Content inference system and electronic device

Abstract

The application discloses a content reasoning system and an electronic device, and relates to the technical field of artificial intelligence. The content reasoning system comprises an input module, an auxiliary module and a reasoning module. Historical reasoning data is introduced for reuse in the content reasoning process, and reasoning execution logic is optimized. The auxiliary module can manage the historical reasoning data, so that when new reasoning source data is received, the corresponding historical reasoning data is matched and called as reasoning reference data, and the situation of building a reasoning path from zero is reduced. In this way, the reasoning module jointly executes reasoning based on the reasoning source data and the reasoning reference data, which is beneficial to reducing repeated calculation and redundant operation, shortening the reasoning path, and thus reducing the overall reasoning operation amount and system resource consumption, and improving the response speed and processing efficiency of content reasoning. The technical problems of complex content reasoning operation amount and poor reasoning timeliness are solved, and the technical effects of improving content reasoning efficiency and reducing reasoning operation amount to improve reasoning performance are achieved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a content reasoning system and an electronic device. Background Technology

[0002] When content reasoning systems acquire source data for reasoning, they can perform full-process reasoning operations on the source data, meaning they typically rely on real-time computation during the reasoning process. However, as the complexity of reasoning tasks increases and the amount of data grows, real-time full-data reasoning can easily prolong response time, resulting in poor timeliness of content reasoning and making it difficult to meet the demands for efficient content reasoning. Summary of the Invention

[0003] This application provides a content reasoning system and an electronic device to at least solve the problems of cumbersome content reasoning computation and poor reasoning timeliness in related technologies.

[0004] This application provides a content reasoning system, comprising: an input module, an auxiliary module, and a reasoning module; the input module is used to acquire reasoning source data; the auxiliary module is connected to the input module and is used to manage historical reasoning data; in response to receiving reasoning source data, the system queries historical reasoning data that matches the reasoning source data as reference source data and generates reasoning reference data from the reasoning process information of the reference source data; the reasoning module is connected to the input module and the auxiliary module and is used to acquire reasoning source data and reasoning reference data, and to perform content reasoning on the reasoning source data in conjunction with the reasoning reference data to obtain target reasoning data.

[0005] This application also provides an electronic device, which includes: a device body and a content reasoning system as described above, wherein the content reasoning system is located on the device body.

[0006] This application optimizes the inference execution logic by reusing historical inference data during the content inference process. The auxiliary module manages this historical inference data, matching and retrieving corresponding historical inference data as inference reference data when receiving new inference source data, reducing the need to build inference paths from scratch. Thus, the inference module jointly executes inference based on both inference source data and inference reference data, reducing redundant calculations and shortening inference paths. This reduces the overall computational load and system resource consumption, improving the response speed and processing efficiency of content inference. Therefore, it solves the technical problems of cumbersome computational load and poor timeliness in content inference, achieving the technical effect of improving content inference efficiency and reducing computational load to enhance inference performance. Attached Figure Description

[0007] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of the structure of an embodiment of the electronic device of this application; Figure 2 This is a schematic diagram of the structure of an embodiment of the reasoning system of this application; Figure 3 This is a schematic diagram of another embodiment of the reasoning system for the content of this application; Figure 4 This is a flowchart illustrating an embodiment of the working principle of the data fusion unit in this application; Figure 5 This is a flowchart illustrating an embodiment of the reasoning method for the content of this application. Detailed Implementation

[0009] 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 of ordinary skill in the art without creative effort are within the protection scope of this application.

[0010] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0011] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0012] The specific application environment architecture or specific hardware architecture on which the execution of the content reasoning system depends is described here.

[0013] The embodiments of this application provide an electronic device. The electronic device will be described in detail below in conjunction with its structural topology and working principle.

[0014] Please see Figure 1 , Figure 1 This is a schematic diagram of the structure of an embodiment of the electronic device of this application.

[0015] In one embodiment, the electronic device includes a device body 11 and a content reasoning system 12.

[0016] The device body 11 serves as the hardware carrier for electronic devices to perform various data processing and computation functions. It can integrate various components and functional units required for the basic operation, data storage and computation scheduling of electronic devices. It can independently complete the basic operation scheduling and resource allocation of electronic devices, while providing an installation carrier and operation support for the content inference system 12.

[0017] The content reasoning system 12 is located on the device body 11. The content reasoning system 12 can serve as a functional unit for the electronic device to achieve intelligent content understanding, content reasoning, and result output. In this embodiment, the content reasoning system 12 can acquire reasoning source data. It queries historical reasoning data that matches the reasoning source data as reference source data, and generates reasoning reference data from the reasoning process information of the reference source data. It then combines the reasoning reference data to perform content reasoning on the reasoning source data to obtain the target reasoning data.

[0018] The content reasoning system 12 will be described in detail below. That is, embodiments of this application also provide a content reasoning system 12. The communication management system will be described in detail below, combining the structural topology and working principle of the content reasoning system 12.

[0019] Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of an embodiment of the reasoning system for the content of this application.

[0020] In one embodiment, the content reasoning system may include an input module 21, an auxiliary module, and a reasoning module 23.

[0021] Input module 21 is used to acquire inference source data. In other words, input module 21 can be considered as the basic receiving unit for the content inference system to achieve external data access and data acquisition.

[0022] The auxiliary module is connected to the input module 21 and is used to manage historical inference data. In response to receiving inference source data, it queries historical inference data that matches the inference source data as reference source data, and generates inference reference data from the inference process information of the reference source data. The auxiliary module is used by the content inference system to realize the storage, maintenance, and reuse of historical inference data. It provides a data interaction path with the input module 21, allowing it to retrieve and match stored historical inference data after receiving inference source data transmitted from the input module 21. It identifies historical inference data that matches the inference source data as reference source data, and uses this reusable inference reference data to provide inference basis for the inference module 23, reducing the burden of repeatedly constructing inference logic and thus lowering the computational load.

[0023] The inference module 23 is connected to the input module 21 and the auxiliary module. It is used to obtain inference source data and inference reference data, and to perform content inference on the inference source data in combination with the inference reference data to obtain the target inference data.

[0024] The inference module 23, as the main execution entity of the content inference system, performs intelligent calculations and outputs inference results. It can connect with the input module 21 and the auxiliary module. The inference module 23 can obtain the inference source data from the input module 21 and the inference reference data from the auxiliary module. Using the inference reference data as reference information for the inference process, it performs content inference on the inference source data. This approach can reduce redundant calculations, shorten the inference path, and ensure the reliability of the target inference data output, thereby optimizing the performance of the content inference model.

