Information Processing & Management最新文献

Multimodal Entity and Relation Extraction (MERE) encompasses tasks, including Multimodal Named Entity Recognition (MNER) and Multimodal Relation Extraction (MRE), aiming to extract valuable information from environments rich in multimodal data. Currently, many research endeavors face various challenges, including the insufficient utilization of emotional information in multimodal data, mismatches between textual and visual content, ambiguous meanings, and difficulties achieving precise alignment across different semantic levels. To address these issues, we propose the Hierarchical Generation of Multi Evidence Alignment Fusion Model for Multimodal Entity and Relation Extraction (HGMAF). This model comprises a hierarchical diffusion semantic generation stage and a multi-evidence alignment fusion module. Initially, we designed different prompt templates for the original text, using the Large Language Model (LLM) to generate corresponding hierarchical textual content. Subsequently, the generated hierarchical content is diffused to obtain images with rich hierarchical semantic information. This stage contributes to enhancing the model's understanding of hierarchical information in the original content. Following this, we design the multi-evidence alignment fusion module, which combines the generated textual and image evidence, fully leveraging information from different sources to improve extraction accuracy. Experimental results demonstrate that our model achieves F1 scores of 76.29 %, 87.66 %, and 87.34 % on the Twitter2015, Twitter2017, and MNRE datasets, respectively. These results surpass the previous state-of-the-art models by 0.29 %, 0.1 %, and 2.77 %. Furthermore, our model demonstrates superior performance in low-resource scenarios, confirming its effectiveness. The related code can be found at https://github.com/lsx314/HGMAF.

Biometrics have copious merits over traditional authentication schemes and promote information management. The demand for large-scale biometric identification and certification booms. In spite of enhanced efficiency and scalability in cloud-based biometrics, they suffer from compromised privacy during the transmission and storage of irrevocable biometric information. Existing biometric protection strategies fatally degrade the recognition performance, due to two folds: inherent drawbacks of uni-biometrics and inevitable information loss caused by over-protection. Hence, how to make a trade-off between performance and protection is an alluring challenge. To settle these issues, we are the first to present a cancelable multi-biometric system combining iris and periocular traits with recognition performance improved and privacy protection emphasized. Our proposed binary mask-based cross-folding integrates multi-instance and multi-modal fusion tactics. Further, the steganography based on a low-bit strategy conceals sensitive biometric fusion into QR code with transmission imperceptible. Subsequently, a fine-grained hybrid attention dual-path network through stage-wise training models inter-class separability and intra-class compactness to extract more discriminative templates for biometric fusion. Afterward, the random graph neural network transforms the template into the protection domain to generate the cancelable template versus the malicious. Experimental results on two benchmark datasets, namely IITDv1 and MMUv1, show the proposed algorithm attains promising performance against state-of-the-art approaches in terms of equal error rate. What is more, extensive privacy analysis demonstrates prospective irreversibility, unlinkability, and revocability, respectively.

Aspect-Controllable Related Papers Recommendation (ACRPR) aims to satisfy users’ fine-grained needs for specific aspects when finding related papers. Existing approaches rely on the segmentation of texts or aspects to independently learn multi-aspect representations of papers. However, different aspects of a paper are guided by its overall theme and interconnected with intrinsic relevance. In light of this, we propose a simple yet effective ACRPR framework called mCTRL, which leverages multiple control codes in Transformer to simultaneously learn multiple aspect-specific paper representations. Specifically, mCTRL incorporates a $\left[\mathrm{CLS}\right]$ control code to capture the overall theme and multiple $\left[\mathrm{ASP}\right]$ control codes to exploit fine-grained aspect information. Additionally, we introduce a hierarchical loss function to balance the overall theme and various aspects of a paper, enabling their mutual enhancement and alignment. Extensive comparative experiments on real-world datasets demonstrate the superiority of our proposed method over previous state-of-the-art approaches. Evaluations are conducted on 5 backbone models and 5 dimensions, which confirm the generalization ability of mCTRL. Moreover, ablation studies and further analyses prove the effectiveness and efficiency of mCTRL and the specialization across aspects of generated embeddings.

