Thailand's diverse ecosystem harbors numerous snake species, often causing venomous snake bites necessitating antivenom treatment. Accurate identification of snake species enhances treatment success. This study classified 12 Venomous Snake with medical importance in Thailand, using 6,555 annotated images from iNaturalist, Nick Wildlife and the Queen Saovabha Memorial Institute. Images were annotated and split into training, testing and validation sets. YOLOv7, YOLOv8 and YOLOv9 models were employed with batch size 16, image size 640, and epochs 50, 100 and 150. YOLOv9m excelled with Precision 0.85, Recall 0.815, outperforming YOLOv9s and YOLOv9c. The results show that, YOLOv9m proved the most effective for accurate detection, especially in complex environments for venomous snake species differentiation.
Effective management of digital assets on blockchains is increasingly important for enhancing transparency and efficiency in areas like Supply Chain Management and ownership of user-generated content on the Web. Blockchains often have a very limited set of native transaction and asset types (usually a cryptocurrency) but allow user-defined transactions and assets implemented as part of programs called smart contracts. However, the semantics of these user-defined components are opaque to the blockchain engine, limiting reuse in transaction and asset modeling, blockchain interoperability and opportunities for automatic optimization of user-defined blockchain transactions. This paper explores the explicit modeling of digital assets and their impact on blockchain performance and usability. In particular, we introduce an approach for enabling a smart asset model and propose an implementation approach using the BigChainDB blockchain database as our context. We present an empirical evaluation that shows significant improvements in transaction latency, throughput, and correctness over traditional smart contract systems.
In the context of search as learning, users achieve their learning goals (target knowledge) through exploratory research. Consequently, it is essential to track the user's knowledge state in order to estimate how close the user is to achieving their learning goals. A recently proposed approach uses knowledge graphs to represent both the user's knowledge and the user's target knowledge. However, some basic pieces of information are not taken into account because they are not explicitly present in the resources during the user's search session. This omission results in incomplete relations between resources and reduces the accuracy of the user's knowledge gain estimation. Our approach, which is based on this foundation, introduces a method for enhancing the user's knowledge gain estimation by considering implied knowledge. In more precise terms, our approach extends the representation of both the user's and the target knowledge by utilising different predicate properties to complete these relations. We show that the newly proposed approach improves the knowledge gain estimation by taking into account an additional aspect of knowledge.
Artificial intelligence (AI) and machine learning have recently become central to decision-making. However, decisions made by AI are often complex and not easily understood. Inferences derived from these black-box models require an explanation, highlighting the importance of explainable AI (XAI). Approximate inverse model explanations (AIME), as a notable method in this domain, derives the approximate inverse operators of black-box models and has been explored extensively for XAI. This method facilitates easily interpretable explanations of model behaviors and estimations while maintaining low computational complexity. However, interpreting these results can be challenging for end users, particularly with complex features. This study introduces explainable-augmented generation (XAG), which combines the strengths of XAI with advanced language models to make AI decisions clearer. XAG uses AIME to explain the decision-making of AI systems and utilizes large language models (LLMs) to translate technical details into easy-to-understand explanations. Well-crafted prompts allow integrating the results of XAI methodologies such as AIME with LLMs, thereby enabling XAG to identify the features that contribute the most to decision-making and interpret the significance of feature importance. This approach enables a nuanced understanding of the high or low importance assigned to certain features within the decision-making process of the model. By merging the robust capabilities of AIME to generate interpretable model insights with the narrative power of LLMs, XAG offers a pioneering approach to demystifying AI operations for non-expert users.
Few-shot Fine-grained Entity Typing (FET) techniques are applied to classify entities recognized in texts into classes from taxonomies using a limited amount of training data. The task supports entity extraction and knowledge graph (KG) construction by classifying novel entities according to a taxonomy of interest. In this paper, we present PROMET (PROmpt-tuning using implicit Mask filling for Entity Typing), a novel parameter-efficient prompt-based approach to few-shot FET that exploits implicit mask filling. The usage of the embedding of the masked tokens avoids the necessity of a manually defined verbalizer to initialize the matrix which maps the predicted token to entity classes. By directly fine-tuning Adapters and a linear layer instead of 1) the whole PLM and 2) the verbalizer matrix, PROMET uses a number of trainable parameters that is two orders of magnitude smaller than existing models, achieving better or comparable performance on benchmark datasets. Finally, we develop PROMET (and modify the state-of-the-art few-shot FET approach) to work in multi-label inference settings, coherently with earlier work in the field and with typing patterns in KGs.
