Lexical and Semantic Features for Cross-lingual Text Reuse Classification: an Experiment in English and Latin Paraphrases

Lexical and Semantic Features for Cross-lingual Text Reuse Classification: an Experiment in English and Latin Paraphrases Maria Moritz1, David Steding2 1Institute of Computer Science, University of G¨ottingen, Germany 2Computer Science Faculty, University of Bremen, Germany maria.moritz@stud.uni-goettingen.de, dsteding@uni-bremen.de Abstract Analyzing historical languages, such as Ancient Greek and Latin, is challenging. Such languages are often under-resourced and lack primary material for certain time periods. This prevents applying advanced natural-language processing (NLP) techniques and requires resorting to basic NLP not relying on machine learning. An important analysis is the discovery and classification of paraphrastic text reuse in historical languages. This reuse is often paraphrastic and challenges basic NLP techniques. Our goal is to improve the applicability of advanced NLP techniques on historical text reuse. We present an experiment of cross-applying classifiers—that we trained for paraphrase recognition on modern English text corpora—on historical texts. We analyze the impact of four different lexical and semantic features, on the resulting reuse-detection accuracy. We find out that—against initial conjecture—word embeddings can help to drastically improve accuracy if lexical features (such as the overlap of similar words) fail. Keywords: cross-lingual classification, collocations, semantic features, word vectors 1. Introduction Linguistic analyses, such as paraphrastic text-reuse discov- ery and classification, typically require advanced natural- language processing (NLP) techniques relying on machine learning. Yet, under-resourced historical languages, such as Ancient Greek and Latin, often lack enough primary ma- terial for certain time periods to properly train machine- learning models (a.k.a., classifiers). Consequently, only basic NLP techniques (e.g., checking similarity thresholds over string- and n-gram-shingles), which are independent from an advanced global knowledge or training experience, are applicable for analyzing historical corpora (B¨uchler et al., 2010). To improve the applicability of advanced NLP techniques, we need to experiment and systematically study the performance of such techniques on historical texts. Specifically, we must understand if—and in what way— ancient languages behave differently than contemporary languages when they are transferred and reused, especially when reused paraphrastically (where the reuse is not a lit- eral copy of the source). In this work, we focus on an important linguistic analysis— the detection and classification of paraphrastic text reuse in historical texts. We study a range of different machine- learning classifiers trained on modern and applied to an- cient text-reuse excerpts. We analyze how the trained clas- sifier models behave different or similar to each other for detecting paraphrastic text reuse. We use two modern English text corpora—one containing about 2700 original-and-reuse pairs, the other containing 2600 compressed news articles (the so called “banner”) to- gether with their headlines—and a data set of Latin Bible reuse with around 1100 pairs of original and reused Bible verse. We test whether three classifiers (K-Nearest Neighbor, De- cision Tree, Support Vector Machine) relying on a handful of lexical features (e.g., no. of repeating words) and seman- tic features (e.g., word vector-based features) can correctly classify the Latin text excerpts as reuse when trained on the modern English data. By identifying features that posi- tively affect the classification, we show their …

Content truncated to 500 words