An Adaptive Method for Determining Literature Relevance Based on Downstream Material Extraction Utility Feedback
By constructing a dual-domain isolated closed-loop feedback architecture and adaptive model optimization, the problems of resource waste and information traversal in material literature screening are solved, achieving efficient literature relevance determination and continuous optimization, and improving the system's resource utilization and stability.
Patent Information
- Application Number
- CN202611614293.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies cannot effectively distinguish between thematic relevance and data extractability of materials and literature, resulting in wasted resources and rigid model capabilities, lack of adaptability, and the risk of information leakage.
We construct a full-link closed-loop feedback architecture based on dual-domain isolation. Through multi-dimensional reward functions and adaptive model evolution strategies, combined with natural language inference and dynamic threshold control, we achieve continuous optimization of the upstream model and downstream feedback mechanism.
It significantly improves resource utilization, filters out pseudo-relevant literature, increases downstream system throughput, possesses continuous evolution capabilities, and ensures system stability and robustness.
Smart Images

Figure CN123115171A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and big data technology, specifically involving the engineering application of natural language processing (NLP), high-throughput literature screening, and reinforcement learning in specific vertical fields, and particularly involving an adaptive method for determining literature relevance based on feedback from downstream material extraction utility. Background Technology
[0002] With the advancement of the Materials Genome Initiative (MGI) and the exponential growth of scientific publications, automatically constructing structured materials property databases (such as band gap, synthesis conditions, crystal structure, etc.) from massive amounts of academic literature has become a key means to accelerate the discovery of new materials. A typical processing pipeline usually includes two sequential core steps: the first step is correlation determination, which quickly determines whether a paper belongs to the target field based on its title and abstract; the second step is full-text information extraction, which involves downloading the full text of the papers determined to be relevant, parsing the PDFs, and using sophisticated deep learning models to extract specific entities and relationships.
[0003] However, in the existing technological system and practical engineering practice, the above process faces severe challenges and unresolved technical pain points, specifically in the following four aspects: The semantic gap between "topic relevance" and "data availability": Existing relevance assessment techniques essentially aim to determine "topic relevance." However, in the ultimate business objective of materials data mining, the truly valuable factor is "data extractability." Current technologies cannot distinguish between a "review article discussing the history of perovskite development" (highly relevant but lacking experimental data) and a "research paper detailing experimental parameters for perovskite synthesis" (relevant and containing high-value data). This semantic gap leads to a large number of "false positives" flowing downstream.
[0004] The enormous waste of downstream computing resources and the bottleneck in system efficiency: Full-text information extraction is an extremely computationally intensive process. Because upstream relevance assessment models lack the ability to perceive "data extractability," a large number of documents that do not contain the target data are incorrectly sent downstream. Statistical data shows that in certain materials science fields, less than 20% of the documents screened by traditional models actually contain extractable structured data. This means that the downstream system is in a "idle" state 80% of the time.
[0005] The lack of a closed-loop feedback mechanism leads to the solidification of model capabilities: Existing literature processing systems typically adopt an "open-loop" architecture. Once the upstream screening model is deployed, its parameters become fixed, making it impossible to learn from the numerous "failure cases" downstream and optimize itself, resulting in the same errors being repeated in the system over a long period of time.
[0006] Data consistency and information traversal risk during training and inference: If the model is trained and selected directly using the full text, a serious "information traversal" problem will be introduced. That is, the model relies on the features of the main text during training, but only the summary can be obtained during online inference. This feature inconsistency will lead to a catastrophic decline in the performance of the model after it is deployed online.
[0007] In summary, there is an urgent need for a new technical solution that, under the premise of strictly limiting input to only metadata, combines advanced key technologies such as semantic processing, syntactic analysis, and natural language query to enable the screening model to "learn" to understand and predict the extractability of full-text data, thereby achieving optimal efficiency across the entire process.
