State-of-the-art semantic role labelling systems require large annotated corpora to achieve full performance. Unfortunately, such corpora are expensive to produce and often do not generalize well across domains. Even in domain, errors are often made where syntactic information does not provide sufficient cues. In this paper, we mitigate both of these problems by employing distributional word representations gathered from unlabelled data. While straight-forward word representations of predicates and arguments improve performance, we show that further gains are achieved by composing representations that model the interaction between predicate and argument, and capture full argument spans.

No abstract

We present a new approach for unsupervised semantic role labeling that leverages distributed representations. We induce embeddings to represent a predicate, its arguments and their complex interdependence. Argument embeddings are learned from surrounding contexts involving the predicate and neighboring arguments, while predicate embeddings are learned from argument contexts. The induced representations are clustered into roles using a linear programming formulation of hierarchical clustering, where we can model task-specific knowledge. Experiments show improved performance over previous unsupervised semantic role labeling approaches and other distributed word representation models.

Linear support vector machine training can be represented as a large quadratic program. We present an efficient and numerically stable algorithm for this problem using interior point methods, which requires only O(n) operations per iteration. Through exploiting the separability of the Hessian, we provide a unified approach, from an optimization perspective, to 1-norm classification, 2-norm classification, universum classification, ordinal regression and -insensitive regression. Our approach has the added advantage of obtaining the hyperplane weights and bias directly from the solver. Numerical experiments indicate that, in contrast to existing methods, the algorithm is largely unaffected by noisy data, and they show training times for our implementation are consistent and highly competitive. We discuss the effect of using multiple correctors, and monitoring the angle of the normal to the hyperplane to determine termination.