[0025] Therefore, this embodiment incorporates historical inference data for reuse during content inference, optimizing the inference execution logic. The auxiliary module manages historical inference data, matching and retrieving corresponding historical inference data as inference reference data when receiving new inference source data, reducing the need to construct inference paths from scratch. Thus, the inference module 23 jointly executes inference based on the inference source data and inference reference data, which helps reduce redundant calculations and operations, shortens the inference path, thereby reducing the overall inference computation and system resource consumption, improving the response speed and processing efficiency of content inference, and ultimately improving content inference efficiency and reducing inference computation to enhance inference performance.

[0026] Please see Figure 3 , Figure 3 This is a schematic diagram of another embodiment of the reasoning system for the content of this application.

[0027] In one embodiment, the auxiliary module includes a vector generation module 221 and a data management module 222, which are connected together.

[0028] The vector generation module 221 is used to obtain the original text of the reasoning data process of the reasoning target and use it as the historical reasoning original text. It converts the historical reasoning original text into a representation vector and transmits the representation vector to the data management module 222.

[0029] The data management module 222 is used to construct historical management files to store historical inference data and its indicator data, and to retrieve reference source data that matches the inference source data from the historical management files. This improves the processing accuracy and retrieval efficiency of historical inference data. By using representation vectors, unstructured text vectorization representation of the historical inference original text can be achieved, improving the accuracy and speed of historical data matching. Furthermore, constructing standardized historical management files facilitates centralized storage and management of historical inference data and corresponding indicator data, ensuring the traceability, reuse, and retrieval of historical inference data. Thus, this embodiment reduces redundant storage and invalid comparisons, further reducing computational overhead in the content inference process, improving the matching efficiency of reference source data, and enhancing the performance of the content inference system.

[0030] Furthermore, the vector generation module 221 includes a single vector unit 2211 and a multi-vector unit 2212, which are connected together.

[0031] The single-vector unit 2211 is used to obtain the original historical reasoning text, ignore stop words in the original text to obtain its substantive words, assign part-of-speech weights to the substantive words based on their part-of-speech matching, and evaluate the vector of the original historical reasoning text based on the part-of-speech weights of the substantive words as its representation vector. In other words, the single-vector unit 2211 can perform stop word filtering, substantive word extraction, and part-of-speech weight assignment on the original historical reasoning text, which can effectively remove invalid information, highlight core semantic features, and make the generated representation vector more consistent with the true meaning of the original reasoning text, thereby improving the accuracy and distinguishability of semantic expression.

[0032] The multi-vector unit 2212 is used to evaluate the attribute weights of the representation vectors. The attribute weights are associated with the confidence factor and / or importance factor of the representation vectors. Multiple representation vectors and their attribute weights are weighted, fused, and normalized to obtain a fused memory vector participating in the matching of reference source data. By introducing attribute weights associated with the confidence factor and importance factor, the multi-vector unit 2212 performs weighted fusion and normalization on multiple representation vectors to obtain a fused memory vector with improved representativeness and stability. Thus, this embodiment can significantly improve the semantic similarity recognition accuracy during historical data matching, reduce false matches and invalid comparisons, further improve the efficiency and accuracy of reference source data retrieval, and reduce redundant computational overhead.

[0033] For example, when the single vector unit 2211 generates a single memory vector, it may include stages such as word segmentation, weighting, averaging, and normalization.

[0034] In the word segmentation and word vector encoding stage, a word segmentation tool can be used to segment the memory original text, that is, the inference original text of the historical inference data, split the memory original text into words, and label the word types (such as nouns / verbs / adjectives / adverbs / prepositions / stop words, etc.).

[0035] Allocate weights and filter according to the adapted word types. Weights can be assigned to each word according to the词性映射表 (词性 mapping table) to distinguish the importance of different word types in semantic representation. For example, nouns and named entities usually carry the core semantics and can be given the highest weight, set to 0.5; verbs express actions or relationships, and the weight can be set to 0.3; other word classes (such as adjectives, adverbs, prepositions, etc.) have lower weights and can be set to 0.2; stop words (such as "的", "是", "在") do not participate in the weighted average and can be filtered in advance and are given the lowest weight, such as 0.

[0036] Obtain and weight word vectors. Use a pre-trained word vector model to obtain the vector representation of each word. Multiply each word vector by its corresponding weight to get the weighted word vector. To perform weighted averaging, that is, after summing the weighted vectors of all words, divide by the total weight of the valid words to get the original memory vector. The specific calculation formula can be as shown in the following example: (w1×分词向量1+...+wm×分词向量m) / 有效词总权重 Formula 1-1 Where, w1 represents the word segmentation weight of the word segmentation vector 1; wm represents the word segmentation weight of the word segmentation vector m, and m is a positive integer.

[0037] Normalize the original memory vector to make it a unit vector for subsequent similarity calculation (such as cosine similarity). Calculate the L2 norm (Euclidean norm) of the original vector, and divide the vector by its length to get the final vector of unit length. Among them, the L2 norm represents the square root of the sum of the squares of the elements of the vector and is used to measure the length or distance of the vector. The specific process description combined with the text expression can be as shown in the following example: "S01: Input the memory original text text; S02: Segment and label the词性 of text to get the list [(word1,pos1),(word2,pos2),...]; S03: Initialize the weight mapping table pos_weight_map={'名词':0.5,'动词':0.3,'形容词':0.2,...}; S04: Initialize the stop word list stop_words=['的','是','在',...]; S05: Initialize the valid word list weighted_words=[]; It should be noted that there is a misspelling in the original text where "词性映射表" is not properly translated. It should be something like "词性 mapping table" as shown in the translation. Also, "分词向量" is left in Chinese in the original and should be translated more precisely, but following the instruction to keep it as is for now.S06: Iterate through each (word, pos): If the word is in stop_words, skip it; Otherwise, retrieve the weight w from pos_weight_map. If it is not defined, use the default weight (e.g., 0.1) to add (word, w) to weighted_words. S07: Initialize the sum_vector to zero vector and the total_weight to 0; S08: Iterate through each (word, w): Get the word vector word_vec of word. sum_vector += word_vec * w total_weight += w; S09: If total_weight > 0, then raw_vector = sum_vector / total_weight; otherwise, return the zero vector. S010: Calculate the L2 norm of raw_vector: norm = sqrt(sum(raw_vector)) 2 )); S011: If norm > 0, then final_vector = raw_vector / norm; otherwise, final_vector = raw_vector. S012: Output final_vector as a single-memory vector.

