Section: New Results
Natural Language Processing
In , we introduce a simple semi-supervised approach to improve implicit discourse relation identification. This approach harnesses large amounts of automatically extracted discourse connectives along with their arguments to construct new distributional word representations. Specifically, we represent words in the space of discourse connectives as a way to directly encode their rhetorical function. Experiments on the Penn Discourse Treebank demonstrate the effectiveness of these task-tailored representations in predicting implicit discourse relations. Our results indeed show that, despite their simplicity, these connective-based representations outperform various off-the-shelf word embeddings, and achieve state-of-the-art performance on this problem.
Along the PhD thesis of Thibault Liétard , we are working on learning a similarity between text entities for the task of coreference resolution. Unlike indirect classification criteria often used in the literature, the similarity function naturally operates on pairs of mentions and several relevant objectives can be considered. For instance, we can learn the parameters of the similarity function such that the similarity of a given mention to its closest antecedent coreferent mention is larger than to any closer non-coreferential antecedent candidate. The resulting similarity scores can then be plugged into a greedy clustering procedure, or used to build a weighted graph of mentions to be clustered by spectral algorithms. For the representations of (pairs of) mentions on which the similarity function is learned, we consider both traditional linguistic features as well as external information about the general context of occurrence of the mentions using word embeddings.
Along the PhD thesis of Mathieu Dehouck , we study the problem of cross-lingual dependency parsing, aiming at leveraging training data from different source languages to learn a parser in a target language. Specifically, this approach first constructs word vector representations that exploit structural (i.e., dependency-based) contexts but only considering the morpho-syntactic information associated with each word and its contexts. These delexicalized word embeddings, which can be trained on any set of languages and capture features shared accross languages are then used in combination with standard language-specific features to train a lexicalized parser in the target language. We evaluate our approach through experiments on a set of eight different languages that are part the Universal Dependencies Project. Our main results show that using such embeddings (monolingual or multilingual) achieves significant improvements over monolingual baselines. The work is submitted.