Cross-modal retrieval, particularly between image and text modalities, is a fundamental task in multimodal learning that seeks to align semantic representations across heterogeneous sources. Existing methods struggle at scale due to excessive parameterization, high-dimensional outputs, and reliance on recall-based metrics that inadequately capture ranking relevance. We propose a cross-modal distillation framework that combines model compression with a ranking-based evaluation metric, Teacher-weighted NDCG. By distilling a 12-layer BEiT3 teacher into an 8-layer student, the framework reduces model size by 25.79 % and computation by 32.95 %, while retaining 94.83 % of the teacher’s Recall@ k and 98.94 % of its Teacher-weighted NDCG@ k . Trained on CC3M and SBUCaption and fine-tuned on Localized Narratives, COCO, and Flickr30K, the student shows strong generalization. Across five benchmarks, including the large-scale Open Images-LN126K, it outperforms lightweight baselines such as TinyCLIP and CLIP-ViT-B/32, delivering superior ranking quality in a compact architecture well suited for real-world retrieval applications.

Twitter is a popular social media platform, where millions of tweets are being generated every day. On Twitter, users can tweet about certain topic during occurrence of events. This results in trending topics, such as #MH370, #MH17, #South Korea ferry etc. When viewing trending topic, not all the tweets are relevant to one self. Therefore, it is important to classify these tweets based on individual preference for better information retrieval. To address this problem, this paper focuses on automated personalization of tweets for popular trending topics. The main objective is to classify the tweets information as “Like” or “Dislike” on a particular topic depending on personal preference by feature selection. For enhancement, topic-related keywords are selected as features for representing a category model from user preferred tweets and other sources like news. Finally, the result of experiment shows promising outcome that a training based on low number of tweets can give quick personalization on next incoming tweets.

The flexibility of XML allows document to be annotated easily. However, these annotations come from different sources like Wordnet thesaurus, POS, DTD, semantic roles etc. These annotations can either be combined in the same document or captured separately in different document. The former, though richer in annotations, may look messy and requires more parsing time. The latter needs control of document consistency. This paper proposes a string-trees model to represent XML document for multiple sources of annotations. This model extends the existing string-tree structure for linguistic content in order to support structured contents of XML document. In this paper, we describe how this model is refined and applied on XML document.

Looking at the amount of structured contents available on the web in recent years, we can be certain that the needs of structured retrieval systems are getting more prominent. In order to access the structured contents, information requests are formulated in the form of query languages supported by the structured retrieval systems. Nevertheless, information requests at this layer cannot be specified directly by end users. Hence, a common solution is to map information needs obtained from the application layer to the query language of a retrieval system automatically. Although rule-based approach can be used to generate query automatically, it is designed for a single target query language, whereas, there are not one but different types of structured queries with different forms of syntax used in a structured retrieval environment.

No abstract available.

No abstract available.

The search that involves structured Web resources like XML data, services is still lagging of its own method and relying on contemporary search systems. This paper presents a method that learns semantics from structured information of these resources. Instead of committing the semantic meaning of resources to strict and formal vocabularies like ontology or data dictionary, we are interested to interpret the meaning based on the natural context of the resources. The semantics are used in search process, i.e. query reasoning and resource selection, to provide better answer in terms of context relevancy and clearer result description

Resembling the desktop of personal computer in which files and applications are stored, MICE3desktop is a Web-based information desktop where individual can keep and manage resources such as links to Web sites and services, as well as Web sources and documents. Resources on the desktop are represented in Web directory structure and annotated with information such as concepts, entities, etc. Categories are used to organize resources in the directory. Each category is annotated with concept learnt from the contents of the category. MICE3desktop allows users to exploit information and knowledge from their resources. Users can retrieve information by navigating the directory structure, look up the resource index or searching the resource contents. The desktop also allows users to effectively manage their resources by enabling them to enhance categories with concept. The concept is further used to assist user in organizing resources into their directory

Categories are used to organize information and knowledge in directory system, folder etc. As the amount of information increase and the types of information diversify, it is common to have more categories created. As the number of categories increases, it becomes more difficult to organize, manage and look up information from existing categories. In this paper, categories are annotated with concept features to facilitate the access, retrieval and sharing of information in the categories. We have observed that training texts is crucial in learning the concept of a category and serves as a good measure to help human to construct the category model. Hence, we present a study on training texts selection and evaluate the effectiveness of training texts, as well as its capability to complement human's knowledge in constructing the category model. Experimental evaluation shows that using training texts approach in category model construction gives promising results in both effectiveness and complement measures

Having broad coverage of search results returned by various search sources, combining and organizing these results in a meaningful way has become a common issue in the field of information retrieval. In this demo paper, we describe our meta search system, MICE, that is able to aggregate and classify search results based on user-customized categories. Categories help user to focus on search results, with respect to the categories concept customized by the user.