MWUs

MWUs, or multiword units, are sequences of two or more words that function as a single unit within a language. They include idioms (kick the bucket), collocations (strong tea, heavy rain), verb–noun patterns (make up, take care), proper names (New York City), and fixed expressions (as a matter of fact). Some MWUs are non‑compositional, meaning their meaning cannot be deduced from the meanings of the individual words; others are conventionalized but still partly compositional.

In linguistics and natural language processing, MWUs are studied because they often carry fixed meanings, idiomatic nuances, or lexicalized patterns that affect fluency and interpretation. They vary across languages and domains, and identifying their boundaries can be challenging. Problems include determining when a sequence counts as an MWU, handling productive yet variable forms, and dealing with morphological and syntactic variation.

MWUs are extracted and analyzed using a range of methods. Statistical measures such as mutual information, t-score, and log-likelihood are used to identify word associations; syntactic patterns help reveal recurrent structures; and supervised or semi-supervised learning on annotated corpora can classify and segment MWUs. Notable resources and initiatives include the PARSEME shared task, which coordinates multilingual MWU annotation, and the Multiword Expressions Bank (MWEB), a repository of annotated MWUs for several languages.

Applications of MWU research include improving machine translation, information retrieval, lexical resources and dictionaries, parsing accuracy, and overall language modeling. Recognizing MWUs reduces ambiguity and improves naturalness in NLP systems. Cross-linguistic studies of MWUs also support multilingual lexicon development and transfer in multilingual NLP applications.