Judith Wagner
Dr. sc. hum.
Multilingual Natural Language Generation from a Medical Concept
Representation
Geboren am 30.07.1963 in Weinheim
Reifeprüfung am 12.05.1982 in Weinheim
Studiengang der Fachrichtung Medizinische Informatik vom WS 1982 bis SS 1988
Vordiplom am 10.01.1985 an der Universität Heidelberg
Diplom am 9.09.1988 an der Universität Heidelberg
Promotionsfach: Medizinische Biometrie und Informatik
Doktorvater: Prof. Dr. rer. biol. hum. R. Haux
A number of compositional Medical Concept Representation systems are being
developed in order to enable re-use and sharing of medical knowledge and data. These
need to be presentable as surface language phrases for use by end-users and
applications. This is achieved by generation of natural language, providing the
translation of conceptual representations to one or several 'natural' languages. It is a
potential way of reconciling two requirements, both of which are essential for many of
today's medical informatics developments, and especially for a computer-based patient
record, but which are sometimes conflicting: effective communication and the sharing
and reuse of medical information.
The main goals of this work were to describe and represent the knowledge necessary
for multilingual generation from a Medical Concept Representation, as well as to
develop a method and to design and implement a tool for multilingual generation,
which can be adapted specifically to the requirements and the language of the medical
domain. Typical medical language has become a telegraphic sublanguage with
particular characteristics and conventions, and with a high density of information
expressed in short and complex phrases: the medical 'jargon'.
This thesis presents a multilingual, semantic-oriented sublanguage approach to natural
language generation. The approach is based on the hypothesis that a conceptual model
of the domain and linguistic knowledge have to be clearly distinguished, and to be
linked by defined structures. The knowledge base includes the domain model, the
linguistic knowledge, and the links between both. The process comprises
transformation operations, selection operations, and the linguistic realisation in a
particular language, with a focus on how to parametise these operations by the
individual language parameters. The link between conceptual and linguistic structures
is then established on two levels: on the one hand, annotations link conceptual entities
(concepts, relations) to linguistic entities (words, syntactic structures). On the other
hand, transformation operations adapt conceptual representations by using definitions
to a level where they can be directly translated to linguistic structures.
This approach has been applied to the generation of noun phrases in several European
languages for conceptual representations of an existing medical concept
representation: the GALEN Common Reference Model. A large-scale experiment has
been done on the Urology part of the French ‘Nomenclature Commune des Actes
Médicaux’. This part includes 522 surgical procedures, which have been conceptually
represented following an overall surgical procedures scheme, and phrases have been
re-generated in different languages. The generation tool has also been used for a
Structured Clinical User Interface collecting information about urinary infections:
information is collected in one language by forms, and can be summarised in phrases
of the same or a different language.
These experiments have demonstrated that the generation tool can be adapted to
different source modelling schemes and to different destination languages, or
sublanguages, of a domain. The introduction of new languages was facilitated by the
generic approach. The generation of results in multiple languages for one domain has
shown to be relatively simple once a domain is covered on the conceptual side. In
addition, the effort for covering a new medical subdomain converges.
The transformation operations have been a central means for bridging distinct ways of
representing the same things conceptually - distinct modelling styles - and distinct, but
different, ways of expressing the same concepts in language - different languages and
language styles. The operations are based on the Conceptual Graphs formalism, and
they are parameterised by the availability of language-specific annotations, language-
independent definitions, and overall generation strategies. They enable not only
bridging of terminological gaps and the lack of exact translation equivalents, but also
the creation of different generation strategies, which allow the style of the generated
language to be adapted to application purposes and users. Together with specific
definitions which mirror a conceptual scheme, they allow for tailoring to the specific
characteristics of a modelling scheme.
The experience demonstrated that it is important to rely on an existing domain model
for generation in order to arrive at an application-relevant domain coverage. It also
showed the importance of the representation of conceptual relations, and of
distinguishing conceptual and linguistic models. The usefulness of Natural Language
Generation tools as a validation tool for complex conceptual modelling in medical
concept representation systems came also to the fore. Finally, it has been shown to be
possible for different applications to rely on a single common conceptual
representation, adapting the generation of surface language to match those
applications.
