Waterway transportation sustainably facilitates global trade through eco-efficient cargo movement, where accurate water level forecasting is critical for ensuring navigational safety and operational continuity. To develop a highly accurate prediction model, it is essential to consider the periodic characteristics of water level data, which often emerge in real-world datasets. This study introduces a novel reservoir state structure based on reservoir computing theory, the Self-attention Hierarchical Kernel Reservoir State Network (SHK-RSN). It employs three primary mechanisms. First, a hierarchical feature extraction method groups training data and extracts high-dimensional features from these groups using the kernel trick in a hierarchical manner. Second, a self-attention weight selection approach is introduced to replace the random weights in the Hierarchical Kernel Reservoir State Network (HK-RSN), improving the rationale for hidden neuron connections and enhancing the interpretability of weight selection. Third, a novel reservoir state structure is proposed to capture periodic information and extract temporal features across periods, enabling the model to capture richer temporal information and identify relationships among periods. Experiments are conducted on one artificial and five real-world time series datasets, with forecast performance evaluated over 1–7 steps. Our proposed model, SHK-RSN, is compared with models based on randomization, the kernel trick, and deep learning. The experimental results demonstrate that SHK-RSN exhibits superior forecasting ability relative to the baselines. It achieves the best Symmetric Mean Absolute Percentage Error (SMAPE) across all datasets in the 1–7 period average among baseline methods, demonstrating a relative improvement of 25.7% to 46.9% over the conventional Echo State Network.

The exponential technological breakthrough of the FinTech industry has significantly enhanced user engagement through sophisticated advisory chatbots. However, large-scale fine-tuning of LLMs can occasionally yield unprofessional or flippant remarks, such as “With that money, you’re going to change the world,” which, though factually correct, can be contextually inappropriate and erode user trust. The scarcity of domain-specific datasets has led previous studies to focus on isolated components, such as reasoning-aware frameworks or the enhancement of human-like response generation. To address this research gap, we present Fin-Solution 2.O, an advanced solution that 1) introduces the multi-turn financial conversational dataset, Fin-Vault, and 2) incorporates a unified model, Fin-Ally, which integrates commonsense reasoning, politeness, and human-like conversational dynamics. Fin-Ally is powered by COMET-BART-embedded commonsense context and optimized with a Direct Preference Optimization (DPO) mechanism to generate human-aligned responses. The novel Fin-Vault dataset, consisting of 1,417 annotated multi-turn dialogues, enables Fin-Ally to extend beyond basic account management to provide personalized budgeting, real-time expense tracking, and automated financial planning. Our comprehensive results demonstrate that incorporating commonsense context enables language models to generate more refined, textually precise, and professionally grounded financial guidance, positioning this approach as a next-generation AI solution for the FinTech sector.

With the increasing use of Retrieval-Augmented Generation (RAG), strong retrieval models have become more important than ever.In healthcare, multimodal retrieval models that combine information from both text and images offer major advantages for many downstream tasks such as question answering, cross-modal retrieval, and multimodal summarization, since medical data often includes both formats.However, there is currently no standard benchmark to evaluate how well these models perform in medical settings.To address this gap, we introduce M3Retrieve, a Multimodal Medical Retrieval Benchmark.M3Retrieve, spans 5 domains,16 medical fields, and 4 distinct tasks, with over 1.2 Million text documents and 164K multimodal queries, all collected under approved licenses.We evaluate leading multimodal retrieval models on this benchmark to explore the challenges specific to different medical specialities and to understand their impact on retrieval performance.By releasing M3Retrieve, we aim to enable systematic evaluation, foster model innovation, and accelerate research toward building more capable and reliable multimodal retrieval systems for medical applications.

Punit Kumar Singh, Nishant Kumar, Akash Ghosh, Kunal Pasad, Khushi Soni, Manisha Jaishwal, Sriparna Saha, Syukron Abu Ishaq Alfarozi, Asres Temam Abagissa, Kitsuchart Pasupa, Haiqin Yang, Jose G Moreno. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

While considerable research has delved into detecting toxic content in text-based data, the realm of video content, particularly in languages other than English, has received less attention. Prior studies have primarily focused on creating automated tools to identify online toxic speech but have often overlooked the crucial next steps of mitigating its impact and discouraging future use. We can discourage social media users from sharing such material by automatically generating interventions that explain why certain content is inappropriate. To bridge this research gap, we propose an innovative task: generating interventions for toxic videos in code-mixed languages which go beyond existing methods focusing on text and images to combat online toxicity. We are introducing a Toxic Code-Mixed Intervention Video benchmark dataset (ToxCMI), comprising 1697 code-mixed toxic video utterances sourced from YouTube. Each utterance in this dataset has been meticulously annotated for toxicity and severity, accompanied by interventions provided in Hindi-English code-mixed languages. We have developed an advanced multimodal framework ToxVI, specifically designed for the task of generating Toxic Video appropriate Interventions, leveraging Large Language Models (LLMs), which comprises three modules - Modality module, Cross-Modal Synchronization module and Generation module. Our experiments demonstrate that integrating multiple modalities from the videos significantly enhances the performance of the proposed task and outperforms all the baselines by a significant margin.

Social media platforms are a double-edged sword: on the one hand, they enable the dissemination of information; but on the other hand, they also provide an avenue for spreading online abuse and harassment, such as hate speech. While significant research efforts are being devoted to detecting online hate speech in the English language, little attention has been paid to the Thai language. In this study, we created a benchmark dataset, calledHateThaiSent, which labels each post with both hate speech and sentiment information. To detect hate speech, we created a multitask model that uses a dual-channel deep learning approach based on FastText and BERT embeddings, with an added capsule network. One channel utilizes pretrained FastText embeddings while the other uses embeddings from the BERT language model. We aimed to answer two research questions: (Q1) Does incorporating sentiment information improves the performance of hate speech detection (HD) in the Thai language? (Q2) What is the comparative effectiveness of two different approaches for sentiment-aware HD in the Thai language: feature engineering versus multitasking? Our proposed approach outperformed other baselines and state-of-the-art models on theHateThaiSentdataset, with overall accuracy/macro-F1 values of 89.67%/89.79%, and 80.92%/80.97% for hate speech and sentiment detection tasks, respectively. We concluded that multitasking is more effective than feature engineering in enhancing the performance of the main task (HD).

