Medical WordNet: A New Methodology for the Construction and Validation of Information Resources for Consumer Health

A consumer health information system must be able to comprehend both expert and nonexpert medical vocabulary and to map between the two. We describe an ongoing project to create a new lexical database called Medical WordNet (MWN), consisting of medically relevant terms used by and intelligible to non-expert subjects and supplemented by a corpus of natural-language sentences that is designed to provide medically validated contexts for MWN terms. The corpus derives primarily from online health information sources targeted to consumers, and involves two sub-corpora, called Medical FactNet (MFN) and Medical BeliefNet (MBN), respectively. The former consists of statements accredited as true on the basis of a rigorous process of validation, the latter of statements which non-experts believe to be true. We summarize the MWN / MFN / MBN project, and describe some of its applications. 1 From WordNet to Medical WordNet WordNet is the principal lexical database used in natural language processing (NLP) research and applications. (Miller, 1995), (Fellbaum, ed., 1998) While WordNet’s current version (2.0) has broad medical coverage, it manifests a number of defects, which reflect both the lack of domain expertise on the part of the responsible lexicographers, and also the fact that WordNet was not built for domain-specific applications. The research community has long been aware of these defects (Magnini and Strapparava, 2001), (Bodenreider and Burgun, 2002), (Burgun and Bodenreider, 2001), (Bodenreider, et al., 2003). Our response is to create Medical WordNet (MWN), a free-standing lexical database designed specifically for the needs of natural-language processing in the medical domain, with the goal of removing the ‘noise’ which is associated with the application of WordNet and similar resources to this specialized domain. MWN’s initial focus is on English single-word expressions as used and understood by nonexperts. We systematically review WordNet’s existing medical coverage by assembling a validated corpus of sentences involving specific medically relevant vocabulary. Input to our validation process includes the definitions of medical terms already existing in WordNet, and also sentences generated via the semantic relations linking such terms in WordNet. In addition, input includes sentences derived from online medical information services targeted to consumers. Our methodology is designed (1) to document natural language sentential contexts for each relevant word sense in such a way that the expressed information can be (2) validated by medical experts and (3) accessed automatically by NLP applications such as information retrieval, machine translation, question-answer systems, and text summarization. A major stumbling block for existing NLP applications is automatic sense disambiguation. An automatic system can detect with high reliability that a given occurrence of a word like feel or dead is a verb or adjective. But it cannot easily determine which of a variety of alternative meanings such polysemous words have in any given context. WordNet’s architecture, designed for representing and distinguishing word senses, has made an important contribution towards a solution of the automatic word sense disambiguation problem. Our corpus of English language sentences relating to medical phenomena is designed to build upon this contribution. The corpus is restricted to grammatically complete, syntactically simple sentences in natural language which have been rated as understandable by non-expert human subjects in controlled questionnaire-based experiments. It is restricted in addition to sentences which are self-contained in the sense that they make no reference to any prior context and do not contain any proper names, or anaphoric elements (like it or he or then) that need to be interpreted with respect to other sentences or some surrounding discourse or context. This corpus is designed to be used initially for purposes of quality assurance of MWN and also to support the population of MWN by yielding new families of words and word senses for inclusion. As will become clear, however, our use of human validators will allow us to extend the usefulness of the corpus in a variety of ways. Thus we can use it to build new sorts of applications for information retrieval in the domain of consumer health. But it also allows new avenues of research in linguistics and psychology, for example in allowing us to explore individual and group differences in medical knowledge and vocabulary, and in understanding non-expert medical reasoning and decision-making. 2 Medical FactNet and Medical BeliefNet To this end, however, we need to exploit our validation data to create two sentential subcorpora, called Medical FactNet (MFN) and Medical BeliefNet (MBN), respectively. MFN consists of those sentences in the corpus which receive high marks for correctness on being assessed by medical experts. MFN is thus designed to constitute a representative fraction of the true beliefs about medical phenomena which are intelligible to non-expert English-speakers. MBN consists of those sentences in the corpus which receive high marks for assent. MBN is thus designed to constitute a representative fraction of the beliefs about medical phenomena (both true and false) distributed through the population of English speakers. The validation process that is involved in the construction of MFN is used to detect errors in the existing WordNet, and also to ensure