By itself, Figure 3 is a simple representation of the content of the document. To allow the reader to make an independent judgement, the full text of this article appears in Appendix A. This list, in combination with the structural information discussed above, can be used at indexing time for IR applications.

The challenge in preparing an abbreviated representation of an article is to identify heuristics which make it possible to represent to the user the sense in which in which an author used an expression in the document, without performing full sense disambiguation. In an important sense, every document can be viewed as forming its own ‘self-contained’ world. A document is written to get across a particular idea or set of ideas. The task of the author, at least in documents intended for public distribution, is to convey to the reader what general knowledge is assumed and to inform the reader of the context so that ambiguous expressions can be easily identified. These references are governed by certain standard conventions.

For example, in an edited document such as a newspaper article, the first reference to a named entity such as a person, place or organization typically uses a relatively full form of the name in a version which is sufficient to disambiguate the reference for the expected audience. Later in the document, the same entity is usually referred to by a shorter, more ambiguous form of the name (Wacholder and Ravin 1997). An article might first refer to Columbia University or, (more formally) Columbia University in the City of New York, and later refer only to Columbia. Without the initial disambiguating reference, Columbia by itself is quite ambiguous. It might be a city (Columbia, MD), a bank (Columbia Savings and Loan) or one of many other entities. Nominator, a module which identifies proper names developed at the IBM TJ Watson Research categorizes them, and links expressions in the same document which refer to the same entity successfully exploited this property of documents (Wacholder et al. 1997). Nominator first builds a list of proper names in each document and then applies heuristics in order to link names which refer to the same entity (e.g., Hillary Clinton and Ms. Clinton, but not Bill Clinton). This technique produces reliable links between references to the same entity in a document.

Common NPs also manifest a pattern of referential linking in documents, although it is more subtle and complicated than the proper name behavior. Any article of more than minimal length contains repeated references to important concepts. In general, when a word appears as a head of an NP in a document, it is used in the same sense throughout the document, especially in articles of newspaper length. Some of the references to the head are elliptical and therefore very ambiguous, at least out of context, but some of the references are usually fuller and therefore more specific and more informative. For example, in the Wall Street Journal article that is used as the primary example throughout most of this paper, the most frequent head of simplex NPs is the workers. Six of the nine references to workers are not preceded by an adjective or noun which delimits the intended sense of workers; however, one of these references is to the more specific asbestos workers. The different references to a concept implicitly or explicitly refer to each other and collectively form an abstract construct that conveys the sense that the author (presumably) intended to convey. (See Kameyama (1997) for a discussion of the importance of establishing all referential links within a document for information extraction applications, so that information about these entities can be merged.) When simplex NPs are clustered by head, NPs with the same head are likely to refer to the same concept, if not to the same entity. For example, in the sentence "Those worker got a pay raise but the other workers did not", the same sense of worker is used in both NPs, but the workers referred to are different.

Sorting by final word of the name is a simplification for proper names which, in contrast to common NPs, do not have a head in the sense that there is a single word which is semantically and semantically the most important. However, clustering proper names by the final word, as if it were the head, is satisfactory for certain kinds of proper names, including human ones. For example, Talcott can reasonably be considered the head of both James Talcott and Dr. Talcott. We are currently in the process of refining the head clustering procedure to handle organizations and other categories of proper names that have different naming conventions (Wacholder et al. 1997). In contrast to common NPs and proper NPs, reference to concepts in the form of pronominal anaphors contribute no references to new entities and therefore will not be discussed in this paper.

The head clustering technique provides a way to situate the entities referred to in the document in the context of related entities so their sense is comprehensible to users who have not actually read a document. The full list of simplex NPs which have these heads appears in Figure 4. Examination of this list suggests that it provides a more explicit representation of the content of the article than does the list in Figure 3.

