Indexing Automatic and Government-Binding Theory
Robert J. Kuhns 205 Walnut Street Brookline, M A 02146 USA
ABSTRACT
This project note describes a systern that receives, parses, indexes, and routes news reports. The core of this ,'mtomatic indexer is a parser based on Govermnent-Binding Theory which derives thematic and binding relationships of arguments of the sentences of slories. These syntactic structures are interpreted by a semantic processor which is linked to conceptual representations of terms from a controlled indexing vocabulary. As a result, the system is capable of indexing news with respect to a large set of let+ms that denote Ihe content of the articles.
This suggests that systems that could operate at a conceptual level would be capable of indexing in ways that cotdd permit highly effective retrieval. It is with the assumption that NLP technology can provide tile means of categorizing text that guide several recent efforts. In particular, each of/Hayes et al. 1988/,/Kuhns 1988/, and/Rau and Jitcobs 1988/ describes systems that characterize news reports with results that could llOt be obtained by keyword methods alone. Since a news analysis system (NAS) was first reported in/Kuhns 1988/, a number of major enhancements to its design and underlying functionality have been incorporated. It is the purpose of this paper to report on tbe current state of NAS and ils COlllpolleI]ts. S C O P E AND OB,IECTIVES A primary design goal behind NAS is to develop a system employing NLP technology that wouh:t be capable of either routing news t'rom electronic news feeds in real-time or indexing news with respect to very large sets of indexing terms (authority files). ~ This vocabulary is broad and ranges over diverse domains, with lerms representing proper names, concrete objects, or abslract relationships. The last type is dependent on the content of the stories and is particularly suited for the syntactic/ semantic techniques of NAS. The form of the indexes that NAS produces is a set of pairs of headings and subheadings or descriptors. The headings are frequently proper names while the descriptors add detailed information to the index by denoting various relationships of entities mentioned in the news reports.
I',ACKGROUNI)
With the rapidly increasing \olume of text being generated. transmitted, processed, and stored, it becomes critical that informalion retrieval and routing be highly efficient, both in time of processing and accuracy. To this end, indexing techniques have become the prinmry focus of much research, a~d 3'el dlese methods have relied on automatic keyword identification from texts. This is not to say that natural hmguagc techniques have not been examined with respect to their relevance for indexing and retrieval (cf./Sparck Jones and Kay 1973/, /Walker, Karlgren, and Kay 1977/, and more recently,/Sahon mid Smith 1989/). It is that most systems rely on the presence or absence of keywords with additional mechanisms such as proximity constraints, statistical weighting, word-stem truncation, and boolean relrieval expressions. However, these methods do ~ot take into account the syntactic and semantic structure inherent in tile text being indexed. That is, they make virtually no use of the fact that it is natural language and not a collection of arbitrary strings of characters that is being processed. Natural language processing (NLP) can make its most valuable cont,'ibutions to those aspects of indexing where the keyword approaches fail, viz., the assignment of terms ~.o text based on their semantic or conceptual content. This involves deriving abstract relationships among conceptual units. For example, consider a story stating: ( 1) China bought 6,000 tonncs of wheat from the United States. [)he plausible categorization of (1) is thai it is about foreign trade. However, the phrase "foreign trade" does not appear in (1), and it is invariably absent fl'om foreign trade stories in general. Furthermore, it is extremely unlikely that such foreign trade stories cottld be retrieved in an efficient manner, i.e., with a tew simple queries. The central issue is that although the particulars (e.g., country names and types o fcommodilies) wuy, the basic lneanings of foreign trade slodes are equivalent at some level, and that this level is wduable for indexing purposes.
ARCHITECTURAL OVERVIEW
Tile architecture of NAS is modular consisting of several main subsystems, viz., a set of preprocessors and template filters, a parser, a lexicon, a semantic interpreter, a set of concept bases, and an indexer (Figure 1). The system is transportable in that it can be interfaced to different news streams and indexing vocabularies. A preprocessor receives a news stream which can be flom a satellite dish link, a direct line, or a text file, and identifies the beginning and ending of stories in addition to their titles, sentences, and words. Since the format of each news feed, e.g., Reuters or Kyodo, is distinct from the others, a single preprocessor will accept only one news feed. For rigidly-formatted articles that are numerical or non-texttlal in form, a template filter, which is an indexing component of low-level routines, categorizes them from lhe title while deriving specifics l'rom the body of tile story.
