Prediction and illusion detection method for large language model content generation in field of science and technology intelligence

Abstract

The invention discloses a prejudice and illusion detection method for large language model content generation in the field of science and technology intelligence, relates to the technical field of large model prejudice treatment, and aims to solve the problem of decision misjudgment caused by large model prejudice and illusion in the field of science and technology. According to the scheme, a specialized support system comprising a scientific fact library and professional prejudice resources is constructed, a three-layer progressive full-link detection traceability architecture is adopted, and semantic feature extraction, prejudice dynamic detection and cause positioning are sequentially realized through data preprocessing, real-time detection and traceability analysis; factual verification is enhanced by combining multi-source evidence chain verification, multi-view analysis and multi-dimensional evaluation. According to the method, accurate identification and full-link traceability of specific prejudice in the science and technology field can be realized, prejudice core causes are defined, targeted optimization suggestions are output, the output credibility and detection stability of a large model are remarkably improved, and high-reliability quality control support is provided for scenes such as science and technology intelligence analysis and scientific research trend prediction.

Description

Technical Field

[0001] This invention relates to the field of large model bias governance technology, and in particular to a method for detecting bias and illusion in the content generation of large language models in the field of scientific and technological intelligence. Background Technology

[0002] With the widespread application of large-scale models in scenarios such as science and technology intelligence analysis, scientific research trend prediction, and technical consulting, the objectivity and authenticity of their generated content directly affect the scientific nature of scientific research decisions. However, current large-scale models for the science and technology field generally suffer from two major problems: first, bias, meaning that the generated content has an unfair or discriminatory tendency; and second, illusion, meaning that the generated content contradicts the truth.

[0003] These two types of problems mainly stem from three core aspects. First, the pre-training data is flawed. The scientific corpora that large models rely on, such as academic literature and patent texts, naturally contain systemic biases. For example, descriptions of "scientists" are often predominantly male, and authoritative databases like Web of Science tend to include academic achievements from Europe and America, leading to an imbalance in the distribution of training data. Second, fine-tuning and human feedback can easily introduce biases during reinforcement learning. Specialized fine-tuning for specific tasks in the scientific field can reduce the model's generalization ability. Coupled with the subjective cognitive biases of the annotators in human feedback, it can easily amplify the model's misjudgments of specific technologies or regional academic achievements. Third, the generation mechanism has limitations. The autoregressive generation method of large models can easily lead to the gradual accumulation of erroneous information, and the model struggles to accurately identify its own capability boundaries. For example, in the analysis of medical-related scientific and technological intelligence, the model may overuse the general conclusion that "coughing and fever are mostly cold symptoms," ignoring key specific information mentioned by users such as difficulty breathing and chest pain, thus outputting incorrect judgments.

[0004] Current large-scale model bias detection technologies have significant shortcomings in adapting to the needs of the scientific and technological field. On the one hand, there is a lack of domain-specific detection solutions. Existing technologies are mostly geared towards general scenarios, failing to integrate authoritative data sources such as scientific literature, patents, and technical standards as detection benchmarks, and lacking dedicated detection logic for issues specific to the scientific and technological field, such as biases in academic data sources and overgeneralization of scientific conclusions. This results in low detection accuracy and a high false positive rate. On the other hand, the detection scope is limited to the generated results, only identifying bias in the final content generated by the large model. It is impossible to trace the propagation path of bias from training data to fine-tuning and then to the generated results, making it difficult to pinpoint the root cause and hindering model optimization. Furthermore, the detection methods are relatively static, often based on one-time verification on a fixed test set, which cannot adapt to the semantic distribution drift during the iteration process of large models, resulting in insufficient robustness. At the same time, there is a lack of multi-source evidence cross-validation mechanisms, making it difficult to distinguish between bias and illusion. There are also no technical means to reverse-map bias patterns to training data or the fine-tuning process, making it impossible to clarify the bias evolution path and resulting in low model correction efficiency.

[0005] Therefore, there is an urgent need for a large-scale model bias control technology that can be deeply adapted to the characteristics of the science and technology field, cover the entire chain from training data to generated content, and support dynamic detection and accurate source tracing, so as to solve the cognitive bias and decision-making misjudgment caused by bias and illusion in science and technology intelligence analysis, and ensure the reliability of large-scale models in the application of science and technology fields. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting bias and illusion in the content generation of large language models in the field of scientific and technological intelligence. It aims to solve the problems of inaccurate identification of bias and illusion, inability to trace the causes of bias, insufficient detection adaptability, and lack of closed-loop optimization schemes in the application of large models in the field of science and technology. It achieves accurate identification of biases specific to the field of science and technology, full-link tracing of causes and quantitative credibility verification, and outputs targeted model optimization suggestions.

[0007] To achieve the above objectives, this invention provides a method for detecting bias and illusion in the content generation of large language models in the field of scientific and technological intelligence, the steps of which are as follows: S1. Construct a professional support system in the field of science and technology that includes a scientific fact base and professional bias resources, and provide detection benchmarks and feature support; S2. A three-layer progressive full-link bias detection and tracing architecture is adopted, which realizes semantic feature extraction, dynamic bias detection and cause tracing in sequence through data preprocessing layer, real-time monitoring layer and tracing analysis layer; S3. Through multi-source evidence chain verification, multi-perspective analysis and multi-dimensional comprehensive evaluation, the fact verification and credibility determination of the large model output are realized. S4. Based on the detection, tracing and evaluation results, output model optimization suggestions for the causes of bias.

[0008] Preferably, the construction of the scientific fact base includes: integrating a scientific and technological literature database, a patent data path, and a technical standard database; using hybrid neural machine translation to achieve multilingual alignment; constructing a domain ontology database based on graph databases and graph neural network embedding methods; and achieving unified semantic representation and association fusion of heterogeneous data through a cross-modal semantic mapping model. The heterogeneous data includes text, icons, and formulas.

[0009] The preferred approach to constructing professional bias resources includes: forming a professional bias dictionary by manually annotating bias scenarios and mining pre-training corpora. The professional bias dictionary contains words related to subject terminology preferences, institutional weighting tendencies, and regional academic resource biases. Features are extracted from three dimensions: semantics, syntax, and domain attributes. A high-dimensional bias feature space with more than 200 dimensions is constructed and dimensionality is reduced using principal component analysis.

[0010] Preferably, in S2, the data preprocessing layer extracts semantic features through the BERT model and generates a concept relationship graph by combining dependency parsing. The real-time monitoring layer adopts a parallel dual-channel detection network. Channel 1 tracks semantic distribution changes through temporal LSTM and sets dynamic thresholds for early warning. Channel 2 detects sensitivity and robustness through a lightweight adversarial classifier model. The source analysis layer uses attention backtracking mechanism and causal intervention analysis to locate the distribution and evolution path of training data corresponding to bias.

[0011] Preferably, the specific process of S3 includes: Multi-source evidence chain verification uses a graph structure retrieval algorithm to retrieve authoritative evidence from a scientific fact base and construct an evidence chain. Multi-perspective analysis achieves cross-validation by comparing the outputs of large models with different architectures; The multi-dimensional comprehensive evaluation system covers four core dimensions: content credibility, logical consistency, domain adaptability, and knowledge novelty. Each dimension constructs a hierarchical indicator tree, uses the analytic hierarchy process to determine the weights, and integrates the results through the fuzzy comprehensive evaluation method.

[0012] The preferred method for constructing the hierarchical indicator tree is as follows: content credibility is decomposed into three secondary indicators: factual accuracy, data traceability, and argumentation completeness; logical consistency is decomposed into three secondary indicators: uncontradictory concepts, unbroken reasoning, and unambiguous expression; domain adaptability is decomposed into three secondary indicators: terminology accuracy, scenario fit, and technological relevance; and knowledge freshness is decomposed into three secondary indicators: data timeliness, results update rate, and standard adaptability.