The deep semantic information mining extracts deep semantic features from textual data and effectively utilizes the world knowledge embedded in these features, so it is widely researched in recommendation tasks. In spite of the extensive utilization of contextual information in prior Point-of-Interest research, the insufficient and non-informative textual content has led to the neglect of deep semantic study. Besides, effectively integrating the deep semantic information into the trajectory modeling process is also an open question for further exploration. Therefore, this paper proposes HyperSE, to leverage prompt engineering and pre-trained language models for deep semantic enhancement. Besides, HyperSE effectively extracts higher-order collaborative signals from global and local hypergraphs, seamlessly integrating topological and semantic information to enhance trajectory modeling. Experimental results show that HyperSE outperforms the strong baseline, demonstrating the effectiveness of the deep semantic information and the model’s efficiency.

This study proposes an innovative model (i.e., SensoryT5), which integrates sensory knowledge into the T5 (Text-to-Text Transfer Transformer) framework for emotion classification tasks. By embedding sensory knowledge within the T5 model’s attention mechanism, SensoryT5 not only enhances the model’s contextual understanding but also elevates its sensitivity to the nuanced interplay between sensory information and emotional states. Experiments on four emotion classification datasets, three sarcasm classification datasets one subjectivity analysis dataset, and one opinion classification dataset (ranging from binary to 32-class tasks) demonstrate that our model outperforms state-of-the-art baseline models (including the baseline T5 model) significantly. Specifically, SensoryT5 achieves a maximal improvement of 3.0% in both the accuracy and the F1 score for emotion classification. In sarcasm classification tasks, the model surpasses the baseline models by the maximal increase of 1.2% in accuracy and 1.1% in the F1 score. Furthermore, SensoryT5 continues to demonstrate its superior performances for both subjectivity analysis and opinion classification, with increases in ACC and the F1 score by 0.6% for the subjectivity analysis task and increases in ACC by 0.4% and the F1 score by 0.6% for the opinion classification task, when compared to the second-best models. These improvements underscore the significant potential of leveraging cognitive resources to deepen NLP models’ comprehension of emotional nuances and suggest an interdisciplinary research between the areas of NLP and neuro-cognitive science.

Anomaly detection in tabular data holds significant importance across various industries such as manufacturing, healthcare, and finance. However, existing methods are constrained by the size and diversity of datasets, leading to poor generalization. Moreover, they primarily concentrate on feature correlations while overlooking interactions among data instances. Furthermore, the vulnerability of these methods to noisy data hinders their deployment in practical engineering applications. To tackle these issues, this paper proposes prototype-oriented hypergraph representation learning for anomaly detection in tabular data (PHAD). Specifically, PHAD employs a diffusion-based data augmentation strategy tailored for tabular data to enhance both the size and diversity of the training data. Subsequently, it constructs a hypergraph from the combined augmented and original training data to capture higher-order correlations among data instances by leveraging hypergraph neural networks. Lastly, PHAD utilizes an adaptive fusion of local and global data representations to derive the prototype of latent normal data, serving as a benchmark for detecting anomalies. Extensive experiments on twenty-six public datasets across various engineering fields demonstrate that our proposed PHAD outperforms other state-of-the-art methods in terms of performance, robustness, and efficiency.

ChatGPT's impressive performance enables users to increasingly apply it to a variety of scenarios. However, previous studies investigating people's perceptions or attitudes towards ChatGPT have not considered the effects of the usage scenario. This paper aims to extract the representative scenarios of ChatGPT, explore differences in user perceptions for each scenario, and provide a policy support model. We extracted five scenarios by collecting 50 open-ended responses from Mturk, including “Scenario 1: Daily life tasks,” “Scenario 2: Enhance efficiency (work and education purposes),” “Scenario 3: Replace manpower (work and education purposes),” “Scenario 4: Browsing and general information seeking,” “Scenario 5: Enjoyment.” Subsequently, we identified four key variables to be tested (i.e., information quality, perceived risk, attitude, and policy support), and classified usage scenarios into different categories according to the perception variables measured via an online survey (n = 514). Finally, we built a model including the four variables and tested it for each scenario. The results of this study provide deep insights into user perceptions towards ChatGPT in distinct scenarios.