Researchers want to automate the extraction of information from charts in paper, that will help do fast reading and compare methods or results efficiency, especially in chemical science research progress. This paper designs and implements a line chart information extraction algorithm using neural networks and other techniques. First, a large dataset of line charts was collected and annotated to provide a foundation for neural network training. Then, the Hough transform line detection algorithm was used to detect the axes, and the line charts were segmented. For each segmented part, different recognition algorithms were designed to identify various elements in the line charts, including axes, line regions, and legends. To validate the effectiveness of the algorithm, experimental tests were conducted in the field of inorganic catalysis, automatically extracting information from line charts in the literature. The experimental results show that the designed algorithm can accurately recognize various elements in line charts and effectively extract experimental data. Compared with traditional manual methods, automated extraction not only saves a considerable amount of time but also improves the accuracy and consistency of data extraction on paper fast reading. In summary, this method provides researchers with an efficient tool that accelerates the acquisition and comparison of experimental data, thereby advancing the progress of related research electronic document.
Pig breeding is a pillar of China's livestock industry. Currently, most farms still rely on empirical feeding strategies, which are either set manually or utilize traditional intelligent control methods such as fuzzy control and expert systems. These methods are heavily dependent on expert experience and extensive historical data, making them susceptible to potential factors or environmental changes that can destabilize the breeding process. To enhance the feeding management of nursery and fattening piglets and improve the feed conversion ratio (FCR) in farms, this paper presents a ConvBiLSTM model for predicting feed intake. Using data mining techniques, the model helps farms determine the optimal feed intake at different growth stages.
Surrogate-assisted models replace the expensive fitness value evaluation by predicting the function value of optimization problems, thus reducing real fitness value evaluations, which becomes the mainstream of solving expensive optimizations. Surrogate-assisted models include classification-based models and regression-based models. Current surrogate-assisted expensive optimization algorithms rely on the surrogate-assisted models. However, the performance of classification-based methods is limited, and the performance of regression-based methods is better, but the computational complexity is high. In this paper, a multi-population surrogate-assisted algorithm based on granular-ball K-nearest neighbors(GBKNN) classification(SAMPGB) is proposed for solving this problem. SAMPGB uses GBKNN algorithm as the surrogate-assisted model to reduce the computational complexity. A multi-population cooperative framework is employed to enhance the performance. SAMPGB divides the whole population into different coarse-grained sub-populations, and each coarse-grained subpopulation is further divided into different fine-grained subpopulations. Coarse-grained subpopulations adopt different evolutionary strategies, while fine-grained subpopulations belonging to the same coarse-grained subpopulation adopt the same evolutionary strategy. Then the results of all the fine-grained subpopulations are integrated as the result of SAMPGB. The effectiveness of SAMPGB is proved through 30-1000 dimensional experiments of 6 widely benchmark functions.
This paper presents a new approach to improving retrieval system performance by integrating Pseudo-Relevance Feedback (PRF) with external knowledge introduced through Large Language Models (LLMs). Query expansion techniques can improve the accuracy of retrieval systems. The study Introducing external knowledge into the query expansion process is also an effective method of data augmentation. Additionally, This paper investigates the integration of externally generated knowledge from LLMs into query expansion within dense retrieval models. It examines the selection of relevant external knowledge and the effective combination of this knowledge with pseudo-relevant document features. The study evaluates the impact of incorporating external knowledge not only for the original query but also from pseudo-relevant documents, assessing its effect on retrieval performance. The experimental results across two datasets and three evaluation metrics demonstrate that the proposed method, which integrates external knowledge with pseudo-relevant document features, significantly improves the accuracy of the retrieval system.
In the context of the burgeoning online education landscape, course recommendation has emerged as a pivotal element in enhancing learning efficacy and user experience. However, traditional course recommendation models often grapple with the challenges of cold start and data sparsity. To address these issues, this paper introduces a meta-learning course recommendation model based on heterogeneous information networks (HIN). The model capitalizes on the intricate structure and interconnections within HIN to learn adaptive recommendation strategies tailored to diverse user segments within the meta-learning framework.Initially, the model encodes various types of nodes and edges to construct a multi-layered HIN. It employs a weighted average approach to aggregate enhanced user embedding representations derived from different meta-paths through multipath aggregation. Subsequently, the model learns initial parameters adapted to different users from a limited set of observed user behavior data. Through the optimization of the model parameters using the model-agnostic meta-learning (MAML) algorithm, personalized course recommendations are achieved.Experimental results have validated that the proposed model outperforms existing educational recommendation models in scenarios of non-cold start and data sparsity. Furthermore, when compared to the MELU model solely employing the MAML algorithm, our enhanced MElU model integrated with multipath aggregation demonstrates a reduction of 6% in both the MAE and RMSE metrics, with a commensurate increase of 4% in the NDCG@5 metric.
