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

Lance A. MillerAAAI 1980

Abstract

In the late 1970s, researchers at IBM sought to address the growing burden of business correspondence through Project EPISTLE (Evaluation, Preparation, and Interpretation System for Text and Language Entities). This early natural language processing (NLP) system was designed to provide automated summaries, grammatical critiques, and stylistic assessments of business letters. By combining a semantic-free syntactic parser with a case-relation mapping strategy, EPISTLE represented a pioneering attempt to bring "intelligence" to the office environment.

Introduction

Long before the advent of modern large language models, the vision of an "intelligent" office assistant was already taking shape. In the late 1970s, Lance A. Miller and his team at IBM Research introduced Project EPISTLE. The project aimed to automate the analysis of business correspondence, a task that was becoming increasingly burdensome as the volume of corporate communication grew. The goal was not merely to store text, but to enable computers to understand, summarize, and critique the nuances of professional writing.

The Problem: Information Overload and Quality Control

The researchers identified two primary bottlenecks in the business workflow of the era:

  1. Information Overload: Executives required a method to quickly digest incoming mail through automated synopses or abstractions.
  2. Quality Control: Outgoing communication needed to maintain high standards of professionalism and tone, yet manual proofreading was inefficient.

The challenge lay in applying general NLP techniques to business prose, which the authors discovered was far more complex and irregular than previously assumed. The system needed to handle the specific linguistic quirks of the corporate world while remaining adaptable to different domains.

System Architecture

The EPISTLE system was built upon three foundational pillars:

1. The NLP Parser

The system utilized a bottom-up parser based on an "augmented phrase structure grammar." A key design choice was keeping the initial parse relatively "semantic-free." By minimizing semantic constraints during the first stage, the researchers aimed to create a more adaptable parser that could function across various business domains, deferring the resolution of ambiguities to later stages of processing.

2. Meaning-Assignment via Case-Relations

To derive meaning, the system mapped sentences to "action-concepts" using a "case-relation" framework. This involved identifying the actor, the action, and the object within a sentence. EPISTLE employed a weighted preference strategy to select the most likely interpretation when a sentence only partially fit the expected case-relation structures.

3. Self-Adaptation

EPISTLE featured an early form of system learning. The system could dynamically adapt its word-sense selection by assigning weights to lexical items and concepts based on their frequency and context. This allowed the system to tune itself to specific industries or even the unique writing styles of individual authors.

Empirical Analysis of Business Prose

The development of EPISTLE was informed by a custom database of 400 business letters used to identify linguistic regularities. This corpus revealed that business writing is not "simple" or "standardized." Instead, it features high frequencies of subject-verb inversions and "grammatical but unsystematic structure deletions."

Key findings from the data included:

  • Structural Regularity: The average business letter consisted of 8 sentences, with an average of 18 words per sentence.
  • Cohesion Drivers: Textual cohesion was largely driven by four categories: lexical repetitions (29%), pronouns (28%), lexical collocations (16%), and nominal substitutions (9%).
  • Regularity of Purpose: While the syntax was complex, the underlying intent of the letters (e.g., requests, notifications) was highly predictable, facilitating automatic theme recognition.

Quantifying Style and Tone

One of the most ambitious aspects of EPISTLE was the attempt to quantify subjective qualities. The researchers defined Style through dimensions such as sentence precision and readability. Tone was assessed through metrics of "personal-ness," "informal-ness," and "strength." This represented an early effort to transform the "feel" of a document into actionable, quantitative data, providing executives with a high-level characterization of an author's intent.

Limitations and Legacy

Despite its innovations, Project EPISTLE faced significant hurdles. The authors noted that their indices for style and tone were "crude indicators" at best. Furthermore, mapping words to a "concept space" did not significantly reduce complexity, as business nuances required an exceptionally rich set of features to distinguish between similar concepts. The system’s generative capabilities were also limited to short, specific formats like information requests.

Project EPISTLE stands as a landmark in the history of NLP, highlighting the enduring challenges of teaching machines to navigate the complexities of human professional communication. It laid the groundwork for future systems that would eventually transform the "paperless office" from a dream into a reality.

Summary updated

BibTeX
@inproceedings{miller1980project,
  author = {Lance A. Miller},
  title = {{Project EPISTLE: A System for the Automatic Analysis of Business Correspondence}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure