Language Engineering

Natural Language Processing has been a scientific enterprise for 40 years. In the last decade and half, however, the interest in Applied Natural Language Processing has increased steadily. The reason for this development can be seen in the advances in various areas of computational linguistics on the one hand and both the availability of inexpensive and powerful computing machinery and the growing demand for various types of natural language functionalities as part of software systems on the other hand. This increase in practical and industrial interest has also been accompanied by substantial funding. And although the natural language community has been more careful (perhaps due to the experience of the ALPAC report in the sixities) than researchers in other fields of Artificial Intelligence—where the creation of unrealistic expectations led to a substantial scepticism followed by a decrease of support—fairly high expectations with respect of the near and mid-term exploitability of natural language products do exist.

In order to meet the growing expectations raised by serious market studies for the remainder of the decade, it is commonly agreed that natural language processing must be based on a sound theoretical basis that allows for a simplified construction, adaptation, or modification of natural language programmes to particular tasks. The task is quite a general one, since to the extent that natural language products are software products, the development of natural language programmes needs to take into account at least the dimensions of software-engineering, ergonomics, and linguistic functionality. Successful commercialization of natural language processing is not a natural consequence of more sophisticated linguistics, but primarily requires an efficient management of the linguistic software development cycle, and a clear conception of the use of natural language as a means of communication in general, and human computer interaction in particular.

Development of natural language systems, apart from the research interest, always worked on the assumption that such systems would be useful in practice. With the arrival of systems and prototypes that can be used, the question needs to be considered as to what makes them usable and what detracts from usability. The initial idea, that the usability of natural language systems was dependent solely on their linguistic capabilities, turned out to be mistaken. Like other systems, they need reasonable diagnostics, legible manuals, and must fit into their environment. As natural language systems, they differ from other systems in how they foreground specific facets of human language behaviour.

Most software development projects follow some type of clearly defined process, refined over many years of practical experience. Different phases of development are carefully tracked and documented according to strict guidelines. Until now the majority of natural language systems have been developed as research projects, that is they have been developed without such rigid conditions. However, now that the market for natural language systems is beginning to open up, there is a growing number of commercially available products. It therefore seems timely to start to discuss how the more traditional approaches to software development can be applied to such technologies. This paper concentrates on one vital stage in any development process, namely the test phase. We will show how current approaches need to be adapted to suit the needs of this type of software and, conversely, how much can be gained by looking at the experiences of testers of more conventional software systems.

Besides the continuing activities of the last 25 years in constructing natural language interfaces by simulating the human natural language capacities on machines and the upcoming of its “natural” alternative at the beginning of the eighties, the graphical interfaces, a lot of recent research activities in computer science, artificial intelligence and information science center around multimodal computer systems. Mixed modalities (e.g. text and graphics) are one of the most favored ideas nowadays in spite of the fact that the area is still in its formative stage (cf. Maybury 1991 as an overview).

The promising results speech recognition achieved in the recent years confront computational linguistics with a new task. The attractivity of speech understanding systems in the market justifies the effort to cope with spoken language computationally. This article is intended to sharpen the awareness of the difficulties that lurk behind the knowledge transfer from text to speech understanding. The comparison of the grammars of spoken and written language, especially the surplus of regular constructions in spoken language are the subject of several linguistic works, reviewed here to set up a sort of syntactic agenda for speech understanding systems. Nothing is said however about the various methods that will help to fulfill this agenda.

As an attribute of the noun language, the adjective human may appear more or less tautological. I want to stress, however, that in this contribution I regard language as one of the biological characteristics of a certain mammalian species, namely of Homo sapiens sapiens. I do so in order to gain a point of view that is as far away as possible from computers as a non-biological, a technical species.

Processing natural language dialogues obviously is much more than simply processing a sequence of more than one sentence. Similarly, processing spoken dialogues is much more than handling just another type of input.

