Retrieval Performance Research Articles

Fine-tuning 3D foundation models for geometric object retrieval

Foundation models, such as ULIP-2 (Xue et al., 2023) recently projected forward the field of 3D deep learning. These models are trained with significantly more data and show superior representation learning capacity in many downstream tasks like 3D shape classification and few-shot part segmentation.A particular characteristic of the recent 3D foundation models is that they are typically multi-modal, and involve image (2D) as well as caption (text) branches. This leads to an intricate interplay that benefits all modalities. At the same time, the nature of the 3D encoders alone, involved in these foundation models is not well-understood. Specifically, there is little analysis on the utility of both pre-trained 3D features provided by these models, or their capacity to adapt to new downstream 3D data. Furthermore, existing studies typically focus on label-oriented downstream tasks, such as shape classification, and ignore other critical applications, such as 3D content-based object retrieval.In this paper, we fill this gap and show, for the first time, how 3D foundation models can be leveraged for strong 3D-to-3D retrieval performance on seven different datasets, on par with state-of-the-art view-based architectures. We evaluate both the pre-trained foundation models, as well as their fine-tuned versions using downstream data. We compare supervised fine-tuning using classification labels against two self-supervised label-free fine-tuning methods. Importantly, we introduce and describe a methodology for fine-tuning, as we found this to be crucial to make transfer learning from 3D foundation models work in a stable manner.

Deep Attention Fusion Hashing (DAFH) Model for Medical Image Retrieval.

To address this, we propose a novel deep learning-based hashing model, the Deep Attention Fusion Hashing (DAFH) model, which integrates advanced attention mechanisms with medical imaging data. The DAFH model enhances retrieval performance by integrating multi-modality medical imaging data and employing attention mechanisms to optimize the feature extraction process. Utilizing multimodal medical image data from the Cancer Imaging Archive (TCIA), this study constructed and trained a deep hashing network that achieves high-precision classification of various cancer types. At hash code lengths of 16, 32, and 48 bits, the model respectively attained Mean Average Precision (MAP@10) values of 0.711, 0.754, and 0.762, highlighting the potential and advantage of the DAFH model in medical image retrieval. The DAFH model demonstrates significant improvements in the efficiency and accuracy of medical image retrieval, proving to be a valuable tool in clinical settings.

Building an Open-Vocabulary Video CLIP Model With Better Architectures, Optimization and Data.

Despite significant results achieved by Contrastive Language-Image Pretraining (CLIP) in zero-shot image recognition, limited effort has been made exploring its potential for zero-shot video recognition. This paper presents Open-VCLIP++, a simple yet effective framework that adapts CLIP to a strong zero-shot video classifier, capable of identifying novel actions and events during testing. Open-VCLIP++ minimally modifies CLIP to capture spatial-temporal relationships in videos, thereby creating a specialized video classifier while striving for generalization. We formally demonstrate that training Open-VCLIP++ is tantamount to continual learning with zero historical data. To address this problem, we introduce Interpolated Weight Optimization, a technique that leverages the advantages of weight interpolation during both training and testing. Furthermore, we build upon large language models to produce fine-grained video descriptions. These detailed descriptions are further aligned with video features, facilitating a better transfer of CLIP to the video domain. Our approach is evaluated on three widely used action recognition datasets, following a variety of zero-shot evaluation protocols. The results demonstrate that our method surpasses existing state-of-the-art techniques by significant margins. Specifically, we achieve zero-shot accuracy scores of 88.1%, 58.7%, and 81.2% on UCF, HMDB, and Kinetics-600 datasets respectively, outpacing the best-performing alternative methods by 8.5%, 8.2%, and 12.3%. We also evaluate our approach on the MSR-VTT video-text retrieval dataset, where it delivers competitive video-to-text and text-to-video retrieval performance, while utilizing substantially less fine-tuning data compared to other methods.