[0038] When the multi-vector unit 2212 generates multi-memory vectors (i.e., fused memory vectors), it can fuse a group of related memory units into a unified semantic vector representation to capture their collective semantic core. In this process, the weight setting for each memory unit can comprehensively consider confidence and importance to ensure the reliability of the generated combined vector in reflecting the core semantics, while also suppressing interference from low-quality or irrelevant memories. The formula for calculating attribute weights can be exemplified as follows: ws = confidence-x×importance-x Formula 2-1 Where ws represents the attribute weight; confidence-x represents the current confidence factor of the representation vector of the historical reasoning text; and importance-x represents the current importance factor of the representation vector of the historical reasoning text.

[0039] The calculation process for the fused memory vector may include weight normalization, weighted summation, and vector normalization. First, the attribute weights can be normalized to ensure their sum is 1, reducing scale differences introduced by varying weight sums. Here, m represents the number of historical inference data points involved in the combination. The single memory vector (representation vector) of each memory unit is multiplied by its corresponding weight and then summed. The original vector obtained from the weighted average is then normalized to transform it into a unit vector, facilitating subsequent calculations such as cosine similarity to ultimately generate the fused memory vector.

[0040] The calculation formula for the fusion memory vector generation process can be exemplified as follows: Equation 2-2 Equation 2-3 Equation 2-4 in, This represents the attribute weights of the i-th representation vector after normalization; Let represent the attribute weights calculated by Equation 2-1 for the i-th representation vector; j represents the current j-th representation vector; m represents the total number of representation vectors generated from the fusion memory vectors; This represents the result of weighted fusion of multiple representation vectors and their attribute weights; This represents the i-th representation vector; Represents the fused memory vector; express The length of the vector.

[0041] Optionally, the data management module 222 includes a data memory unit 2221 and a data retrieval unit 2222.

[0042] Data memory unit 2221 is used to store historical management files. These files include time information, data type information, the original historical reasoning text, and its representation vector. Indicator data includes at least one of confidence factor, importance factor, and frequency of use weight. Data types include fact type, opinion type, and method type, with the initial confidence factors decreasing sequentially for fact type, method type, and opinion type.

[0043] The data retrieval unit 2222 is used to access the data memory unit 2221 in response to obtaining inference source data, to filter preliminary data from the historical management file by keyword matching, to evaluate the semantic similarity between the preliminary data and the inference source data to filter reference source data, and to update the indicator data of the reference source data according to the usage of the reference source data.

[0044] Therefore, the data memory unit 2221 can improve the orderly storage of historical reasoning data, which is beneficial to the efficiency and reliability of retrieving relevant information from historical reasoning data. For example, the data retrieval unit 2222 can adopt a two-level retrieval mechanism of first keyword filtering and then semantic similarity evaluation, which can improve the efficiency of narrowing the search scope and also help ensure the accuracy of matching results, significantly improving the efficiency and accuracy of reference source data filtering. Dynamically updating indicator data based on the actual usage of reference source data enables adaptive optimization and continuous iteration of historical data, thereby further reducing invalid searches and redundant calculations, increasing the reuse value of historical data, and enhancing the stability and intelligence of the content reasoning system.

[0045] In layman's terms, this embodiment forms a structured storage of historical reasoning data through historical management files, which may include historical reasoning text, representation vectors, and information such as time features and / or indicator data and / or data type information.

[0046] As explained above, indicator data can include confidence factors, importance factors, and usage frequency weights. The following example illustrates the calculation principle of indicator data.

[0047] The confidence factor can be dynamically adjusted based on at least one dimension, including usage feedback and time decay. Optionally, different initial confidence factors can be configured for historical inference data of different data types. Taking a confidence factor range of [0,1] as an example, the following examples are provided with specific data.

[0048] The initial confidence factor for fact-based historical inference data can be 0.9; the initial confidence factor for method-based historical inference data can be 0.7; and the initial confidence factor for opinion-based historical inference data can be 0.8.

[0049] After historical reasoning data is invoked and used in the content reasoning process, the confidence factor of the historical reasoning data can be adjusted based on the correctness of the reasoning results.

[0050] Specifically, when the usage of the reference source data is successful, a positive gain is assigned to the reference source data's metrics. When the usage of the reference source data is abnormal, a negative gain is assigned to the reference source data's metrics. And / or, when the unused duration of historical inference data exceeds a duration threshold, its metrics are assigned a negative gain.

[0051] The following example illustrates the calculation formula for updating the confidence factor based on usage feedback.

[0052] When the reference source data is used successfully, the memory results of the reference source data can be considered correct and effectively accelerate content reasoning. The calculation formula for its confidence factor update can be exemplified as follows: confidence-n =min(1.0,confidence+0.01×(success_count / usage_count)) Formula 3-1 Where confidence-n represents the updated confidence factor; min() represents the minimum function; confidence represents the confidence factor to be updated; success_count represents the number of times the usage was successful; usage_count represents the total number of times the content was used for reasoning; and 0.01 represents the confidence factor update coefficient.

[0053] When the usage of the reference source data is anomalous, it can be assumed that the memory of the reference source data leads to errors in content inference, thereby accelerating invalidation. The formula for calculating the update of its confidence factor can be exemplified as follows: confidence-n=max(0.0,confidence-0.05×(failure_count / usage_count)) Formula 3-2 Here, max() represents the function that takes the largest value; failure_count represents the number of times the exception occurred.

[0054] Furthermore, if outdated or contradictory content is found during the periodic verification of historical inference data, the formula for updating the confidence factor can be illustrated as follows: confidence-n=max(0.0,confidence-0.1) Formula 3-3 The following example demonstrates the calculation formula for updating the confidence factor when the unused duration of historical inference data exceeds a duration threshold, using a duration threshold of 30 days as an example: confidence-n=confidence×(1-0.1×days_unused / 30) Formula 3-4 Here, days_unused represents the unused time.

[0055] The importance factor, as an indicator of the contribution of memory units to reasoning acceleration, can range from [0,1]. The importance factor can be influenced by one or more of the following: frequency of use, effectiveness of use, and user marking.

[0056] For example, the initial value of the importance factor for historical inference data can be set to 0.5.

[0057] The following is an example of the calculation formula for updating the importance factor based on usage feedback.

[0058] When the reference source data is used successfully, the formula for updating its importance factor can be exemplified as follows: importance-n=min(1.0,importance+0.005×success_count / usage_count) Formula 4-1 Where importance-n represents the updated importance factor; importance represents the importance factor to be updated; and 0.005 represents the importance factor update coefficient.