that the coverage of the natural language medical lexicon in MWN is of a scientific level sufficient to allow MWN technology to work in tandem with terminology and ontology systems designed for use by experts. Both MFN and MBN inherit from MWN the formal architecture of the Princeton WordNet. (Fellbaum, ed., 1998) However, we enhance this architecture to maximize its usefulness in medical information retrieval. Compiling MFN and MBN in tandem allows systematic assessment of the disparity between lay beliefs and vocabulary as concerns medical phenomena and the corresponding expert medical knowledge. The ultimate goal of our work on MFN is to document the entirety of the medical knowledge that can be understood by average adult consumers of healthcare services in the United States today. If the methodology for the creation and validation of the corpus here described proves successful, then we believe that the preconditions for the realization of this much larger goal will have been established. Responses from NLP researchers and from online information providers to our initial work on MFN/MBN convinces us that this realization would have considerable significance for the management and retrieval of consumer health information in the future. 3 Background and Motivation Recent studies of the use of computer-based tools for consumer health information retrieval point to a mismatch between existing tools and the non-expert language used by most consumers – the language used not only by patients but also by family members, advisors, administrators, lawyers, and so forth, and to some degree also by nurses and physicians. (Slaughter, 2002), (C. A. Smith, et al., 2002), (Tse, 2003), (Tse and Soergel, 2003), (McCray and Tse, 2003), (Zeng, et al., in press) Where the usage of medical terms by professionals is at least in principle subject to control by standardization efforts, the highly contextually dependent usage of medical terms on the part of lay persons is much more difficult to capture in applications – and this in spite of the fact that it is in many ways simpler than expert usage. The taxonomies reflecting popular lexicalizations in all domains are indeed much less elaborate at both the upper and lower levels than in the corresponding technical lexica. (Medin and Atran, eds., 1999) Thus there are no popular terms linking infectious disease and mumps, so that in the popular medical taxonomy of diseases the former immediately subsumes the latter. The popular medical vocabulary naturally covers only a small segment of the encyclopedic vocabulary of medical professionalsm, and it lexicalizes mainly at the level of taxonomic orders. Popular medical terms (flu) are often fuzzier than technical medical terms. Many popular terms also cover a larger range of referent types than do technical terms; others may cover only part of the extension of their technical counterparts. We hypothesize, however, that with few exceptions the focal meanings (Berlin and Kay, 1969) of expert and non-expert terms will be identical. Constructing MFN and MBN allows us to test this and related hypotheses in a systematic way. 4 Mismatches in Doctor-Patient Communication The skills of a physician in general practice comprise the ability to acquire relevant and reliable information through communication with patients through the use of non-expert language and to convey diagnostic and therapeutic information in ways tailored to the individual patient. Since the physician, too, is a member of the wider community of non-experts, and continues to use the non-expert language for everyday purposes, one might assume that there are no difficulties in principle keeping him from being able to formulate medical knowledge in a vocabulary that the patient can understand. As (Slaughter, 2002) and (C. A. Smith, et al., 2002) have shown, however, there are limits to this competence. The former examines dialogue between physicians and patients in the form of question-answer pairs, focusing especially on the relations documented in the UMLS Semantic Network. Only some 30% of the relations used by professionals in their answers directly match the relations used by consumers in formulating their questions. An example of one such questionanswer pair is taken from (Slaughter, p. 224): Question Text: My seven-year-old son developed a rash today that I believe to be chickenpox. My concern is that a friend of mine had her 10-day-old baby at my home last evening before we were aware of the illness. My son had no contact with the infant, as he was in bed during the visit, but I have read that chickenpox is contagious up to two days prior to the actual rash. Is there cause for concern at this point? Answer Text: (a) Chickenpox is the common name for varicella infection. [...] (b) You are correct in that a person with chickenpox can be contagious for 48 hours before the first vesicle is seen. [...] (c) The fact that your son did not come in close contact with the infant means he most likely did not transmit the virus. (d) Of concern, though, is the fact that newborns are at higher risk of complications of varicella, including pneumonia. [...] (e) There is a very effective means to prevent infection after exposure. A form of antibody to varicella called varicella-zoster immune globulin (VZIG) can be given up to 48 hours after exposure and still prevent disease. Such examples illustrate also that there are lexically rooted mismatches in communication (which may in part reflect legal and ethical considerations) between experts and non-experts. Professionals often do not re-use the concepts