Open-Domain Question Answering (ODQA) has emerged as a critical research area in the field of natural language processing (NLP) in artificial intelligence (AI). Existing approaches primarily follow two paradigms for evidence collection: retrieve-then-read methods often struggle to acquire comprehensive and diverse evidence, and generate-then-read approaches often produce documents that lack contextual accuracy and relevance. We introduce an innovative framework named Expansion Generation and Verification (EGV), derived from the core processes of generating, evaluating, and verifying evidence. EGV encompasses six key stages: generation expansion, expansion evaluation, document re-ranking, re-ranking evaluation, answer generation, and answer verification. This framework effectively integrates the strengths of both retrieval-based and generative evidence collection methodologies. Experimental evaluations on widely-used benchmarks, including NQ, WebQ, and TriviaQA, demonstrate that EGV achieves state-of-the-art performance in both answer accuracy and evidence quality. These results under-score EGV's potential to significantly advance ODQA research and its practical applications.
Since the Covid-19 pandemic, sexual technology sales have increased all over the world, especially for devices allowing users to connect with each other or the internet. The convergence of sexual technology with pervasive computing and artificial intelligence (AI) introduces a new era of human-robot interactions. Sexual technology, ranging from smart sex toys to humanoid sex robots, now incorporates pervasive computing principles and/or AI to enhance intimate experiences. These devices leverage ubiquitous computing capabilities, adaptability, and context awareness to increase sexual satisfaction and integrate into users' lives. The integration of these technologies into intimate interactions raises questions on how these changes will affect different Webs of Life, including but not limited to, the Web of People, the Web of Things and the Web of Health. From health and cybersecurity risks to the reshaping of human-human and human-robot interactions, this paper addresses the challenges at the intersection of sexual technologies, pervasive computing, AI and human-machine interactions. Emphasizing the importance of monitoring changes and increasing cybersecurity to create safe, consensual, and secure intimate experiences.
Information overload is increasingly common due to the overwhelming number of online products and services offered online. Web shops use recommender systems to provide personalized suggestions, which are effective for users with historical data but face challenges with new users due to the coldstart problem. It is useful to recommend items that offer a small but representative insight into new users' preferences, so that future recommendations can be improved. This research tests two error-based active learning strategies, which are especially useful because they are expected to both produce informative training points as well as reduce the predictive error of the model: the Ychange method and a modified version of the CV-based method called the error-change method. These methods rank items for predicting new user preferences, with proposed conservative, moderate, and risky versions. Results are compared with a random selection method and the PopGini method, which was the top-performing strategy in a research with a similar set-up. The moderate version of the error-change method significantly outperforms the other considered methods, while the Y-change method still outperforms PopGini.
The United States has seen a significant increase in unsolved violent crime cases since the 1980s. Currently, around 250,000 unsolved homicide cases are on record, with an estimated 6,000 new cases added annually. Many of these crime cases possess valuable biological evidence but remain unsolved due to the lack of a reference DNA profile from a person of interest. In such cases, prediction of the visual appearance of a person from a DNA sample might generate invaluable investigative leads. Therefore, in this study, we explored the decision-making abilities of Large Language Models (LLMs) to establish connections between an input of a subject's structured genomic data and produce a textual output that characterizes the phenotypic attributes of that subject using its general knowledge and the effects of fine-tuning. By leveraging large amounts of LLMs' pre-training with genomic sequence data, supervised fine-tuning can be used to guide LLMs toward the specific downstream task of predicting subject-specific phenotypic characteristics. We used genomic and the corresponding phenotypic data to train the GPT 3.5-Turbo and GPT 4o-mini models. Both models exhibited reasonable performance for the prediction of numerical phenotypic traits, like height and weight, but performed poorly with categorical traits like eye color and hair texture, due to a limited dataset. Following human evaluation, we concluded that both models could construct nonrandom relationships between genomic data and observable phenotypic attributes, like skin color and biogeographic ancestry, to produce understandable textual descriptions. This study works to evaluate LLMs' ability to use general knowledge and benefit from fine-tuning to successfully translate complex genomic data into descriptions of a phenotype. This provides the premise for future work relevant to further integrating generative AI into forensic science applications.
