Using Web Corpora for the Automatic Acquisition of Lexical-Semantic Knowledge

This article presents two case studies to explore whether and how web corpora can be used to automatically acquire lexical-semantic knowledge from distributional information. For this purpose, we compare three German web corpora and a traditional newspaper corpus on modelling two types of semantic relatedness: (1) Assuming that free word associations are semantically related to their stimuli, we explore to which extent stimulus– associate pairs from various associations norms are available in the corpus data. (2) Assuming that the distributional similarity between a noun–noun compound and its nominal constituents corresponds to the compound’s degree of compositionality, we rely on simple corpus co-occurrence features to predict compositionality. The case studies demonstrate that the corpora can indeed be used to model semantic relatedness, (1) covering up to 73/77% of verb/noun–association types within a 5-word window of the corpora, and (2) predicting compositionality with a correlation of ρ = 0.65 against human ratings. Furthermore, our studies illustrate that the corpus parameters domain, size and cleanness all have an effect on the semantic tasks.