In addressing challenges within the field of Natural Language Processing (NLP), supervised fine-tuning is an efficient technique that allows pre-trained Large Language Models to adapt to specific tasks. This is especially crucial for low-resource languages, such as Tibetan, where the demand for high-quality fine-tuning datasets is particularly pronounced. This paper introduces the Tibetan Instruction-Following Dataset (TIFD), comprising 11,535 JSON objects, each with four attributes: a unique identifier, instructions, input, and output. These attributes correspond to the task’s unique ID, the instruction describing the task, supplementary input for the task instruction, and the response generated by GPT-4. In the data preprocessing phase, unsuitable data for Tibetan fine-tuning were first filtered out. To ensure the diversity of instructions, the Tibetan instructions were effectively vectorized using LaBSE, followed by the selection of diverse Tibetan instruction data using a method based on the K-Center-Greedy algorithm. After preliminary English-Tibetan machine translation of responses generated by GPT-4, a web program was developed to facilitate Tibetan proofreaders in referencing the original English and corresponding Chinese translations for rectifying errors in words, grammar, and structure in the Tibetan data. Subsequently, five Tibetan language professionals were invited to manually proofread the data, checking and modifying the data format and Tibetan content, and removing erroneous and meaningless tasks, thereby ensuring the quality and reliability of the TIFD. Using TIFD to supervise the fine-tuning of the Tibetan large language model TiLamb has enhanced TiLamb’s Tibetan dialogue and instruction-following capabilities. The TIFD serves as a vital resource for Tibetan data, playing a significant role in the development of Tibetan NLP technology and in the effective customization of Large Language Models (LLMs).

The Data Market Management Strategy project proposes a comprehensive framework to harness AI technologies for optimizing data-driven decision-making processes. This framework, illustrated as an integrated ecosystem, underscores the importance of data and model reuse through a structured marketplace environment. However, challenges such as data standardization, interoperability, and privacy concerns remain prevalent in current data markets. For instance, many data platforms still suffer from data silos and inconsistent metadata standards, making it difficult for researchers to efficiently access and reuse data across sectors.Addressing these issues, the proposed system integrates a data market and a model marketplace, facilitating seamless information exchange through the Taiwan Computing Cloud (TWCC). Within this ecosystem, users can generate new models, upload, and share data, contributing to a dynamic and continuously evolving repository. The system enables users to access diverse datasets via standardized APIs and develop advanced models within modular containers such as Jupyter Notebooks.The model marketplace serves as a critical hub, supporting AI model sharing, refinement, and lifecycle management, fostering an environment where data and models are continuously reused. By emphasizing interdisciplinary collaboration, the framework enhances resource utilization, mitigates redundant efforts, and accelerates the development of novel AI solutions. The proposed approach aligns with global trends in federated learning, data privacy-preserving techniques, and open AI model hubs (e.g., Hugging Face, TensorFlow Hub), ensuring ethical and secure data practices. Ultimately, the framework promotes scalable AI-powered applications, contributing to a more sustainable future in data management.

Tibetan medicine, as one of the oldest medical systems in the world, is an integral part of traditional Chinese medicine. However, its knowledge is primarily preserved in traditional formats, limiting its digitization, structuring, and intelligent analysis. In this paper, we construct a structured dataset for Tibetan medicine, covering 1,962 diseases, 125 causes, 3,919 symptoms, and 995 treatment methods. Existing Named Entity Recognition (NER) methods typically overlook the uniqueness of the Tibetan medicine domain and fail to fully account for the differences between Tibetan medicine and general domains. To better adapt the model to this specific domain, we introduced a domain-aware prompt fine-tuning strategy. Experimental results show that our model achieved an F1 score of 93.33% in the Tibetan medicine domain, with an improvement of at least 5.2% compared to the baseline model. Ultimately, we constructed a Tibetan medicine disease knowledge graph consisting of 7,139 entity nodes and 9,300 entity relationships, providing a structured representation of Tibetan medical knowledge. We also performed data analysis using similarity calculation and topological link prediction algorithms, advancing its intelligent analysis process.

