In the preface to Proceedings of COLING 2000, Martin Kay notes that when David Hays coined the expression Computational Linguistics back in the ’sixties, his intention was “to provide a more solid theoretical foundation for work on machine translation.” And, although this meaning has always lurked in the background, the term right from the beginning could not be restricted solely to machine translation―fortunately. Scholars from diverse disciplines―mathematics, linguistics, logic, information theory, statistics, the humanities―came together to pool their knowledge to attain a common goal: to create a machine that could translate from one language to another; or even more ambitiously: to enable a computer to behave like a human language user. The first COLING conference, in 1965 in New York, revealed this idea of an international, interdisciplinary forum for discussion. In following years, COLING became a truly international event, carefully orchestrated under the auspices of the International Committee on Computational Linguistics (ICCL), which “exists for the sole purpose of arranging” the conference every two years. “The members of the ICCL represent only themselves; certainly not the countries or institutions they come from” and they are independent from any particular scientific organization. (See Martin Kay at www.dcs.shef.ac.uk/research/ilash/iccl/.)
On that basis, COLING conferences have become a well-established occasion for the exchange of ideas and experiences in the field. COLING- conferences have tried to cover all aspects of the field and to reflect the main directions of its participants, such as machine translation, machine-aided translation, information retrieval and information extraction, grammar formalisms, parsing, semantic models, summarization, generation, natural language understanding, and question/answering systems. Innovations and extraordinary ideas were always welcomed.
Over the years, Computational Linguistics, partly supplemented by Artificial Intelligence, has become a regular academic discipline in many countries all over the world. Computational Linguistic research has also take a leading role in language engineering and language technology. The world-wide evolution of Computational Linguistics, on the other hand, has been influenced by several more or less ‘external’ conditions:
First of all, the rapid development of computer power has disposed of the most serious initial obstacle to natural language processing: Since speed of processing and memory size are no longer insurmountable hindrances, researchers are no longer forced to work with restricted small- scale vocabularies or knowledge bases (i.e., the so-called ‘toy’ systems). Both rule-based processing assisted by large lexica and the tour-de-force application of statistical processes and learning algorithms based on large- scale corpora have become widely available. Many research issues previously considered to be nearly impossible for machines to resolve (such as POS assignment) have received robust and replicable solutions with the stochastic methods supported by large corpora. With the past decade dominated by stochastic approaches and their impressive results, computational linguists are now looking back to integrating rule-based knowledge. Combinations of rule-directed deep analysis with robust statistically based shallow analysis systems are now under consideration and may help to bring language processing systems to previously unforeseen levels of efficiency and application.
A second development that has had an enormous and persistent influence on the development of Computational Linguistics concerns the availability and re-usability of language resources. Most importantly, affordable, fast computing power is a prerequisite to the wide application of large-scale language resources in Computational Linguistics. The first step has been the PC revolution, which facilitated the construction and availability of language resources in languages other than English. Once computing became available and accessible for many different languages, the collection of language resources followed. These resources in turn have provided the basis for natural language processing in those languages. The next step was a rising demand for language resources, when the issue of re-usability became central. On the one hand, the building of language resources is very labor-intensive, hence re-using language resources would save time and money. On the other hand, carefully constructed language resources can be used in different applications, and subsets of language resources can be combined under different criteria to form a different language resource. In other words, re-usable language resources can create added value. It is also crucial to note that linguistic acts are transient and that language varies with time and location. Hence some language resources are no longer replicable because of the change of contexts or the loss of native speakers. Thus, re-using language resources makes both budgetary and theoretical sense.
The re-usability of language resources depends crucially on the standardization of formats and tools. The worldwide endeavors to standardize the format of data and annotations started from the very beginning of Computational Linguistic research. Success came with SGML in the late ’eighties and—for Computational Linguistics—with the Textual Encoding Initiative (TEI) in the early ’nineties. But the implications of standardization go far beyond facilitation of information exchange. Standardization also makes international resource sharing possible. Textual corpora, lexica, grammars, and tools can all be shared on an international level nowadays regardless of the languages involved. Thus, it is no longer necessary to construct separate dictionaries (or grammars, or textual resources) for every new project. Instead, linguistic knowledge is provided in a theory-independent manner and in an international distributable standard. Standardization, in connection with the development of the World Wide Web, promoted these efforts and led to more international coöperation than ever before. In addition to this, network-based projects like WordNet and the recent Semantic Web Vision demonstrate that there are common structures between languages and cultures that may allow computerization of language-independent representation of knowledge, which is in turn accessible in a multi-lingual environment.
The third point concerns the introduction of multimodality into the scope of Computational Linguistics. The rapid technological progress of the last fifteen years has made it possible to incorporate communication channels other than written text, i.e., spoken and visual channels. Computational Linguistics for more than 40 years has remained focused on orthographically rendered language. Indeed, there was significant research in speech recognition and speech production, called Digital Signal Processing even in the ’sixties. However, as the name suggests, these studies typically do not attempt to use linguistic structure and hence there was no real interaction with Computational Linguistics, although there were a few studies concerning the relationship between written and spoken language, such as phonological programs simulating the idea of minimal pair analysis. However, these studies all started with and focused on annotated phonetic input/output. Nowadays, since modern computers have multimedia capacities—including audio and visual—it is possible to integrate digital signal processing with the so-called ‘higher’ levels of understanding, syntax and semantics. We now are able to use parameters of spoken language like word stress and intonation patterns for the description and disambiguation of sentence and discourse structures. In the near future, the visual channel will become available for transcribing gestures and face expressions. This information will be used for structural analysis and for generation.
Last but not least is the wide accessibility of the web. Before the World-Wide Web became the dominant medium of information exchange, the applications of natural language technology were limited. It was difficult to imagine that Computational Linguistic technology would apply to many languages in the world and could be used daily by the general public. The Web as an information infrastructure drew users with no technological background to the computer. These users had common needs to access information using daily language, which in turn, redirected the focus of human language technology back to natural language processing. Although Computational Linguistics has yet to prove that it delivers the optimal solution for cross-lingual and cross-modal information access problems, its does clearly identify and define the relevant issues. The dominance of the web also brings a renewed focus to semantics and the relation between language and information in Computational Linguistics.
COLING Conferences are organized to encourage new developments and ideas, regardless of theoretical, geographic, political, or commercial interest. Conference presentations (papers and project notes screened by a rigorous formal reviewing system) are published in the conference proceedings. In addition to the presentations, COLING conferences have fostered a pleasant, yet intellectually stimulating environment that allows for lively discussion. Important in this context are plenary lectures and panel discussions where new ideas and developments may be presented. A more recent tradition (just a little more than two decades old) is the sponsoring of pre-COLING and post-COLING events, such as tutorials and workshops. At these events, particular problems and trends within the field can be discussed at greater leisure.