[0059] When the usage of the reference source data is abnormal, the formula for calculating its importance factor update can be exemplified as follows: importance-n=max(0.0,importance-0.02×failure_count / usage_count) Formula 4-2 As explained earlier, importance factors can be influenced by user marking. Historical inference data marked by users can be identified as important data, and therefore, their importance factor values, such as 0.9, can be assigned to the target data.

[0060] The following is an example of a calculation formula for updating the importance factor based on usage frequency. The calculation formula can be expressed as follows: importance-n=importance×(1-0.05×min(days_used_last,30) / 30) Formula 4-3 Where 0.05 represents the importance factor frequency decay coefficient; days_used_last represents the number of days since the last time the historical inference data was used for content inference.

[0061] Furthermore, the data retrieval unit 2222 can be considered as being implemented through effective boundary conditions. Specifically, effective boundary conditions can be considered as a set of constraint rules attached to each memory unit, used to clearly define the scope of application and the prerequisites for the memory to take effect. This helps prevent the memory from being misused in irrelevant or unsuitable input scenarios, thereby ensuring the accuracy of the reasoning results while accelerating reasoning using historical information. Through boundary conditions, the content reasoning system can achieve more accurate memory matching and filtering.

[0062] In detail, before calculating the vector similarity between the source inference data and historical inference data, the memory retrieval stage can use boundary conditions to quickly filter all memories. Memories that pass all boundary condition checks enter the similarity matching stage, significantly improving retrieval efficiency. In the strategy decision-making stage, an inference strategy can be selected based on similarity and confidence, and the boundary conditions of the most relevant memories can be checked again to ensure they are fully satisfied. In the result feedback stage, after each inference iteration, regardless of whether historical inference data has been used, the matching status of the boundary conditions can be recorded, serving as the data basis for subsequently updating the confidence, importance, and boundary conditions themselves. The structured representation of boundary conditions can use data structures, such as the "valid_bounds" field, stored as a key-value dictionary, containing sub-category boundary conditions such as semantic similarity thresholds, required keyword lists, time validity windows, and data text length ranges.

[0063] Specifically, the semantic similarity threshold can be set such that the vector similarity between the inference source data and the historical inference data must be higher than a certain semantic similarity value (e.g., 0.8) before historical inference data can be selected as a candidate.

[0064] A list of required keywords can constrain the input text to contain one or more specific keywords listed, ensuring topic relevance.

[0065] The time validity window can define that historical inference data is valid between a specified start time and end time (e.g., from "2023-09-01" to "2024-12-31") to accommodate information with explicit time sensitivity.

[0066] The input text length range can constrain the length of historical inference data input (such as the number of characters or words) to be within the range defined by the minimum and maximum values, allowing it to participate in the calculation of whether to be a candidate, in order to filter out unconventional queries that are too long or too short.

[0067] The data retrieval unit 2222 can extract keywords from the reasoning source data as a primary index, use the primary index to perform a coarse screening of historical reasoning data, and use semantic similarity as a secondary index to select reference source data.

[0068] Specifically, when a new memory unit, i.e., historical reasoning data, is created and stored in the memory bank, i.e., the historical management file, the content reasoning system extracts a set of representative keywords from its original text. Specifically, word segmentation and词性 analysis can be performed, that is, the memory original text is segmented, and the词性 of each word is identified (such as nouns, verbs, adjectives, etc.). Then keyword screening is carried out. The content reasoning system can preferentially select nouns and named entities as index keywords. And filtering and deduplication are performed to remove common meaningless words (stop words, such as "of", "and", "is"), and the screened keywords are standardized (such as unified conversion to lowercase) to generate a unique keyword list.

[0069] The dictionary structure of the "inverted index table" can be maintained in the memory device. Using the keyword as the key (Key), and the list of the IDs of all memory units containing this keyword as the value (Value). The inverted index table can be implemented in the memory with a dictionary (hash table) data structure.

[0070] Among them, the key (Key) can be considered as a standardized keyword string. Standardization includes: conversion to lowercase, removal of punctuation, stemming (such as "running" being reduced to "run"), etc., to ensure the consistency of the index.

[0071] The value (Value) can be considered as an ordered list of integers, where each integer is a MemoryUnitID (memory unit ID) that uniquely identifies a certain memory unit. The list remains ordered (for example, it can be in ascending order) for efficient set operations.

[0072] Index update can add memories. That is, when a new memory is added, its ID can be uniformly added to the ID list corresponding to each of its keywords; when a memory is deleted or updated, when the memory is deleted or the content is modified, the inverted index table can be updated synchronously to ensure the consistency between the index and the content of the memory bank.

[0073] Semantic vector similarity calculation can perform refined sorting based on the rough screening results of the first-level index to ensure semantic relevance.

[0074] Vectorization can be used, specifically a pre-trained semantic encoding model, to convert the user-input query text and the candidate memory text filtered by the first-level index into high-dimensional semantic vectors. Next, similarity calculation and sorting are performed, calculating the cosine similarity between the query vector and each candidate memory vector. This value can range from -1 to 1, with higher values ​​indicating greater semantic similarity. Candidate memories are then sorted in descending order based on similarity, and a set similarity threshold (e.g., 0.75) is set to ensure that only historical inference data with similarity higher than this threshold is allowed to be output to the inference module 23.

[0075] The following is a simplified explanation of the multi-level retrieval workflow. Specifically, when a user initiates a query, the query is parsed, and the same word segmentation and keyword extraction process as when storing memories is performed on the user's input text, resulting in a set of query keywords. A primary coarse screening is performed, where these query keywords are searched in the "inverted index table," and all matching memory IDs are merged to form a preliminary candidate memory set. This step can eliminate the vast majority of irrelevant memories within milliseconds. A secondary fine ranking is performed, where the memories in the preliminary candidate set are vectorized and their similarity is calculated. The refined list of highly relevant memories that meet the similarity threshold is then output in descending order of similarity.

[0076] Please continue reading. Figure 3 In one embodiment, the inference reference data includes candidate data. The auxiliary module includes a data management module 222 and an inference management module 223 as described above.

[0077] The reasoning management module 223, the data management module 222, and the input module 21 are connected to each other.

[0078] The data management module 222 is used to store historical inference data and its indicator data.

[0079] The reasoning management module 223 is used to access the data management module 222 in response to obtaining the reasoning source data, so as to obtain historical reasoning data that meets the preset conditions of similarity with the reasoning source data as candidate data.

[0080] In other words, the data management module 222 and the inference management module 223 facilitate the decoupling of historical data storage management and inference scheduling control functions. Simultaneously, the inference management module 223 responds to the inference source data, accurately retrieving historical inference data from the data management module 222 that meets preset similarity criteria as candidate data. This provides a relatively reliable reference for content inference, further enhancing the effectiveness of inference reference data and reducing the risk of invalid data interfering with the inference process, thereby improving the accuracy and efficiency of content inference.

