Model credibility evaluation method and device, electronic equipment and medium

Abstract

The embodiment of the application applies a model credibility evaluation method and device, electronic equipment and medium, relates to the technical field of artificial intelligence, and a specific implementation manner of the method comprises: obtaining an activation vector generated by a preset language model when reasoning a target sentence; based on a pre-constructed sentence truth value direction vector and a semantic polarity direction vector, performing projection processing on the activation vector to obtain a feature vector of the activation vector in the sentence truth value direction and the semantic polarity direction for representing the authenticity determination result of the target sentence and the semantic sentence type; based on the feature vector and a true-false classifier pre-trained based on a constructed sample sentence set, analyzing the reasoning credibility of the preset language model to obtain a credibility evaluation result. The model internal cognition is mined, the semantic sentence interference correction is considered, the misjudgment probability is effectively reduced, and the accuracy and interpretability of the credibility evaluation result are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a model credibility evaluation method, apparatus, electronic device, and medium. Background Technology

[0002] In recent years, Large Language Models (LLMs) have been widely used in various fields such as dialogue systems, content generation, and task execution. However, as the capabilities of these models continue to improve, LLMs may generate false information that does not conform to the facts, and in some cases, even output statements that are known to be untrue.

[0003] Currently, in related technologies, the credibility assessment of models is mainly divided into three categories: the first is the fact base verification method, which judges the authenticity by comparing the generated content with the fact database or search engine; the second is the consistency detection method, which assesses the credibility of the model by comparing the differences in the model's output multiple times; and the third is the contrastive learning method, which constructs complementary question-answer sample pairs of true and false, and judges the authenticity of the answers based on the probability complementarity characteristics.

[0004] However, the relevant technologies rely solely on the output text to determine the model's credibility, which is a black-box detection method and makes it difficult to explain the basis for the judgment. Furthermore, they are susceptible to interference from rhetoric and linguistic ambiguity, have insufficient ability to handle complex semantic scenarios, and have a high probability of misjudgment, thus failing to meet the model credibility evaluation requirements in high-reliability scenarios. Summary of the Invention

[0005] This application provides a model credibility assessment method, apparatus, electronic device, and medium to solve the problems in related technologies where model credibility is assessed solely based on output text, the judgment logic lacks interpretability, and is prone to misjudgment.

[0006] In a first aspect, embodiments of this application provide a method for evaluating the credibility of a model. The method includes: obtaining activation vectors generated by a preset language model when reasoning about a target statement; the activation vectors characterize the inherent cognitive results formed by the preset language model after semantic understanding of the target statement; projecting the activation vectors based on pre-constructed truth-value direction vectors and semantic polarity direction vectors to obtain feature vectors of the activation vectors in the truth-value direction and semantic polarity direction; the feature vectors characterize the truth-based judgment result of the target statement and the semantic sentence type of the target statement; and analyzing the reasoning credibility of the preset language model based on the feature vectors and a truth / false classifier to obtain a credibility evaluation result. The truth / false classifier is pre-trained based on a constructed set of sample statements.

[0007] In some embodiments, the projection processing of activation vectors based on pre-constructed statement truth direction vectors and semantic polarity direction vectors to obtain feature vectors of activation vectors in the statement truth direction and semantic polarity direction includes: determining the mean of target sample activation vectors corresponding to the activation vectors based on at least one set of topic statements in the sample statement set, wherein the mean of target sample activation vectors is the mean of sample activation vectors of at least one sample activation vector in the set of topic statements corresponding to the target statement; centering the activation vectors based on the mean of target sample activation vectors to obtain updated activation vectors corresponding to the target statement; projecting the updated activation vectors based on the statement truth direction vectors to obtain truth direction projection values ​​of the updated activation vectors along the statement truth direction; projecting the updated activation vectors based on the semantic polarity direction vectors to obtain polarity direction projection values ​​of the updated activation vectors along the semantic polarity direction; normalizing the truth direction projection values ​​based on the number of lexical units corresponding to the target statement to obtain updated truth direction projection values, and integrating the updated truth direction projection values ​​and polarity direction projection values ​​to obtain feature vectors.

[0008] In some embodiments, analyzing the reasoning credibility of a preset language model based on feature vectors and a true / false classifier to obtain a credibility evaluation result includes: inputting the feature vectors into the true / false classifier to generate a credibility probability value corresponding to the preset language model; if the credibility probability value is greater than or equal to a preset probability threshold, determining that the credibility result is that the reasoning result of the preset language model on the target statement is credible; if the credibility probability value is less than the preset probability threshold, determining that the credibility result is that the reasoning result of the preset language model on the target statement is unreliable.

