<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-1302">
  <Title>Towards a Road Map on Human Language Technology: Natural Language Processing</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Scope of this Document
</SectionTitle>
      <Paragraph position="0"> One of the items on ELSNET's agenda for the period 2000-2002 is to develop views on and visions of the longer-term future of the field of language and speech technologies and neighboring areas, also called ELSNET's Road Map for Human Language Technologies. As a first step in this process, ELSNET's Research Task group is organizing a series of brainstorming workshop with a number of prominent researchers and developers from our community. The first one of these workshops took place in November 2000 under the general motto &amp;quot;How will language and speech technology be used in the information world of 2010? Research challenges and infrastructure needs for the next ten years&amp;quot;. The second one was coorganized in July 2001 by ELSNET and MITRE as part of ACL-2001 and had the somewhat more specific orientation on &amp;quot;Human Language Technology and Knowledge Management (HLT-KM)&amp;quot;. This workshop brought together more than 40 researchers from industry and academia and covered a considerable range of topics related to KM and HLT in general.</Paragraph>
      <Paragraph position="1"> This paper aims at summarizing and organizing material from both workshops, but concentrates on applications and technologies that involve NLP, i.e. the processing of written natural language, as speech-related technologies and new models of interactivity have already been covered in documents presented around the first workshop. In the discussion of question answering and summarization, vision papers and roadmaps compiled by researchers in the US and published by NIST have been taken as an additional source of inspiration.</Paragraph>
      <Paragraph position="2"> The Growing Need for Human Language Technology Natural language is the prime vehicle in which information is encoded, by which it is accessed and through which it is disseminated. With the explosion in the quantity of on-line text and multimedia information in recent years there is a pressing demand for technologies that facilitate the access to and exploitation of the knowledge contained in these documents.</Paragraph>
      <Paragraph position="3"> Advances in human language technology will offer nearly universal access to on-line information and services for more and more people, with or without skills to use computers.</Paragraph>
      <Paragraph position="4"> These technologies will play a key role in the age of information and are cited as key capabilities for competitive advantage in global enterprises.</Paragraph>
      <Paragraph position="5"> Extraction of knowledge from multiple sources and languages (books, periodicals, newscasts, satellite images, etc.) and the fusion into a single, coherent textual representation requires not only an understanding of the informational content of each of these documents, the removal of redundancies and resolution of contradictions. Also, models of the user are required, the prior knowledge that can be assumed, the level of abstraction and the style that is appropriate to produce output that is suitable for a given purpose.</Paragraph>
      <Paragraph position="6"> More advanced knowledge management (KM) applications will be able to draw inferences and to present the conclusions to the user in condensed form, but let the user ask for explanations of the internal reasoning. In order to find solutions for problems beyond a static pool of knowledge, we need systems that are able to identify experts, who have solved similar problems. Again, advanced NLP capabilities will be required to appraise the aptitude of candidates from documents authored by them or describing prior performance.</Paragraph>
      <Paragraph position="7"> But also outside of KM, sophisticated applications of NLP will emerge over the next years and decades and find their way into our daily lives. The range of possibilities is almost unlimited. An important group of applications is related to electronic commerce, i.e. new methods to establish and maintain contact between companies and their customers. Via mobile phones, e-mail, animated web-based interfaces, or innovative multi-channel interfaces, people will want to make use of all kinds of services related to buying and selling goods, home-banking, booking of journeys, and the like. Also in the area of electronic learning a considerable growth is expected within the coming years.</Paragraph>
      <Paragraph position="8"> Multilinguality Whereas English is still the predominant language on the WWW, the fraction of non-English Web pages and sites is steadily increasing. Contrasting earlier apprehensions, the future will probably present ample opportunities for giving value to different languages and cultures.</Paragraph>
      <Paragraph position="9"> However, the possibility to collect information from disparate, multilingual sources also provides considerable challenges for the human user of these sources and for any kind of NLP technology that will be employed.</Paragraph>