[0081] Furthermore, the inference management module 223 may also include a data matching unit 2231 and a mode selection unit 2232, which are connected together.

[0082] The data matching unit 2231 is used to evaluate the similarity between historical reasoning data and reasoning source data, and to filter historical reasoning data that meet the preset similarity conditions as candidate data.

[0083] The pattern selection unit 2232 is used to acquire indicator data and similarity of candidate data to evaluate auxiliary reasoning patterns that match the indicator data and similarity. The indicator data is then used to participate in the reasoning of the target data according to the corresponding auxiliary reasoning pattern. Thus, by evaluating the auxiliary reasoning patterns of candidate data based on their indicator data and similarity, and by differentiating the auxiliary reasoning patterns of candidate data with varying degrees of credibility, the depth of candidate data participation in the content reasoning process can be differentiated, which helps ensure the reliability of auxiliary content reasoning. This improves the rationality and effectiveness of historical data reuse, further enhancing the performance of the content reasoning system.

[0084] For example, the preset conditions include that the similarity of the historical reasoning data is higher than the similarity of the target number of other historical reasoning data.

[0085] The following provides detailed examples of the working principles of candidate data participating in content reasoning in various auxiliary reasoning modes, and takes the implementation of selecting auxiliary reasoning modes based on confidence factors and similarity in indicator data as an example.

[0086] For example, auxiliary reasoning modes may include a first mode, a second mode, and a third mode.

[0087] When the confidence factor is in the first confidence interval and the similarity is in the first similarity interval, the first mode is selected, and the reasoning result of the candidate data is applied when reasoning the target reasoning data.

[0088] When the confidence factor is within the second confidence interval and the similarity is within the second similarity interval, the second mode is selected to verify the reliability of the candidate data. In response to the candidate data passing the reliability verification, the inference result of the candidate data is applied when inferring the target inference data. The reliability verification includes at least one of cross-validation and boundary checks.

[0089] When the confidence factor is in the third confidence interval and the similarity is in the third similarity interval, the third mode is selected to extract the intermediate layer activation value of the candidate data and inject it as pre-computed data into the inference module 23, so that the inference module 23 can integrate the pre-computed data to perform content inference.

[0090] Otherwise, the mode selection unit 2232 discards the candidate data.

[0091] Among them, the confidence factor values ​​of the first confidence interval, the second confidence interval, and the third confidence interval decrease in sequence, and the similarity values ​​of the first similar interval, the second similar interval, and the third similar interval decrease in sequence.

[0092] By employing the first, second, and third modes, along with a discard mechanism, the adaptive scheduling capability and resource utilization of content reasoning can be significantly improved, enabling refined and hierarchical control of the reasoning strategy. Each auxiliary reasoning mode is sequentially mapped to confidence factors and similarity value ranges, allowing for differentiated processing strategies based on the quality grading of historical candidate data. In high-confidence, high-similarity scenarios, the reasoning results can be directly used, significantly improving reasoning efficiency. In medium-confidence and similarity scenarios, reliability verification is automatically introduced, balancing reasoning efficiency and result security. In low-confidence and similarity scenarios, intermediate layer activation values ​​are extracted as pre-computed data and injected into the reasoning module 23, achieving lightweight reuse of historical computational information and reducing redundant computation. Discarding candidate data with conflicting confidence and similarity values, as well as low-confidence and low-similarity data, reduces interference with the reasoning process.

[0093] In layman's terms, it dynamically selects reasoning strategies based on memory similarity and confidence, enabling computation skipping and intermediate result injection; the reasoning process can include similar memory matching, strategy selection, and result verification, thus achieving dynamic computation skipping.

[0094] For example, for similar memory matching, a vector of input content, i.e., the inference source data, can be generated as a data vector, which can be logically matched with a single memory vector; the cosine similarity between the input vector and all memory vectors in the memory bank is calculated (cosine similarity = vector dot product / (input vector length × memory vector length)); sorted in descending order of similarity, the top 5 most similar memories are selected as candidates.

[0095] If the confidence factor is greater than 0.95 and the similarity is greater than 0.9, it can be considered that the confidence factor is in the first confidence interval and the similarity is in the first similarity interval. The calculation can be skipped completely, and the reasoning result in memory can be used directly. If the confidence factor is greater than 0.8 and the similarity is greater than 0.8, it can be considered that the confidence factor is in the second confidence interval and the similarity is in the second similarity interval. Therefore, the memory result can be quickly verified first (such as cross-checking, boundary checking, etc.). If the verification is successful, the memory result is used; otherwise, the complete reasoning is executed. If the confidence factor is greater than 0.7 and the similarity is greater than 0.7, it can be considered that the confidence factor is in the third confidence interval and the similarity is in the third similarity interval. The activation values ​​of the intermediate layer in memory can be extracted and injected into the model reasoning process as a pre-computation part. Reasoning can be carried out from the intermediate layer to realize memory-enhanced hybrid reasoning. The results are output after verification.

[0096] In other cases, the inference module 23 performs complete inference, that is, inference calculations are performed only based on the inference source data.

[0097] In hybrid reasoning, memory injection can be used to significantly improve reasoning efficiency. For example, pre-stored key-value vectors from intermediate layers of reasoning models such as Transformer (a neural network model) can be directly injected into the corresponding layer of the current reasoning process, thereby effectively reducing redundant calculations by reusing historical computation results.

[0098] Specifically, when performing inference tasks, memory units highly similar to the current input can be retrieved from the memory bank. If a high-quality memory vector (i.e., the output vector of a specific layer of the Transformer) is successfully matched, this pre-computed memory vector can be directly assigned as the output of the corresponding layer of the model. This allows the model to skip the complete computation process of that layer and all layers before it, and continue to perform inference operations in subsequent layers, thereby significantly reducing the overall computation time.

[0099] Specifically, please refer to the following: Figure 3 as well as Figure 4 , Figure 4 This is a flowchart illustrating an embodiment of the working principle of the data fusion unit 2233 of this application.

[0100] In one embodiment, the inference management module 223 further includes a data fusion unit 2233.

[0101] The data fusion unit 2233 is connected to the inference module 23 and is used to use candidate data whose similarity to the inference source data reaches a similarity threshold as fused data. The enhancement vector obtained based on the fused data is assigned as the output of the corresponding model layer of the inference module 23.

[0102] Furthermore, the data fusion unit 2233 includes an input terminal IN, a computing layer, and an output terminal OUT connected in sequence.