[0009] In some embodiments, the method includes: dividing at least one sample statement into topics to obtain at least one set of topic statements; based on the statement authenticity of at least one topic statement in a first set of topic statements, performing truth / false labeling on at least one topic statement to obtain at least one statement truth / false label corresponding to at least one topic statement, wherein the first set of topic statements is any set of at least one set of topic statements; based on the semantic polarity of at least one topic statement, performing semantic polarity labeling on at least one topic statement to obtain at least one semantic polarity label corresponding to at least one topic statement, wherein semantic polarity is used to indicate whether at least one topic statement is an affirmative statement or a negative statement; inputting at least one topic statement into a preset language model for inference to obtain at least one sample activation vector corresponding to at least one topic statement; and constructing a set of sample statements based on at least one sample activation vector, at least one statement truth / false label, and at least one semantic polarity label.

[0010] In some embodiments, based on pre-constructed statement truth direction vectors and semantic polarity direction vectors, the activation vectors are projected to obtain feature vectors of the activation vectors in the statement truth direction and semantic polarity direction. Before this, the method includes: determining the mean of a first sample activation vector of at least one sample activation vector corresponding to a first set of topic statements; constructing a first linear combination relationship containing an initial statement truth direction vector and an initial semantic polarity direction vector for each activation vector in the at least one sample activation vector based on the mean of the first sample activation vector, at least one statement truth / false label, and at least one semantic polarity label; and fitting the at least one first linear combination relationship based on a preset fitting algorithm to obtain the truth direction vector and polarity direction vector.

[0011] In some embodiments, the reasoning credibility of a preset language model is analyzed based on feature vectors and a true / false classifier to obtain a credibility evaluation result. Prior to this, the method includes: centering at least one sample activation vector based on the mean of a first sample activation vector to obtain at least one updated sample activation vector; projecting at least one updated sample activation vector based on the statement truth direction vector and semantic polarity direction vector to obtain at least one sample feature vector corresponding to the at least one updated sample activation vector; normalizing at least one sample feature vector based on the number of lexical units corresponding to each topic statement in at least one topic statement to obtain at least one updated sample feature vector; and training the initial true / false classifier using at least one updated sample feature vector as input to the initial true / false classifier and at least one statement true / false label as a supervision signal to obtain the true / false classifier.

[0012] Secondly, embodiments of this application provide a model credibility evaluation apparatus, the apparatus comprising:

[0013] The acquisition unit is used to acquire the activation vector generated by the preset language model when reasoning about the target statement. The activation vector is used to represent the internal cognitive result formed by the preset language model after completing the semantic understanding of the target statement. The projection unit is used to project the activation vector based on the pre-constructed truth direction vector and semantic polarity direction vector of the statement to obtain the feature vector of the activation vector in the truth direction and semantic polarity direction of the statement. The feature vector is used to characterize the truth judgment result of the target statement and the semantic sentence type of the target statement. The analysis unit is used to analyze the reasoning credibility of a pre-defined language model based on feature vectors and a true / false classifier to obtain credibility evaluation results. The true / false classifier is pre-trained based on a constructed set of sample sentences.

[0014] Thirdly, embodiments of this application provide an electronic device, including: a processor and a memory for storing a computer program capable of running on the processor, wherein, when the processor runs the computer program, it performs the method described in any embodiment of the first aspect.

[0015] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods described in any embodiment of the first aspect.

[0016] Fifthly, embodiments of this application provide a computer program product including a computer program that, when executed by a processor, performs the method described in any embodiment of the first aspect.

[0017] This application provides a model credibility evaluation method that obtains activation vectors generated by a preset language model when reasoning about a target statement. Based on pre-constructed truth direction vectors and semantic polarity direction vectors, the activation vectors are projected to obtain feature vectors in the truth direction and semantic polarity direction. Based on these feature vectors and a truth / false classifier, the credibility of the preset language model's reasoning is analyzed. This process overcomes the black-box limitation of evaluating model credibility solely based on output text. By mining the model's internal cognitive representations and considering semantic and sentence structure interference correction, it effectively reduces the probability of misjudgment and improves the accuracy and interpretability of the credibility evaluation results.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are merely embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort, and this application can be applied to other similar scenarios based on the provided drawings.