(48) the workers

(83) workers

(100) the workers

(104) any asbestos workers

(144) 160 workers

(152) Workers

(161) Workers

(169) those workers

(2) asbestos

(41) asbestos

(43) no asbestos

(67) asbestos

(115) asbestos

(118) asbestos

(141) asbestos

(143) cancer-causing asbestos

(3) Kent cigarette filters

(27) Micronite cigarette filters

(70) the filters

(72) filter

(74) the filters

(112) the cigarette filters

(147) the Kent filters

(160) filters

(10) researchers

(20) researchers

(47) the researchers

(61) researchers

(92) the researchers

(1) The asbestos fiber

(28) needle-like fibers

(35) More common chrysotile fibers

(57) acetate fibers

(58) the dry fibers

(12) crocidolite

(25) crocidolite

(116) crocidolite

(129) crocidolite

(151) the crocidolite

(103) paper factory

(145) a factory

(150) the factory

(165) the factory

(9) 30 years

(29) a year

(39) years

(178) 35 years

(56) James A. Talcott

(59) Dr. Talcott

(97) Dr. Talcott

(122) Dr. Talcott

(5) cancer deaths

(88) 18 deaths

(99) lung cancer deaths

(84) asbestos-related diseases

(96) asbestos-related diseases

(171) asbestos-related diseases

(24) Kent cigarettes

(51) the Kent cigarettes

(73) 9.8 billion Kent cigarettes

Figure 4

For example, filter is the head of eight simplex NPs. Four of these have adjective and nominal premodifiers: Micronite cigarette filters, Kent cigarette filters, the cigarette filters and Kent filters. In the absence of other references to specific kinds of filters, the correct and accurate generalization is that the kinds of filters discussed in this document are cigarette filters, rather than coffee filters or oil filters. Asbestos workers and cancer-causing asbestos, the most specific NPs with the head workers and asbestos respectively, as measured by number of content words preceding the head, accurately characterizes the property of the workers and of asbestos that is most important for this document. Similarly, the most specific simplex NP suggests that the type of factory under discussion is a paper factory.

For the head fiber, there are five different premodifiers. While it is impossible to determine from the list here which of these types of fiber are the same and which are different, the variety of premodifiers suggests that types of fibers are being discussed in this document.

For researchers and crocidolite, this technique provides no further specification of information, but merely does the same thing that a count of the occurrence of these strings in the document would yield, along with the additional information that these words are repeatedly used as heads in the document and therefore are more likely to be candidate significant topics than a word like Kent which is used five times, but only as a modifier.

The technique proposed in this paper is a general purpose one that can be used in a variety of ways to identify significant topics in a document. In the long run, the practical value of this technique will be judged by its utility in NLP applications such as run-time indexing for information retrieval, automatic summarization and back-of-the-book indexing.

However, an initial evaluation, as well as useful suggestions for refining the technique, has been obtained from human users. In the evaluation, three articles were presented to five individuals; none had any experience in NLP (though one was a professional librarian and indexer). The judges uniformly ranked the list of clustered NPs was ranked most highly, with an average rating of 3.15; the keyword list was ranked second, with an average rating of 2.45, and the list of word sequences was ranked last, with an average rating of 1.92.

The evaluation was conducted as follows. In order to determine whether the list of significant topics output by the headed clustering technique conveys the sense of a document, judges were asked to compare it to two other kinds of output: a list of keywords of frequency of more than one and a list of repeated sequences of words. The keyword list was chosen because it has become a standard in many NLP applications. It therefore establishes a baseline for comparison, even though lists of keywords are not generally used to represent document aboutness. The list of repeated word sequences is similar in its use of repeated phrases, except that it uses a variation of relies notions of technical terms and technical prominence rather than on repeated heads.

The evaluation was conducted as follows. A list of simplex NPs clustered by head was output by LinkIT for each of the three articles. The list included all clusters whose head occurred in the document as a head of more than one simplex NP; duplicates were removed. For wsj_0003, 10 of the 32 simplex NPs considered significant because their heads occurred more than once are shown in Figure 5. The number in brackets is the frequency of occurrence.

workers [4]

the workers [2]

any asbestos workers [1]

160 workers [1]

those workers [1]

asbestos [7]

no asbestos [1]

cancer-causing asbestos [1]

Kent cigarette filters [1]

Micronite cigarette filters [1]

the filters [2]

filter [1]

the cigarette filters [1]

the Kent filters [1]

filters [1]

asbesto [14]

work [11]

filt [8]

canc [7]

research [6]

cigaret [6]

make [5]

lorillard [5]

kent [5]

fibe [5]

The list of repeated sequences of words was output by termer, an implementation by Min Yen-Kan of Katz and Justeson’s technical term algorithm. All word sequences which occurred more than once in the document were listed and capitalization was added where appropriate. There were six repeated word pairs in this document, all of which are listed in Figure 7.

Kent cigarette [4]

cigarette filter [3]

Dr. Talcott [3]

cancer death [2]

lung cancer [2]

U.S. [2]

Figure 7

Although care was taken to make the lists as equivalent as possible, not all the differences could be balanced out while maintaining faithfulness to the reliability of the method. For example, the repeated sequence method consistently produces a shorter list than does the clustered NP technique, and keyword technique produces the longest list.

The judges were asked to study the three lists, compare them to each other and to the text of the article and then rank each one on a scale of 1 to 5, where 1 indicated that the list provided no idea of the content of the article and 5 indicated that the list provided an excellent idea of the article content. The results of the evaluation are shown in Figure 8.

