A Method to Quantify Corpus Similarity and its Application to Quantifying the Degree of Literality in a Document

Comparing and quantifying corpora is a key issue in corpus based translation and corpus linguistics, for which there is still a notable lack of measures. This makes it difficult for a user to isolate, transpose, or extend the interesting features of a corpus to other NLP systems. In this work we address the issue of measuring similarity between corpora. We suggest a scale between two user chosen corpora on which any third given corpus can be assigned a coefficient of similarity, based on the cross-entropy of statistical N-gram character models. A possible application of this framework is to quantify similarity in terms of literality (or conversely, orality). To this end we carry out experiments on several well-known corpora in both English and Japanese language, and show that the defined similarity coefficient is robust in terms of language and model order variations. Comparing it to other existing similarity measures shows similar performance while extending widely the range of application to electronic data written in languages with no clear word segmentation. Whithin this framework we further investigate the notion of homogeneity in the case of a large multilingual resource.