[0013] Preferably, in S4, the model optimization suggestions include: data source supplementation and bias filtering schemes to address the deficiencies of pre-training data, sample balancing and feedback mechanism optimization strategies to address fine-tuning biases, and output verification enhancement and capability boundary labeling measures to address the limitations of the generation mechanism.

[0014] A bias and illusion detection system for content generation using large language models in the field of science and technology intelligence, including: The support system module is used to build and store scientific fact bases and professional bias resources, and to provide detection benchmarks and feature support. The three-layer progressive full-link bias detection and tracing architecture includes a data preprocessing unit, a real-time monitoring unit, and a tracing analysis unit, which respectively perform semantic feature extraction, dynamic bias detection, and cause tracing operations. The verification and evaluation module is used to perform multi-source evidence chain verification, multi-perspective analysis, and multi-dimensional comprehensive evaluation. The optimization suggestion output module is used to generate and output model optimization suggestions based on the detection, tracing and evaluation results.

[0015] Preferably, the support system module has built-in data source integration unit, multilingual alignment unit, ontology library construction unit, cross-modal mapping unit, as well as dictionary compilation unit, feature extraction unit, and feature dimensionality reduction unit; The data source integration unit, multilingual alignment unit, ontology library construction unit, and cross-modal mapping unit work together to complete the integration of the scientific fact library, multilingual alignment, ontology library construction, and heterogeneous data fusion; the scientific fact library is stored in a graph database, supporting efficient retrieval and real-time updates of heterogeneous data; The dictionary compilation unit, feature extraction unit, and feature dimensionality reduction unit work together to complete the compilation of a professional bias dictionary, feature extraction, and dimensionality reduction of the high-dimensional feature space.

[0016] In the preferred three-layer progressive end-to-end bias detection and tracing architecture, The data preprocessing unit incorporates a BERT encoding subunit and a syntactic analysis subunit, which respectively implement semantic feature extraction and dependency parsing to generate a concept relationship graph. The real-time monitoring unit incorporates a temporal LSTM semantic tracking subunit and an adversarial classification detection subunit, enabling parallel semantic distribution tracking and early warning, as well as model output sensitivity and robustness detection. The source analysis unit incorporates an attention backtracking subunit and a causal intervention analysis subunit, which work together to locate the training data corresponding to bias and analyze the evolution path.

[0017] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) By constructing a specialized support system for the science and technology field, it has achieved accurate adaptation and identification of biases specific to science and technology scenarios. The scientific fact base integrates authoritative resources such as scientific and technological literature, patents, and technical standards, and combines cross-modal semantic mapping and domain ontology modeling to provide benchmark support for detection that fits the characteristics of the science and technology field; the professional bias resources are designed for typical scenarios in the science and technology field such as subject terminology preferences and regional academic resource biases, which effectively solves the problems of insufficient adaptability and ambiguous identification of general bias detection technology in professional scenarios, making bias detection more in line with the actual application needs of science and technology intelligence analysis, scientific research trend prediction, etc.

[0018] (2) Based on a three-layer progressive full-link detection and source tracing architecture, the complete bias tracking and cause localization from training data to generated content is realized. The hierarchical collaboration of data preprocessing, real-time detection and source tracing analysis, combined with techniques such as attention backtracking and causal intervention analysis, can accurately pinpoint whether the bias originates from defects in pre-training data, fine-tuning bias, or limitations in the generation mechanism, breaking the limitation of traditional techniques that can only detect bias at the output end. At the same time, the solution provides targeted and implementable model optimization suggestions, forming a closed loop of "detection-source tracing-optimization", which can achieve bias correction without large-scale model reconstruction, greatly improving the efficiency and practicality of model iteration.