[0103] The input terminal IN is used to input fused data and its similarity and confidence factors, and then transmits it to the computation layer.

[0104] The computation layer is used to embed and fuse the representation vector of the fused data with the word embedding vector and position encoding within the layer to obtain the enhanced vector, which is then assigned a value at the output terminal OUT.

[0105] Optionally, the embedding fusion representation assigns adjustment weights to the representation vector of the fusion data to obtain an adjustment vector, and then superimposes the adjustment vector with the word embedding vector and the position encoding.

[0106] The adjustment weights are obtained by fitting the similarity of the fused data, the confidence factor, and the preset level coefficients of the computation layer.

[0107] The overall architecture of memory-enhanced inference is as described in the data fusion unit 2233. A memory fusion mechanism is introduced at the input stage (IN), where the memory vector is synchronously fused with the word embedding vector and positional encoding vector at each layer of the Transformer's input processing stage. Its basic formula can be illustrated as follows: Augmented vector = word embedding vector + positional encoding + α × representation vector of fused data (Formula 5-1) Here, α represents the adjustment weight, which can be considered a parameter that is dynamically adjusted based on experience, determined by both memory similarity and confidence. The formula for calculating α can be illustrated as follows: α = Similarity × Confidence Factor × β (Equation 5-2) Where β represents the preset layer coefficient of the computation layer. During the memory vector generation process, a fusion intensity coefficient can be pre-calculated for each layer as the preset intensity weight of that layer, i.e., the preset layer coefficient. Smaller values ​​(such as 0.3) can be set for shallower layers, and larger values ​​(such as 0.7) can be set for deeper layers to achieve progressive memory fusion.

[0108] Optionally, to ensure the stability and reliability of the system, the data range of α can be pre-constrained. For example, it can include a lower limit constraint and an upper limit constraint; for instance, the lower limit constraint means that if α < 0.1, the system determines that the memory relevance is insufficient and will not perform memory fusion; the upper limit constraint means that α is usually limited to between 0.5 and 0.8 to prevent memory information from excessively covering the original input and to ensure that the model maintains the necessary self-reasoning ability.

[0109] In layman's terms, before inference begins, a set of hierarchical memory vectors adapted to each Transformer layer can be dynamically generated based on the similarity and confidence of the current input and memory, combined with the aforementioned hierarchical parameters. These vectors are then injected layer by layer during subsequent inference. Each vector is a different semantic representation of the same memory content and is specifically adapted to the functional characteristics and input structure of the corresponding Transformer layer, enabling multi-layer injection of the same memory. Although the memories received by each Transformer layer originate from the same content, a set of layer adaptation parameters can be configured for that memory during the memory vector generation stage to achieve differentiated information injection between different levels. In specific implementation, a unique layer identifier vector (such as "LayerID Embedding") is assigned to each Transformer layer (e.g., layers 1 to N, where N is a positive integer) through layer identifier embedding technology. Its dimension is consistent with the memory vector. This identifier vector is fused with the memory vector during the generation process to form a layer-aware memory vector, enabling the same memory to have a vector representation that matches the hierarchical function in different layers. This allows for hierarchical differentiation of the same memory during the memory vector generation stage, enabling the injection of memory information into different Transformer layers with appropriate strength and form, thereby significantly improving the accuracy of memory enhancement and overall reasoning efficiency.

[0110] Please continue reading. Figure 3 In one embodiment, the auxiliary module further includes an auxiliary update module 224, which maintains a verification task file. The verification task file stores the verification status and time information of historical inference data. The verification status is associated with the usage frequency and conditions of the historical inference data, so as to update the indicator data of the historical inference data using the verification status.

[0111] The auxiliary update module 224 is set up to maintain the verification task file, recording and monitoring the verification status and time information of historical inference data. This facilitates the management of the actual usage frequency and status of historical inference data. Furthermore, dynamically updating the indicator data of historical inference data based on the verification status improves the alignment of indicator data with data accuracy, reliability, and current applicability. This helps reduce the interference of long-unused or degraded historical data on the inference process. Ultimately, this ensures the consistent reliability of historical data used in content inference, enhances the rationality of reference source data selection and auxiliary inference mode selection, further strengthens the stability, adaptability, and overall inference performance of the content inference system, and ultimately improves the effectiveness and timeliness of historical inference data.

[0112] It can be assumed that the auxiliary update module 224 can be used to control the auxiliary module and the content reasoning system, continuously monitor, verify, optimize and update the memory units in the historical management file, and ensure that the memory content always remains accurate, timely and effective.

[0113] The auxiliary update module 224 can consist of two sub-modules: a periodic verification sub-module and a feedback and update sub-module. The periodic verification sub-module is used to actively detect the effectiveness and accuracy of memory; the feedback and update sub-module is used to collect and process usage feedback, analyze the reasons for failure, update memory, etc.

[0114] The periodic verification submodule can perform periodic memory verification. Specifically, it can verify task scheduling and verification methods.

[0115] A priority queue mechanism can be used to schedule verification tasks. Upon system startup, a priority queue for verification tasks is first established, allowing higher-priority tasks to be processed first. The system maintains a verification task table, recording the verification status and timestamp of each memory unit. Priority is given to verifying the following types of memories: frequently used but recently unverified memories, memories with high confidence but limited supporting evidence, and important memories with an increasing failure rate.

[0116] The following examples illustrate the verification process using specific numerical values.

[0117] Memory that is frequently used but not recently verified can be defined as one that has been used more than 50 times and has not been verified within 7 days. When verifying a memory, the usage count (usage_count) and the last verified time (last_validated_time) are obtained. The time difference between the current time and the last verified time is calculated. If the usage_count is greater than or equal to 50 and the time difference is greater than or equal to 7 days (i.e., 604,800 seconds), a base priority score of 10 points is awarded. This is adjusted based on usage frequency: 1 point is added for every 50 uses, up to a maximum of 5 points.

[0118] A memory with high confidence but limited supporting evidence can be one with a confidence level above 0.8 but used less than 5 times. To verify it, we obtain the memory's current confidence factor (confidence-x) and usage count (usage_count); we determine if confidence-x is greater than or equal to 0.8 and usage_count is less than or equal to 5; if the conditions are met, a base score of 8 points is awarded; adjustments are made based on confidence level: 2 points are added if confidence-x is between 0.8 and 0.9, and 4 points are added if confidence-x is ≥ 0.9.

