Resumen de New approaches to interactive multimedia content retrieval from different sources
Julián Moreno Schneider
Interactive Multimodal Information Retrieval systems (IMIR) increase the capabilities of traditional search systems with the ability to retrieve information in different types (modes) and from different sources. The increase in online content while diversifying means of access to information (phones, tablets, smart watches) encourages the growing need for this type of system. In this thesis a formal model for describing interactive multimodal information retrieval systems querying various information retrieval engines has been defined. This model includes formal and widespread definition of each component of an IMIR system, namely: multimodal information organized in collections, multimodal query, different retrieval engines, a source management system (handler), a results management module (fusion) and user interactions. This model has been validated in two stages. The first, in a use case focused on information retrieval on sports. A prototype that implements a subset of the features of the model has been developed: a multimodal collection that is semantically related, three types of multimodal queries (text, audio and text + image), six different retrieval engines (question answering, full-text search, search based on ontologies, OCR in image, object detection in image and audio transcription), a strategy for source selection based on rules defined by experts, a strategy of combining results and recording of user interactions. NDCG (normalized discounted cumulative gain) has been used for comparing the results obtained for each retrieval engine. These results are: 10,1% (Question answering), 80% (full text search) and 26;8% (ontology search). These results are on the order of works of the state of art considering forums like CLEF. When the retrieval engine combination is used, the information retrieval performance increases by a percentage gain of 771,4% with question answering, 7,2% with full text search and 145,5% with Ontology search. The second scenario is focused on a prototype retrieving information from social media in the health domain. A prototype has been developed which is based on the proposed model and integrates health domain social media user-generated information, knowledge bases, query, retrieval engines, sources selection module, results' combination module and GUI. In addition, the documents included in the retrieval system have been previously processed by a process that extracts semantic information in health domain. In addition, several adaptation techniques applied to the retrieval functionality of an IMIR system have been defined by analyzing past interactions using decision trees, neural networks and clusters. After modifying the sources selection strategy (handler), the system has been reevaluated using classification techniques. The same queries and relevance judgments done by users in the sports domain prototype will be used for this evaluation. This evaluation compares the normalized discounted cumulative gain (NDCG) measure obtained with two different approaches: the multimodal system using predefined rules and the same multimodal system once the functionality is adapted by past user interactions. The NDCG has shown an improvement between -2,92% and 2,81% depending on the approaches used. We have considered three features to classify the approaches: (i) the classification algorithm; (ii) the query features; and (iii) the scores for computing the orders of retrieval engines. The best result is obtained using probabilities-based classification algorithm, the retrieval engines ranking generated with Averaged-Position score and the mode, type, length and entities of the query. Its NDCG value is 81,54%.