[0020] Figure 1 A flowchart illustrating the first model credibility assessment method provided in this application embodiment; Figure 2 A flowchart illustrating the second model credibility evaluation method provided in this application embodiment; Figure 3 A schematic diagram of a model credibility evaluation device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. The described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be noted that the terms "system," "device," "unit," and / or "module" used in this application are methods of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they can be replaced by other expressions.

[0023] Hereinafter, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include one or more of that feature.

[0024] Large Language Models (LLMs) are widely used in various scenarios such as dialogue systems, content generation, and task execution. The behavior of models outputting "false information" or "knowingly false statements" has gradually attracted attention. Especially after the model has a certain ability to reason about strategies and simulate motivations, it may produce statements that it knows to be false, which is the so-called "AI (Artificial Intelligence) lying" situation.

[0025] Currently, in related technologies, the credibility assessment of models mainly falls into three categories: First, the fact base verification method, which compares the content generated by the model with external fact databases or search engines to determine whether it conforms to objective facts; second, the consistency detection method, which compares the differences in the model's output under multiple contexts or different parameter settings to infer whether it outputs inconsistent or unreliable content; and third, the contrastive learning method, such as the Contrast Consistent Search (CCS) method, which constructs contrast pairs, where for each question, the probability of the model answering true and the probability of answering false are complementary, thereby distinguishing between true and false answers.

[0026] However, the relevant technologies still have the following limitations: 1) Relying on output text and lacking access to internal structure: The relevant techniques only rely on the final output text of the model for analysis, without accessing or exploring the internal structure of the model. This is a "black box method" and the judgment criteria lack interpretability.

[0027] 2) Weak ability to handle complex semantic structures: Related technologies are easily misled by the rhetorical devices and linguistic ambiguity of the model, and have insufficient ability to handle complex semantic structures, which can easily lead to misjudgments.

[0028] 3) Unable to detect “motivated lies”: Related technologies have difficulty capturing the strategic expressions of models during the reasoning stage and have difficulty identifying motivated false outputs with subjective intentions, that is, they cannot judge “motivated lies”.

[0029] To address the aforementioned issues, this application proposes a model credibility evaluation method. By extracting internal activation vectors from the model's reasoning process and combining them with the truth-value direction and semantic polarity direction to complete a projection representation, a two-dimensional feature vector representing the model's "inner beliefs" and "sentence structure" is obtained. Based on a pre-trained truth / false classifier, the method analyzes the model's inherent cognitive tendencies using these feature vectors and evaluates the credibility of the model's reasoning based on the analysis results. This process directly captures the model's inherent encoding of statement truthfulness by accessing its internal activation states, making the judgment more interpretable. Simultaneously, the introduction of polarity direction compensation compensates for the influence of complex semantic structures on the activation vectors, significantly improving the evaluation accuracy and generalization ability in complex semantic scenarios.

[0030] The following section provides a detailed description of a model credibility assessment method provided in this application, with reference to the accompanying drawings.

[0031] Figure 1 A flowchart of a model credibility evaluation method provided in an embodiment of this application is shown. Figure 1 As shown, the model credibility assessment method includes steps 101-103.

[0032] Step 101: Obtain the activation vector generated by the preset language model when reasoning about the target statement. In the embodiments of this application, the preset language model (hereinafter referred to as the model) is a large language model that has been trained, has a transformer structure, and is used for sentence reasoning and content generation.

[0033] The target statement is the natural language statement to be tested for credibility and authenticity. It is either pre-generated by the model or provided by the user. The specific method of obtaining the target statement is not limited in the embodiments of this application.

[0034] An activation vector is a numerical representation generated by neurons within a model when processing an input statement; it is a high-dimensional feature vector output by the intermediate layers of a neural network. It encodes the model's understanding of the input statement, including semantic, syntactic, and truth / false information. In other words, the activation vector represents the model's internal cognitive outcome after completing a semantic understanding of the target statement.

[0035] Optionally, for the extraction position of the activation vector, this application selects the residual flow activation vector of the intermediate layer (such as the 12th layer) in the process of the model processing the sentence, and extracts it at the position corresponding to the last token of the target sentence. This position represents the final "conclusion" of the model's understanding of the entire target sentence.