[0019] (3) Through a triple mechanism of multi-source evidence chain verification, multi-perspective cross-verification, and multi-dimensional comprehensive evaluation, the consistency and credibility of the facts output by the large model are significantly enhanced, effectively distinguishing between bias and illusion, and avoiding misjudgments caused by cognitive bias in scientific and technological decision-making. The parallel design of dynamic threshold warning and adversarial detection allows the detection system to adapt to the semantic distribution changes during the iteration process of the large model, ensuring the stability of detection in long-term applications; the hierarchical evaluation system realizes the quantitative judgment of credibility and accurate risk warning, helping users to quickly judge the reference value of the model output, and providing strong support for highly reliable decision-making in the field of science and technology.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating an embodiment of the bias and illusion detection method for generating large language model content in the field of scientific and technological intelligence according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example The hardware environment is configured with a high-performance CPU, GPU, and large-capacity memory and storage devices to meet the needs of large-scale model inference, feature extraction, and database retrieval. This includes: It features dual Intel Xeon Platinum 8480C processors (48 cores and 96 threads per processor, 2.0GHz), providing powerful parallel computing capabilities to meet the high computing demands of multimodal data processing, model inference, and feature extraction; it is equipped with four NVIDIA A100 80GB GPUs, supporting FP16 / FP32 mixed-precision computing to accelerate deep learning tasks such as BERT model encoding, temporal LSTM tracking, and adversarial classifier training; and it has 512GB of DDR5 4800MHz memory to ensure the memory bandwidth and capacity requirements for large-scale corpus loading, graph database retrieval, and multi-model parallel inference.

[0026] The storage resources adopt an all-flash storage architecture, configured with a 4TB NVMe SSD local cache + a 20TB distributed file storage system. The local cache is used to store core data of the scientific fact base, professional bias dictionary and intermediate results of the model that are frequently accessed. The distributed storage is used to archive complete scientific and technological literature, patent data and training logs. The read and write speed can reach 2GB / s, and it supports efficient read and write and real-time updates of heterogeneous data.

[0027] The software environment utilizes mainstream operating systems, deep learning frameworks, graph databases, natural language processing toolkits, and data processing tools. These include: The system utilizes Ubuntu 22.04 LTS to provide a stable operating environment. Python 3.9 is installed, along with the PyTorch 2.0 deep learning framework and TensorFlow 2.10 to assist model training. The Hugging FaceTransformers library is used to load pre-trained BERT models and various large-scale model architectures. Neo4j 5.0 graph database is selected to store the domain ontology and heterogeneous data relationships of the scientific fact base, supporting graph structure retrieval and complex semantic queries. NLTK 3.8 and spaCy 3.5 natural language processing toolkits are integrated for tasks such as dependency parsing and term extraction. Pandas, NumPy, and Scikit-learn are used for data preprocessing, feature engineering, and evaluation metric calculation.

[0028] Methods for detecting biases and illusions in content generation using large language models in the field of science and technology intelligence, such as... Figure 1 As shown, the steps are as follows: S1. Building a professional support system in the field of science and technology. The Science Fact Repository integrates authoritative resources such as the Web of Science Core Collection, CNKI Academic Literature Repository, USPTO Patent Database, and ISO / IEC Technical Standards Repository, covering 12 sub-fields of science and technology, including agriculture, medicine, and engineering technology. It has accumulated more than 8 million documents, more than 5 million patents, and more than 100,000 technical standards, with a data time span from 2010 to 2024. It includes heterogeneous data such as text, charts, and formulas to ensure the comprehensiveness and freshness of knowledge coverage.

[0029] Through multilingual alignment, domain ontology library construction, and cross-modal semantic mapping, unified semantic representation and association fusion of heterogeneous data are achieved.

[0030] Multilingual alignment employs a hybrid neural machine translation model (a fusion of Transformer architecture and traditional statistical translation model) to perform alignment processing for six mainstream scientific and technological languages, including English, Chinese, German, and Japanese, thereby solving the semantic consistency problem of multilingual scientific and technological documents. The domain ontology is constructed using a graph neural network embedding method (GraphSAGE), resulting in a domain ontology containing over 100,000 core concepts and over 500,000 concept relationships. Concepts cover subject classifications, technology categories, research methods, and institution names, while relationships include hierarchical relationships, causal relationships, and application relationships. The ontology is stored in the graph database Neo4j, supporting efficient concept retrieval and association reasoning based on the Cypher query language.

