The LILOG Project (1985–1991) by IBM Germany

Throughout the entire duration of the project, the results of the LILOG project were published without any restrictions in a technical report series, at numerous national and international conferences, as well as in journals and books. In addition to the scientific output documented, among others, in the final report [ 4 ], the success of a project like LILOG can also be measured by its broader impact: Significant emphasis was placed during the project on ensuring that participating researchers and students could attain advanced academic qualifications within the scope of the project. By the end, approximately 65 diploma theses, 26 doctoral dissertations, and 3 postdoctoral theses (Habilitationen) had been completed. Moreover, ten members of the IBM LILOG group were appointed to professorships during the project, with another twelve receiving offers after its conclusion. LILOG members taught at numerous universities, and around 40 monographs, collected volumes, and essays were published annually and made accessible to the scientific community through the IKBS Technical Report Series. In addition to internal LILOG workshops, two public IBM symposia were organized by LILOG. Moreover, LILOG members contributed to the organization of numerous national and international conferences and summer schools. A further outcome was the transfer of scientific results into IBM’s development divisions. LILOG research contributed to IBM products such as full-text search (from 1993), data and knowledge mining (from 1993), Knowledge-Based System Security Manager (from 1989), Prolog-based production planning (from 1989), and an environmental impact assessment system for the Expo 2000 development in Hanover (from 1992). These contributions significantly enhanced the functionality of the respective products. LILOG proved to be scientifically more successful than initially anticipated. It not only enabled a large volume of interdisciplinary research—reflected in conference papers, monographs, theses, and dissertations—but also brought together a generation of scholars focused on language understanding. This contributed to a revitalization of computational linguistics in Germany. Following LILOG, several German universities (e.g., Stuttgart, Tübingen, Heidelberg, Erlangen-Nürnberg, Osnabrück) began offering Bachelor’s and Master’s programs in computational linguistics. These programs benefited from the resources, infrastructure, and personnel developed during the LILOG project. They adopted an interdisciplinary focus (combining computer science, linguistics, psychology, and cognitive science) and gradually evolved toward data science and natural language processing (NLP). It should also be mentioned that the LILOG project revealed certain limitations of symbolic natural language and knowledge processing: A logic-based grammar approach cannot be fully comprehensive due to the inherent variability of language, and defining grammars for broad analysis and generation is highly resource-intensive. A logic-based grammar approach cannot be fully comprehensive due to the inherent variability of language, and defining grammars for broad analysis and generation is highly resource-intensive. Background knowledge, context, pragmatics, and application-specific information are difficult—if not impossible—to be formalized completely at the symbolic level for general large-scale text understanding. Despite their considerable scope, the implemented knowledge bases only covered narrowly defined domains together with the background knowledge necessary for those domains. Therefore, the semantic analysis of natural language in general was shown to be limited by incomplete contextual and pragmatic knowledge. Background knowledge, context, pragmatics, and application-specific information are difficult—if not impossible—to be formalized completely at the symbolic level for general large-scale text understanding. Despite their considerable scope, the implemented knowledge bases only covered narrowly defined domains together with the background knowledge necessary for those domains. Therefore, the semantic analysis of natural language in general was shown to be limited by incomplete contextual and pragmatic knowledge. The inference engines’ runtime complexity was very high relative to the available computing power at the time. The inference engines’ runtime complexity was very high relative to the available computing power at the time. However, the final report [ 4 ] demonstrates that the project made substantial progress toward developing a theory of text understanding. Much of this theoretical work could also be implemented in the operational system LEU/2 proving that the concepts developed in the project carried a long way towards the understanding of “text understanding”.