[0119] Important memories with an increasing failure rate can be those with an importance greater than 0.7 and a failure rate exceeding 30%. Validation involves obtaining the memory's current importance factor (importance-x); calculating the failure rate (failure_count / max(1, usage_count); and obtaining recent failure trends (e.g., the number of failures in the last 10 uses). If importance-x is greater than or equal to 0.7, the failure rate is greater than or equal to 0.3, and the recent failure trend is upward, the memory meets the criteria and receives a base score of 12 points (i.e., high priority). Adjustments are made based on the severity of the failure rate: 2 points are added for a failure rate between 0.3 and 0.5, and 4 points are added for a failure rate ≥ 0.5.

[0120] Different priority scores are assigned to each case, with higher scores indicating higher priority for verification. The priority scores of all memories can be recalculated every hour, for example.

[0121] In layman's terms, the implementation process of verification task scheduling can be considered as traversing all memory units in the memory bank; applying the above rules to each memory to calculate the total score; updating the priority queue; retrieving the top P tasks from the priority queue (P is the batch size, for example, 5 by default); allocating verification resources to each task and executing the verification process in parallel; and recording the verification results, which may include success / failure status, verification time, types of problems found, and suggested remedial measures.

[0122] The verification method can use a Transformer model with the same memory vector. Based on the original text, full-text calculation is performed to infer the corresponding alignment vector, and then the alignment vector is semantically compared with the representation vector. The specific formula for calculating the semantic relevance index, cosine similarity, can be illustrated as follows: similarity=(A·B) / (||A||×||B||) Formula 6-1 Where, similarity represents semantic similarity; A represents the alignment vector; B represents the stored representation vector; ||A|| represents the length of the alignment vector; and ||B|| represents the length of the stored representation vector. Thus, the range of cosine similarity is [-1, 1], and the semantic similarity should be (0.7, 1.0).

[0123] Cosine similarity measures the directional similarity between the comparison vector and the representation vector: cosine similarity is 1 when their directions are exactly the same; 0 when their directions are perpendicular (or orthogonal); and -1 when their directions are opposite. A semantic similarity greater than 0.85 indicates the memory is correct and verification is successful. Otherwise, verification fails. For failed verifications, the confidence level is immediately reduced (e.g., by 0.1), and the memory is marked for further verification.

[0124] When inference is successfully accelerated using reference source data, its confidence (e.g., increasing the confidence factor by 0.01 while keeping the confidence factor below 1.0) and importance (e.g., increasing the importance factor by 0.005 while keeping the importance factor below 1.0) can be improved, and the successful context can be recorded. When using reference source data leads to inference errors, reduce its confidence level (e.g., reduce the confidence factor by 0.05 and control the confidence factor to be no lower than 0.0) and importance (e.g., reduce the importance factor by 0.02 and control the importance factor to be no lower than 0.0); and analyze the cause of the error.

[0125] At the same time, preset elimination rules can be implemented. For example, when the failure rate of historical inference data is higher than 50% and the confidence factor is less than 0.5, it can be directly deleted from the memory, i.e., the history management file. The importance factor of historical inference data that has not been used for more than 30 days can also be decayed to reduce its importance, for example, by reducing the importance factor by 0.1.

[0126] Furthermore, the auxiliary module may also include an incremental update unit. The incremental update unit is connected to the auxiliary update module 224 and the data management module 222. The incremental update unit can be used to monitor the generation status of the target inference data in real time. When the target inference data meets the preset update conditions, it extracts the key inference original text, representation vector, and corresponding indicator data from the current inference process and writes them incrementally into the historical management file. This incremental update unit enables the incremental writing of historical data, reducing the need for a full reconstruction of the historical management file and lowering the computational load and resource consumption during the data update process.

[0127] Please continue reading. Figure 3 In one embodiment, the content reasoning system further includes an exchange module 24 and a storage module 25, the storage module 25 being used to store data during the content reasoning process.

[0128] The exchange module 24 is connected to the auxiliary module, the inference module 23 and the storage module 25 respectively, so as to enable data interaction among the three components through the exchange module 24.

[0129] For example, storage module 25 may include storage devices and memory devices, etc.

[0130] The inference module 23 may include one or more of a graphics processing unit (GPU), a central processing unit (CPU), and a data processing unit (DPU).

[0131] The descriptions of features in the embodiments corresponding to the communication management device can also be applied to the descriptions of the embodiments corresponding to the content reasoning method, and will not be repeated here. Embodiments of this application provide a communication management method, and the communication management method will be described in detail below in conjunction with its execution flow.

[0132] Please see Figure 5 , Figure 5 This is a flowchart illustrating an embodiment of the reasoning method for the content of this application.

[0133] S101: Obtain inference source data.

[0134] In this embodiment, the inference source data represents the data content that forms the basis of content inference. It can be text, images, or multimodal data, etc., without strict limitation. Among them, multimodal data refers to a dataset composed of two or more different modalities of data. Data modalities can include text, images, audio, video, sensor data, etc.

[0135] S102: Query historical inference data that matches the inference source data as reference source data, and generate inference reference data from the inference process information of the reference source data.

[0136] In this embodiment, after obtaining the inference source data, historical inference data matching the inference source data can be queried and used as reference source data. It is easy to understand that historical inference data is used to characterize data information generated by similar or related inference tasks previously completed by the system. The inference process information corresponding to the reference source data is used as inference reference data to provide reference information for content inference based on the inference source data.

[0137] S103: Combine the inference reference data with the inference source data to perform content inference to obtain the target inference data.

[0138] In this embodiment, after generating inference reference data, the inference source data can be combined with the inference reference data to perform content inference on the inference source data. By reusing effective logic and / or intermediate results from historical inference processes, it is beneficial to reduce repetitive and redundant calculations, simplify the inference path, and thus reduce system computation and resource consumption while ensuring the accuracy of inference results, thereby improving the response speed and overall efficiency of content inference.

[0139] As can be seen from the above, this embodiment introduces the querying and reuse of historical reasoning data, transforming the matched historical reasoning process into reasoning reference data to participate in the current reasoning. During content reasoning, it eliminates the need to perform full, repetitive calculations on the reasoning source data, thereby improving content reasoning efficiency and the overall system reasoning performance.

[0140] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms, or of course by hardware.

[0141] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above-described content reasoning method embodiments.

[0142] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described content reasoning method embodiments at runtime.

[0143] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0144] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above-described content reasoning method embodiments.

[0145] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above-described content reasoning method embodiments.

[0146] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0147] The content reasoning system and electronic device provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only intended to help understand the method and core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A content reasoning system, characterized in that, The content reasoning system includes: The input module is used to acquire inference source data; An auxiliary module, connected to the input module, is used to manage historical inference data; in response to receiving the inference source data, it queries the historical inference data that matches the inference source data as reference source data, and generates inference reference data from the inference process information of the reference source data. The inference module, connected to the input module and the auxiliary module, is used to acquire the inference source data and the inference reference data, and to perform content inference on the inference source data in combination with the inference reference data to obtain the target inference data.