This study investigates the efficacy of a hybrid recommendation model for personalized meal plans, integrating Knowledge Base Question Answering (KBQA), Information Retrieval (IR), and Recommendation techniques. It utilizes the hybrid model to consider both different dietary preferences and nutritional requirements. In addition, it tries to bridge the gap between the recommender itself and its effectiveness in the real world by offering interfaces for integrating persuasion via explanation and gamification. The findings contribute to extending knowledge about the development of food recommendation systems in constrained contexts. The system can address health awareness by considering user-defined constraints. However, in big use cases, it has issues with its scalability. Future work involves refining the data generation processes and exploring non-KBQA models for broader scalability and adaptability.
The unreliability of Large Language Models (LLMs) owing to Machine Hallucination necessitates a shift towards Retrieval Augmented Generation (RAG), to ensure the veracity of LLM generated content. Inspired by Knowledge Graph based RAG techniques and their advantages of reduced token cost, improved computational performance and knowledge discovery, we present a “Knowledge Web” Structure and its associated techniques for Information Storage and Retrieval. We also explore the structure's synergy with a Fact Finding LLM Agent, capable of generating answers for multi-hop questions from challenging datasets such as the MuSiQue dataset. While still in the experimental phase, our initial results are promising, indicating the potential efficacy of our approach compared to existing RAG techniques.
As the rate of web data creation and storage continues to rise, so does the need for effective multi-document summarization techniques. Computer-generated summaries that can effectively reflect the original data in a readable format saves the need for manual human labor. In the field of online product reviews specifically, a single product can have hundreds to thousands of reviews, and an average shopper is often to only read a handful of them. With effective summarization tools, these shoppers could be given a single generated summary that condenses the content of the review set into a single paragraph or two. Many efforts have been made to fulfill this need for multi-document summarization; however, most require complex data graphs, structures, or language models. Our research stresses simplicity in the extractive algorithm, making the process easier to understand and implement. In this paper, we propose four different versions of multi-document extractive summarizers based on KL-Divergence, TF-IDF, and Diversity scoring. These extractive summarizers are then measured against each other, along with several top-notch summarizers, in terms of effectiveness at expressing relevant content and linguistic quality. The results of these tests show a significant advantage for our summarizers, thus promoting our summarizers as a powerful-yet-simple process ready for use in product review summarization.
In recent years, artificial intelligence (AI) has attained or even surpassed the level of the human brain in fields such as machine translation and computer vision. However, AI systems still face limitations due to challenges in effectively managing uncertain information and a lack of interpretability. Recent developments in brain science and discoveries in cognitive mechanisms offer potential solutions to address these issues. Inspired by the brain's dual-pathway uncertainty-processing mechanism, we designed a lung nodule diagnosis model that integrates convolutional neural networks (CNNs) and transformers, simulating the regulatory effects of dopamine and kainate receptors on uncertainty. We introduce two regularization terms to reduce diagnostic uncertainty for lung nodules. Experimental results on the LIDC-IDRI and LNDb datasets indicate that this model not only performs well but also provides visual explanations that are more consistent with domain knowledge.
Music can serve as a mirror of social conditions, and studying social phenomena requires proper computational tools to evaluate the discourses in song lyrics on a large scale. Brazilian Funk (BRFunk) is a popular genre that emerged from the favelas in Rio de Janeiro, regarded as controversial for merging stories of resilience and social criticism, with themes such as violence, drugs, and women objectification. These themes portray the social conditions of daily life in Brazilian slums (favelas). In this paper, we propose a novel topic modeling approach that explores the power of Large Language Models (LLM) and pre-trained models to extract insights from song lyrics, using BRFunk as a case study. The generative power of LLM is leveraged through prompt engineering to summarize lyrics excerpts into themes in an iterative process to create a reliable distribution of themes. Pre-trained models are then deployed to condensate the themes into a non-redundant and cohesive set of topics using BERTopic, and through BERT fine-tuned models that classify new lyrics excerpts according to these topics. Using the top 100 BRFunk songs of 2023, we illustrate how this large-scale computational strategy is an efficient ally for understanding social phenomena and cultural movements.
In social networks, some individuals or groups create numerous fake identities to increase their control or influence, which severely undermines the integrity and credibility of the net-work. Therefore, identifying these anomalous users is particularly crucial. Past research has mainly focused on supervised or semi-supervised node detection, analyzing the behavior of individual users. However, in real-world social platforms, anomalous users often operate in an organized manner. To address this, we explore the real-time behavioral patterns of anomalous groups from the perspective of event propagation. Then, by using the Leiden algorithm for group detection and combining permanence methods with anomaly detection, we have constructed a model named LAGEP (Leiden-Based Model for Detecting Anomalous Groups in Event Propagation) for identifying anomalous groups. Through the analysis of a large amount of data collected from real social platforms, we found that LAGEP improves the F1 score by 5.1% compared to the best baseline method.