and relations made explicit in the questions put to them by consumers. In our example, the questioner requests a yes/no-judgment on the possibility of contagion in a 10-day-old baby. In fact, however, only section (c) of the answer responds to this question, and this in a way which involves multiple departures from the type of nonexpert language which the questioner can be presumed to understand. Rather, physicians expand the range of concepts and relations addressed (for example through discussion of issues of prevention, etc.). In all cases, the information source, whether it be a primary care physician or an online information system, must respond primarily with generic information (i.e. with information that is independent of case or context), and this is so even where requests relate to specific and episodic phenomena (occurrences of pain, fever, reactions to drugs, etc.). (Patel, et al., 2002) In our example, all sections except for (c) are of this generic kind. They contain answers in the form of contextindependent statements about causality, about types of persons or diseases, about typical or possible courses of a disease. MFN is accordingly designed to map the generic medical information which non-experts are able to understand. 5 Non-Expert Language in Online Communication Understanding patients requires both explicit medical knowledge and tacit linguistic competence dispersed across large numbers of more or less isolated practitioners. This is not a problem so long as this knowledge is to be applied locally, in face-to-face communication with patients. However, as a result of recent developments in technology, including telemedicine and internet-based medical query systems, we now face a situation where such dispersed, practical (human) knowledge does not suffice. (Ely, et al., 2000) and (Jacquemart and Zweigenbaum, 2003) have shown that clinical questions are expressed in a small number of different syntactic-semantic patterns (about 60 patterns account for 90% of the questions). Such yes/no questions are of the forms: Do hair dyes cause cancer?, Can I use aspirin to treat a hangover? With the right sort of information resource, questions such as these can easily be transformed automatically into statements providing correct answers: Hair dyes can cause bladder cancer, Aspirin doesn’t help in case of a hangover , and these answers can be linked further to relevant and authoritative sources. MEDLINEplus is described in its online documentation as a source of medical information for both experts and non-experts ‘that is authoritative and up to date.’ Enquirers can use MEDLINEplus like a dictionary, choosing health topics by keywords. Alternatively, they can use the system’s search feature to gain access to a database of relevant online documents selected for reliability and accessibility on the basis of preestablished criteria. Table 1 shows the problems that can arise when a system fails to take account of the special features of the knowledge and vocabulary of typical non-expert users. Here success in finding the needed information depends too narrowly on the precise formulation of the query text. Thus tremble and trembling call forth different responses (one lists caffeine, the other phobias), even though the terms in question differ only in a morphological affix that does not involve a meaning distinction. Such problems are characteristic of information services of this kind. Experienced internet users are of course familiar with the limitations of search engines, and so they are able to manipulate their query texts in order to get more and better results. Even experienced users, however, will not be able to overcome the arbitrary sensitivities of an information system, and the latter cannot have the goal of bringing non-experts’ ways of using language into line with that of the system. 6 Corpusand Fact-Based Approaches to Information Retrieval (Patel, et al., 2002) make clear that if a medical information system is to mediate between experts and non-experts, then it must rest on an understanding of both expert and non-expert medical vocabulary. But terms, or word forms, are not always associated with word meanings in a clear-cut and unambiguous fashion; and the problem of polysemy is compounded when different speaker populations are involved. A lexical database must represent all and only the meanings of each given term in such a way that these meanings can be clearly discriminated and mapped onto word occurrences in natural text and speech. Achieving these ends is one of the hardest challenges facing both theoretical and applied linguistic science today. It is generally agreed that the meanings of highly polysemous terms cannot be discriminated without consideration of their contexts (e.g., Pustejovsky, 1995). People manage polysemy without apparent difficulties; but modeling human speakers’ capacity for lexical disambiguation in automatic language processing tasks is hard. The idea underlying the present proposal draws on currently emerging NLP methodologies that harness the ability of powerful and fast computers to store and manipulate both lexical databases and large collections of text collections or corpora. The strategy is to train automatic systems on large numbers of semantically annotated sentences that are naturally used and understood by human beings, and to exploit standard pattern-recognition and statistical techniques for purposes of disambiguation. Words and the representation of their senses, stored in lexical databases, can be linked for this purpose to specific occurrences in corpora.