PhD Oral Exam - Majid Laali, Computer Science
When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.
Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.
An important aspect of natural language understanding and generation involves the recognition and processing of discourse relations. Building applications such as text summarization, question answering and natural language generation needs human language technology beyond the level of the sentence. To address this need, large scale discourse annotated corpora such as the Penn Discourse Treebank (PDTB; Prasad et al., 2008a) have been developed.
Manually constructing discourse resources (e.g. discourse annotated corpora) is expensive, both in terms of time and expertise. As a consequence, such resources are only available for a few languages. In this thesis, we propose an approach that automatically creates two types of discourse resources from parallel texts: 1) PDTB-style discourse annotated corpora and 2) lexicons of discourse connectives. Our approach is based on annotation projection where linguistic annotations are projected from a source language to a target language in parallel texts.
Our work has made several theoretical contributions as well as practical contributions to the field of discourse analysis. From a theoretical perspective, we have proposed a method to refine the naive method of discourse annotation projection by filtering annotations that are not supported by parallel texts. Our approach is based on the intersection between statistical word-alignment models and can automatically identify 65% of unsupported projected annotations. We have also proposed a novel approach for annotation projection that is independent of statistical word-alignment models. This approach is more robust to longer discourse connectives than approaches based on statistical word-alignment models.
From a practical perspective, we have automatically created the Europarl ConcoDisco corpora from English-French parallel texts of the Europarl corpus (Koehn, 2009). In the Europarl ConcoDisco corpora, around 1 million occurrences of French discourse connectives are automatically aligned to their translation. From the French side of Europarl ConcoDisco, we have extracted our first significant resource, the FrConcoDisco corpora. To our knowledge, the FrConcoDisco corpora are the first PDTB-style discourse annotated corpora for French where French discourse connectives are annotated with the discourse relations that they signalled. The FrConcoDisco corpora are significant in size as they contain more than 25 times more annotations than the PDTB. To evaluate the FrConcoDisco corpora, we showed how they can be used to train a classifier for the disambiguation of French discourse connectives with a high performance. The second significant resource that we automatically extracted from parallel texts is ConcoLeDisCo. ConcoLeDisCo is a lexicon of French discourse connectives mapped to PDTB discourse relations. While ConcoLeDisCo is useful by itself, as we showed in this thesis, it can be used to improve the coverage of manually constructed lexicons of discourse connectives such as LEXCONN (Roze et al., 2012).