2. The content reasoning system according to claim 1, characterized in that, The auxiliary module includes a vector generation module and a data management module, which are connected together. The vector generation module is used to obtain the original text of the reasoning process of reasoning the target reasoning data and use it as the historical reasoning original text, convert the historical reasoning original text into a representation vector, and transmit the representation vector to the data management module. The data management module is used to construct a historical management file to store the historical inference data and its indicator data; and to retrieve the reference source data that matches the inference source data from the historical management file.

3. The content reasoning system according to claim 2, characterized in that, The vector generation module includes single-vector units and multi-vector units, which are connected together. The single vector unit is used to obtain the original historical reasoning text, ignore stop words in the original historical reasoning text to obtain its substantive word segmentation, assign part-of-speech weights to the substantive word segmentation and evaluate the vector of the original historical reasoning text based on the part-of-speech weights of the substantive word segmentation as its representation vector. The multi-vector unit is used to evaluate the attribute weights of the representation vector; wherein the attribute weights are associated with the confidence factor and / or importance factor of the representation vector; and the multiple representation vectors and their attribute weights are weighted, fused, and normalized to obtain a fused memory vector that participates in the matching of the reference source data.

4. The content reasoning system according to claim 2, characterized in that, The data management module includes a data memory unit and a data retrieval unit; The data memory unit is used to store the historical management file; wherein, the historical management file includes the time information, data type information, the original text of the historical inference data, and its representation vector; the indicator data includes at least one of confidence factor, importance factor, and usage frequency weight; the data type includes fact type, opinion type, and method type, and the initial confidence factors of the fact type, the method type, and the opinion type decrease sequentially; The data retrieval unit is used to access the data memory unit in response to obtaining the inference source data, to filter preliminary data from the historical management file by keyword matching, to evaluate the semantic similarity between the preliminary data and the inference source data to filter the reference source data, and to update the indicator data of the reference source data according to the usage of the reference source data.

5. The content reasoning system according to claim 4, characterized in that, When the reference source data is used successfully, a positive gain is assigned to the reference source data's metrics; when the reference source data is used abnormally, a negative gain is assigned to the reference source data's metrics; and / or, When the unused duration of the historical inference data exceeds the duration threshold, its indicator data is given a negative gain.

6. The content reasoning system according to claim 1, characterized in that, The inference reference data includes candidate data; The auxiliary module includes a data management module and an inference management module, and the inference management module, the data module, and the input module are connected to each other respectively; The data management module is used to store the historical reasoning data and its indicator data; The reasoning management module is used to access the data management module in response to obtaining the reasoning source data, so as to obtain historical reasoning data that meets the preset conditions of similarity with the reasoning source data as the candidate data.

7. The content reasoning system according to claim 6, characterized in that, The reasoning management module includes a data matching unit and a pattern selection unit, which are connected together. The data matching unit is used to evaluate the similarity between the historical inference data and the inference source data, and to filter the historical inference data whose similarity meets the preset conditions as the candidate data; The mode selection unit is used to obtain the indicator data and the similarity of the candidate data, to evaluate the auxiliary reasoning mode that matches the indicator data and the similarity, and to make the indicator data participate in the reasoning of the target reasoning data according to the corresponding auxiliary reasoning mode.

8. The content reasoning system according to claim 7, characterized in that, The preset conditions include the similarity of the historical reasoning data being higher than the similarity of the target number of other historical reasoning data.

9. The content reasoning system according to claim 7, characterized in that, The indicator data includes confidence factors; the auxiliary reasoning modes include a first mode, a second mode, and a third mode; When the confidence factor is in the first confidence interval and the similarity is in the first similarity interval, the first mode is selected, and the reasoning result of the candidate data is applied when reasoning the target reasoning data. When the confidence factor is in the second confidence interval and the similarity is in the second similarity interval, the second mode is selected to verify the reliability of the candidate data; in response to the candidate data passing the reliability verification, the inference result of the candidate data is applied when inferring the target inference data; wherein, the reliability verification includes at least one of cross-validation and boundary checking; When the confidence factor is in the third confidence interval and the similarity is in the third similarity interval, the third mode is selected to extract the intermediate layer activation value of the candidate data and inject it as pre-computed data into the inference module, so that the inference module can integrate the pre-computed data to perform content inference. Otherwise, the mode selection unit discards the candidate data; The confidence factor values ​​of the first confidence interval, the second confidence interval, and the third confidence interval decrease sequentially, as do the similarity values ​​of the first similarity interval, the second similarity interval, and the third similarity interval.

10. The content reasoning system according to claim 6, characterized in that, The reasoning management module also includes a data fusion unit; The data fusion unit is connected to the inference module and is used to take candidate data whose similarity to the inference source data reaches a similarity threshold as fused data; and to assign the enhancement vector obtained based on the fused data as the output of the corresponding model layer of the inference module.

11. The content reasoning system according to claim 10, characterized in that, The data fusion unit includes an input terminal, a computing layer, and an output terminal connected in sequence. The input terminal is used to input the fused data and its similarity and confidence factors, and transmit them to the computing layer; The computation layer is used to embed and fuse the representation vector of the fused data with the word embedding vector and position encoding within the layer to obtain the enhanced vector, which is then assigned a value at the output terminal.

12. The content reasoning system according to claim 11, characterized in that, The embedding fusion representation assigns adjustment weights to the representation vector of the fused data to obtain an adjustment vector, and then superimposes the adjustment vector with the word embedding vector and the position encoding; The adjustment weights are obtained by fitting the similarity of the fused data, the confidence factor, and the preset level coefficients of the calculation layer.

13. The content reasoning system according to claim 1, characterized in that, The auxiliary module also includes an auxiliary update module, which is used to maintain a verification task file. The verification task file stores the verification status and time information of the historical inference data. The verification status is associated with the usage frequency and usage of the historical inference data, so as to update the indicator data of the historical inference data using the verification status.

14. The content reasoning system according to claim 1, characterized in that, The content reasoning system also includes an exchange module and a storage module, wherein the storage module is used to store data during the content reasoning process; The exchange module is connected to the auxiliary module, the inference module, and the storage module respectively, so as to enable data interaction among the three components through the exchange module.

15. An electronic device, characterized in that, The electronic device includes: Equipment body; The content reasoning system as described in any one of claims 1 to 14 is located on the device body.

Images