Search

We present a method for automatically converting the Dutch Lassy Small treebank, a phrasal dependency treebank, to UD. All of the information required to produce accurate UD annotation appears to be available in the underlying annotation. However, we also note that the close connection between POS-tags and dependency labels that is present in UD is missing in the Lassy treebanks. As a consequence, annotation decisions in the Dutch data for such phenomena as nominalization and clausal complements of prepositions seem to differ to some extent from comparable data in English and German. Because the conversion is automatic, we can now also compare three state-of-theart dependency parsers trained on UD Lassy Small with Alpino, a hybrid Dutch parser which produces output that is compatible with the original Lassy annotations.

The paper introduces a methodological innovation as well as a practical innovation. Firstly, two scenarios are introduced to compare accurate, but slow parsers on the one hand, with faster, but less accurate parsers on the other hand. Secondly, a corpus-based technique is described to improve the efficiency of wide-coverage high-accuracy parsers. By keeping track of the derivation steps which lead to the best parse for a very large collection of sentences, the parser learns which parse steps can be filtered without significant loss in parsing accuracy, but with an important increase in parsing efficiency. Experimental results with the Alpino parser for Dutch indicate that the technique yields much faster parsers that perform with almost the same level of accuracy. An interesting characteristic of our approach is that it is self-learning, in the sense that it uses unannotated corpora.

We propose a simple, yet effective, Word Sense Disambiguation method that uses a combination of a lexical knowledge-base and embeddings. Similar to the classic Lesk algorithm, it exploits the idea that overlap between the context of a word and the definition of its senses provides information on its meaning. Instead of counting the number of words that overlap, we use embeddings to compute the similarity between the gloss of a sense and the context. Evaluation on both Dutch and English datasets shows that our method outperforms other Lesk methods and improves upon a state-of-theart knowledge-based system. Additional experiments confirm the effect of the use of glosses and indicate that our approach works well in different domains.

Most state-of-the-art parsers take an approach to produce an analysis for any input despite errors. However, small grammatical mistakes in a sentence often cause parser to fail to build a correct syntactic tree. Applications that can identify and correct mistakes during parsing are particularly interesting for processing user-generated noisy content. Such systems potentially could take advantage of linguistic depth of broad-coverage precision grammars. In order to choose the best correction for an utterance, probabilities of parse trees of different sentences should be comparable which is not supported by discriminative methods underlying parsing software for processing deep grammars. In the present work we assess the treelet model for determining generative probabilities for HPSG parsing with error correction. In the first experiment the treelet model is applied to the parse selection task and shows superior exact match accuracy than the baseline and PCFG. In the second experiment it is tested for the ability to score the parse tree of the correct sentence higher than the constituency tree of the original version of the sentence containing grammatical error.