Travel recommendation aims to infer travel intentions of users by analyzing their historical behaviors on Online Travel Agencies (OTAs). However, crucial keywords in clicked travel product titles, such as destination and itinerary duration, indicating tourists’ intentions, are often overlooked. Additionally, most previous studies only consider stable long-term user interests or temporary short-term user preferences, making the recommendation performance unreliable. To mitigate these constraints, this paper proposes a novel Keywords-enhanced Contrastive Learning Model (KCLM). KCLM simultaneously implements personalized travel recommendation and keywords generation tasks, integrating long-term and short-term user preferences within both tasks. Furthermore, we design two kinds of contrastive learning tasks for better user and travel product representation learning. The preference contrastive learning aims to bridge the gap between long-term and short-term user preferences. The multi-view contrastive learning focuses on modeling the coarse-grained commonality between clicked products and their keywords. Extensive experiments are conducted on two tourism datasets and a large-scale e-commerce dataset. The experimental results demonstrate that KCLM achieves substantial gains in both metrics compared to the best-performing baseline methods. Specifically, HR@20 improved by 5.79%–14.13%, MRR@20 improved by 6.57%–18.50%. Furthermore, to have an intuitive understanding of the keyword generation by the KCLM model, we provide a case study for several randomized examples.

Long prompt leads to huge hardware costs when using transformer-based Large Language Models (LLMs). Unfortunately, many tasks, such as summarization, inevitably introduce long documents, and the wide application of in-context learning easily makes the prompt length explode. This paper proposes a Self-Compressor (SelfCP), which adopts the target LLM itself to compress over-limit prompts into dense vectors on top of a sequence of learnable embeddings (memory tags) while keeping the allowed prompts unmodified. Dense vectors are then projected into memory tokens via a learnable connector, allowing the same LLM to understand them. The connector and the memory tag are supervised-tuned under the language modeling objective of the LLM on relatively long texts selected from publicly accessed datasets involving an instruction dataset to make SelfCP respond to various prompts, while the target LLM keeps frozen during training. We build the lightweight SelfCP upon 2 different backbones with merely 17M learnable parameters originating from the connector and a learnable embedding. Evaluation on both English and Chinese benchmarks demonstrate that SelfCP effectively substitutes 12$\times$ over-limit prompts with memory tokens to reduce memory costs and booster inference throughputs, yet improving response quality. The outstanding performance brings an efficient solution for LLMs to tackle long prompts without training LLMs from scratch.

Using the user’s past activity across different domains, the cross-domain recommendation (CDR) predicts the items that users are likely to click. Most recent studies on CDR model user interests at the item level. However because items in other domains are inherently heterogeneous, direct modeling of past interactions from other domains to augment user representation in the target domain may limit the effectiveness of recommendation. Thus, in order to enhance the performance of cross-domain recommendation, we present a model called Cross-domain Recommendation based on Intent Disentanglement and Contrast Learning (IDC-CDR) that performs contrastive learning at the intent level between domains and disentangles user interaction intents in various domains. Initially, user–item interaction graphs were created for both single-domain and cross-domain scenarios. Then, by modeling the intention distribution of each user–item interaction, the interaction intention graph and its representation were updated repeatedly. The comprehensive local intent is then obtained by fusing the local domain intents of the source domain and the target domain using the attention technique. In order to enhance representation learning and knowledge transfer, we ultimately develop a cross-domain intention contrastive learning method. Using three pairs of cross-domain scenarios from Amazon and the KuaiRand dataset, we carry out comprehensive experiments. The experimental findings demonstrate that the recommendation performance can be greatly enhanced by IDC-CDR, with an average improvement of 20.62% and 25.32% for HR and NDCG metrics, respectively.