The judges’ preference for keywords over technical terms was surprising, given the claim made above that phrases are more informative than keywords. However, in informal discussion, judges confirmed that the coherent expressions in the clustered NP list and the repeated word sequence list were more meaningful than the stems and isolated words in the keyword list. However, the fact that the repeated sequence list was significantly shorter than the other two made it less helpful than the other two and was responsible for the relatively low scores that this list received. This suggests that better results might be obtained from a list in which the list of clustered NPs is further filtered to include only simplex NPs with content-bearing modifiers; for example, instead of the simplex NPs whose head is workers shown in Figure 7, only asbestos workers would be listed.

This evaluation suggests that the head clustering method does in fact produce a set of plausible signficant topics.

In conclusion, it appears that the head clustering technique is a promising one for a variety of applications. Moreover, since head clustering has a grammatical basis, the method discussed in this paper is in principle domain general. In fact, the code for recognizing simplex NPs in Wall Street Journal articles did not have to be modified in order to handle abstracts of National Science Foundation grant applications, a quite different genre and domain than newspaper articles. The method described in this paper therefore merits further study.

We plan to take this research in several directions. First we are exploring the applicability of head clustering to other types of documents and to documents that are longer than newspaper articles or proposal abstracts. Second, we are undertaking qualitative and quantitative analysis of the significant topics identified by the method

described in this paper and evaluation of their usefulness, in comparison with other techniques for identifying significant topics. Finally, LinkIT output is being used in a variety of research applications.

A form of asbestos once used to make Kent cigarette filters has caused a high percentage of cancer deaths among a group of workers exposed to it more than 30 years ago, researchers reported.

The asbestos fiber, crocidolite, is unusually resilient once it enters the lung, with even brief exposures to it causing symptoms that show up decades later, researchers said. Lorillard Inc., the unit of New York-based Loews Corp. that makes Kent cigarettes, stopped using crocidolite in its Micronite cigarette filters in 1956.

Although preliminary findings were reported more than a year ago, the latest results appear in today's New England Journal of Medicine, a forum likely to bring new attention to the problem.

A Lorillard spokeswoman said, "This is an old story. We're talking about years ago before anyone heard of asbestos having any questionable properties. There is no asbestos in our products now."

Neither Lorillard nor the researchers who studied the workers were aware of any research on smokers of the Kent cigarettes. "We have no useful information on whether users are at risk," said James A. Talcott of Boston's Dana-Farber Cancer Institute. Dr. Talcott led a team of researchers from the National Cancer Institute and the medical schools of Harvard University and Boston University.

The Lorillard spokeswoman said asbestos was used in "very modest amounts" in making paper for the filters in the early 1950s and replaced with a different type of filter in 1956. From 1953 to 1955, 9.8 billion Kent cigarettes with the filters were sold, the company said.

Among 33 men who worked closely with the substance, 28 have died -- more than three times the expected number. Four of the five surviving workers have asbestos-related diseases, including three with recently diagnosed cancer. The total of 18 deaths from malignant mesothelioma, lung cancer and asbestosis was far higher than expected, the researchers said.

"The morbidity rate is a striking finding among those of us who study asbestos-related diseases," said Dr. Talcott. The percentage of lung cancer deaths among the workers at the West Groton, Mass., paper factory appears to be the highest for any asbestos workers studied in Western industrialized countries, he said.

The finding probably will support those who argue that the U.S. should regulate the class of asbestos including crocidolite more stringently than the common kind of asbestos, chrysotile, found in most schools and other buildings, Dr. Talcott said.

The U.S. is one of the few industrialized nations that doesn't have a higher standard of regulation for the smooth, needle-like fibers such as crocidolite that are classified as amphobiles, according to Brooke T. Mossman, a professor of pathlogy at the University of Vermont College of Medicine. More common chrysotile fibers are curly and are more easily rejected by the body, Dr. Mossman explained.

In July, the Environmental Protection Agency imposed a gradual ban on virtually all uses of asbestos. By 1997, almost all remaining uses of cancer-causing asbestos will be outlawed.

About 160 workers at a factory that made paper for the Kent filters were exposed to asbestos in the 1950s. Areas of the factory were particularly dusty where the crocidolite was used.

Workers dumped large burlap sacks of the imported material into a huge bin, poured in cotton and acetate fibers and mechanically mixed the dry fibers in a process used to make filters. Workers described "clouds of blue dust" that hung over parts of the factory, even though exhaust fans ventilated the area.

"There's no question that some of those workers and managers contracted asbestos-related diseases," said Darrell Phillips, vice president of human resources for Hollingsworth & Vose. "But you have to recognize that these events took place 35 years ago. It has no bearing on our work force today."