Learning Language Using a Pattern Recognition Approach

A pattern recognition algorithm is described that learns a transition net grammar from positive examples. Two sets of examples-one in English and one in Chinese-are presented. It is hoped that language learning will reduce the knowledge acquisition effort for expert systems and make the natural language interface to database systems more transportable. The algorithm presented makes a step in that direction by providing a robust parser and reducing special interaction for introduction of new words and terms. We are developing a natural language interface to an expert system for message processing. Both the expert system and its natural language component take a knowledgebased approach. A learning mechanism has been implemented in order to facilitate knowledge acquisition. We believe that learning will extenuate the bottleneck associated with natural language processing. The basic method for acquiring knowledge is through learning by positive example. Up to this point we have successfully developed an algorithm that learns a simple transition net grammar. It also categorizes words by parts of speech. The categories produced are similar to the ones in a standard dictionary. The syntax that is learned recognizes word phrases but makes no attempt to discover dependencies between phrases. This algorithm is a robust parser. It does not need to prompt the user for information about new words such as conjugation rules or part of speech. The fact that the parser can deal with patterns it has never seen before makes it useful in applications that mix languages, such as Thti work waa done while at Planning Research Corporation, McLean, Vqania. natural language used with tables, shorthand, acronyms, or another natural language. There are four basic principles to the algorithm. 0 The part of speech of a word can be determined by the parts of speech of the preceding and following words: one word before and one word after. 0 A transition net is created by connecting the categories of individual words as they appear in input sentences. 0 Only one instance of each category is allowed in a transition net. 0 A word can be put into more than one category. The initial experiment used two passes on the sample sentences. The first pass created two lists for each word. One list called PREVIOUS contained all the words that preceded this word in the examples by one word. The second list was called NEXT and contained all the words that immediately followed by one word. Another way to say this is that we looked at all possible sequential triples of words. For example, for the following sentences: The party was held yesterday. The meeting was held today. the PREVIOUS list for “was” is (party meeting). The NEXT list for “was” is (held). The syntax categories were determined by comparing the PREVIOUS and NEXT lists of every word to the PREVIOUS and NEXT lists of every other word. If the PREVIOUS lists of two words intersected, and if the NEXT lists of the two words intersected, then the two words were placed in the same category. The number of words in the intersection was controlled by a parameter. 64 THE AI MAGAZINE Spring, 1985 AI Magazine Volume 6 Number 1 (1985) (© AAAI)