Project EPISTLE: A System for the Automatic Analysis of Business Correspondence

The developing system described here is planned to provide the business executive with useful applications for the computer processing of correspondence in the office environment. Applications will include the synopsis and abstraction of incoming mail and a variety of critiques of newly-generated letters, all based upon the capability of understanding the natural language text at least to a level corresponding to customary business communication. Successive sections of the paper describe the Background and Prior Work, the planned System Output, and Implementation. I. BACKGROUND AND PRIOR WORK We conclude from these behavioral findings that there are indeed extensive regularities in the characteristics of business letters, determined primarily by the purpose objectives. It is these constraints that most strongly indicate to us the feasibility of developing automatic means for recognizing content-themes and purposes from the letter text (as well as the converse, generating letter text from information about purposes). Other analyses have been undertaken to estimate the linguistic complexity and regularities of the texts. The average letter appears to contain 8 sentences, with an average of 18 words each; in the 400 letter-bodies there are roughly 57900 words and 4500 unique words totaL An ongoing hand analysis of the syntactic structure of sentences in a 50-letter sample reveals a relatively high frequency of subject-verb inversions (about 1 per letter) and complex lengthy complementizers (l-4 per letter). These features, along with very frequent noun phrase and sentence coordination, accompanied by a wide variety of grammatical but unsystematic structure deletions, indicate an exceptionally high level of grammatical complexity of our texts. With respect to overall text syntax we have analyzed 10 letters for text cohesion, using a modification of Halliday and Hasan’s coding scheme E41; 82 percent of the instances of cohesion detected were accounted for by 4 categories: lexical repetitions (29%), pronouns (28%), nominal substitutions (9%, e.g., “one”, “same”), and lexical collocations (words related via their semantics, 16%). In an extension of this discourse structure analysis we are analyzing 50 letters, coding all occurrences of functional nouns in terms of (1) the grammatical case function served and (2) the cohesive relation to prior nouns. Preliminary results indicate consistent patterns of case-shift and type of cohesion as a function of the pragmatic and content themes. The results of these linguistic analyses will help determine the strategy ultimately adopted for selecting surface parses and meaning interpretations.