A Technique for Managing the Lexicon in a Natural Language Interface to a Changing Data Base

1. Treating unknown terms a& semantically ambiguous

In co-oP, selecting an appropriate DB image for unknown terms is reduced to the problem of selecting a sense (meaning) for each word in the input query. The various . senses that a word can take on may be regarded as· semantic ambiguities. Associated with each entry in the lexicon for terms that make reference is a set of definitions of the possible senses that term can take on in the domain of the DB. For example, a term like "orders" might have a list of senses designating room service orders, purchase orders, reception desk assignments, etc. Temporary lexical entries are created for unknown terms that designate as their potential senses . each field in the DB that could reasonably occur as a DB image in a query. Thus, an unknown term like "William Smith" might be treated as a lexical item that could have the DB image of either an employee name, a supplier name, a guest name, etc. The word sense disam­ biguation routines then proceed to select a sense for . unknown terms as they. do for other terms, as detailed below.

2. Predictive value of selectional restrictions

~ particular word sense is selected using two heur'istics. The first exploits the predictive value .of the words in the immediate syntactic context .of the ambiguous term to constrain the set of potential senses. Specifically, verbs and prepositions are .marked in the lexicon (when appropriate) to indicate the more likely subjects and objects they can take. If a subject or object has one or more senses that are consistent with these predict.ions, the alternative senses are eliminated from further consideration. In effect, simple selectional restrictions are used to reduce the semantic ambiguities.

For example, a word like "who" may refer to employees, suppliers, or guests; in the context of the query "Who works in the maintenance department?", the verb "work" may predict em­ ployees as a subject, and the possibility of "who" referring to suppliers or guests can be eliminated. Similarly, "William Smith" can be resolved to be a hotel guest, by the predictive value of the verb "checked out". Another potentially useful technique (not implemented) would be to take advantage of verb case roles, or semantic memory structures to help predict categories for the unknown term. As none of these richer knowledge structures were present in CO-OP, no such approach was adopted.

3. Schema distance and semantic relatedness

The above heuristic will not always result in an .unambiguous selection. Indeed, a verb like "is" may have no predictive value at all. Remaining ambiguities are resolved by choosing the sense whose DB image results in the shortest distance through the DB schema to the DB image of the surrounding referring terms in the query*. This distance metric is a rough measure of semantic relatedness - semantically related terms tend to have DB images that appear close to each other in the DB schema. Thus "Who is the maintenance department?" may still select employees as the sense of 'who' because the DB image of employees may be closer in the schema to departments than suppliers (perhaps linked through inventory) or guests (perhaps linked through orders or com­ plaints) . Similarly, the unknown term "mainten­ ance" (assuming it appears solely as a DB value, if at all) can be correctly identified as a --------------------

* A network like structure (such as CODASYL ) is required for the operation of this heuristic.

department name, through the close association in the schema of the "DEPT-NAMEt' field with the

DB image of thli! more speaific .. term. "department".

These heuristics, while admittedly rather crude, tend to operate acceptably when the DB schema accurately models the structure of the real world. When CO-OP makes an incorrect interpre­ tation, its mistake is made clear to the user via a paraphrase of the query (3), presented tc the user for approval. _ Experience .indicates that it is easy to correct these misinterpreta­ tions by rephrasing the query in more specific terms, for example, "Which employees are in the maintenance department?"

D. Examples

Some examples of interpretations of unknown terms actually produced by CO-OP are reprodeced below "Q: " (below stands for query, "P:" for paraphrase (responses are deleted in the interest of brevity), commentary is in square brackets<[]> . The first examples are taken from a DB at the National Center for Atmospheric Research in Boulder, Colorado. The DB contains information on their users, divisions, projects, and compu­

ter resource utilization.

• Q: Which projects in oceanography does NASA Headquarters sponsor?

P: (I am assuming that 'NASA Headquarters' is a SPONSOR NAME.) (I am assuming that 'oceanogra­ phy' is an AREA OF INTEREST.) Which projects does NASA Headquarters sponsor? Look for pro­ jects that are in oceanography.

[This query contained the unknown terms 'oc~an­ ography' and 'NASA Headquarters'. The system has correctly identified the first because the preposition 'in' predicts a division, super­ division, or area-of-interest as an object in this domain, and area-of-interest is the nearest was identified as a SPONSOR NAM& because it is

to projects in the DB schema. 'NASA Headquarters' the subject of the verb 'sponsor' .J

• Q: Which projects are advised by Thomas Harris?
• P: (I am assuming that 'Thomas Harris' is an ADVISOR NAME.) Which projects are advised by Thomas Harris?

[co-oP is capable of recognizing that 'Thomas Harris' is the logical subject of advise, and hence is probably an advisor, although the passive form is maintained throughout the balance of the processing for other purposes4]

[The following examples are drawn from the ONRODA DB, a DB about military equipment and logistics created as part of the Operational Decision Aids Program of t -he Office of Naval Research at the University of Pensylvania*.]

• ~; What is the fire rate of the M-61?
• P: (I am assuming that 'M-61' is a GUN NAME.) Display the fire rate of the M-61.

tHere, 'M-61' has been identified as a GUN NAME because it is the nearest field in the DB schema to the 'FIRE RATE' field that could reasonably contain an unknown term. The seman­ tic relatedness (schema distance) heuristic is the only one invoked in this case~

• Q: Which F-Ss ca~ry strut curve radar?
• P: (I .am assuming that 'F-Ss' is a SHIP NAME.) (I am assuming that 'strut curve' is a RADAR NAME.) Which F-Ss carry strut curve radar?

[Because of insufficient contextual cues, CO-OP incorrectly identifies 'F-Ss' as a SHIP NAME (it is actually a type of aircraft) - radar can appear on ships, subs, or aircraft in the DB, and the selection defaults to ships. By contrast, 'strut curve' is correctly identified as a type of radar, because it is parsed as modifying 'radar'. The proper interpretation can be induced by re-phrasing the question in a more specific form, as follows.]

• Q: Which F-5 aircraft carry strut curve radar?