disregarding semantics and the compositional nature of language.With a generalized description of statistical learning this paper tries to exhibit the disadvantages of replacing even pre-existing causation with co-relation as the final judgment in human reasoning.Since any statistical result always yields certain scale and dimension,which are purely preferential,we urge for a joint effort from logicians and data scientist to uncover the human input previously unrecognized in the learning process in language processing and alike.