As a computational scientist, I have a strong commitment to the practical implementation of theoretical results and findings in concrete NLP applications.
As such, the Mental Salience Framework (MSF), that I've developed in my PhD thesis, is a framework for the development and application of salience metrics in Natural Language Generation: If the grammar provides alternative realization candidates for a given piece of information, the MSF defines how salience metrics can be applied to predict the contextually contextually most adequate realization. I've practially implemented such salience-based preferences for the choice of referring expressions, the assignment of grammatical roles and word order preferences in the sentence planning module for the PoliBox system. Together with Manfred Stede, I've also built a text planning module for PoliBox where the discourse structure (and hence, the hierarchical organization and the choice of discourse connectives) were derived from primitive pieces of information. Subsequent research on discourse phenomena was chiefly concerned with the evaluation of design decisions made in these systems.
In the context of Natural Language Understanding, the MSF can be employed to the spotting of `relevant' or `important' expressions: It specifies a distinction between two dimensions of salience and formalizes their interaction. If one dimension, say, backward-looking salience, is known, and deviations from packaging preferences according to this dimension of salience are encountered, then the other dimension (i.e., forward-looking salience, or importance) must be involved. The MSF also has a potential application in Machine Translation: Information packaging preferences differ across the languages of the world, and the salience-based approach can be applied in the pre- and postediting steps of MT, e.g., in order to derive contextual appropriate word order. Both applications, however, have not yet been systematically explored.
In a number of papers, the OLiA ontologies, originally designed to allow users to explore corpora with multiple layers of heterogeneous, unknown or complex annotations, have also been shown to be practically applicable for various NLP tasks. The ontologies allow to abstract from string-based representations of linguistic annotations, and to operate on conceptual representations of linguistic annotations: In ensemble combination architectures, the OLiA ontologies make it possible to combine linguistic analyses produced by tools with different tagsets; in NLP pipeline systems, e.g., UiMA, the OLiA ontologies can be employed for the tool-independent specification of interfaces between different modules; further, the OLiA ontologies provide a natural integration of NLP analyses with Semantic Web techniques. They contribute to the portability and scalability of linguistic annotations and the NLP tools that they produce.