In this paper the question of what is important for the acceptability of natural language interfaces is addressed. I will argue that while linguistic coverage is an important factor for the acceptance of Natural Language Interfaces (NLIs) it is not the only one. At least as important are features which are consequences of the use of natural language as communication medium. These features can be subsumed under the terms portability, habitability and robustness. Portability concerns the ease of transferring the NLI to different hardware/software environments, different domains, different applications and different natural languages. Good habitability secures ease of use and a short training time for the end-user. Robustness is a key factor in both maintainability and ease of use of an NLI. I will discuss these features in detail and show current solutions as well as their limitations.

Mapping natural language into a formal language with well defined semantics is prerequisite but not enough for practical applications such as “front ends” to database systems or other applications. We have to translate the “logical form” produced by the semantics module into a formal language understood by the application. Many applications allow a programmatic interface, and some applications (such as databases) define an interface language, e.g. SQL which has become a standardised language. For communication with knowledge bases, a standard communication language— KIF “Knowledge Interchange Format”—is emerging ([7]), making it possible to query or update knowledge bases written in different representation languages.

The following paper describes the development of a tool which verifies controlled grammars. First, the development context and the application is described. Then, the definition of the task, and the architecture of the system are given: It consists of a syntactic analyser, a repository of ill-formed structures, a matcher, and an output formulator. Finally, system test results are presented, and some problems to be solved are explained; also, the application environment is sketched.

From a commercial point of view, Natural Language (NL) systems can be divided into two categories:

Natural Language Interface Systems

By a technology I understand a set of technical solutions for a particular set or natural class of problems.

What does it mean for a human-machine interaction paradigm to be natural? Certainly it means to take into account what man has developed filogenetically as his devices for interacting with his similars: natural language in the first place. We are well aware that natural language is not only the main vehicle for such communication, but also that it has the dual role of being the device through which we organize our thought. This makes natural language communication a window on the mind as many human scientists have told us so convincingly, but also implies that natural language communication does not need “cognitive transducers”. If often we do need some conscious phase in organizing what we want to communicate, this depends on planning our communicative actions, deciding what we want to say and optimizing our rhetoric strategies; it does not depend much on the lower level of expressing ourselves through words. So, for human-machine interfaces, natural language has an unmatchable potential. On the other side we are facing an extraordinary revolution that touches all levels of our life. Computers and telecommunication are really part of our ecological system.

Spelling checkers, natural language interfaces and machine translation systems (MT) have something in common: they belong to the group of natural language processing systems and there is no doubt, that the theoretical basis of these systems is called computational linguistics. This basis has its roots in linguistics, i.e. in formal linguistics, in mathematics and computer science. Furthermore, there are influences from sociology, psychology and the interdisciplinary fields among these five. Computer science with its subfields hardware and software is the only candidate which could contribute to a technology.

In a small sector of the software market, there are products available that deal with written or spoken natural language. Among these software products are machine translation systems, database access systems, spelling checkers, retrieval systems etc. The astonishing fact about most of the products is that they are not based on techniques that are currently under discussion in the field of computational linguistics. This may be due to the necessary time lag between research, development and marketing. In some cases it seems to be a more substantial discrepancy where products are constructed without taking the scientific efforts sincerely into account.

In general, terminology specialists have little to contribute to the topic “natural language” (that is the concept proper of “general language” with regard to “universal language” including so-called “sublanguages”) as far as “general language” is concerned; they can only listen and show their interest. However, as soon as issues of specialized languages with their high amount of terminological units are addressed, they are very well in a position to contribute considerably. This was the actual focus of the critical remarks I made on the occasion of the seminar and which I would like to sum up in this paper. However, I would like to stress that the issues addressed below were not intended to delimitate against each other the disciplines concurring in this field, but rather to find a sensible integration of their theories and methods.

As linguists, we know our problems in syntax, morphology or semantics and we try to solve them with the best scientific methods available. We also know, approximatively, what would count as a scientific solution. Scientific solutions are the necessary prerequisite for the development of language industries, but their availability is not sufficient at all. In this sense research in pure linguistics is only partially relevant for the development of language industries.