Egocentric Action Recognition (EAR) aims to identify fine-grained actions and interacted objects from first-person videos, forming a core task in egocentric video understanding. Despite recent progress, EAR remains challenged by limited data scale, annotation quality, and long-tailed class distributions. To address these issues, we propose REAR, a Retrieval-augmented framework for EAR that leverages external third-person (exocentric) videos as auxiliary knowledge—without requiring synchronized ego-exo pairs. REAR adopts a dual-branch architecture: one branch extracts egocentric representations, while the other retrieves semantically relevant exocentric features. These are fused via a cross-view integration module that performs staged refinement and attention-based alignment. To mitigate class imbalance, a class-adaptive selector dynamically adjusts retrieval depth based on class frequency, and independent classifiers are trained with logit-adjusted cross-entropy. Extensive experiments across three benchmarks demonstrate that REAR achieves state-of-the-art performance, with significant gains in object recognition and tail-class accuracy. The source code is publicly available at https://github.com/zou-y23/REAR.

Abstract Data fusion has demonstrated its effectiveness in enhancing information retrieval across various studies. However, advanced fusion methods typically require a dataset with extensive relevance judgments to train optimal model weights, necessitating labor-intensive and costly manual efforts. This study explores efficient methods for generating training data to facilitate affordable relevance judgments and improve fusion model quality. Experiments conducted on six datasets from TREC’s Precision Medicine and Deep Learning tracks reveal that with careful sampling design, near-optimal fusion weights can be achieved using only 5% of the documents compared to the full TREC judgments. This translates to a dataset comprising 20 queries and 500 relevance-judged documents in total. The findings highlight the potential for sophisticated fusion techniques to become more accessible to researchers and practitioners, delivering substantial performance improvements with minimal judgment effort and cost.

Text matching aims to establish the matching relationship between two texts. It is an important operation in some information retrieval related tasks such as question duplicate detection, question answering, and dialog systems. Bidirectional long short term memory (BiLSTM) coupled with attention mechanism has achieved state-of-the-art performance in text matching. A major limitation of existing works is that only high level contextualized word representations are utilized to obtain word level matching results without considering other levels of word representations, thus resulting in incorrect matching decisions for cases where two words with different meanings are very close in high level contextualized word representation space. Therefore, instead of making decisions utilizing single level word representations, a multi-level matching network (MMN) is proposed in this paper for text matching, which utilizes multiple levels of word representations to obtain multiple word level matching results for final text level matching decision. Experimental results on two widely used benchmarks, SNLI and Scaitail, show that the proposed MMN achieves the state-of-the-art performance.

Query Performance prediction aims to evaluate the effectiveness of the results returned by a search system in response to a query without any relevance information. In this paper, we propose a method that considers both magnitude and variance of scores of the ranked list of results to measure the performance of a query. Using six different TREC test sets, we compare our predictor with three of the state-of-the-art techniques. The experimental results show that our method is very competitive. Pairwise comparisons with each of the three other methods show that our predictor performs better in more data sets.

In recent years, researchers have investigated search result diversification through a variety of approaches. In such situations, information retrieval systems need to consider both aspects of relevance and diversity for those retrieved documents. On the other hand, previous research has demonstrated that data fusion is useful for improving performance when we are only concerned with relevance. However, it is not clear if it helps when both relevance and diversity are both taken into consideration. In this short paper, we propose a few data fusion methods to try to improve performance when both relevance and diversity are concerned. Experiments are carried out with 3 groups of top-ranked results submitted to the TREC web diversity task. We find that data fusion is still a useful approach to performance improvement for diversity as for relevance previously.

Data fusion is currently used extensively in information retrieval for various tasks. It has proved to be a useful technology because it is able to improve retrieval performance frequently. However, in almost all prior research in data fusion, static search environments have been used, and dynamic search environments have generally not been considered. In this article, we investigate adaptive data fusion methods that can change their behavior when the search environment changes. Three adaptive data fusion methods are proposed and investigated. To test these proposed methods properly, we generate a benchmark from a historic T ext RE trieval Conference data set. Experiments with the benchmark show that 2 of the proposed methods are good and may potentially be used in practice.

One of the most popular techniques of generating classifier ensembles is known as stacking which is based on a meta-learning approach. In this paper, we introduce an alternative method to stacking which is based on cluster analysis. Similar to stacking, instances from a validation set are initially classified by all base classifiers. The output of each classifier is subsequently considered as a new attribute of the instance. Following this, a validation set is divided into clusters according to the new attributes and a small subset of the original attributes of the instances. For each cluster, we find its centroid and calculate its class label. The collection of centroids is considered as a meta-classifier. Experimental results show that the new method outperformed all benchmark methods, namely Majority Voting, Stacking J48, Stacking LR, AdaBoost J48, and Random Forest, in 12 out of 22 data sets. The proposed method has two advantageous properties: it is very robust to relatively small training sets and it can be applied in semi-supervised learning problems. We provide a theoretical investigation regarding the proposed method. This demonstrates that for the method to be successful, the base classifiers applied in the ensemble should have greater than 50% accuracy levels.

Data fusion in information retrieval has been investigated by many researchers and quite a few data fusion methods have been proposed, but why data fusion can bring improvement in effectiveness is still not very clear. In this paper, we use a geometric probabilistic framework to formally describe data fusion, in which each component result returned from an information retrieval system for a given query is represented as a point in a multiple dimensional space. Then all the component results and data fusion results can be explained using geometrical principles. In such a framework, it becomes clear why quite often data fusion can bring improvement in effectiveness and accordingly what the favourable conditions are for data fusion algorithms to achieve better results. The framework can be used as a guideline to make data fusion techniques be used more effectively.

Abstract The aim of this research is twofold. On the one hand, high accuracy retrieval has been a concern of the information retrieval community for some time. We aim to investigate this issue via data fusion. On the other hand, the correlation among component results has been proven harmful to data fusion, but it has not been taken into account in data fusion algorithms. In the hope of achieving better performance, we propose a group of algorithms to eliminate the effect of uneven correlation among component results by assigning different weights to all component results or their combinations. Then the linear combination method or a variation is used for fusion. Extensive experimentation is carried out to evaluate the performances of these algorithms with six groups of component results, which are the top 10 systems submitted to Text REtrieval Conference (TREC) 6, 7, 8, 9, 2001, and 2002. The experimental results show that all eight data fusion methods involved outperform the best component system on average. Therefore, we demonstrate that the data fusion technique in general is effective with accurate retrieval results. The experimental results also demonstrate that all six methods presented in this article are effective for eliminating the effect of uneven correlation among component results. All of them outperform CombSum and five of them outperform CombMNZ on average.

No abstract available.

This paper proposes an adptive approach for data fusion of information retrieval systems, which exploits estimated performances of all component input systems without relevance judgement or training. The estimation is conducted prior to the fusion but uses the same data as fusion applies. The experiment shows that our algorithms are competitive with, and often outperform CombMNZ, one of the most effective algorithms in use.