[0036] Specifically, the target sentence, either pre-generated by the model or provided by the user, is input into a preset language model. The model processes the sentence word by word, performing a complete sentence reasoning and parsing process. The activation vectors corresponding to the last token position of the target sentence in the intermediate residual flow network layer of the model are extracted.

[0037] Step 102: Based on the pre-constructed statement truth direction vector and semantic polarity direction vector, the activation vector is projected to obtain the feature vectors of the activation vector in the statement truth direction and semantic polarity direction.

[0038] In the embodiments of this application, the statement truth direction vector is a pre-learned internal feature direction vector that linearly distinguishes between true and false statements, reflecting the direction of distinction between "true statements" and "false statements" in the activation space of the model. In this direction, the model's activation vector has significant linear separability between "true" and "false" statements.

[0039] The semantic polarity direction vector is a pre-learned modified feature direction vector adapted to complex semantic sentence structure differences (such as affirmative and negative sentences). It reflects the encoding differences in the activation space of the model. The semantic polarity direction vector is used to supplement the ability to separate true and false statements under changes in semantic polarity (affirmative / negative), correct the separation failure caused by negative sentences, and eliminate the interference of semantic sentence structure in true and false judgment.

[0040] The feature vector is a two-dimensional feature data obtained by projecting and normalizing the activation vector. It integrates the model's true / false cognition with semantic sentence features and is used to characterize the authenticity judgment result of the target sentence and the semantic sentence type of the target sentence.

[0041] Specifically, by centering the activation vector and performing bidirectional projection operations on the truth value direction and semantic polarity direction of the statement, and by normalizing the vector to remove high-dimensional redundant information, the model obtains the true and false cognitive features and semantic sentence structure features of the target statement, and integrates them to generate a quantifiable and classifiable standardized feature vector.

[0042] Step 103: Based on feature vectors and true / false classifiers, analyze the reasoning credibility of the preset language model to obtain credibility evaluation results.

[0043] In the embodiments of this application, the true / false classifier is pre-trained based on a constructed set of sample sentences. It is a binary classification model (such as logistic regression) pre-trained based on a labeled sample dataset, used to identify the true / false tendency of the inherent cognition of the preset language model and output the model's credibility probability. The set of sample sentences is a pre-constructed structured sample dataset containing multiple topics, semantic polarity annotations, and true / false sentence annotations.

[0044] Reasoning credibility refers to the degree of matching between the internal cognition of the pre-set language model and objective facts and logical rules. It is used to determine whether the model has false reasoning or deliberate lying behavior.

[0045] Specifically, the standardized feature vectors are input into a pre-trained true / false classifier. Through the supervised learning logic of the classifier, the true cognitive tendency inside the model is analyzed, the probability value of the model's reasoning credibility is quantified, and finally, an evaluation conclusion on whether the model's reasoning result is credible or unreliable is output.

[0046] In summary, the method for evaluating model credibility proposed in this application involves obtaining the activation vectors generated by a pre-defined language model when reasoning about a target statement; projecting these activation vectors onto pre-constructed truth direction vectors and semantic polarity direction vectors to obtain feature vectors of the activation vectors in the truth direction and semantic polarity direction; and analyzing the reasoning credibility of the pre-defined language model based on these feature vectors and a truth / false classifier to obtain the credibility evaluation result. This process overcomes the black-box limitation of evaluating model credibility solely based on output text. By mining the internal cognitive representations of the model and considering semantic and sentence structure interference correction, it effectively reduces the probability of misjudgment of model credibility and improves the accuracy and interpretability of the credibility evaluation results.

[0047] As one possible approach, before using a true / false classifier to evaluate the credibility of a pre-defined language model, it is necessary to first construct a sample dataset and then construct a truth value direction vector and a semantic polarity direction vector based on the sample dataset, thereby training the true / false classifier.

[0048] The process for constructing the sample dataset is as follows: 1) Divide at least one sample statement into topics to obtain at least one set of topic statements.

[0049] In the embodiments of this application, the sample statements are a large number of natural language statements with objective truth or falsehood standards collected manually or semi-automatically. The sample statements should cover multiple topics (such as city affiliation, animal and plant classification, language translation, scientific common sense, etc.) and try to control the consistency of language style to eliminate interference from non-truthful factors.

[0050] Specifically, based on the topics to which a large number of sample statements belong, the sample statements are divided into several topic statement sets, and the sample statements (topic statements) in each topic statement set belong to the same topic.

