We formulate the problem of nonprojective dependency parsing as a polynomial-sized integer linear program. Our formulation is able to handle non-local output features in an efficient manner; not only is it compatible with prior knowledge encoded as hard constraints, it can also learn soft constraints from data. In particular, our model is able to learn correlations among neighboring arcs (siblings and grandparents), word valency, and tendencies toward nearlyprojective parses. The model parameters are learned in a max-margin framework by employing a linear programming relaxation. We evaluate the performance of our parser on data in several natural languages, achieving improvements over existing state-of-the-art methods. 1

This paper describes our contribution to the semantic role labeling task (SRL-only) of the CoNLL-2009 shared task in the closed challenge (Hajič et al., 2009). Our system consists of a pipeline of independent, local classifiers that identify the predicate sense, the arguments of the predicates, and the argument labels. Using these local models, we carried out a beam search to generate a pool of candidates. We then reranked the candidates using a joint learning approach that combines the local models and proposition features. To address the multilingual nature of the data, we implemented a feature selection procedure that systematically explored the feature space, yielding significant gains over a standard set of features. Our system achieved the second best semantic score overall with an average labeled semantic F1 of 80.31. It obtained the best F1 score on the Chinese and German data and the second best one on English. 1

Thesis for a diploma in computer science, 30 ECTS credits

While the overwhelming majority of information extraction efforts in the biomedical domain have focused on the extraction of simple binary interactions between named entity pairs, some recently published corpora provide complex, nested and typed event annotations that aim to accurately capture the diversity of biological relationships. We present the first machine learning approach for extracting such relationships, utilizing both a graph kernel and a novel, task-specific feature set. We show that relationships can be predicted with 77 % F-score, or 83 % if their type and direction is disregarded. Using both gold standard and generated parses, we determine the impact of parsing on extraction performance. Finally, we convert our predicted complex relationships to binary interactions, recovering binary annotation with 62 % F-score, relating the new method to the large body of work available on binary interactions. 1

We formulate the problem of nonprojective dependency parsing as a polynomial-sized integer linear program. Our formulation is able to handle non-local output features in an efficient manner; not only is it compatible with prior knowledge encoded as hard constraints, it can also learn soft constraints from data. In particular, our model is able to learn correlations among neighboring arcs (siblings and grandparents), word valency, and tendencies toward nearlyprojective parses. The model parameters are learned in a max-margin framework by employing a linear programming relaxation. We evaluate the performance of our parser on data in several natural languages, achieving improvements over existing state-of-the-art methods. 1

LTAG-spinal is a novel variant of traditional

We present a unified view of two state-of-theart non-projective dependency parsers, both approximate: the loopy belief propagation parser of Smith and Eisner (2008) and the relaxed linear program of Martins et al. (2009). By representing the model assumptions with a factor graph, we shed light on the optimization problems tackled in each method. We also propose a new aggressive online algorithm to learn the model parameters, which makes use of the underlying variational representation. The algorithm does not require a learning rate parameter and provides a single framework for a wide family of convex loss functions, including CRFs and structured SVMs. Experiments show state-of-the-art performance for 14 languages. 1

Informática. N. S. was supported in part by Qatar NRF NPRP-08-485-1-083. E. X. was supported by AFOSR

M. F. and P. A. were supported by the FET programme (EU

Training structured predictors often requires a considerable time selecting features or tweaking the kernel. Multiple kernel learning (MKL) sidesteps this issue by embedding the kernel learning into the training procedure. Despite the recent progress towards efficiency and scalability of MKL algorithms, the structured output case remains an open research front. We propose a new family of online proximal algorithms able to tackle many variants of MKL and group-LASSO, and for which we show regret, convergence, and generalization bounds. Experiments on handwriting recognition and dependency parsing illustrate the success of the approach. 1