Next: Summary and Future Up: A unified approach to Previous: Semantically Rich Information

Related Work

While semantic networks from AI were influential in the development of the KEYNET model, its indexing technique is fundamentally based on the vector space model for IR [Sal89]. From the point of view of IR, KEYNET can be regarded as a mechanism for unifying a number of different IR mechanisms, and the KEYNET system can be regarded as a high-performance, distributed search engine that can be applied effectively to vector-based retrieval from a corpus of annotated documents. From this point of view, keynets serve the same role as subject classification categories, keywords and properties such as author, title or date of publication.

Despite their reputation in IR circles as cumbersome, inefficient and suitable only for small databases, at least one IR researcher has used knowledge-based indices successfully [FHL+91]. Fuhr et al's AIR/X performs automatic indexing of documents using terms (descriptors) from a restricted vocabulary. Probabilistic classification determines indexing weights for each descriptor using rule-based inference. KEYNET differs from AIR/X in using a form of semantic network as part of the retrieval algorithm rather than in the extraction of suitable terms to be used for indexing. The extraction of terms (or in our case clasps) from a corpus of textual information objects is an important problem. One of the projects related to KEYNET is an effort to automate the extraction of keynets from biological research papers, in particular the Materials &Methods sections of such papers. See [BFH+93b][BFH+93a][BFFP93]. However, such extraction is independent of the architecture and algorithm employed for retrieval (and isn't even possible for non-textual information objects).

After building a system similar to AIR/X, Jacobs [Jac93] determined that ``the combination of statistical analysis and natural language based categorization is considerably better than either alone.'' His paper describes an automated set of statistical methods for pattern acquisition that operate inside a knowledge-based approach for news categorization (an area closely related to document classification and other IR tasks). Like AIR/X, Jacobs' system does not employ semantic networks in the IR engine. Another difference between KEYNET and Jacobs' system is that KEYNET is designed for a corpus of documents in a single subject area, where it is feasible to develop a subject-specific ontology. Developing an ontology for heterogeneous textual documents is a formidable task, many orders of magnitude larger than is feasible with current technology.

The EDS TemplateFiller system [SMHC93] applies Message Understanding (MUC) text-filtering techniques to the generation of knowledge frames for one or a few specific subject areas from entire texts (computer product announcements). TemplateFiller fills in slots for frames that exist in a predefined schema of templates, ignoring subjects that are not in the schema.

There are many other MUC-style systems; in fact, there is an annual competition among them. Such a system can automate the construction of the content labels for a collection of specialized textual documents. In a project related to the KEYNET project, we are building a MUC-style system for biological research papers [BFH+93a].

A structural model of IR was developed by a project at the University of Western Ontario[Lu90]. This model uses case relations as the structure. Case relations are a major component of case grammars which are a tool proposed by linguistic theorists and developed by computational linguists for natural language processing. The term ``case'' here is a refinement and generalization of well-known grammatical relationships such as ``subject,'' ``object'' and ``indirect object.'' Similarity of a query to a document is measured using a form of structural similarity. One conclusion of the study was that the proposed IR mechanism does not improve retrieval effectiveness. Although this system has some superficial similarity to KEYNET it differs in a number of important respects. The most important difference is that the Western Ontario system uses ``surface'' syntactic structures while KEYNET uses conceptual structures. Another important difference is that the Western Ontario system combines the two tasks of knowledge extraction from text with IR of the resulting knowledge structures. The first task is known to be very difficult, with the best such systems (the MUC-style systems discussed above) achieving only about 50%accuracy, and even this requires that the documents be restricted to a specialized topic area. Accuracy is much lower when general documents are being analyzed.

Several families of databases for semantic networks have been developed. Such databases are often called knowledge-base systems. Some of the best known of these are: Conceptual Dependency, ECO, KL-ONE, NETL, Preference Semantics, PSN and SNePs (see [Leh92]). All of these support link types, frame systems and so on, but few if any explicitly concern themselves with performance measures familiar to work in IR. Hence it is not surprising that these techniques have acquired a reputation for being cumbersome, inefficient and suitable only for small databases. The KEYNET system shows that it is possible to use a limited form of semantic network model in a high-performance IR system.

Some other examples of knowledge-based query modification systems include systems primarily for information retrieval such as those in [QF93][Har92][CD90][GS93] as well as systems designed for database systems such as the KNOWIT system of Sølvberg, Nordbø and Aamodt [SNA92] and the cooperative query answering system in [CC92]. All of these are front-end query modification systems added to an IR or database system. Such query modification techniques can also be used with a KEYNET system; in fact, a keynet ontology is very well suited to the support of such techniques; and the high-performance of a KEYNET search engine is useful for supporting the much larger queries generated by query modification techniques. However, such techniques are fundamentally a front-end for the actual search engine.



Next: Summary and Future Up: A unified approach to Previous: Semantically Rich Information

kenb@ccs.neu.edu
Fri Jan 20 21:43:28 EST 1995