Syntactic pre-processing in single-word prediction for disabled people

Many disabled persons in the world are unable to communicate verbally with those around them. This may be because they have motor-control di culties and so cannot operate the muscles which actually produce the speech. Conversely, it may be a neurophysiological problem, either congenital or as a result of trauma, resulting in an inability of the brain to create the speech in the rst place. Either way the di culties inherent with such problems are immense, and anything which can be used to help the situation can bring welcome relief. Many people have no verbal communication skills at all and as such need some form of prosthetic aid known as an Augmentative and Alternative Communication (AAC) device. There are many forms of device from very basic pointing boards to complex computer systems and a device will be chosen to suit the needs and abilities of an individual. However, people with such di culties often have their problems compounded by not being able to use the rest of their bodies properly either. If one cannot use one's voice box because of motor-control di culties then it is very likely that one will not be able to co-ordinate limbs properly. As such, using a complex device can be very di cult, and needless to say time consuming. For instance, typing on a keyboard can become an almost impossible task. The most common method of operating an AAC device is with a single switch, selecting the letters one-by-one as the system scans through them on a screen or lightboard. This produces prohibitively slow communication rates and as a consequence, even using an ordinary AAC device, users cannot join in with a normal conversation. A method is required which will improve the rate at which users can participate in a conversation using an AAC device. Prediction systems are amongst methods used to facilitate such improvements. They attempt to predict the user's intended next word using various methods of computation. Many methodologies employ statistical modelling of the mode of usage of words within the user's language, bearing no weight of choice upon the grammatical status of the sentence. Thus many of the predictions they produce are grammatically incorrect and as such wasteful to the user. Using grammar modelling a system can rst narrow down the range of words which could follow the current one to those which are grammatically correct. When the more normal statistical methods are overlaid on this method then the hit rate (the rate at which the system predicts a word correctly) will be higher.