Graph and Neural Network-Based Intelligent Conversation System

Although neural conversation models are effective in learning how to produce fluent responses, their primary challenge lies in knowing what to say to make the conversation contentful and non-vacuous. We present a new end-to-end approach to contentful neural conversation that jointly models response generation and on-demand machine reading. The key idea is to provide the conversation model with relevant long-form text on the fly as a source of external knowledge. The model performs QA-style reading comprehension on this text in response to each conversational turn, thereby allowing for more focused integration of external knowledge than has been possible in prior approaches. To support further research on knowledge-grounded conversation, we introduce a new large-scale conversation dataset grounded in external web pages (2.8M turns, 7.4M sentences of grounding). Both human evaluation and automated metrics show that our approach results in more contentful responses compared to a variety of previous methods, improving both the informativeness and diversity of generated output.

Data-driven, knowledge-grounded neural conversation models are capable of generating more informative responses. However, these models have not yet demonstrated that they can zero-shot adapt to updated, unseen knowledge graphs. This paper proposes a new task about how to apply dynamic knowledge graphs in neural conversation model and presents a novel TV series conversation corpus (DyKgChat) for the task. Our new task and corpus aids in understanding the influence of dynamic knowledge graphs on responses generation. Also, we propose a preliminary model that selects an output from two networks at each time step: a sequence-to-sequence model (Seq2Seq) and a multi-hop reasoning model, in order to support dynamic knowledge graphs. To benchmark this new task and evaluate the capability of adaptation, we introduce several evaluation metrics and the experiments show that our proposed approach outperforms previous knowledge-grounded conversation models. The proposed corpus and model can motivate the future research directions.

Dynamic scheduling is a complicated problem with high research value in the production field, because it requires a fast response to the dynamic disturbance event. In this paper, knowledge graph is combined with reinforcement learning to solve this problem. In the proposed method, knowledge graph model is constructed by a series of steps, including ontology-based knowledge modeling, knowledge extraction, and knowledge fusion. In the knowledge graph model, the state information of the shop floor is extracted from the production data, including the information about machine, job, and process. On this basis, reinforcement learning model is designed, which is composed of machine assignment agent and job sequence agent. Thus, the real-time scheduling decisions can be made rapidly under the dynamic production environment. The effectiveness of the proposed method is verified on a case study of aviation component production workshop. Compared with other methods, the proposed method obtains the best results.

Timely and accurate decision-making is critical during gas tunnel emergencies, yet relevant knowledge is typically fragmented across unstructured reports, manuals, and case records, hindering rapid access. To address this challenge, a knowledge recommendation approach was proposed that synergistically combines knowledge graph (KG) and large language model (LLM). This method first constructs a domain-specific KG from heterogeneous emergency documents using prompt-optimized LLMs. During inference stage, it retrieves the most contextually relevant subgraph via semantic similarity-based vector search and re-ranks candidates to enhance precision. This retrieved knowledge dynamically augments the LLM, grounding its responses in verified domain facts and reducing hallucinations. Evaluations show that approach substantially outperforms both standalone LLMs and alternative architectures in accuracy, completeness, and coherence. The method proposed in the study improves the speed and quality of emergency response for gas tunnels, offering a structured and interpretable approach to knowledge-driven decision support in this specific high-risk context.

Machine Reading Comprehension (MRC) has garnered significant attention in the field of natural language processing. While the maritime domain lacks openly available MRC task models and corresponding datasets, resulting in a deficiency within existing models to provide accurate answers tailored to the maritime domain. In response to these issues, our work uses hand annotation and data augmentation to create the OurShipSet machine reading comprehension dataset, which is tailored for the maritime industry. Additionally, we propose the KGAT-MRC model, which integrates relevant knowledge triplets from the domain to enhance the understanding of text semantics. In order to mitigate the issue of noise introduced by incorporating knowledge triplets, this paper embeds the most relevant knowledge triplets to the questions, ensuring that the impact of noisy questions on the model is minimized through computation-based selection. Experimental results demonstrate that the proposed approach achieves an EM (Exact Match) score of 70.69% and an F1 score of 87.94%, representing improvements of 4.22% and 2.47% respectively over existing methods.

Carbonate bioreef formations serve as crucial hydrocarbon reservoirs, and their accurate identification bears significant implications for oil and gas exploration. Moreover, the precise and refined delineation of prograding body structures aids in the comprehensive analysis of stratigraphic geologic configurations. We develop the knowledge graph and geologic strata interpolation constraints (KGGSICs) model for the intricate identification of carbonate bioreefs and prograding body structures. Furthermore, we assess our KGGSIC-Unet architecture on the Dengying Formation Sections 3-4 carbonate bioreefs and prograding bodies in the Moxi area of the Sichuan Basin. Experimental results indicate that the KGGSIC enhances the predictive performance of the U-Net and realizes the precise and refined segmentation of carbonate bioreefs and prograding body structures. In addition, through a meticulous geologic study of the area, we synthesize the 2D profile identification results to achieve the precise and refined identification of carbonate bioreefs and prograding bodies.

Investigation of a Recipe Recommendation System for Patients with Underlying Medical Conditions Based on Knowledge Graph and Large Language Model

With the booming development of online education, the number of course resources has shown a geometric growth, and the surge in course resources has also brought challenges in management and personalized learning. As an important component of online course resources, question banks can provide students with a large number of practice questions to help them identify and reinforce their knowledge. However, current question bank systems often lack intelligent and personalized push mechanisms. Knowledge tracking is one of the important means to achieve smart education. This article is based on a knowledge graph and uses the knowledge tracking improvement model KMAKT. The TransR graph embedding algorithm is used to represent the course knowledge graph embedding, and the algorithm is used to obtain the degree of student knowledge point mastery. The deep neural network (DNN) attention network is used to determine the influence of precursor knowledge, and the multi-layer neural network is used to obtain the predicted output of exercise answer results. The exercise recommendation system, taking Java Web application development course as an example, has achieved accurate and personalized exercise recommendations. The experimental results have verified the effectiveness and practicality of the system.