Summary This paper proposes a new color interpolation method which can be used in embedded devices for IoT system. In this work, we use regression approach for generating and designing filters to restore color image. The filters are designed with four sizes, 5x5 training filter, 7x7 training filter, 9x9 training filter, and 11x11 training filter. The obtained filters are tested in 25 LC dataset to assess the performance. Experimental results inform that the proposed filters provide outstanding performance when they are compared with conventional methods. As compared with the other methods, the proposed filters produce the best average interpolation performance both objectively and visually.

BACKGROUND: The number of porcine Single Nucleotide Polymorphisms (SNPs) used in genetic association studies is very large, suitable for statistical testing. However, in breed classification problem, one needs to have a much smaller porcine-classifying SNPs (PCSNPs) set that could accurately classify pigs into different breeds. This study attempted to find such PCSNPs by using several combinations of feature selection and classification methods. We experimented with different combinations of feature selection methods including information gain, conventional as well as modified genetic algorithms, and our developed frequency feature selection method in combination with a common classification method, Support Vector Machine, to evaluate the method's performance. Experiments were conducted on a comprehensive data set containing SNPs from native pigs from America, Europe, Africa, and Asia including Chinese breeds, Vietnamese breeds, and hybrid breeds from Thailand. RESULTS: The best combination of feature selection methods-information gain, modified genetic algorithm, and frequency feature selection hybrid-was able to reduce the number of possible PCSNPs to only 1.62% (164 PCSNPs) of the total number of SNPs (10,210 SNPs) while maintaining a high classification accuracy (95.12%). Moreover, the near-identical performance of this PCSNPs set to those of bigger data sets as well as even the entire data set. Moreover, most PCSNPs were well-matched to a set of 94 genes in the PANTHER pathway, conforming to a suggestion by the Porcine Genomic Sequencing Initiative. CONCLUSIONS: The best hybrid method truly provided a sufficiently small number of porcine SNPs that accurately classified swine breeds.

It is always desirable to be able to manage level of water in river, dam, and reservoir. Models have been constructed for predicting the level of these bodies of water, and good models can help increase the effectiveness of water management. Presently, the model that is employed by the Hydrographic Department of the Royal Thai Navy for predicting the level of water in Chao Phraya river is a harmonic method of tidal modeling. This model can predict the overall trend well but with high individual prediction error. Many machine learning algorithms for making predictions have also been introduced in recent years. Therefore, it was attempted in this study to compare the prediction performance of several machine learning models to that of the Royal Thai Navys model. These models were the following: linear regression, kernel regression, support vector regression, k-nearest neighbors, and random forest. The data input into these models were water level time series data of past 24, 48, and 72 hours measured at the Royal Thai Navy headquarters station, Phra Chulachomklao Fort, thirteen other stations along the river, and the output were predictions for the next 24 hours. It was found that all of the machine learning techniques were able to achieve better performances than that of the harmonic method of tidal modeling. The support vector regression model with Radial basis function kernel and 72-hour past time series data yielded prediction results with the least errors, at 0.117 m and 0.116 m for the water levels at the Royal Thai Navy headquarters station and Phra Chulachomklao Fort, respectively.

We apply Elastic-net Multiple Kernel Learning (MKL) to the MDL Drug Data Report (MDDR) database for the problem of drug screening. We show that combining a set of kernels constructed from fingerprint descriptors, can significantly improve the accuracy of prediction, against a Support Vector Machine trained on each kernel separately. To the best of our knowledge, this is the first application of MKL to the MDDR database for drug screening.

In order to help users navigate an image search system, one could provide explicit information on a small set of images as to which of them are relevant or not to their task. These rankings are learned in order to present a user with a new set of images that are relevant to their task. Requiring such explicit information may not be feasible in a number of cases, we consider the setting where the user provides implicit feedback, eye movements, to assist when performing such a task. This paper explores the idea of implicitly incorporating eye movement features in an image ranking task where only images are available during testing. Previous work had demonstrated that combining eye movement and image features improved on the retrieval accuracy when compared to using each of the sources independently. Despite these encouraging results the proposed approach is unrealistic as no eye movements will be presented a-priori for new images (i.e. only after the ranked images are presented would one be able to measure a user's eye movements on them). We propose a novel search methodology which combines image features together with implicit feedback from users' eye movements in a tensor ranking Support Vector Machine and show that it is possible to extract the individual source-specific weight vectors. Furthermore, we demonstrate that the decomposed image weight vector is able to construct a new image-based semantic space that outperforms the retrieval accuracy than when solely using the image-features.

In this paper we predict the relevance of images based on a lowdimensional feature space found using several users' eye movements. Each user is given an image-based search task, during which their eye movements are extracted using a Tobii eye tracker. The users also provide us with explicit feedback regarding the relevance of images. We demonstrate that by using a greedy Nyström algorithm on the eye movement features of different users, we can find a suitable low-dimensional feature space for learning. We validate the suitability of this feature space by projecting the eye movement features of a new user into this space, training an online learning algorithm using these features, and showing that the number of mistakes (regret over time) made in predicting relevant images is lower than when using the original eye movement features. We also plot Recall-Precision and ROC curves, and use a sign test to verify the statistical significance of our results.