Expanding Small Corpora to Aid People with Communication Impairment

Abstract

Difficulties in the communication of people with various movement and cognitive disorders may be alleviated by means of pictorial symbols. Automatic transformation of symbol sequences to natural language is of high importance. Performing this task by defining all valid sentences manually would require a large amount of work. We show that a small initial seed corpus is sufficient, which can be expanded automatically by generating candidate sentences and filtering them using N-gram statistics from a much larger corpus. The method is evaluated on a seed corpus containing dialogues, collected from an English language learning website. The ratio of useful sentences in the expanded corpus is 3-4 times bigger than in the set of unfiltered candidate sentences. We also use a manually constructed corpus for further evaluation. To demonstrate the practical applicability of the method, we have implemented a sentence production prototype that performs the transcription of symbol sequences to natural language. The system produces new and meaningful sentences and thus it can considerably decrease the size of the corpus needed, while it can increase the variability of sentences.