Ambiguity is widespread in natural language, so that a single piece of text can potentially be interpreted in different ways by different readers.
We are investigating how to model noxious ambiguity, which is about the relationship between text and a group of readers: text displays noxious ambiguity when it is interpreted differently by different readers. Ambiguity is innocuous if different readers interpret it in the same way, even if different interpretations are possible.
Our approach uses a collection of heuristics and distribuitional measures to predict which passages in text are likely to give rise to misunderstandings, because they exhibit noxious ambiguity. So far, by validating the heuristics against collections of human judgements, we have been able to show that noxious apmbiguity does occur in text (specifically software requirements).
We are currently focussing on expanding the range of linguistic phenomena which can lead to noxious ambiguity, and the heuristics needed to recognise it.