Where Understanding Meets Learning: Getting at Some Meaning in Texts

Nevertheless, considerable progress has been achieved in two strands of research that try to approximate human language processing in computational models in rather different ways. The older one of the two lines of research is dedicated to intellectually encoding the implicit linguistic knowledge a human language user needs for producing and understanding utterances. The languages for encoding are logic-based grammar formalisms. The younger research paradigm attempts to develop machine learning techniques for the automatic acquisition of the needed knowledge. Usually, the results of such learning experiments are statistical language models that do neither reflect the linguistic insights into language nor do they comply with the psychological findings on human sentence processing. As opposed to large-scale hand-crafted grammars, such statistical language models developed for one application are rarely usable for any other application. However, there are also learning approaches that result in rule sets that are transparent and reusable. These learning methods can be utilized for the extraction of some selected relevant types of information from large volumes of texts.
In my presentation I will present examples for progress in both linguistic methods and machine learning techniques from our own research. Recent progress in deep linguistic processing has been achieved through augmenting intellectual grammar development by data-intensive automatic learning methods. I will then demonstrate that the two major paradigms can be combined in fruitful and insightful ways for the difficult task of extracting complex semantic relations from texts. Finally, I will argue that a very demanding but highly promising research program is dedicated to the combination of the most ambitious grammar-based approaches with pragmatic but nevertheless scientifically systematic learning methods.