Exploring the Synergy of Web Usage Data and Content Mining for Personalized Effectiveness

In light of the exponential growth of web data and user volume, individuals are increasingly overwhelmed by information overload on the internet. Addressing this challenge, our study focuses on enhancing web information retrieval and presentation by leveraging web data mining techniques to uncover intrinsic relationships within textual, linkage, and usability data. Specifically, we aim to improve the performance of web information retrieval and presentation by analysing web data features. Our approach centres on web usage mining to identify usage patterns and integrate this knowledge with user profiles for personalized content delivery. Personalization, tailored to user’s characteristics and behaviours, serves to enhance engagement, conversion, and long-term commitment to websites. The objective of our research is to develop a web personalization system that enables users to access relevant website content without the need for explicit queries. This paper presents an extensive survey of various approaches proposed by researchers in the field of web personalization. It highlights the diverse methodologies and techniques employed to enhance user experience and engagement on the web. The paper identifies key challenges that require urgent attention to advance the field of web personalization.

Integrating Depth Measurements From Gaging Stations With Image Archives for Spectrally Based Remote Sensing of River Bathymetry

AbstractRemote sensing can be an effective tool for mapping river bathymetry, but the need for direct measurements to calibrate image‐derived depth estimates impedes broader application of this approach. One way to circumvent the need for field campaigns dedicated to calibration is to capitalize upon existing data. In this study, we introduce a framework for Bathymetric Mapping using Gage Records and Image Databases (BaMGRID). This workflow involves retrieving depth measurements made during gaging station site visits, downloading archived multispectral images, and then combining these two data sets to establish a relationship between depth and reflectance. We developed a processing chain that involves using application programming interfaces to obtain both depth measurements made during site visits and images centered on the gage and then linking depth to reflectance via an optimal band ratio analysis (OBRA) algorithm modified for small sample sizes. Applying this workflow to selected gages within two river basins indicated that depth retrieval from multispectral satellite images could be highly accurate, but with variable results from one image to the next at a given site. High resolution aerial photography was less conducive to bathymetric mapping in one of the basin considered. Of the four predictors of depth retrieval performance we evaluated (mean and standard deviation of depth, width, and an index of water clarity), only width was consistently significantly correlated with OBRA R2 (p < 0.026). Currently, BaMGRID is best‐suited for site‐by‐site analysis to support practical applications at the reach scale; continuous, basin‐wide mapping of river bathymetry will require additional research.

Reusable CS-Ca@PEI/CuMnO2 Hydrogel Beads for Peroxymonosulfate-Activated Degradation of Congo Red.

Metal oxides can activate peroxymonosulfate (PMS) for the catalytic degradation of organic dyes. However, achieving high catalytic efficiency, structural stability, ease of recovery, and recyclability remains challenging for both research and practical applications. To address these requirements, a bimetallic oxide, CuMnO2, was synthesized using a simple hydrothermal approach and was encapsulated to create hydrogel beads, CS-Ca@PEI/CuMnO2. Subsequently, CS-Ca@PEI/CuMnO2 was used to activate PMS and establish a solid-liquid heterogeneous oxidation system (CS-Ca@PEI/CuMnO2/PMS) for the degradation of Congo red (CR). The effects of various parameters such as different systems, catalyst dosages, initial pH values, PMS concentrations, temperatures, and anion types on the catalytic degradation properties of CS-Ca@PEI/CuMnO2 for CR were systematically evaluated. The results indicated that CS-Ca@PEI/CuMnO2 has exceptional degradation capacity, achieving 91.0% degradation of CR at pH 7. After three degradation cycles, the catalyst maintained an 86.9% degradation efficiency compared to its original performance, highlighting its robust structural stability. The presence of reactive radicals, specifically 1O2 and •O2-, were confirmed through quenching experiments, X-ray photoelectron spectroscopy (XPS), and electron paramagnetic resonance spectroscopy (EPR). Liquid chromatography-tandem mass spectrometry (LC-MS) revealed ten proposed intermediates in the catalytic degradation process. Due to its exceptional catalytic performance, structural durability, recyclability, and ease of retrieval, the catalyst shows great potential for effectively removing organic pollutants from industrial wastewater.