[0008] The prior art CN107247780A discloses a patent document similarity measurement method based on knowledge ontology, which includes extracting core technical solutions based on the structural characteristics, positional features, and keyword features of patent documents; constructing a patent classification number subject term word relationship model; constructing a domain dictionary based on the classification number subject term word relationship model and using it to segment and remove stop words from the core technical solutions; combining the subject term word relationships with TF-IDF as the initial weight of TextRank words to extract keywords and weights; training a FastText model to generate word vectors; and calculating the EMD distance based on keywords, word weights, and word vectors to obtain the semantic distance. This invention solves the problem of low similarity caused by traditional patent document similarity measurement methods not fully considering the structural characteristics, domain characteristics, word relationship characteristics, and inconsistent semantic approximations of patent texts. However, this invention lacks adaptive capability and has low accuracy in relevance discrimination. The prior art CN116401340A discloses a method and system for querying and comparing standard documents. This method includes preprocessing user-input keywords and retrieved documents, calculating the similarity between the preprocessed keywords and documents, using the similarity as an objective function, constructing a deep neural network model based on the similarity, optimizing the learning rate of the deep neural network model using a first algorithm, optimizing the gradient estimation of the deep neural network model using a second algorithm, training the optimized deep neural network model with the preprocessed text, retaining documents with a similarity greater than 0.9 until all documents have been traversed, and outputting the comparison results. This method not only improves comparison accuracy but also has good interpretability and can be directly applied to standard document query and comparison systems. However, the evaluation of this method is too subjective, and it also suffers from poor adaptability and ineffective feedback. Summary of the Invention
[0009] To address the aforementioned problems, this invention proposes an adaptive method for determining literature relevance based on feedback from downstream material extraction utility. This method solves the technical problem that existing literature screening models suffer from low overall system efficiency, significant resource waste, and an inability to continuously evolve their capabilities due to their inability to perceive downstream "data extractability." To achieve the above solution, this invention provides the following technical solution: An adaptive method for determining literature relevance based on feedback from downstream material extraction utility includes the following steps: S1: Construct a full-link closed-loop feedback architecture based on dual-domain isolation, and establish information isolation and signal bridging between the metadata decision domain and the full data execution domain; S2: Based on the dual-domain isolated full-link closed-loop feedback architecture constructed in step S1, establish a multi-dimensional reward function model based on extractability utility to quantify the execution results of the downstream full data execution domain into continuous scalar reward signals. S3: Implement a phased adaptive model evolution strategy to continuously optimize the upstream correlation determination model based on the scalar reward signal described in step S2; S4: Construct a dynamic threshold control mechanism with circuit breaking and rollback mechanisms to ensure the online operation of the system.
[0010] Further, step S1 includes: S11: Divide the metadata determination domain. This domain only maintains and processes the title, abstract, and author metadata information of the document, and forces the model deployed in this domain to only receive the metadata as input to prevent information from crossing over. Furthermore, in step S11, the input of the relevance determination model deployed in the metadata determination domain is physically restricted to include only the title, abstract, author and year information of the paper, so as to prevent the model from relying on words or data formats that are only in the text during the screening stage.
[0011] S12: Divide the full data execution domain. This domain is responsible for receiving candidate documents selected by the decision domain, performing full-text download, parsing and information extraction, and acting as a black-box environment, only outputting quantified utility feedback signals to the decision domain.
[0012] Furthermore, in step S12, the full data execution domain is equipped with an interconnected utility evaluator and a signal bridge. The utility evaluator is used to quantify the extraction result into the scalar reward signal, and the signal bridge is used to unidirectionally transmit the scalar reward signal back to the metadata decision domain without transmitting any text content or attribute values.
[0013] Further, step S2 includes: S21: Define the base gain and penalty for the extraction state. If the downstream extraction result is empty, a base negative reward is given; if the extraction result exists, the attribute richness is calculated based on the ratio of the number of effective attributes extracted to the total number of target attributes, and a base positive gain is given. Further, in step S21, the base negative reward is -1.0, the attribute richness is the ratio of the number of extracted effective attributes N_props to the total number of target attributes N_target, and the base positive gain is calculated according to... α The calculation is performed using *log(1+N_props), where... α The weighting coefficient is used; the scalar reward signal is the sum of all reward items normalized to the interval [-1, 1] by the Sigmoid function.
[0014] S22: Calculate the confidence and quality weights of the extraction results, comprehensively evaluate the confidence score of each extracted attribute and whether it has a corresponding text segment anchor in the original text as evidence, and weight the base gain accordingly. Furthermore, in step S22, for each extracted attribute, if the attribute has a corresponding text segment anchor point in the original text as evidence, then the quality gain of that attribute is... β *conf, if the text fragment anchor of this attribute is missing, the quality gain of this attribute is β *0.1, where β is the weighting coefficient, and conf is the confidence score for this attribute.