1
397
Template
Filter
~1
I
ndexes
Textual Stories
Lexicon
I
Words
[Parser
Structures ~
I
i
%c~Ttl
I
to.co,
~
i
Terms
An Architectural Overview of NAS
Figure 1
Figure 2 is an actual example (from Reuters) that has been indexed by the ternplate filter. The system outputs the company name with a descriptor "3rd Quarter Earnings," as well as the current and cumulative earnings or losses. @R101903647 /&ACTMEDIA INC <ACTM.0> 3RD QTR LOSS WESTHAMPTON BEACH, N.Y., Oct 19, Reuter Shr loss 14 cts vs profit 10 cts Net loss 1,674,000 vs profit 1.207,000 Revs 26.7 mln vs 19.1mln Nine Months Shr loss 19 cts vs profit 34 cts Net loss 2,280,000 vs profit 4,080,000 Revs 71.6 mln vs 59.8 rain Reuter indexes: Company Name Descriptor Subject Descriptor Subject Descriptor ACTMEDIA INC 3RD Quarter Earnings Net loss 1,674,000 Current Earnings Net loss 2,280,000 Cumulative Earnings (2) Agent (COMPANY) Predicate (INTRODUCE) Theme (PRODUCT). This structure denotes that a product introduction is one where a company introduces (or, synonymously, releases) a product. In short, it is a list of typed nodes. A report will be characterized as a product introduction story if it contains a sentence some of whose grammatical components (e.g., agent, predicate, theme) can be associated to the corresponding nodes of{ 2), Suppose, l:or example, a news item reports
{3)
Alpha Corp said it plans to release a new workstation in Japan.
The parser, in accordance with GB principles, produces: (4) Agent (Alpha Corp) Predicate (said) Proposition Agent (Alpha Corp) Predicate (plans) Proposition Agent (Alpha Corp) Predicate (release) Theme (a new workstation) In Numerical (Japan)
A Story and Its Index
Figure 2
In contrast, textual stories require grammatical processing and these are sent to the parser and semantic interpreter. The parser which relies on the principles of Government-Binding (GB) Theory (/Chomsky 1981/3, outputs predicate-argument structure of each sentence of a sto W.a In doing so, the parser identi ties empty categories, viz., PROs, traces, and variables, and thematic relations, and resolves antecedent and anaphor and pronominal bindings. It should be noted that the parser is interfaced to a lexicon of over 17,000 items that was developed by analyzing strings (words) from a newswim. The size of the lexicon is sufficient for news processing. The semantic interpreter maps the grammatical structures onto conceptual representations or filters stored in a concept base. For instance, a representation for "Product Introduction" is
The parser binds the pronominal it, the agent of plans, to A~)ha Cor~tx, the subject or agent of the matrix clause. The parser also detects an empty category, viz., PRO, in the embedded sentence (proposition) with release as the verb and binds the pronominal
to it.
Since bound arguments share the same semantic features, the semantic interpreter determines that the agent of release in (4) is of type COMPANY. In other words, PRO inherits the property of COMPANY from the agent of the matrix sentence via the intermediate pronominal k. It also determines that the predicate release is synonymous with introduce and the theme workstation is a product. With the arguments typed and membership of the predicate within a synonym class known, the semantic processor can match the corresponding nodes of the most deeply embedded clause of (4) with (2), and thus determines that the sentence is about a product introduction. Associated with each conceptual filter is a set of indexing procedures that are invoked
398
2
by the indexing mechanism when a conceptual filter is satisfied. These thnctions are integrated with databases containing the indexing vocabulary and they identify specific information about a story including company, personal, and product names, and descriptors indicating specific relationships, Figure 3 illustrates the corporate and personal name identification capabilities and the level of "understanding" as reflected by the subheadings. @ R080100252 /&ALCO HEALTH<AAIIS.0> CHIEF EXECUTIVE RETIRES VALLEY FORGE, Pa., Aug 1, Renter - Alco Health Services Corp said Ray B. Mundt has been named acting chairman and chief executive officer, succeeding John H. Kennedy, who is retiring. indexes: Company Name Descriptor Personal Name Descriptor Personal Name Descriptor Alco Health Services Corp Officials and Employees Kennedy, John H. Retirement Mundt, Ray B. Select ion",Appointment
parser. (/Kuhns 1986/describes an earlier implementation of GB Theory.) From a research perspective, a parser based on linguistic theory and applied to "real-world" text helps identify the boundaries or interface conditions between core and peripheral aspects of the theory. In other words, since GB is a model of core grammar and language contains marginal or marked conslructions, a GB-based parser nmst co-routine principles of the theory with hmguage-specific rules in order to have wide coverage. (/Fomita 1988/makes a similar observation.) Thus, it is this combination of a psychologically-real theory and an application using fl'ee text that may provide insight into the human sentence processing mechanism.
REFERENCES Chomsky, N., ( 1981 ) Lectures on Government and Binding, Foris Publications, Dordrecht, Holland. ttayes, P.J., L.E. Knecht, and M.J Cellio, (1988), "A News Categorization System," in Proceedings o1' the Second Conference on AEplied Natural Langu~,~c Processing, Austin, Texas. Kuhns, R.J., (1986), "A PROLOG Implementation of Government-Binding Theory," in Proceedings of the Xlth International Conference on Computational Lin~t_fistic~, Bonn, West Germany. Kuhns, R.J., (1988), "A News Analysis System," in Proceedings of the Xllth International Conference on Computational Lin,,,,uistics...