[0031] The aligned text is segmented, part-of-speech tagging is performed, and entity recognition is conducted to extract core information such as subject-specific terminology, technical parameters, and research conclusions. Semantic features are extracted from text data using the BERT model, visual features from chart data are extracted using CNN and converted into structured data using OCR technology, and feature vectors from formula data are generated through LaTeX parsing and mathematical symbol embedding. A cross-modal semantic mapping model (an improved CLIP architecture) maps the feature vectors of the three types of data to a unified semantic space, achieving the fusion of heterogeneous data.

[0032] The professional bias resource is formed by manually annotating common bias scenarios in the agricultural field and combining them with pre-trained corpus mining to form a professional bias dictionary. Features are extracted from three dimensions: semantics, syntax, and domain attributes, and a high-dimensional bias feature space is constructed and then subjected to dimensionality reduction processing.

[0033] S2, Three-layer progressive full-link detection and traceability architecture operation The data preprocessing layer extracts semantic features from the input and output texts of the large model using the BERT model, and generates a concept relationship graph by combining dependency parsing. An improved BERT model based on RoBERTa is used to semantically encode the input and output texts of the large model, generating a 768-dimensional semantic feature vector. During the pre-training stage, specialized tasks such as technical term masking and technical relationship prediction are added to enable the model to more accurately capture the semantic information of technical texts. Based on the spaCy dependency parser, core nouns, verbs, adjectives, and other components in the text are identified, and syntactic relations such as subject-verb, verb-object, and modifier-head are extracted. Combined with a domain ontology, syntactic relations are mapped to concept association relations, generating a concept relationship graph containing 10-100 nodes and 20-200 edges. Nodes represent domain concepts, and edges represent the semantic association strength between concepts, providing structured semantic support for subsequent bias detection.

[0034] The real-time monitoring layer employs a parallel dual-channel detection network: Channel 1 uses a temporal LSTM to track changes in semantic distribution and sets dynamic threshold warnings. It constructs a two-layer temporal LSTM network with the input being a sequence of semantic feature vectors generated by the data preprocessing layer and the hidden layer dimension set to 512. By tracking the dynamic changes in semantic distribution, it captures the evolution trend of bias tendencies. An adaptive dynamic threshold algorithm is adopted to adjust the warning threshold in real time based on historical detection data and domain bias feature distribution. When the semantic distribution deviates from the normal range and exceeds the threshold, a bias warning is triggered.

[0035] Channel 2 uses a lightweight adversarial classifier to detect the sensitivity and robustness of the model output. A lightweight adversarial classifier (an improvement on the MobileNet architecture) is designed, taking the semantic feature vector and concept graph features of the large model's output text as input, and outputting a sensitivity score (0-1) and a robustness score (0-1). Interference samples are generated through adversarial training (such as replacing domain terms or adjusting sentence structure) to test the stability of the model output. When the sensitivity score > 0.6 or the robustness score < 0.4, potential bias is identified.

[0036] The detection results of the two channels are fused using a weighted voting method. The weight of the temporal LSTM channel is set to 0.6, and the weight of the adversarial classifier channel is set to 0.8. When the fusion score is greater than 0.5, it is judged as biased; otherwise, it is judged as unbiased. The source analysis layer uses attention backtracking mechanism, training log analysis and causal intervention analysis to locate the distribution and evolution path of training data corresponding to bias and clarify the causes of bias.

[0037] The attention backtracking mechanism is based on the attention weight matrix of a large model to trace the source of bias-related semantic features and locate the text segments that play a key role in the formation of bias in the pre-training data and fine-tuning samples. Through attention weight visualization, it shows the propagation path of bias features from input to output and clarifies which training data or fine-tuning samples have the main influence on bias generation.