[0015] S23: Introduce strong penalties for illusion and conflict. Use an auxiliary natural language reasoning model to check whether there is a logical contradiction between the extracted attribute values and the original evidence. If a contradiction is found, a strong negative penalty is applied in the reward calculation.
[0016] Further, in step S23, the natural language reasoning model is used to verify whether the extracted attribute values logically imply the original text evidence. If a contradiction is determined, the applied negative penalty term is denoted as... δ , δ The value of is greater than or equal to 2.0, and δ The absolute value is greater than the basic positive gain.
[0017] Further, step S3 includes: S31: Perform supervised fine-tuning cold start, using high-quality manually labeled seed datasets accumulated in history to fine-tune the pre-trained language model and form a baseline policy model; S32: Implement offline preference alignment, and construct large-scale preference pair data based on historical runtime logs. d w , d l ),in d w Samples with high downstream utility reward values d l For samples with low downstream utility reward values, the direct preference optimization algorithm is used to optimize the model so that its probability distribution is tilted towards highly extractable samples. Furthermore, in step S32, the preference pair data is constructed by sampling from a dynamically maintained experience replay pool; when constructing the preference pair data, hard negative example mining is performed, that is, samples whose model prediction scores are higher than a set score but whose downstream extraction actually fails are selected as samples with low downstream utility reward values.
[0018] S33: Enable online adaptive evolution. After the system goes online, it continuously collects real-time extraction feedback. For samples whose model prediction scores are in the preset hesitation range, it adopts the ε-greedy strategy to explore and allow them to pass, collect their real feedback, and perform incremental training regularly.
[0019] Further, step S4 includes: S41: Implement dynamic threshold adjustment and introduce PID control algorithm. When the idling rate exceeds the set threshold, the threshold for correlation determination is automatically raised; otherwise, the threshold is lowered. Furthermore, in step S41, the idling rate of the downstream extraction system is used as the controlled variable and a target idling rate is preset, and the truncation threshold for correlation determination is set. τ As a control variable, the PID control algorithm outputs an increment based on the difference between the idling rate and the target idling rate to adjust the cutoff threshold. τ The output of the PID control algorithm is smoothed using a sliding window, and the cutoff threshold... τ It is restricted to a preset safety boundary [0.5, 0.9].
[0020] S42: Deploy an abnormal circuit breaker mechanism. When a model update is detected that causes a sharp decline in the quality of downstream data, the system will automatically trigger the circuit breaker, roll back the correlation determination model to the model weight of the previous stable version, and lock the threshold.
[0021] Furthermore, in step S42, the moving average of the illusion rate and the manual rejection rate, which reflect the downstream data quality, is calculated in real time. If the increase of the illusion rate and the manual rejection rate both exceed a set ratio within a preset time period, the current model version is determined to be invalid and a circuit breaker is triggered. After the circuit breaker is triggered, the traffic is switched to the model weight of the previous stable version, and the threshold adjustment function of the dynamic threshold control mechanism is locked, and conservative threshold operation is forced until manual intervention is carried out.
[0022] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention can significantly improve resource utilization, effectively filter more than 40%-60% of "spurious relevance" literature, and increase the throughput of downstream systems by more than 2 times.
[0023] 2. The dual-domain isolation architecture proposed in this invention fundamentally solves the problem of predicting the value of full-text data when only metadata is input, and effectively solves the problem of information traversal.
[0024] 3. This invention enables the upstream model to automatically adapt to changes in the downstream task through a closed-loop feedback mechanism, eliminating the need for large-scale manual relabeling, saving labor costs, and possessing continuous evolution capabilities.
[0025] 4. The built-in dynamic threshold and circuit breaker rollback mechanism of this invention ensures long-term stable operation of the system and has high robustness and security. Attached Figure Description
[0026] Figure 1This is a flowchart of the adaptive literature relevance determination method based on downstream material extraction utility feedback, as described in this invention. Figure 2 This is a schematic diagram of the end-to-end closed-loop system architecture based on dual-domain isolation of the present invention.
[0027] Figure 3 This is a schematic diagram of the information flow and gradient backhaul mechanism based on the data isolation wall of the present invention.
[0028] Figure 4 This is a flowchart illustrating the downstream feedback utility quantification and reward function construction of the present invention.
[0029] Figure 5 This is a schematic diagram of the online threshold adaptive adjustment logic based on PID control theory of the present invention.
[0030] Figure 6 A schematic diagram of the algorithm flow for the Offline Preference Alignment (DPO) training process. Detailed Implementation
[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0032] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0033] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.
