A semantic indexing approach of multimedia documents content based partial transcription
Issam Bendib, Mohamed Ridda Laouar · 2018
The complexity of searching and indexing spoken document retrieval depends closely on content structure and access strategies. However, Spoken Document Retrieval (SDR) approaches based on the transcription of content by LVCSR systems must treat the impact of recognition errors on the performance of the information retrieval system. Currently, several multimedia resources (spoken document) are unexploited, because the errors generated in the transcription process, decrease the performance of the systems of searches for spoken documents. Our contribution in this paper is the proposition a novel approach semantic based indexing content of multimedia documents by using the results of a partial automatic transcription to ride out research and indexing problems on these resources. In this context, the main hypothesis is developed around the following question: is it possible to identify and retrieve spoken document by using a semantic content indexing system based on his partial transcription? Also, does the partial content transcription of a spoken document is sufficient for his indexing? First, we are interested to look for a linguistic syntactic representation of the whole of the content spoken document. Then, it will be used to define the most discriminating indexing terms for his content. Nevertheless, in our approach, we have an assumption that the use of segments of a spoken document with an efficient semantic enrichment process instead of the whole spoken document content is sufficient. Also, with this strategy, we can resolve OOV problems, recognition errors and technical terms. Finally, in validation and experimentation phase, we tested the different modules proposed on the TED-LIUM corpora. The results obtained are interesting and encourage us to lead off our future perspectives.