Lightweight Document Matching for Help-Desk Applications

AN EARLY GOAL FOR AI SYSTEMS was to provide a response when a user typed a problem description into a computer. With help desks, this classic problem emerges in modern form. One example is when a user enters a complete problem description and expects a program to provide help, often in the form of relevant documents from a database. The information-retrieval community has extensively addressed such problems. Most versions of solutions involve text-searchbased systems that accept as input a query or limited-length textual phrase, and produce as output a list of potentially relevant documents.1 Unfortunately, many such systems require substantial storage and computing resources. Also, the document representations and document-matching algorithms they employ can be complicated. Search engines are one embodiment of these systems. Typically, they have an index of most of their stored documents’ words, to which they match words from a query. A search algorithm attempting to identically match many input words will not likely find any documents for an exact match. In contrast, a document matcher accepts an entire, new document as input, so the query can have hundreds of words. In this article, we describe a completely automated Java-based document matcher that accepts an unlimited-length textural structure as input and employs a fast matching algorithm to produce, like a search engine, a ranked list of relevant documents. Our approach requires minimal processing and storage, and is therefore suitable for installation in restricted environments, such as Javacompatible mobile or small desktop computers (the document matcher can run on a large server, as well, but the approach we take here is effective even when resources are relatively scarce). Empirical results show that despite its lightweight algorithms, the method effectively fulfills its predictive-performance goals.