[0038] Causal intervention analysis: Using a causal graph model, a causal relationship network is constructed from training data to fine-tuning process to generation mechanism to output result. Through intervention analysis, such as removing specific training data or adjusting fine-tuning parameters, the contribution of each link to bias is quantified. Training log analysis: Retrieve detailed logs from the large model training process, including pre-training data batches, fine-tuning parameter settings, and changes in the loss function. Combined with the source tracing results, analyze key nodes in the training process that may lead to bias, such as sample distribution imbalance during the fine-tuning stage and unreasonable learning rate settings. Finally, output a bias cause report to clarify whether the bias stems from defects in the pre-training data, fine-tuning bias, or limitations in the generation mechanism.

[0039] S3, Multi-source Validation and Multi-dimensional Evaluation Multi-source evidence chain verification uses a graph structure retrieval algorithm to retrieve authoritative evidence from a scientific fact base to construct an evidence chain and determine the accuracy of the output conclusion. Multi-perspective analysis selects mainstream large-scale models with different architectures to generate responses, comparing the consistency of the target model's output with the core viewpoints of other models.

[0040] A multi-dimensional comprehensive assessment includes: Based on the pre-defined hierarchical indicator tree, each secondary indicator is quantitatively scored (0-10 points): Content credibility: factual accuracy (scoring based on the results of evidence chain verification), data traceability (checking whether the data citations are marked with sources and whether the sources are authoritative), and argumentation completeness (assessing whether the conclusions are supported by sufficient arguments). Logical consistency: No contradictions in concepts (check whether the use of concepts is consistent), no gaps in reasoning (assess whether the reasoning process is coherent), and no ambiguity in expression (check whether the language is clear and accurate). Domain adaptability: Terminology accuracy (assessing whether domain terminology is used correctly), scenario fit (checking whether the output content meets the specific technological scenario requirements), and technology relevance (assessing whether the technical description is relevant to actual applications). Knowledge freshness: data timeliness (checking whether the data is the latest data in the last 3 years), results update rate (assessing whether it covers the latest research results), and standard compatibility (checking whether it meets the latest technical standards).

[0041] The Analytic Hierarchy Process (AHP) was used to determine the weights of each dimension and secondary indicator, with content credibility weighted at 0.35, logical consistency at 0.25, domain suitability at 0.25, and knowledge novelty at 0.15. The weights of the secondary indicators were determined based on expert scores. The fuzzy comprehensive evaluation method was used to integrate the scores of each indicator to generate a final credibility score (0-10 points). A score ≥8 points was considered satisfactory, 6-8 points was considered to require optimization, and <6 points was considered unsatisfactory.

[0042] S4. Output Model Optimization Suggestions Based on the source analysis and evaluation results, this paper proposes data source supplementation and bias filtering schemes to address the deficiencies in pre-training data, sample balancing and feedback mechanism optimization strategies to address fine-tuning bias, and output verification enhancement and capability boundary labeling measures to address the limitations of the generation mechanism. The optimized model shows significant improvements in bias identification accuracy, output factual consistency, and credibility.

[0043] Pre-training data defect optimization: To address the problem of unbalanced data distribution, a data source supplementation scheme is proposed (such as increasing the proportion of academic achievements from developing countries and literature from niche disciplines), and the type, quantity and quality requirements of the supplementary data are clarified; a bias filtering algorithm is designed to screen the pre-training data based on a professional bias dictionary and high-dimensional bias features, and remove text segments with obvious bias tendencies.

[0044] Fine-tuning bias optimization: To address the sample balance problem, a sample expansion and resampling scheme is proposed to increase the number of minority class samples, such as those for special technical scenarios and niche research directions. The SMOTE algorithm is used to generate synthetic samples to make the proportion of each class of samples more balanced. The human feedback reinforcement learning (RLHF) mechanism is optimized to standardize the behavior of annotators and reduce the impact of subjective cognitive bias on the fine-tuning process.

[0045] The optimization suggestions were applied to the pre-training data update, fine-tuning of parameters, and improvement of generation mechanism of the large model. After retraining the model, the same detection and evaluation process was used for testing. The bias identification accuracy, credibility score, false alarm rate and other indicators before and after the iteration were compared. If the indicators did not meet the expectations, the source tracing results were re-analyzed and the optimization suggestions were adjusted until the model performance met the requirements.

