Using syntactic and semantic information in a word prediction aid

An initial attempt has been made to include syntactic and semantic information in programs for word prediction. A study has begun in which 1500 words in a Swedish lexicon have been marked with semantic categories. Graphical semantic overviews and sketches of four texts have been composed using this classification. These graphical descriptions appear to be faithful to the subject matter. Small changes in prediction priority depending on semantic and syntactic information have not as yet resulted in keystroke savings. However, these are only the frrst of a number of planned uses of this information. INTRODUCTION Two word prediction programs have been developed since 1983 as speaking and writing aids [1]. One program, which is being used by persans experiencing extreme difficulties in movement, predicts a likely candidate word from the beginning letter(s) of a word. The candidate is chosen either from a variable-size frequency ranked lexicon or from a Iist of words previously used, thus taking both frequency and recency into account. The lexicon is updated during a session, so that it becomes individualized. Output can be presented as synthesized speech, and can be saved for later use or editing. Another output file is a work log which lists all key choices and the time elapsed between keystrokes. The program has been distributed to 14 locations in one of four languages (English, Swedish, Danish, Norwegian). It has recently been rewritten as a predicting editor which will be tested during the coming year. A new version will also be written to serve as a resident program which can be used with other programs. A second program is now in trial form through consultation with · several language therapists engaged in aphasia rehabilitation. With this program, a word is accessed through optional features. The program does not presuppose the ability to spell a word, but rather uses any (non-semantic) information that the person has about a word to "access" it. The word, oralist of word predictions, can then be synthesized for the user to hear and choose among. Lexical adaptation and output files are also included, as with the first system described. SEMANTIC INFORMATION A study has begun in which 1500 words in the 10,000word Swedish lexicon have been marked with semantic categories. The words correspond to the standard international Blissymbols, and are marked with the categories assigned by the Easter Seal Communica'tions Institute [2]. Semantic information is being used to optimize ranking in the list of predicted words. The main categories into which the words are classified (called Level 1 hereafter) are: 1) The World we Live in; 2) Living Things; 3) Being Alive; 4) How we View the World; and 5) Living Together. Several of these categories are divided into two or more subcategories (Level 2). Category 3, Being Alive, is divided into a) Things we Do and b) Things we Need. Category 4, How we View the World, is divided into a) Our Sensesand b) Classification. Category 5, Living Together, is divided into four subcategories, a) Communicaiton, b) Transportation, c) Occupations and d) Recreation. A further subcategorization divides the categories (or subcategories) into between three and eight (Level 3) subcategories. An example is given by the four subcategories of Living Together: Occupations, which is divided into (1) Primary Industriesand Manufacturing, (2) Commerce, (3) Govemment and (4) Community Services. Words that belang to more than one category (or subcategory) are Iabelied as "general" words in that category (or subcategory). Using this categorization a "semantic overview" and a "semantic sketch" can been constructed. Such an overview has been computed for four short texts. Parts of these overviews will be shown below. A sketch will also be shown for one of the texts. Scmantic Content of Test. Texts Text 1 is from a Swedish story about a boy who is made small enough to ride on the back of a goose [3]. The passage taken here describes a troll he met on his joumey who loved playing in the wind, and who could control the weather. The general semantic overview for this text is shown in Figure 1. 5 10 6 13 13 5 #words