[0051] 2) Based on the statement authenticity of at least one topic statement in the first topic statement set, mark the at least one topic statement as true or false to obtain at least one statement true or false label corresponding to at least one topic statement. The first topic statement set is any topic statement set in the at least one topic statement set.

[0052] In the embodiments of this application, for each type of topic statement, an equal number of true statements and false statements are constructed based on the truthfulness of each statement. For example: "The chemical formula of water is H2O." (True) "The sun is a satellite of the earth." (False) The data format must be uniformly set to natural language sentences, usually ending with a period, to facilitate subsequent input into a preset language model for processing and extraction of activation vectors at standard positions.

[0053] For each topic statement, a clear "statement truth / false label" (T) is assigned based on objective facts, defined as: if the statement is true, then t = +1; if the statement is false, then t = 1.

[0054] 3) Based on the semantic polarity of at least one topic statement, perform semantic polarity annotation on at least one topic statement to obtain at least one semantic polarity label corresponding to at least one topic statement. The semantic polarity is used to indicate whether at least one topic statement is an affirmative statement or a negative statement.

[0055] In the embodiments of this application, to enhance the model's generalization ability, semantic polarity diversity needs to be introduced, namely, positive statements (affirmative sentences) and negative statements (negative sentences). Before semantic polarity annotation, the original topic statements can be "negated" using procedural templates or manual methods, that is, constructing a negative version for each positive statement. Specific methods include inserting negative words (such as "no," "not," "not," etc.) into the original topic statements or using equivalent negative expressions.

[0056] For example, the original topic statement "Mars is a planet" (true) is changed to "Mars is not a planet" (false); the original topic statement "The sun revolves around the earth" (false) is changed to "The sun does not revolve around the earth" (true).

[0057] For each topic statement with a negative version, maintain consistent "true / false labels" to ensure the training data remains accurate and usable. Add a "semantic polarity label" (P) to each topic statement: if the statement is affirmative, then p = +1; if the statement is negative, then p = 1 / 2. 1. Ultimately, a diverse training dataset is formed, containing three-dimensional labels (topic, semantic polarity, truth or falsehood), which helps in the subsequent construction of a generalizable "truth orientation".

[0058] 4) Input at least one topic statement into the preset language model for inference and obtain at least one sample activation vector corresponding to at least one topic statement.

[0059] In the embodiments of this application, after the dataset is prepared, these topic sentences need to be input into the specified large language model one by one, and the intermediate representations (i.e. activation vectors) of the model in the process of processing sentences need to be extracted in order to analyze the encoding method of the model's "inner state" for true and false concepts.

[0060] Specifically, each topic statement is treated as a complete input unit and fed into the model. The intermediate layer residual flow activation vector is extracted from the position corresponding to the last token of the topic statement. Each topic statement corresponds to one activation vector (i.e., a sample activation vector). Where i represents the topic category and j represents the statement number. Finally, the following metadata is recorded for each sample statement: : Sample activation vector Statement truth / false tags Semantic polarity label (negation / affirmation).

[0061] 5) Construct a set of sample statements based on at least one sample activation vector, at least one statement truth / false label, and at least one semantic polarity label.

[0062] The sample activation vectors, truth and false labels, and semantic polarity labels of all sample sentences are combined to form a structured training feature set, which is used to learn two key direction vectors that can reflect "truth and false judgment" and "semantic polarity", namely the truth value direction vector and the semantic polarity direction vector.

[0063] Understandably, for each sample statement, if the statement is true, then... = +1; if the statement is false, then = 1. If the statement is affirmative, then = +1; If the statement is negative, then = 1. Among them, the true / false label of the statement is used to indicate the position where the activation vector should be distributed in the true / false space; the semantic polarity label is used to help eliminate the spurious separation factor introduced by the sentence structure (such as negation) in subsequent learning.

[0064] The process for constructing the truth value direction vector and semantic polarity direction vector of a statement is as follows: 1) Determine the mean of the first sample activation vectors corresponding to at least one sample activation vector of the first topic statement set.

[0065] In embodiments of this application, the first set of topic statements is any topic statement in at least one set of topic statements. The mean of the sample activation vectors of each topic statement in each set of topic statements is calculated to obtain the mean of the sample activation vectors for each set of topic statements.

[0066] 2) Based on the mean of the first sample activation vector, at least one statement truth value label and at least one semantic polarity label, construct a first linear combination relationship for each activation vector in at least one sample activation vector, which includes the initial statement truth value direction vector and the initial semantic polarity direction vector.

