Computational extraction of formulaic sequences from corpora

Abstract

We describe a new algorithm for the extraction of formulaic language from corpora. Entitled MERGE (Multi-word Expressions from the Recursive Grouping of Elements), it iteratively combines adjacent bigrams into progressively longer sequences based on lexical association strengths. We then provide empirical evidence for this approach via two case studies. First, we compare the performance of MERGE to that of another algorithm by examining the outputs of the approaches compared with manually annotated formulaic sequences from the spoken component of the British National Corpus. Second, we employ two child language corpora to examine whether MERGE can predict the formulas that the children learn based on caregiver input. Ultimately, we show that MERGE indeed performs well, offering a powerful approach for the extraction of formulas.