      <Paragraph position="10"> One of the major challenges is lexical complexity. There will be about 200 different languages on the web and thus about 40.000 potential language pairs for translation. Clearly, it will not be possible to build bilingual dictionaries that are comprehensive both in the number of language pairs and in the coverage of application domains. Instead, multilingual vocabularies need to provide mappings into language independent knowledge organization structures, i.e. common systems of concepts linked by semantic relations. However, the definition of such an &amp;quot;interlingua&amp;quot; will be difficult in cases in which languages make distinctions of different granularity.</Paragraph>
      <Paragraph position="11"> Research Trends and Challenges The field of human language technology covers a broad range of activities with the goal of enabling people to communicate with machines using natural communication skills.</Paragraph>
      <Paragraph position="12"> Although NLP can help to facilitate knowledge management, it requires a large amount of specialized knowledge by itself. This knowledge may be encoded in complex systems of linguistic rules and descriptions, such as grammars and lexicons, which are written in dedicated grammar formalisms and typically require many person-years of development effort. The rules and entries in such descriptions interact in complex ways, and adaptation of such a sophisticated system to a new text style or application domain is a task that requires a considerable amount of specialized manpower.</Paragraph>
      <Paragraph position="13"> One way to cope with the difficulties in the acquisition of linguistic knowledge was to restrict attention to shallower tasks, such as looking for syntactic &amp;quot;chunks&amp;quot; instead of a full syntactic analysis. Whereas this has proven rather successful for some applications, it obviously severely limits the depth to which the meaning of a document or utterance is taken into account.</Paragraph>
      <Paragraph position="14"> Another approach was to shift attention towards models of linguistic performance (what occurs in practice, instead of what is principally possible) and to use statistical or machine learning methods to acquire the necessary parameters from corpora of annotated examples.</Paragraph>
      <Paragraph position="15"> These data-driven approaches offer the possibility to express and exploit gradual distinctions, which is quite important in practice. They are not only easier to scale and adapt to new domains, their algorithms are also inherently robust, i.e. they can deal, to a certain extent, gracefully with errors in the input.</Paragraph>
      <Paragraph position="16"> Statistical parsers, trained on suitable tree banks, now achieve more than 90% precision and recall in the recognition of syntactic constituents in unseen sentences from English financial newspaper text.</Paragraph>
      <Paragraph position="17"> However, a lot of work remains to be done, and it is not obvious how the success of corpus-driven approaches can be enlarged along many dimensions simultaneously. One challenge is that analysis methods need to work for many languages, application domains and text types, whereas the manual annotation of large corpora of all relevant types will not be economically feasible. Another challenge is that, other than syntax, many additional levels of analysis will be required, such as the identification of word sense, the reference of expressions, structure of argumentation and of documents, and the pragmatic role of utterances. Often, the theoretical foundation that is required before the annotation of corpora can begin is still lacking.</Paragraph>
      <Paragraph position="18"> One could say that for corpus-driven approaches the issue of scalability of the required resources shows up again, albeit in a somewhat different disguise. Hence, research in NLP will have to address this issue seriously, and find answers to the question how better tools and learning methods can reduce the effort of manual annotation, how annotated corpora of a slightly different type could best be re-used, how data-driven acquisition processes can exploit and extend existing lexicons and grammars, and finally how analysis levels for which the theoretical basis is still under development could be advanced in a data-driven way.</Paragraph>
      <Paragraph position="19"> Structure of this Document The remainder of this document is structured as follows. In Chapter 2 we describe a number of prototypical applications and scenarios in which NLP will play a crucial role. Whereas each of these scenarios is discussed mainly from a user's perspective, we also give indications, which technological requirements must be met to make various levels of sophistication of these applications possible. In Chapter 3, the technologies that have been mentioned earlier are discussed in more detail, and we try to indicate which levels of functionality may be expected within the timeframe of this study. These building blocks are then put into a tentative chronological order, which is displayed in Chapter 4. Finally, Chapter 5 gives some general recommendations about beneficial measures concerning the infrastructure for the relevant research.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>