[0067] In the embodiments of this application, in order to extract the general truth direction vector and semantic polarity direction vector, a structure containing the initial statement truth direction vector is constructed. and the initial semantic polarity direction vector The linear combination relationship, where the initial truth direction vector and the initial semantic polarity direction vector refer to the direction vectors that have not yet been learned.

[0068] Assuming the sample activation direction in the pre-defined language model It consists of the following three parts:

[0069] Part One The mean of the sample activation vectors for the topic statement set i is used to represent the general language expression bias for that topic; Part Two Indicates the main direction of true and false information in the activation space; Part Three It indicates the direction of offset correction for true / false information in affirmative / negative contexts; Right now, It is modeled as a linear combination of "the mean of the topic activation vector + the truth / falseness direction of the statement + the semantic polarity direction". In other words, for a statement, its position in the activation space can be determined by its topic context. Authenticity and semantic polarity The decision is made by a combination of three factors.

[0070] 3) Based on the preset fitting algorithm, at least one first linear combination relationship is fitted to obtain the truth value direction vector and semantic polarity direction vector of the statement.

[0071] In the embodiments of this application, the preset fitting algorithm adopts the least squares method, and the optimization objective is to solve the parameters by minimizing the vector restoration error.

[0072] Specifically, in order to obtain the two key direction vectors in the above linear combination relationship... and Define it as the optimal solution that minimizes the following loss function:

[0073] The significance of this optimization objective is to find two directions such that the combination of them results in... It can most accurately restore the original activation vector, thereby extracting the two main axes of "true / false" and "polarity".

[0074] Through the learned truth direction vector of statements and semantic polarity direction vector We will construct a practically deployable true / false classifier. This classifier can map the activation vectors of any input statement in a large language model to a two-dimensional space and determine whether it is true or false based on the projection result. The training process of the true / false classifier is as follows: 1) Based on the mean of the first sample activation vector, center at least one sample activation vector to obtain at least one updated sample activation vector.

[0075] In the embodiments of this application, before training, the activation vector of each sample is first... Perform centralized processing, subtracting the mean of the sample activation vectors of its respective topic. :

[0076] in, Sample activation vector The corresponding updated sample activation vector.

[0077] 2) Based on the truth direction vector and semantic polarity direction vector of the statement, project the activation vector of at least one updated sample to obtain at least one sample feature vector corresponding to the activation vector of at least one updated sample.

[0078] In the embodiments of this application, the activation vector is projected onto the direction vectors of the statement truth direction and the semantic polarity direction to obtain: the projection value along the statement truth direction. and projection values ​​along the semantic polarity direction .

[0079] in, This represents the transposed updated sample activation vector. Following the method described above, the activation vector of each sample statement is mapped to a two-dimensional feature vector ( , ), representing the internal representation of the sample statement in terms of "true or false" and "polarity".

[0080] 3) Based on the number of lexical units corresponding to each topic statement in at least one topic statement, normalize at least one sample feature vector to obtain at least one updated sample feature vector.

[0081] In the embodiments of this application, the number of tokens represents the total number of tokens contained in the topic sentence (or sample sentence) after word segmentation, and is used to characterize the length of the sentence. To avoid the activation vector in direction T becoming too long as the sample sentence is longer... G The projection intensity on the surface tends to weaken, so the projection value along the truth value direction of the statement needs to be length normalized.

[0082] Specifically, the number of tokens for each topic statement is calculated and denoted as . ; Projection value along the truth value direction of the statement Normalization is performed as follows:

[0083] Along the direction of semantic polarity The projection retains the original values. The final sample feature vector (i.e., the updated sample feature vector) for each topic statement is:

[0084] This transformation ensures that the projection intensity of the true and false concepts onto the data scale remains consistent regardless of the length of the statement, thereby improving the robustness of the true and false classifier.

[0085] 4) Train the initial true / false classifier by using at least one updated sample feature vector as input and at least one statement true / false label as supervision signal.

[0086] In the embodiments of this application, based on the two-dimensional feature vector constructed above... A supervised classifier is trained to evaluate the reliability of the model.

