Multi-word units (and tokenization more generally): a multi-dimensional and largely information-theoretic approach
Abstrakt
It has been argued that most of corpus linguistics involves one of four fundamental methods: frequency lists, dispersion, collocation, and concordancing. All these presuppose (if only implicitly) the definition of a unit: the element whose frequency in a corpus, in corpus parts, or around a search word are counted (or quantified in other ways).
Usually and with most corpus-processing tools, a unit is an orthographic word. However, it is obvious that this is a simplifying assumption borne out of convenience: clearly, it seems more intuitive to consider because of or in spite of as one unit each rather than two or three.
Some work in computational linguistics has developed multi-word unit (MWU) identification algorithms, which typically involve co-occurrence token frequencies and association measures (AMs), but these have not become widespread in corpus-linguistic practice despite the fact that recognizing MWUs like the above will have a profound impact on just about all corpus statistics that involve (simplistic notions of) words/units. In this programmatic proof-of-concept paper, I introduce and exemplify an algorithm to identify MWUs that goes beyond frequency and bidirectional association by also involving several well-known but underutilized dimensions of corpus-linguistic information: frequency: how often does a potential unit (like in_spite_of) occur?; dispersion: how widespread is the use of a potential unit?; association: how strongly attracted are the parts of a potential unit?; entropy: how variable is each slot in a potential unit? The proposed algorithm can use all these dimensions and weight them differently.
I will (i) present the algorithm in detail, (ii) exemplify its application to the Brown corpus, (iii) discuss its results on the basis of several kinds of MWUs it returns, and (iv) discuss next analytical steps.