Learning Word Association Norms Using Tree Cut Pair Models

We consider the problem of learning co-occurrence information between two word categories, or more in general between two discrete random variables taking values in a hierarchically classified domain. In particular, we consider the problem of learning the `association norm' defined by A(x,y)=p(x, y)/(p(x)*p(y)), where p(x, y) is the joint distribution for x and y and p(x) and p(y) are marginal distributions induced by p(x, y). We formulate this problem as a sub-task of learning the conditional distribution p(x|y), by exploiting the identity p(x|y) = A(x,y)*p(x). We propose a two-step estimation method based on the MDL principle, which works as follows: It first estimates p(x) as p1 using MDL, and then estimates p(x|y) for a fixed y by applying MDL on the hypothesis class of {A * p1 | A \in B} for some given class B of representations for association norm. The estimation of A is therefore obtained as a side-effect of a near optimal estimation of p(x|y). We then apply this general framework to the problem of acquiring case-frame patterns. We assume that both p(x) and A(x, y) for given y are representable by a model based on a classification that exists within an existing thesaurus tree as a `cut,' and hence p(x|y) is represented as the product of a pair of `tree cut models.' We then devise an efficient algorithm that implements our general strategy. We tested our method by using it to actually acquire case-frame patterns and conducted disambiguation experiments using the acquired knowledge. The experimental results show that our method improves upon existing methods.