[0087] Specifically, the two-dimensional feature vector As input, the statement truth / false labels The values ​​∈{True, False} are converted into binary labels as supervision signals. Logistic regression is used for training, and the classifier is trained on a balanced training set. Cross-validation or hold-out methods are used to evaluate its generalization ability on different topics and statements of different polarities. The output is a probability value p∈[0,1]. A threshold (e.g., 0.5) is set to evaluate the model's credibility. For example, when the output probability value p>0.5, it indicates that the model's reasoning is credible, and when p<0.5, it indicates that the model's reasoning is uncredible. Finally, a true / false classifier that can accurately distinguish the model's inherent true / false cognition is obtained.

[0088] In summary, this application constructs a standardized sample sentence set with multiple themes, dual semantic polarities, and truth / false labels. It then uses the least squares method to fit truth direction vectors and semantic polarity direction vectors that can adapt to complex sentence structures. Furthermore, it optimizes features by combining data centralization and length normalization, ultimately training a high-performance truth / false classifier. This allows for the automated evaluation of the credibility of the output content of a pre-defined language model. This approach overcomes the limitations of traditional black-box evaluation, avoids misjudgment interference caused by rhetoric, semantic ambiguity, and sentence structure, accurately uncovers inherent cognitive biases in the model's reasoning stage, and thus accurately identifies motivated lies, improving the credibility and security of artificial intelligence systems.

[0089] As one possible implementation method, Figure 2 A flowchart of the second model credibility evaluation method provided in the embodiments of this application is shown. Figure 2 based on Figure 1 The illustrated embodiment further defines steps 102 and 103. Figure 2 In the illustrated embodiment, step 102 includes steps 201, 202, 203, 204, and 205, and step 103 includes steps 206, 207, and 208, as follows. Figure 2 As shown, the method includes the following steps: Step 201: Based on at least one set of topic statements in the sample statement set, determine the mean value of the target sample activation vector corresponding to the activation vector.

[0090] Step 202: Based on the mean of the target sample activation vector, the activation vector is centered to obtain the updated activation vector corresponding to the target statement.

[0091] In the embodiments of this application, the mean of the target sample activation vector is the mean of the sample activation vectors of each sample in the set of topic sentences to which the target sentence belongs. Specifically, based on the topic to which the target sentence belongs, the mean of the corresponding sample activation vector is selected as the mean of the target sample activation vector from the mean values ​​of each topic calculated during the training phase.

[0092] The updated activation vector is obtained by subtracting the mean of the target sample activation vectors from the activation vector of the target statement, thus eliminating topic bias, as follows:

[0093] in, The activation vector for the target statement. ∈ d The activation vector is a high-dimensional representation of the target statement within the model; To update the activation vector; The mean of the activation vectors of the target sample.

[0094] Step 203: Based on the truth direction vector of the statement, perform projection processing on the update activation vector to obtain the truth direction projection value of the update activation vector along the truth direction of the statement.

[0095] Step 204: Based on the semantic polarity direction vector, project the updated activation vector to obtain the polarity direction projection value of the updated activation vector along the semantic polarity direction.

[0096] In the embodiments of this application, the truth direction vector of the statement is pre-constructed during the training phase. and semantic polarity direction vector The updated activation vector is subjected to vector projection operation to extract the true direction projection value and the polarity direction projection value, as follows:

[0097]

[0098] in, and These are the true value projected in the direction of the polarity and the polarity projected in the direction of the polarity, respectively. This represents the updated activation vector after transposition.

[0099] Step 205: Based on the number of lexical units corresponding to the target sentence, normalize the truth direction projection value to obtain the updated truth direction projection value, and integrate the updated truth direction projection value and the polarity direction projection value to obtain the feature vector.

[0100] In the embodiments of this application, in order to eliminate numerical errors caused by the length of statements, the projection values ​​of the true value direction need to be normalized.

[0101] Specifically, the number of tokens in the target statement is counted, and the truth-direction projection value is normalized to obtain the updated truth-direction projection value, as follows:

[0102] in, The number of terms in the target statement.

[0103] The polarity direction projection value remains unchanged. The updated truth direction projection value is combined with the polarity direction projection value to form the final two-dimensional feature vector of the target statement. .Should The vector is the standard input used for judging whether a classifier is true or false.

[0104] Step 206: Input the feature vector into the true / false classifier to generate the confidence probability value corresponding to the preset language model.

[0105] In the embodiments of this application, the feature vector of the target statement is... The input is fed into a pre-trained true / false classifier, and the output is a true / false prediction (i.e., a confidence probability value). The specific method is as follows: Will The input is fed into the true / false classifier (logistic regression model). The true / false classifier will output a probability value between 0 and 1, which is the confidence probability value, as follows:

[0106] in, This represents the logistic regression function. The confidence probability value is... The larger the value, the closer the model's internal cognition is to objective facts, and the higher the credibility of the reasoning.

