Background

From DELPH-IN

Jump to: navigation, search

The most serious shortcoming of today's language technology is the lack of methods that get at the real contents of text and speech, systems that come close to language understanding. Therefore the central and most ambitious problem in computational linguistics has always been the realization of ‘deep’ linguistic processing involving an accurate mapping between written and spoken utterances on the one side and useful semantic representations on the other.

Modern linguistics has been able to provide theories and formalisms for the specification of grammars that express this mapping in a declarative and transparent way. Computational linguistics has contributed elaborate platforms and tools for grammar development. A few large-scale grammars have been designed that exhibit sufficient accuracy and coverage for real application tasks. However, these encouraging developments were seriously hampered by a lack of methods for language analysis that fulfill the minimal requirements in efficiency, robustness, and specificity. This simply means that all systems working with these grammars have been too slow and too brittle for real applications. Furthermore, they have not been able to manage the vast ambiguity in natural language, i.e. they could not select among large numbers of linguistically correct analyses.

Yet the most serious problem has been time and space efficiency. If an NLP system cannot process everyday sentences in a reasonable amount of time on everyday hardware, it is not suited for most applications. Moreover, there was no chance to improve coverage and solve the issues of robustness and specificity if researchers had to wait for hours for a sentence to process. The efficiency problem was so bad that many promising lines of research such as the large-scale grammar development in a large EU-funded international project and many other academic and industrial projects ended without leaving any applicable results. The situation seemed hopeless since all laboriously achieved gains in efficiency were almost immediately offset by efficiency losses due to increases in coverage or sophistication of the grammars.

When Verbmobil, the largest project ever in speech and language technology, adopted deep linguistic processing on the basis of HPSG as one of the central methods for speech analysis in real time translation of spoken face-to-face dialogues, this decision faced considerable criticism from both inside and outside the consortium. Why should the slowest and most complex processing method be employed in a system that strives for real-time processing? The decision could only be maintained because in the hybrid Verbmobil architecture, deep processing was just one of several processing methods and could therefore always be preempted by an analysis from a faster processing module.

Interestingly, it was the immense pressure for efficiency in this speech application that caused two members of the consortium, the LT Lab at DFKI and CSLI at Stanford University, to join forces in developing completely new strategies for performance research in deep linguistic processing. The methodological basis of the effort was the systematic, sophisticated and very detailed measurement of all relevant performance data for each version of a parser. For each parser and each parsed sentence a database record was created that contained all data on numbers and sizes of successful and unsuccessful, complete and partial results, and on overall time and space consumption. Preconditions for comparable results were the use of the same grammars and the same test corpora. The test corpora had to be annotated by the correct results and linked to previous performance data. The novel engineering platform produced detailed diagnostic reports and complex multidimensional comparisons between alternative systems.

At this time the Stanford HPSG group had already independently initiated a collaboration among several North American research groups, focusing on the joint production of a wide-coverage generic resource grammar of English. The resulting grammar was called LinGO English Resource Grammar (ERG), and the name LinGO Laboratory has since been adapted to refer to the project group at Stanford and its affiliate partners in joint projects.

Later the Natural Language Processing Lab at Tokyo University joined the collaboration. A number of new methods were developed by the three sites and tested in many combinations. In the end, it was a synthesis of methods reached by a true scientific and technological cross-fertilization process that brought about the breakthrough in the fight for efficiency. The overall run-time efficiency gain accomplished was a factor of around 2000. Space consumption could be reduced by several orders of magnitude. These gains were complemented by developments in computer hardware. Sentences can now be analysed in fractions of the time needed for real-time speech applications. Larger texts can be analyzed in a few seconds. Thus HPSG parsing now meets the speed and working memory requirements for a wide range of applications.

This breakthrough led to increased interest in HPSG processing in several areas of research and in industry. More theoreticians and practitioners of grammar have shown an interest in using the software for grammar development. The first industrial applications are being developed. However, the efficiency problem has not been the only obstacle for the wide applicability of deep linguistic processing. The notorious lack of robustness of deep processing has not yet been overcome, nor has the specificity problem found a satisfactory solution.

The main partners of the successful cooperation decided to enter into a new phase of collaboration. Stanford, Saarbrücken and Tokyo have remained core partners of the collaboration. However, for this phase the partnership has been broadened by additional groups or centers that have joined the collaboration: Cambridge University, the University of Sussex, and the University of Trondheim.

The current research takes place in three areas

  • Area One: Robustness, Disambiguation and Specificity of HPSG Processing
  • Area Two: An application of HPSG Processing to Information Extraction
  • Area Three: Multilingual Grammar Engineering

DELPH-IN partners very actively pursue these areas in an ensemble of funded projects and many other forms of exchange and collaboration

Personal tools