With the explosive growth of online courses, per-sonalized learning path recommendations have gained much attention in E-learning area. Studies aim to provide customized guidance that cater to the needs of each learner. However, generating an appropriate learning path remains a challenge due to the complex cognitive relationships among learning items and learners' different abilities. Existing methods often overlook the impacts of cognitive structure and mastery level on the outcomes of customized instructions, leading to cognitive confusion problem. To address this issue, we propose a model named Data-Driven Learning Path based on Knowledge Graph and Tracing Model (LP-KAT), which combines knowledge graph (KG) and knowledge tracing model (KT) to store the relationships among learning items and reflect learners' abilities through the abundant data. The guidance is separated into concept and practice identification; besides, through reinforcement learning, our model can suggest a sequence of practices for different learners. Specifically, we first utilize a knowledge tracing model to track learners' evolving mastery levels at each learning step. Then we design a concept selector that employs knowledge graph navigation algorithm and reinforcement learning to determine which concept to learn next. Finally, we utilize a practice selector to identify practice question based on the relationship between mastery level and difficulty of practice. We evaluate the proposed model against KNN, GRU4Rec, MCS and DQN using a real-world dataset, and the results show an improvement of more than 30 % in the effectiveness of concept mastery level and nearly 20% better than GRU4Rec across various learning path lengths.

The significant shift in consumer financial complaints points out the need for intelligent systems that can reliably categorize complaints and identify fraudulent submissions. This study's hybrid NLP framework uses a lightweight Knowledge Graph (KG) and Large Language Model (LLM) embeddings to improve false complaint detection and complaint classification. Two different classification models compare them: a plain LLM model that uses sentence-transformer embeddings (all-MiniLML6-v2) and a hybrid LLM+KG model which fuses domainspecific keyword mappings (β = 0.25) with embedding similarity (α = 0.75). The hybrid approach achieves higher contextual accuracy and interpretability, according to the results, especially for checking account, credit card, and mortgage categories. Three methods—a supervised hybrid model, a KG-based heuristic, and an embedding-based heuristic—are assessed for false complaint detection; the supervised hybrid approach produces the best results (F1-score 0.80, 100% recall). The accuracy, dependability, and interpretability of automated financial complaint analysis are all positively impacted by the hybrid LLM+KG classification and supervised hybrid detection.

Adversarial training and decoding strategies for end-to-end neural conversation models

The development of a computer-aided design (CAD) system for neutralizing emissions from energy source enterprises and production enterprises requires the creation of an appropriate knowledge base. This article discusses the construction of such a knowledge base and its application in CAD system for neutralizing emissions from energy source enterprises and production enterprises. Using development tools such as Protégé (Java), PyCharm (Python) and Navicat (MySQL), a knowledge base and a client platform for the CAD system for neutralizing emissions from energy source enterprises and production enterprises were developed. A key technical framework for building the CAD system for neutralizing emissions from energy source enterprises and production enterprises knowledge base is developed, covering the main aspects of knowledge acquisition, representation and dynamic updating. The accuracy and efficiency of knowledge retrieval were significantly enhanced through ontological modeling, which enabled the analysis of associations within diverse datasets. During the development of the knowledge base, logical inference rules were created and transformed into a knowledge graph model. This approach, combined with the dynamic characteristics of the knowledge graph, facilitates continuous updating of the knowledge base. The practical value of applying the knowledge base in complex scenarios was validated through its implementation in the system for neutralizing emissions from energy source enterprises and production enterprises. The developed knowledge base is an integral component of the CAD system for neutralizing emissions from energy source enterprises and production enterprises, aimed at addressing emission-related challenges. The CAD system for neutralizing emissions from energy source enterprises and production enterprises knowledge base solves the key problems of knowledge fusion, data migration and reasoning validity, and provides theoretical support and technical path for building an intelligent and scalable CAD system for neutralizing emissions from energy source enterprises and production enterprises knowledge service system.

Neural conversation models are known to generate appropriate but non-informative responses in general. A scenario where informativeness can be significantly enhanced is Conversing by Reading (CbR), where conversations take place with respect to a given external document. In previous work, the external document is utilized by (1) creating a context-aware document memory that integrates information from the document and the conversational context, and then (2) generating responses referring to the memory. In this paper, we propose to create the document memory with some anticipated responses in mind. This is achieved using a teacher-student framework. The teacher is given the external document, the context, and the ground-truth response, and learns how to build a response-aware document memory from three sources of information. The student learns to construct a response-anticipated document memory from the first two sources, and the teacher's insight on memory creation. Empirical results show that our model outperforms the previous state-of-the-art for the CbR task.

LUKE-Graph: A Transformer-based Approach with Gated Relational Graph Attention for Cloze-style Reading Comprehension

Transformer operation and maintenance is very important for the stable operation of the power system, at present, the staff has accumulated a large number of transformer operation and inspection reports, how to effectively use the above data to achieve the prediction and solution of transformer faults is the key to improve the efficiency of transformer operation and maintenance. This paper constructs a transformer intelligent fault diagnosis system that integrates knowledge graph and large language model. Through the fault text preprocessing, knowledge graph construction and the invocation of the large language model to build the transformer intelligent fault diagnosis system, and finally through the entity relationship extraction effect comparison and Q&A system subjective and objective performance evaluation, it is verified that the system can be targeted to solve the information correlation weakness and decision-making generation efficiency in the process of transformer operation and maintenance, and significantly improve the health management level of the transformer.