[0046] This application can be applied to the following scenarios: (1) International policy and cross-border cooperation consulting scenario: It is applicable to scenarios that require neutral analysis of international affairs and the feasibility of cross-border scientific and technological cooperation, such as the policy formulation of international organizations and the evaluation of multinational companies' technology cooperation. It can accurately identify the geopolitical bias of large models, avoid deviations in cooperation decisions due to bias towards the views of specific countries, and ensure the objectivity and fairness of the analysis conclusions.

[0047] (2) Research trend prediction and academic resource allocation scenario: Adapt to the needs of research management departments and academic institutions for research hotspot analysis, research resource allocation planning, and local problem research value assessment. In response to the bias of academic data sources caused by the inclusion preferences of databases such as Web of Science, it corrects the overestimation of research hotspots in developed countries and the underestimation of local problems in developing countries, and helps to rationally allocate research resources and attach importance to local research.

[0048] (3) Decision support scenarios in professional fields: covering scenarios with high requirements for the accuracy of professional conclusions, such as medical diagnosis and consultation, engineering and technical solution recommendation, and public health early warning. It can effectively detect the bias of overgeneralization of scientific conclusions, avoid the application of general models to special scenarios (such as complex disease diagnosis and special technical needs), reduce the risk of professional decision-making caused by simple generalization, and ensure the pertinence and reliability of professional advice.

[0049] (4) Scientific and technological achievement evaluation and academic evaluation scenarios: Applicable to scenarios such as scientific research achievement review, academic award selection, and technology transformation value assessment. It can identify the misjudgment of achievement value caused by the deviation of academic data sources, avoid the unfair evaluation caused by the advantage of European and American academic achievements inclusion, and ensure that scientific research achievements of different regions and different types are objectively evaluated.

[0050] The remaining technical features in the above embodiments can be flexibly selected by those skilled in the art to meet different specific practical needs according to actual circumstances. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims. In the above description, numerous specific details have been set forth to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to implement the present invention. In other instances, to avoid obscuring the present invention, well-known techniques, such as specific construction details, operating conditions, and other technical conditions, have not been specifically described.

[0051] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting bias and hallucination of content generation of large language models in the field of science and technology information, characterized in that, The steps are as follows: S1, construct a scientific fact library and a professional bias resource technology field professional support system, provide detection benchmark and feature support; S2, adopt a three-layer progressive full-link bias detection traceability architecture, realize semantic feature extraction, bias dynamic detection and cause tracing through data preprocessing layer, real-time monitoring layer and traceability analysis layer; S3, through multi-source evidence chain verification, multi-angle analysis and multi-dimensional comprehensive evaluation, realize fact verification and credibility judgment of large model output; S4, based on the detection, tracing and evaluation results, output model optimization suggestions for bias causes.

2. The method of claim 1, wherein the method is characterized in that: The construction of the scientific fact library includes: integrating the technology literature library, the patent data library and the technical standard library, using hybrid neural machine translation to realize multi-language alignment, using graph database and graph neural network embedding method to construct domain ontology library, and using cross-modal semantic mapping model to realize unified semantic representation and correlation fusion of heterogeneous data including text, icon and formula.

3. The method of claim 2, wherein the method is characterized in that: The construction of professional bias resources includes: forming a professional bias dictionary through artificial annotation of bias scenarios and pre-training corpus mining, which contains discipline terminology preference, institutional weight tendency and regional academic resource bias related vocabulary, extracting features from three dimensions of semantics, syntax and domain attributes, constructing 200+dimensional high-dimensional bias feature space and using principal component analysis for dimension reduction.