[0034] In this embodiment of the invention, an adaptive method for determining literature relevance based on feedback from downstream material extraction utility is described, as detailed in the appendix. Figure 1 ~Appendix Figure 6 Specifically, it includes: S1: Construct a full-link closed-loop feedback architecture based on dual-domain isolation, and establish information isolation and signal bridging between the metadata decision domain and the full data execution domain; S2: Based on the dual-domain isolated full-link closed-loop feedback architecture constructed in step S1, establish a multi-dimensional reward function model based on extractability utility to quantify the execution results of the downstream full data execution domain into continuous scalar reward signals. S3: Implement a phased adaptive model evolution strategy to continuously optimize the upstream correlation determination model based on the scalar reward signal described in step S2; S4: Construct a dynamic threshold control mechanism with circuit breaking and rollback mechanisms to ensure the online operation of the system.
[0035] In a specific embodiment, to address the problems of strong subjectivity, low accuracy, low efficiency, lack of effective feedback mechanisms, insufficient adaptability, and risks related to data consistency and information traversal in existing literature relevance determination methods, the literature relevance determination engineer adopted an adaptive literature relevance determination method based on feedback from downstream material extraction utility. First, a dual-domain isolated full-link closed-loop feedback architecture is constructed to establish information isolation and signal bridging between the metadata determination domain and the full data execution domain. Second, based on the constructed dual-domain isolated full-link closed-loop feedback architecture, a multi-dimensional reward function model based on extractability utility is established to quantify the execution results of the downstream full data execution domain into continuous scalar reward signals. Third, a phased adaptive model evolution strategy is implemented to continuously optimize the upstream relevance determination model using the scalar reward signals. Finally, a dynamic threshold control mechanism with circuit breaking and rollback mechanisms is constructed to ensure the system's online operation. Example
[0036] In one embodiment, as shown in the appendix Figure 1 Appendix Figure 2 Appendix Figure 3 As shown, to address the issues of low accuracy and information smuggling risks in current document relevance assessment methods, document relevance assessment engineers have constructed a full-link closed-loop feedback architecture based on dual-domain isolation. Based on this, information isolation and signal bridging are established between the metadata assessment domain and the full data execution domain. Specific steps include: S11: Define the metadata decision domain. In this domain, deploy a lightweight text encoder based on the Transformer architecture, such as DistilBERT or compressed LLaMA. Its input is physically restricted to only include the title, abstract, author and year information of the paper. This prevents "information traversal", that is, it prevents the model from relying on high-frequency words or specific data formats that are only in the text during the screening stage. This ensures that the model can learn the implicit feature of "what kind of abstract implies that the text contains high-quality data". S12: Divide the full data execution domain and establish a signal bridge. This domain is responsible for receiving the "candidate queue" output from the decision domain and performing computationally intensive full-text download, PDF parsing, and structured information extraction (MIE). To achieve closed-loop operation, an effective evaluator and a signal bridge are deployed in this domain. The evaluator is responsible for quantifying the complex extraction results into a scalar reward signal and safely sending the signal back to the metadata decision domain as the sole basis for model optimization. This mechanism does not send back any specific text content or attribute values. Example
[0037] In one embodiment, as shown in the appendix Figure 1 Appendix Figure 4 As shown, to address the issues of low accuracy and lack of dynamic feedback mechanisms in current literature relevance assessment methods, the literature relevance assessment engineer has established a multi-dimensional reward function model based on extractability utility, building upon a dual-domain isolated end-to-end closed-loop feedback architecture. This model specifically includes: S21: Define the basic reward and penalty for the extraction state, and perform structured parsing on the downstream extraction results. If the extraction result is empty, assign a basic negative reward value of -1.0 as a penalty for "determining relevance but having no output"; if the extraction result exists, calculate the "attribute richness" based on the ratio of the number of extracted valid attributes to the total number of target attributes, and assign a basic positive reward, for example, using... α The calculation is performed using *log(1+N_props), where... α Here, N_props is the number of extracted attributes, and N_target is the total number of target attributes; S22: Introduce multi-dimensional quality weighting. For non-empty extraction results, calculate positive returns based on the confidence score of each attribute value and whether an evidence anchor can be found in the original text ("evidence completeness"). If an evidence anchor exists, the return is... β *conf; If evidence is missing, a weighting reduction will be applied, with the benefit being ( β *0.1), where βHere, 'conf' represents the weighting coefficient, and 'conf' represents the confidence score. S23: Set strong penalties for illusions and conflicts, and introduce an auxiliary Natural Language Inference (NLI) model to automatically verify whether the extracted attribute values are logically implied with the original evidence; if a contradiction is found (e.g., the original text states "insulator," but the extracted result is "conductor"), then apply a strong negative penalty to the reward calculation. δ The penalty value is usually set to ≥2.0, which is much larger than the base reward, to ensure the authenticity of the reward signal. Finally, the total reward value is normalized to the range of [-1,1] by functions such as Sigmoid. Example
