A Parallel Algorithm for Statistical Multiword Term Extraction from Very Large Corpora

Multi-word Relevant Expressions (REs) can be defined as sequences of words (n grams) with strong semantic meaning, such as "ice melting" and "Ministère des Affaires Étrangères", useful in Information Retrieval, Document Clustering or Classification and Indexing of Documents. The need of extracting REs in several languages led research on statistical approaches rather than symbolic methods, since the former allow language-independence. Based on the assumption that REs have strong cohesion between their consecutive n-grams, the LocalMaxs algorithm is a language independent approach that extracts REs. Apart from its good precision, this extractor is time-consuming, being inoperable for Big Data if implemented in a sequential manner. This paper presents the first parallel and distributed version of this algorithm, achieving almost linear speedup and sizeup when processing corpora up to 1 billion words, using up to 54 virtual machines in a public cloud. This parallel version of the algorithm explores the statistical knowledge of the n grams in the corpus, to promote the locality of the references.