Focusing on the construction and application of research entity datasets based on scientific literature, it emphasizes the crucial role of data quality in enhancing natural language processing applications in scientific literature mining, addressing the gaps in detailed descriptions and evaluations of datasets in existing research, and promoting advancements in scientific literature entity recognition technology. Through extensive online research and literature review, 22 open-source research entity datasets were analyzed, focusing on the dataset lifecycle’s collection, annotation, release, and application stages. A set of quality assessment methods for research entity datasets is proposed, along with a discussion of the challenges in dataset construction and strategies for efficient use. The study emphasizes the importance of high-quality datasets for natural language processing applications, proposing evaluation methods and sharing strategies intended to serve as a reference for constructing, selecting, and using scientific literature entity datasets.

The scarcity of Tibetan handwriting datasets has hindered the applications of prevailing artificial intelligence models in the Tibetan language. As Khyug-yig is the most common writing style found in the daily lives of Tibetan people, this study proposes a methodology to construct a Tibetan handwritten Khyug-yig dataset to support further research in Tibetan fields. This approach starts with filtering the textual content of writings from multiple sources, encompassing news, medicine, and Buddhism, establishing a corpus of frequent Tibetan words. These words were organized into forms and assigned to 63 Tibetan writers across diverse institutions, including Changdu City’s Sixth Senior High School, Tibet University, and a local calligraphy association. The collected handwriting forms were then processed through scanning, cropping, image preprocessing, grouping, and labeling. As a result, a Tibetan handwriting dataset with 9,874 unique-word images written in Khyug-yig style was constructed, overcoming the limitation of existing Tibetan handwriting datasets, while achieving calligraphic diversity and precise labeling.

Although neural methods have been comprehensively applied in different fields, symbolic based logic reasoning is still the main choice for numerous applications based on knowledge graphs. To enhance the efficiency of knowledge graph reasoning, researchers studied how to design parallel algorithms for reasoning, and take advantage of high-performance architectures, like neural networks. Although parallel algorithms and architectures improve the performance of reasoning to some degree, the task of reasoning is essentially bounded by its computational complexity, i.e., the PT<sc>IME</sc>-Completeness or higher complexities. This means that the task of reasoning is not parallelly tractable. In this work, we investigate the parallel tractability of knowledge graph reasoning from the perspective of parallel complexity. We concentrate on knowledge graphs that are Datalog rewritable. We aim to capture the parallelly tractable classes of knowledge graphs, for which, the task of reasoning falls in the NC complexity. To this end, we employ the computational model of <italic>Boolean circuit</italic> to formalize knowledge graph reasoning and further obtain all the theoretical results. We then use the results to analyze DHL (<italic>Description Horn Logic</italic>), a fragment of description logic. We give the properties that ensure the parallel tractability of DHL reasoning. One can utilize our results to check the parallel tractability of real knowledge graphs. In addition, the Boolean circuits proposed in this paper can also be used to construct neural networks to perform knowledge graph reasoning.

Frame-semantic parsing (FSP) aims to extract frame-semantic structures from text. The taskinvolves three subtasks: target identification (TI), frame identification (FI) and frame-semanticrole labeling (FSRL). Most of the previous studies predominantly focus one or two of the threesubtasks, and limited studies focus on constructing a unified framework to systematically solvethe whole FSP task. In addition, how to efficiently utilize lexical units knowledge and improvethe span representations for FSP is still an open question. Therefore, this work presents a novelunified framework for FSP based on marker attention, in which special markers are strategically inserted into the input sequence of the encoder to incorporate the lexical units knowledgeand frame information, as well as to enhance the span representation through attention mechanisms, without using any additional complicated module. Experiments on two public benchmarks demonstrate the effectiveness of our proposed framework.

