A generalized disambiguation algorithm for weighted finite automata and its application to NLP tasks

Abstract

We present a disambiguation algorithm for weighted finite tree automata (FTA). This algorithm converts ambiguous FTA into equivalent non-ambiguous one where no two accepting paths labeled with the same tree exists. The notion of non-ambiguity is similar to that of determinism in the automata theory, but we show that disambiguation is applicable to the wider class of weighted automata than determinization. We conduct experiments on Natural Language Processing (NLP) tasks, and also show that disambiguated automata become much smaller than determinized automata in practice.