Multimodal Document Alignment : Feature-based Validation to Strengthen Thematic Links

In this paper, we present a validation approach of detected alignment links between dialog transcript and discussed documents, in the context of a multimodal document alignment framework of multimedia events (meetings and lectures). The validation approach consists in an entailment process of the detected alignment links. This entailment process exploits several features, from the structural level of aligned documents to the linguistic level of their tokens. The implemented entailment strategies were evaluated on several multimodal corpora. The obtained results prove that the choice of the relevant entailment strategy depends on the types of documents that are available in the corpus, on their content, and also on the nature of the corpus.