SP500215 DR-LINK: A System Update for TREC-2 chapter E. Liddy S. Myaeng National Institute of Standards and Technology D. K. Harman concept-relation[OCRerr]oncept triples based on the nile. At the same time, it converts the NV into its verb form. In this way, we can allow for a match between a CG fraginents generated from a phrase coutaning verb and another fragment generated from a noun phrase containing the corresprxiding nonlimeized verb. For example, the NV Handler converts the sentence fragment the company's investigation of the incident... into [investigate] -> (AGEN[OCRerr]I) -> [company] [investigate] -> [OCRerr]ATIENl) -> [incident]. This process is much more than a sophisticated way of performing stemming in that we canonicalize concept-relationconcept triples rather thanjust concept nodes. For NV processing, 15,053 case frames have been generated for 1,593 noniinalized verbs. Most of the case frames for NVs were automatically generated from the corresponding verb case frames. This process was also facIlitated by identifying potential NVs from [DOCE. No explicit testing of the impact of NVs in information retrieval has been done yet although we have convinced ourselves with anecdotal evidence that this would improve the retrieval performance. More semantic processing of noininalized verbs in determiing the relations to the surrounding constituents is on the future research agenda. More rigorous study on the impact of NVs on information retrieval should be done, too. 2.1.3. Noun Phrase (NP[OCRerr] and Prepositional Phrase (?P) Handler The noun phrases that are not handled by the complex-noininal handler or by the nominalized verb handler are analyzed so that the head noun is connected to the concepts outside the noun phrase (e.g. a verb concept in the CF Handler). In addition, this module identifies individual concepts corresponding to adjectives and other nouns in a cornpound noun and connects them with CHARACIERI[OCRerr]C, ATIRIBIJIE, or LINK relations. LINK is the most generic relation in our system. Once noun phrases are handled this way, this module handles prepositional phrases by connecting the head noun concept of the noun phrase to the preceding constituent (e.g. a verb or a noun). The preposition attachanent problem is a difficult one, and the current implementation takes the simple- minded approach with general relations such as LINK, which can match with many of other semantically more specific relations. Our preliminary analysis indicates that this approach correctiy handles about 75% of the prepositional phrase cases in the Wall Street Journal collection. More acuurate and finer-level processing will be done with more semantically oriented rules that check the semantic restrictions and use more specific relations. The role of this handier will be diininlshed when we process phrasal verbs as part of the CF handler, for which we have constructed case frames. 2.1.4. Ad-hoc Haddl[OCRerr] This module looks for lexical patterns not covered by any of the other special handlers discussed above. Its processing is also driven by itsown knowledge base of patterns to infer relations between concepts. For example, a sentence fragment bought the item for the purpose of satisfying... contains a pattern ..... for the (ADJ) purpose of [NP] in the knowledge base, and hence results in a triple 95