Vision-language joint representation learning for sketch less facial image retrieval

The traditional sketch-based facial image retrieval (SBFIR) framework assumes that a high-quality facial sketch has been prepared prior to the retrieval task. However, drawing such a sketch requires considerable skills and is time consuming, resulting in limited applicability. Sketch less facial image retrieval (SLFIR) framework aims to break these barriers through human–computer interaction during the sketching process. The primary challenges for the SLFIR problem can be noted that initial sketches (at early sketching) contain only local details and exhibit significant differences among users, resulting in poor performance at early stages and weak generalization abilities in practical testing. In this study, we developed a vision–language pretraining model to align the representation of facial images and their associated semantics. Based on this framework, we proposed a method for learning joint representations by fusing sketches with prior semantics, thereby enriching the information of initial sketches. Specifically, (1) we developed a series of well-designed operations to improve the quality of facial image–text pairs in the LAION-Face dataset; we trained a facial vision–language pretraining (FVLP) model to align the facial image and its semantics at the feature level. (2) subsequently, using FVLP as the backbone, we designed a convolutional attention module to fuse the multiscale features extracted from image encoder. This facilitated the learning of a multimodal representation crucial for the final retrieval process. In experiments, our proposed method achieved state-of-the-art performance at early stages on two public datasets; moreover, it exhibited good generalization capabilities during practical testing. Thus, our method significantly outperforms other baselines in terms of early retrieval performance. Codes are available at: https://github.com/ddw2AIGROUP2CQUPT/FVLP.

Two-stage zero-shot sparse hashing with missing labels for cross-modal retrieval

Recently, zero-shot cross-modal hashing has gained significant popularity due to its ability to effectively realize the retrieval of emerging concepts within multimedia data. Although the existing approaches have shown impressive results, the following limitations still need to be solved: (1) Labels in large-scale multimodal datasets in real scenes are usually incomplete or partially missing. (2) The existing methods ignore the influence of features-wise low-level similarity and label distribution on retrieval performance. (3) The representation ability of dense hash codes limits its discriminative potential. To solve these issues, we introduce an effective cross-modal retrieval framework called two-stage zero-shot sparse hashing with missing labels (TZSHML). Specifically, we learn a classifier through the partially known labeled samples to predict the labels of unlabeled data. Then, we use the reliable information in the correctly marked labels to recover the missing labels. The predicted and recovered labels are combined to obtain more accurate labels for the samples with missing labels. In addition, we employ sample-wise fine-grained similarity and cluster-wise similarity to learn hash codes. Therefore, TZSHML ensures that more samples with similar semantics are clustered together. Besides, we apply high-dimensional sparse hash codes to explore richer semantic information. Finally, the drift and interaction terms are introduced into the learning of the hash function to further narrow the gap between different modalities. Extensive experimental results demonstrate the competitiveness of our approach over other state-of-the-art methods in zero-shot retrieval scenarios with missing labels. The source code of this paper can be obtained from https://github.com/szq0816/TZSHML.

Differentiable modeling for soil moisture retrieval by unifying deep neural networks and water cloud model