4. The bias and hallucination detection method for large language model content generation in the field of scientific and technical intelligence according to claim 3, characterized in that: In S2, the data preprocessing layer extracts semantic features through BERT model, and generates concept relation graph combining with dependency syntax analysis; The real-time monitoring layer adopts a parallel double-channel detection network, channel one tracks semantic distribution changes through time series LSTM and sets a dynamic threshold alarm, and channel two detects the sensitivity and robustness of model output through a lightweight adversarial classifier; The traceability analysis layer locates the training data distribution and evolution path corresponding to the bias through attention backtracking mechanism and causal intervention analysis.

5. The bias and hallucination detection method for large language model content generation in the field of scientific and technical intelligence according to claim 4, characterized in that: The specific process of S3 includes: Multi-source evidence chain verification retrieves authoritative evidence in the scientific fact library through graph structure retrieval algorithm to construct evidence chain; Multi-angle analysis realizes cross-validation through output comparison of different architecture large models; The multi-dimensional comprehensive evaluation system covers four core dimensions of content credibility, logical consistency, domain adaptability and knowledge freshness, each dimension constructs a hierarchical index tree, determines the weight by analytic hierarchy process and fuses the results by fuzzy comprehensive evaluation method.

6. The method of claim 5, wherein the method is characterized in that: The construction method of hierarchical index tree is: content credibility is decomposed into fact accuracy, data traceability and argument completeness; logical consistency is decomposed into concept contradiction, reasoning fault and expression ambiguity; domain adaptability is decomposed into terminology accuracy, scene fit degree and technology correlation; knowledge freshness is decomposed into data timeliness, achievement update degree and standard adaptability.

7. The method of claim 6, wherein the method further comprises: determining a bias and hallucination of the generated content of the large language model in the field of science and technology intelligence. In S4, model optimization suggestions include: data source supplement and bias filtering scheme for pre-training data defects, sample balancing and feedback mechanism optimization strategy for fine-tuning bias, output verification reinforcement and ability boundary labeling measures for limitations of generation mechanism.

8. The system for detecting bias and hallucination of content generation of large language models in the field of science and technology information according to any one of claims 1-7, wherein, It includes: A support system module for building and storing a scientific fact base and a professional bias resource, providing detection benchmarks and feature support; A three-layer progressive full-link bias detection traceability architecture, including a data preprocessing unit, a real-time monitoring unit, and a traceability analysis unit, which respectively perform semantic feature extraction, bias dynamic detection, and cause traceability operations; A verification and evaluation module for performing multi-source evidence chain verification, multi-perspective analysis, and multi-dimensional comprehensive evaluation; An optimization suggestion output module for generating and outputting model optimization suggestions based on detection, traceability, and evaluation results.

9. The system for detecting bias and hallucination in content generation by large language models in the field of scientific and technical intelligence according to claim 8, characterized in that: The support system module includes a data source integration unit, a multi-language alignment unit, an ontology library construction unit, a cross-modal mapping unit, a dictionary compilation unit, a feature extraction unit, and a feature dimension reduction unit; The data source integration unit, the multi-language alignment unit, the ontology library construction unit, and the cross-modal mapping unit work together to complete the integration, multi-language alignment, ontology library construction, and heterogeneous data fusion of the scientific fact base. The scientific fact base is stored in a graph database, supporting efficient retrieval and real-time updating of heterogeneous data; The dictionary compilation unit, the feature extraction unit, and the feature dimension reduction unit work together to complete the professional bias dictionary compilation, feature extraction, and high-dimensional feature space dimension reduction.

10. The system for detecting bias and hallucination in content generation by large language models in the field of science and technology intelligence according to claim 8, characterized in that: In the three-layer progressive full-link bias detection traceability architecture, The data preprocessing unit includes a BERT encoding subunit and a syntax analysis subunit, which respectively implement semantic feature extraction and dependency syntax analysis to generate concept relationship graphs; The real-time monitoring unit includes a time series LSTM semantic tracking subunit and an adversarial classification detection subunit, which work in parallel to realize semantic distribution tracking and early warning, and to detect the sensitivity and robustness of model output; The traceability analysis unit includes an attention backtracking subunit and a causal intervention analysis subunit, which work together to locate the training data corresponding to the bias and analyze the evolution path.

Images