Research interests
Acquisition and Representation of Word and Phrase Meaning
One of my central interests is the search for representations of meaning that can be acquired from corpora (or at least, language users without a linguistics degree) but are still able to account for phenomena of language use and understanding. I've started out by looking at vector spaces, which represent word meaning simply by recording frequent co-occcurrences with other words and are naturally graded, an interesting property with respect to semantic phenomena like vagueness (EP07). I developed a framework for building vector spaces from dependency graphs. which beat surface-based vector spaces on tasks such as synonymy detection and prediction of priming effects (PL03, PL07). The next step was the use of dependency-based semantic space models for the representation of selectional preferences. The resulting models can rival the performance of deep semantic models in predicting human plausibility ratings (PPE07).
The chief challenge in this line of research is to construct models that are able to model the meaning not only of individual words, but of whole predicate-argument combinations -- or, of words in context, which turns out to be a closely related question. I have proposed a model for word meaning in (local) context which works by combining the word's context-independent lexical meaning with the expectations of its context for the position the word occupies (EP08, EP09).
Semantic Processing for NLP Applications and Textual Entailment
An interesting recent development is the paradigm of "Textual Entailment" which tries to cast the central semantic processing needs of NLP applications in terms of a unified task (PD10). In 2008/09, I've worked on the development of semantics-based methods for the evaluation of machine translation output. I was able to show that robustness is not a big issue, and that vanilla entailment features taken from a state-of-the-art Textual Entailment Recognition system (PMMRYM08) can predict human judgments of MT quality significantly better than surface-based methods. Our system came out among the top contenders in the WMT 2009 evaluation task (PGJM09). An extended version including other datasets and the application of textual entailment to MT training is here: PGCJM09.
Knowledge Induction from Multilingual Corpora
Another one of my topics of interest is the corpus-driven induction of crosslingual or multilingual knowledge. In my PhD thesis, I have taken a first step in this direction by investigating the cross-lingual projection of frame-semantic information (coarse-grained word sense for predicates plus semantic roles). Starting from the available English resource (i.e., FrameNet), I use annotation projection in parallel corpora to produce frame-semantic predicate classes and role-annotated corpora for new languages, e.g. using graph-based tree alignment for role projection (PL05a, PL05b). The models are supposed to be language-independent, but my evaluation has concentrated on English as the source and German (PL06) and French (PP07) as the target languages. For an extended version that subsumed much of the above work, see here: PL09.
An important methodological issue is how well the analysis of some linguistic expression can be used to describe its translation (a.k.a. the cross-lingual parallelism of the analysis). In practice, parallelism is never perfect. However, I have found that even when individual frames are imparallel, the translational equivalence can often be modelled using paraphrases based on frame groups (PE05). An aspect I find particularly interesting is the modification of lingusitic annotation in translation and ways to describe this process (P07). Stay tuned for a journal version (PE10).
Another more recent study that brings together different strands of my research is the cross-lingual investigation of selectional preferences. We have been able to show that very simple, syntax-less vector space models support the induction of selectional preferences for new languages (PP10), and that there is interesting food for thought regarding the usefulness of different cross-lingual lexical-semantic relations.
Corpus-based Semantic Lexicon Development
Even though the paradigm of inducing semantic lexical-knowledge from
corpora is our best shot at large-scale resource building, it has its
own share of problems. There may be fundamental disagreement on what
semantic dimensions to use for the description of meaning, be it with
respect to frameworks for semantic role annotation (EEKP06) or
general-purpose semantic verb classifications (CEPS08). Also, the
representation of knowledge across multiple layers of linguistic
analysis (e.g., syntax and semantics) requires answers to questions
about granularity, reliability, and integrated querying (BPSFH08).
Some other
work addresses more specifically issues that arise in the context of
semantic role annotation: Reliability (EKPP03, BEFKPP06), a graphical
user interface that support annotation (BEFKP06), and a
lightweight XML representation for syntactic and semantic annotation
(EP06).
Semantic Role Labelling
I've also done some work on semantic role labelling: less on its properties as a Machine Learning task, and more on its application and implicit linguistic generalizations. I have built a role labelling system, Shalmaneser, which is a freely available for research purposes (EP06, download page). On the linguistics side, I have worked on linked the performance of labelling models to the lexical-semantic properties of predicates, arguing for a combination of syntactic and lexical features (PB04, EP05). To combat the ever-present sparsity issues, I have developed a semi-supervised approach to SRL that uses verb data to label nouns (PPS08).