Machine learning has been widely used in high-spatial-resolution surface soil moisture (SSM) retrieval studies, but in recent years, this purely data-driven retrieval method has been controversial due to its lack of physical interpretability and generalization ability. Physical retrieval models based on the theory of radiative transfer equations respect physical laws, but their retrieval accuracy is usually affected by many insufficient accurate inputs, the complex model structure, and parameter adjustment method. In order to explore the retrieval method of unifying these two types of models, in this study, a differentiable model (DM) was constructed to realize the soil moisture retrieval at 10 m resolution based on Sentinel data. The differentiable soil moisture retrieval model takes the water cloud model (WCM) as the skeleton, and united the WCM and neural networks by implementing differentiable programming of the WCM in a machine learning platform. The differentiability makes the retrieval model trained by the gradient descent method the same as the neural network, which allows the retrieval model to be physical while obtaining more accurate retrieval results. Luan River Basin, Shandian River Basin, Maqu, and Lake Tahoe study areas with various land cover types and climate types were selected for model evaluation, and the performances of DM were close to that of the random forest model with Pearson correlation coefficient (R) of 0.747, 0.853, 0.838 and 0.792 in four study areas, respectively. While in the assessment of extrapolation capability of retrieval models, the DM showed its strong generalization ability and retrieval performance that exceeded that of the other retrieval models, with R of 0.786, unbiased root mean square error (ubRMSE) of 5.523 vol% and bias of 0.054 vol%. The DM synthesizes the advantages of both physical and machine learning models while providing high-resolution SSM estimates with acceptable accuracy for the study areas. This study creates favorable conditions for the realization of large-scale soil moisture retrieval with high resolution and high accuracy, and provides new ideas for the combination of machine learning and physical knowledge in other retrieval studies.

Synergistic retrieval of mangrove vital functional traits using field hyperspectral and satellite data

Mangroves play a pivotal role in the attainment of United Nations Sustainable Development Goals. Leaf equivalent water thickness (EWT) and leaf mass per area (LMA) are vital functional trait parameters for monitoring heath status of mangroves. However, their inherent spectral response characteristics remain unclear, which presents a challenge for high-precision estimating EWT and LMA using remote sensing images. In this study, we attempted to synergistic retrieval of mangrove EWT and LMA using field hyperspectral and satellite data. Specifically, we proposed a novel Fractional-order assisted Spectral angle Analysis (FSA) approach to accurately identify sensitive spectral domains of EWT and LMA for three mangrove species based on in situ full-spectrum hyperspectral (350–2500 nm) data. We constructed a new spectral matching index (SMI) to screen the satellite-based sensitive spectral bands based on the field-based spectral domains, and further verified their feasibility of estimating EWT and LMA using three optical satellite sensors (Sentinel-2A, SDGSAT-1, and OHS-3D). Finally, we explored the effectiveness of combining SAR and thermal infrared (TIR) images with sensitive spectral bands on improving retrieval performance in estimating EWT and LMA, respectively. Results indicated that FSA approach could capture the sensitive spectral domains of EWT (1125–1868 nm) and LMA (1419–1999 nm) for three mangrove species. The SMI could effectively realize the spectral matching between field-based sensitive domains and satellite-based bands, and perform the synergistic retrieval of mangrove EWT (R2 = 0.65) and LMA (R2 = 0.70). Intriguing finding is that the additional TIR bands significantly enhanced the EWT and LMA inversion accuracy (R2 increased by 9.8 %–68.3 %) when compared to SAR images. We confirmed that EWT displayed an inverse trend of LMA with the flooding frequency increasing. Our works provide a scientific basis for protection and sustainable management of mangroves.

Video and audio are images: A cross-modal mixer for original data on video–audio retrieval

Cross-modal retrieval has become popular in recent years, particularly with the rise of multimedia. Generally, the information from each modality exhibits distinct representations and semantic information, which makes feature tends to be in separate latent spaces encoded with dual-tower architecture and makes it challenging to establish semantic relationships between modalities, resulting in poor retrieval performance. To address this issue, we propose a novel framework for cross-modal retrieval consisting of a cross-modal mixer, a masked autoencoder for pre-training, and a cross-modal retriever for downstream tasks. Specifically, we first adopt cross-modal mixer and mask modeling to fuse the original modality and eliminate redundancy. Then, an encoder–decoder architecture is applied to achieve a fuse-then-separate task in the pre-training phase. We feed masked fused representations into the encoder and reconstruct them with the decoder, ultimately separating the original data of two modalities. We use the pre-trained encoder in downstream tasks to build the cross-modal retrieval method. Extensive experiments on 2 real-world datasets show that our approach outperforms previous state-of-the-art methods in video–audio matching tasks, improving retrieval accuracy by up to 2×. Furthermore, we prove our model performance by transferring it to other downstream tasks as a universal model.