Abstract Background: During the COVID-19 pandemic, clinical trial recruitment cannot be carried out due to travel restrictions, transmission risks and other factors, resulting in the stagnation of a large number of ongoing or upcoming clinical trials. Objective: An intelligent screening app was developed using artificial intelligence technology to rapidly pre-screen potential patients for phase I solid tumor drug clinical trials. Methods: A total of 429 screening process records were collected from 27 phase I solid tumor drug clinical trials at the First Affiliated Hospital of Bengbu Medical College from April 2018 to May 2021. Features of the experimental data were analyzed, and the collinearity (principal component analysis) and strong correlation (χ2 test) among features were eliminated. XGBoost, Random Forest, and Naive Bayes were used to sort the weight importance of features. Finally, the pre-screening models were constructed using classification machine learning algorithm, and the optimal model was selected. Results: Among the 429 screening records, 33 were data generated by repeated subject participation in different clinical trials, and of the remaining 396 screening records, 246 (62.12%) were screened successfully. The gold standard for subject screening success is the final judgment made by the principal investigator (PI) based on the clinical trial protocol. A Venn diagram was used to identify the important feature intersections of machine learning algorithms. After intersecting the top 15 characteristic variables of different feature screening models, 9 common variables were obtained: age, sex, distance from residence to the central institution, tumor histology, tumor stage, tumorectomy, the interval from diagnosis/postoperative to screening, chemotherapy, and ECOG (Eastern Cooperative Oncology Group, ECOG) score. To select the optimal subset, the 9 important feature variables were expanded to 12 and 15 feature subsets, and the performance of different feature subsets under different machine learning models was validated. The results showed that optimal performance, accuracy and practicability were achieved using XGBoost with the 12 feature subset. The final model could accurately predict the screening success rates in both internal (AUC =3D 0.895) and external (AUC =3D 0.796) validation, and has been transformed into a convenient tool to facilitate its application in the clinical settings. Subjects with a probability exceeding or equaling to the threshold in the final model had a higher probability to be successfully screened. Conclusion: Based on the optimal model, we created an online prediction calculator and visualization app -- ISSP (Intelligent Screening Service Platform), which can rapidly screen patients for phase I solid tumor drug clinical trials. ISSP can effectively solve the problem of space and time interval. On the mobile terminal, it realizes the matching between clinical trial projects and patients, and completes the rapid screening of clinical trial subjects, so as to obtain more clinical trial subjects. As an auxiliary tool, ISSP optimizes the screening process of clinical trials and provides more convenient services for clinical investigators and patients.

ABSTRACT Knowledge graphs are employed in several tasks, such as question answering and recommendation systems, due to their ability to represent relationships between concepts. Automatically constructing such a graphs, however, remains an unresolved challenge within knowledge representation. To tackle this challenge, we propose CtxKG, a method specifically aimed at extracting knowledge graphs in a context of limited resources in which the only input is a set of unstructured text documents. CtxKG is based on OpenIE (a relationship triple extraction method) and BERT (a language model) and contains four stages: the extraction of relationship triples directly from text; the identification of synonyms across triples; the merging of similar entities; and the building of bridges between knowledge graphs of different documents. Our method distinguishes itself from those in the current literature (i) through its use of the parse tree to avoid the overlapping entities produced by base implementations of OpenIE; and (ii) through its bridges, which create a connected network of graphs, overcoming a limitation similar methods have of one isolated graph per document. We compare our method to two others by generating graphs for movie articles from Wikipedia and contrasting them with benchmark graphs built from the OMDb movie database. Our results suggest that our method is able to improve multiple aspects of knowledge graph construction. They also highlight the critical role that triple identification and named-entity recognition have in improving the quality of automatically generated graphs, suggesting future paths for investigation. Finally, we apply CtxKG to build BlabKG, a knowledge graph for the Blue Amazon, and discuss possible improvements.