[0038] In one embodiment, as shown in the appendix Figure 1 Appendix Figure 6 As shown, to address the shortcomings of current literature relevance assessment methods, such as insufficient adaptability and lack of feedback mechanisms, the literature relevance assessment engineer implemented a phased adaptive model evolution strategy, specifically including: S31: Perform supervised fine-tuning (SFT) cold start. In the initial deployment phase of the system, SFT is performed using a small-scale, high-quality dataset of about 2,000 manually labeled data points. This can be supplemented with rule-based filters such as a list of required keywords to enable the pre-trained language model to have basic domain knowledge and form a reliable baseline policy model. S32: Implement offline preference alignment. The system maintains a dynamic "experience replay pool." During the offline training phase, preference pairs are sampled from the replay pool and constructed. d w , d l ),in d w (winner) is the sample with a high downstream reward value. d l (Loser) refers to samples with low reward values; the system uses a "difficult-to-reward example mining" mechanism to prioritize and select samples that "have high model prediction scores but fail to be extracted in practice" as... d l The Direct Preference Optimization (DPO) algorithm is used to maximize the likelihood difference between winning and losing samples, forcing the model to correct its overconfident misconceptions. S33: Enable online adaptive evolution. After the system goes online, in order to cope with the drift of data distribution over time, it will continuously collect real-time extraction feedback. For samples whose model prediction scores are in the preset "hesitation interval" [0.4, 0.6], exploratory strategies such as ε-greedy will be used to allow a small proportion of samples to pass through, collect their real feedback, and use it for the regular incremental training of the model. Example
[0039] In one embodiment, as shown in the appendix Figure 1 Appendix Figure 5 As shown, to ensure the reliability and sustainable operation of the adaptive literature relevance assessment method based on feedback from downstream material extraction utility, the literature relevance assessment engineer constructed a dynamic threshold control mechanism with circuit breaker and rollback mechanisms, specifically including: S41: Implement dynamic threshold adjustment based on PID control. The "Empty Extraction Rate" of the downstream extraction system is used as the controlled variable, with a preset target idle rate of 20%. The "truncation threshold τ" for correlation determination is used as the control variable. The system employs a PID (Proportional-Integral-Derivative) control algorithm. When the downstream idle rate is detected to increase, the controller outputs a positive increment, raising the threshold τ; conversely, it lowers the threshold. To prevent oscillations, this mechanism sets a sliding window-based smoothing process and an absolute safety boundary of [0.5, 0.9] to ensure the threshold does not deviate to extreme values. S42: Deploy an abnormal circuit breaker and rollback mechanism. The monitoring service calculates the moving average of the "illusion rate" and "manual rejection rate" in real time. Once these two indicators surge by more than 30% within 5 minutes, the system will determine that the current model version is invalid or that a serious data drift has occurred. At this time, the circuit breaker is immediately triggered, performing two core actions: 1) Switching traffic back to the stable model weight of the previous version; 2) Locking the adaptive threshold adjustment function, forcing the use of a conservative high threshold of 0.9 to prioritize data quality until manual intervention is required for investigation.
[0040] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0041] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. An adaptive method for determining literature relevance based on feedback from downstream material extraction utility, characterized in that, Includes the following steps: S1: Construct a full-link closed-loop feedback architecture based on dual-domain isolation, and establish information isolation and signal bridging between the metadata decision domain and the full data execution domain; S2: Based on the dual-domain isolated full-link closed-loop feedback architecture constructed in step S1, establish a multi-dimensional reward function model based on extractability utility to quantify the execution results of the downstream full data execution domain into continuous scalar reward signals. S3: Implement a phased adaptive model evolution strategy to continuously optimize the upstream correlation determination model based on the scalar reward signal described in step S2; S4: Construct a dynamic threshold control mechanism with circuit breaking and rollback mechanisms to ensure the online operation of the system.
2. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 1, characterized in that, Step S1 includes: S11: Divide the metadata determination domain. This domain only maintains and processes the title, abstract, and author metadata information of the document, and forces the model deployed in this domain to only receive the metadata as input to prevent information from crossing over. S12: Divide the full data execution domain. This domain is responsible for receiving candidate documents selected by the decision domain, performing full-text download, parsing and information extraction, and acting as a black-box environment, only outputting quantified utility feedback signals to the decision domain.
3. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 2, characterized in that, In step S11, the input of the relevance determination model deployed in the metadata determination domain is physically restricted to include only the title, abstract, author and year information of the paper, so as to prevent the model from relying on words or data formats that are only in the text during the screening stage.
4. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 2, characterized in that, In step S12, the full data execution domain is equipped with an interconnected utility evaluator and a signal bridge. The utility evaluator is used to quantify the extraction result into the scalar reward signal, and the signal bridge is used to unidirectionally transmit the scalar reward signal back to the metadata determination domain without transmitting any text content or attribute values.
5. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 1, characterized in that, Step S2 includes: S21: Define the base gain and penalty for the extraction state. If the downstream extraction result is empty, a base negative reward is given; if the extraction result exists, the attribute richness is calculated based on the ratio of the number of effective attributes extracted to the total number of target attributes, and a base positive gain is given. S22: Calculate the confidence and quality weights of the extraction results, comprehensively evaluate the confidence score of each extracted attribute and whether it has a corresponding text segment anchor in the original text as evidence, and weight the base gain accordingly. S23: Introduce strong penalties for illusion and conflict. Use an auxiliary natural language reasoning model to check whether there is a logical contradiction between the extracted attribute values and the original evidence. If a contradiction is found, a strong negative penalty is applied in the reward calculation.
6. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 5, characterized in that, In step S21, the base negative reward is -1.0, the attribute richness is the ratio of the number of extracted effective attributes N_props to the total number of target attributes N_target, and the base positive gain is calculated according to... α The calculation is performed using *log(1+N_props), where... α The weighting coefficient is used; the scalar reward signal is the sum of all reward items normalized to the interval [-1, 1] by the Sigmoid function.
7. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 1, characterized in that, Step S3 includes: S31: Perform supervised fine-tuning cold start, using high-quality manually labeled seed datasets accumulated in history to fine-tune the pre-trained language model and form a baseline policy model; S32: Implement offline preference alignment, and construct large-scale preference pair data based on historical runtime logs. d w , d l ),in d w Samples with high downstream utility reward values d l For samples with low downstream utility reward values, the direct preference optimization algorithm is used to optimize the model so that its probability distribution is tilted towards highly extractable samples. S33: Enable online adaptive evolution. After the system goes online, it continuously collects real-time extraction feedback. For samples whose model prediction scores are in the preset hesitation range, it adopts the ε-greedy strategy to explore and allow them to pass, collect their real feedback, and perform incremental training regularly.
8. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 7, characterized in that, In step S32, the preference pair data is constructed by sampling from a dynamically maintained experience replay pool; when constructing the preference pair data, hard negative example mining is performed, that is, samples with model prediction scores higher than a set score but actual downstream extraction failures are selected as samples with low downstream utility reward values.
9. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 1, characterized in that, Step S4 includes: S41: Implement dynamic threshold adjustment and introduce PID control algorithm. When the idling rate exceeds the set threshold, the threshold for correlation determination is automatically raised; otherwise, the threshold is lowered. S42: Deploy an abnormal circuit breaker mechanism. When a model update is detected that causes a sharp decline in the quality of downstream data, the system will automatically trigger the circuit breaker, roll back the correlation determination model to the model weight of the previous stable version, and lock the threshold.
10. The adaptive method for determining literature relevance based on downstream material extraction utility feedback as described in claim 9, characterized in that, In step S42, the moving average of the illusion rate and the manual rejection rate, which reflect the downstream data quality, is calculated in real time. If the increase of the illusion rate and the manual rejection rate both exceed the set ratio within a preset time period, the current model version is determined to be invalid and the circuit breaker is triggered. After the circuit breaker is triggered, the traffic is switched to the model weight of the previous stable version, and the threshold adjustment function of the dynamic threshold control mechanism is locked, and conservative threshold operation is forced until manual intervention is carried out.
Citation Information
Patent Citations
Patent literature similarity measurement method based on ontology
CN107247780A