VTHSC-MIR: Vision Transformer Hashing with Supervised Contrastive learning based medical image retrieval

In past few years, deep learning based medical image analysis technologies have significantly improved computer-assisted tasks like detecting, diagnosing, and predicting medical outcomes. The monitoring and diagnosis of ailments such as cancer and COVID-19 rely primarily on retrieving medical images. As the medical databases size is increasing rapidly it causes difficulty in managing and querying them. The main challenge lies in achieving superior retrieval performance while handling the intricacies of medical imaging datasets. To tackle this, we propose a novel end-to-end trainable framework for medical image retrieval using a vision transformer hashing with Supervised Contrastive (VTHSC) learning. Leveraging the self-attention mechanism of transformers and supervised contrastive loss, our model surpasses existing hashing-based retrieval methods. Focusing on the most recent COVID-19 and breast cancer datasets, we aim to efficiently retrieve relevant medical images from large datasets. Using the ImageNet-pretrained ViT as its backbone network, the VTHSC model incorporates a hashing head combined with a supervised contrastive loss and employs joint loss optimization. The VTHSC model is subsequently fine- tuned for a retrieval task, employing four different hashing frameworks: Deep Supervised Hashing (DSH), GreedyHash, Improved Deep Hashing Network (IDHN), and Deep Polarized Network (DPN). Our model achieves outstanding Mean Average Precision (MAP) scores of 98.9 for the BreakHis dataset and 96.03 for the COVID dataset. Notably, the VTHSC model surpasses several competing hashing-based retrieval methods by a substantial gain in terms of performance across various metrics such as precision, recall, and Top-6 retrieved images retrieval performance on the two benchmark medical image datasets.

Enhancing Medical Image Retrieval with UMLS-Integrated CNN-Based Text Indexing.

In recent years, Convolutional Neural Network (CNN) models have demonstrated notable advancements in various domains such as image classification and Natural Language Processing (NLP). Despite their success in image classification tasks, their potential impact on medical image retrieval, particularly in text-based medical image retrieval (TBMIR) tasks, has not yet been fully realized. This could be attributed to the complexity of the ranking process, as there is ambiguity in treating TBMIR as an image retrieval task rather than a traditional information retrieval or NLP task. To address this gap, our paper proposes a novel approach to re-ranking medical images using a Deep Matching Model (DMM) and Medical-Dependent Features (MDF). These features incorporate categorical attributes such as medical terminologies and imaging modalities. Specifically, our DMM aims to generate effective representations for query and image metadata using a personalized CNN, facilitating matching between these representations. By using MDF, a semantic similarity matrix based on Unified Medical Language System (UMLS) meta-thesaurus, and a set of personalized filters taking into account some ranking features, our deep matching model can effectively consider the TBMIR task as an image retrieval task, as previously mentioned. To evaluate our approach, we performed experiments on the medical ImageCLEF datasets from 2009 to 2012. The experimental results show that the proposed model significantly enhances image retrieval performance compared to the baseline and state-of-the-art approaches.

Disentangled Retrieval and Reasoning for Implicit Question Answering.

To date, most of the existing open-domain question answering (QA) methods focus on explicit questions where the reasoning steps are mentioned explicitly in the question. In this article, we study implicit QA where the reasoning steps are not evident in the question. Implicit QA is challenging in two aspects. First, evidence retrieval is difficult since there is little overlap between a question and its required evidence. Second, answer inference is difficult since the reasoning strategy is latent in the question. To tackle implicit QA, we propose a systematic solution denoted as DisentangledQA, which disentangles topic, attribute, and reasoning strategy from the implicit question to guide the retrieval and reasoning. Specifically, we disentangle the topic and attribute information from the implicit question to guide evidence retrieval. For answer reasoning, we propose a disentangled reasoning model for answer prediction based on retrieved evidence as well as the latent representation of the reasoning strategy. The disentangled framework empowers each module to focus on a specific latent element in the question, and thus, leads to effective representation learning for them. Experiments on the StrategyQA dataset demonstrate the effectiveness of our method in answering implicit questions, improving performance in evidence retrieval and answering inference by 31.7% and 4.5%, respectively, and achieving the best performance on the official leaderboard. In addition, our method achieved the best performance on the challenging EntityQuestions dataset, indicating the effectiveness in improving general open-domain QA tasks.

Comparison of a Reduced Needle With a Thin Tip and a Standard Thin Needle (Same Width Overall) for Oocyte Retrieval.

At our facility, oocyte retrieval had previously been performed with a 20-gauge standard needle that is uniformly thin overall (tSN); but recently, we have instead started using reduced needles, with a 20-gauge tip and 17-gauge body (RN). Until now, there have been comparisons between RN and thick standard needles, but there have been no comparisons between RN and tSN. The purpose of this study was to compare oocyte retrieval outcomes using RN with tSN. Information on oocyte retrieval was extracted from the medical records of 304 cycles performed at our facility from January 2020 to December 2023. The oocyte retrieval outcomes of the two types of needles were compared retrospectively with respect to age, anti-Müllerian hormone (AMH), procedure time, additional sedatives, number of follicles punctured, number of oocytes retrieved, number of oocytes fertilized, oocyte recovery rate, and fertilization rate. When AMH ≥ 1.2 ng/mL, the procedure time was 9.3 ± 3.7 and 12.1 ± 4.6 minutes in the RN and tSN groups, respectively (P< 0.001), and the need for additional sedatives was also significantly different: 54.0% in the RN group and 78.5% in the tSN group (P= 0.002). The oocyte recovery rate was significantly different between the RN and tSN groups at 65.3% and 61.2%, respectively (P= 0.046), and the fertilization rate was significantly different between the RN and tSN groups at 56.8% and 66.8%, respectively (P< 0.001). There were no significant differences by age, AMH, number of follicles punctured, number of oocytes retrieved, or number of oocytes fertilized. Without diminished ovarian reserve, RN reduced procedure time and the need for additional sedatives compared to tSN. In addition, the number of oocytes fertilized per oocyte retrieval remained the same, indicating that oocyte retrieval performance was not affected.

Reactivation strength during cued recall is modulated by graph distance within cognitive maps.

Declarative memory retrieval is thought to involve reinstatement of neuronal activity patterns elicited and encoded during a prior learning episode. Furthermore, it is suggested that two mechanisms operate during reinstatement, dependent on task demands: individual memory items can be reactivated simultaneously as a clustered occurrence or, alternatively, replayed sequentially as temporally separate instances. In the current study, participants learned associations between images that were embedded in a directed graph network and retained this information over a brief 8 min consolidation period. During a subsequent cued recall session, participants retrieved the learned information while undergoing magnetoencephalographic recording. Using a trained stimulus decoder, we found evidence for clustered reactivation of learned material. Reactivation strength of individual items during clustered reactivation decreased as a function of increasing graph distance, an ordering present solely for successful retrieval but not for retrieval failure. In line with previous research, we found evidence that sequential replay was dependent on retrieval performance and was most evident in low performers. The results provide evidence for distinct performance-dependent retrieval mechanisms, with graded clustered reactivation emerging as a plausible mechanism to search within abstract cognitive maps.

Reactivation strength during cued recall is modulated by graph distance within cognitive maps

Declarative memory retrieval is thought to involve reinstatement of neuronal activity patterns elicited and encoded during a prior learning episode. Furthermore, it is suggested that two mechanisms operate during reinstatement, dependent on task demands: individual memory items can be reactivated simultaneously as a clustered occurrence or, alternatively, replayed sequentially as temporally separate instances. In the current study, participants learned associations between images that were embedded in a directed graph network and retained this information over a brief 8 min consolidation period. During a subsequent cued recall session, participants retrieved the learned information while undergoing magnetoencephalographic recording. Using a trained stimulus decoder, we found evidence for clustered reactivation of learned material. Reactivation strength of individual items during clustered reactivation decreased as a function of increasing graph distance, an ordering present solely for successful retrieval but not for retrieval failure. In line with previous research, we found evidence that sequential replay was dependent on retrieval performance and was most evident in low performers. The results provide evidence for distinct performance-dependent retrieval mechanisms, with graded clustered reactivation emerging as a plausible mechanism to search within abstract cognitive maps.

Mapping method for complex mechatronic equipment relationship data to contextual semantic-constrained knowledge graphs

Complex mechatronic equipment’s complete life cycle information model includes multidimensional data on structure, qualities, behavior, and limitations. The data are diverse, derived from multiple sources, and include both streaming and non-streaming cyber-physical data. The product data system’s structured relational model, which is represented by the XBOM (X Bill of Material), is primarily where this data is stored. However, with the growing amount of equipment information, this relatively isolated and redundant architecture poses challenges. Within relational models, the growing intricacy of inner-outer key joins impedes effective data mining and retrieval for dynamic digital twin business applications. In response, this study proposes a knowledge graph transformation method that takes into account the semantic restrictions of the business application context to translate relational data to entities and semantic expressions. The method builds a knowledge context constraint meta-model by classifying data according to context semantics into one-hop entities, multi-hop entities, and abstract classes. Additionally, an ontology-based data extraction transformation method and an ontology-based mapping algorithm based on relational schema transformation are introduced in this article. This strategy significantly minimizes knowledge redundancy and enhances retrieval performance, as demonstrated by experimental verification through the development of a relational data mapping tool software.

A novel re-ranking architecture for patent search

Patent search presents unique challenges due to the intricate structure and specialized terminology embedded in patent documents. While neural models have been successfully applied in various information retrieval (IR) tasks, these inherent complexities have hindered their effectiveness in patent search. To address these challenges, we propose a novel re-ranking architecture that effectively handles long, structured patent documents and leverages AI models to interpolate lexical and semantic signals of relevance. Additionally, the architecture incorporates query-specific weights for the final re-ranking process. To address partial relevance between patent sections our method effectively models the relevance relationships between different sections of patent documents. We calculate lexical and semantic signals of relevance from each document section and feed them as input features to AI models that estimate a combined relevance score. Finally, we compute query-specific weights to determine the relative contributions of lexical and semantic relevance for the final re-ranking. Extensive experiments on the CLEF-IP dataset demonstrate that our method outperforms several baselines, achieving substantial and statistically significant improvements in retrieval performance. We further assess the adaptability of our method using the MSMARCO dataset, where it exhibits limited performance, indicating its suitability for domain-specific patent research.

A Cross-Language Information Retrieval Method Based on Multi-Task Learning

This study introduces a novel Cross-Language Information Retrieval (CLIR) method employing multi-task learning and soft parameter sharing to enhance neural retrieval models' feature extraction across languages. The approach integrates an interaction-based neural retrieval model with a semantic-based text classification model, exchanging hidden vectors for richer feature representation. Experimental results across four language pairs—English-Chinese, English-Arabic, English-French, and English-German—demonstrate significant performance improvements. The proposed method achieved the highest Mean Average Precision (MAP) scores: 0.419 for EN-ZH, 0.403 for EN-AR, 0.427 for EN-FR, and 0.441 for EN-DE, surpassing other models like BM25, BPNRM, KNRM, KNRM-Trans, and KNRM-Embed. This research underscores the potential of multi-task learning for CLIR, showcasing improved retrieval performance through semantic information and knowledge transfer.