[0107] Step 207: If the credibility probability value is greater than or equal to the preset probability threshold, determine that the credibility result is the credibility of the preset language model's reasoning result for the target statement.

[0108] Step 208: If the credibility probability value is less than the preset probability threshold, determine that the credibility result is that the inference result of the preset language model on the target statement is unreliable.

[0109] In the embodiments of this application, the preset probability threshold, which is a pre-set confidence judgment threshold, can be configured according to actual needs and is not limited in the embodiments of this application. This application uses 0.5 as an example. Based on the confidence probability value output by the true / false classifier in step 206, it is determined whether the reasoning result of the preset language model for the target statement is credible.

[0110] Specifically, the generated confidence probability value is compared with a preset probability threshold. When the confidence probability value is greater than or equal to the preset probability threshold (i.e., ... When the internal cognition of the preset language model is consistent with objective facts, the model's reasoning result is considered reliable. When the reliability probability value is less than a preset probability threshold (i.e., ...), the model's internal cognition is considered consistent with objective facts, and the model's reasoning result is considered reliable. When the model fails to generate a statement, it is determined that the underlying cognition of the pre-defined language model in this inference is biased, and the generated statement is a false output (i.e. the model output is a lie), and the model's inference result is unreliable.

[0111] In summary, this application eliminates topic semantic bias by performing topic mean centering on the model activation vector corresponding to the target sentence; it projects the corrected activation vector onto the pre-constructed truth direction and semantic polarity direction, and performs scale normalization in conjunction with the sentence word length to construct a standardized two-dimensional feature vector, avoiding interference caused by changes in the length of the input target sentence, which can further improve the robustness and practical applicability of the true / false classifier; then, it outputs a credibility probability value through the trained true / false classifier, and automatically judges the credibility status of the model's reasoning results based on the credibility probability value, effectively avoiding evaluation interference caused by sentence structure and sentence length, accurately capturing the model's inherent cognitive bias, and realizing refined, interpretable, and automated detection of the credibility of large language model reasoning.

[0112] To implement the above embodiments, this application also provides a model credibility evaluation device. Figure 3 This is a schematic diagram of the structure of a model credibility evaluation device 300 provided in an embodiment of this application. Figure 3 As shown, the device includes: The acquisition unit 310 is used to acquire the activation vector generated by the preset language model when reasoning about the target statement. The activation vector is used to represent the internal cognitive result formed by the preset language model after completing the semantic understanding of the target statement. Projection unit 320 is used to project the activation vector based on the pre-constructed statement truth direction vector and semantic polarity direction vector to obtain the feature vector of the activation vector in the statement truth direction and semantic polarity direction. The feature vector is used to characterize the truth judgment result of the target statement and the semantic sentence type of the target statement. Analysis unit 330 is used to analyze the reasoning credibility of a preset language model based on feature vectors and a true / false classifier to obtain credibility evaluation results. The true / false classifier is pre-trained based on a set of sample sentences constructed.

[0113] In some embodiments, the projection unit 320 is configured to: determine the mean of the target sample activation vectors corresponding to the activation vectors based on at least one set of topic sentences in the sample sentence set, wherein the mean of the target sample activation vectors is the mean of the sample activation vectors of at least one sample activation vector in the set of topic sentences corresponding to the target sentence; center the activation vectors based on the mean of the target sample activation vectors to obtain the updated activation vectors corresponding to the target sentence; project the updated activation vectors based on the truth direction vector of the sentence to obtain the truth direction projection value of the updated activation vectors along the truth direction of the sentence; project the updated activation vectors based on the semantic polarity direction vector to obtain the polarity direction projection value of the updated activation vectors along the semantic polarity direction; normalize the truth direction projection value based on the number of lexical units corresponding to the target sentence to obtain the updated truth direction projection value, and integrate the updated truth direction projection value and the polarity direction projection value to obtain the feature vector.

[0114] In some embodiments, the analysis unit 330 is configured to: input the feature vector into the true / false classifier to generate a credibility probability value corresponding to a preset language model; if the credibility probability value is greater than or equal to a preset probability threshold, determine that the credibility result is that the inference result of the preset language model on the target statement is credible; if the credibility probability value is less than the preset probability threshold, determine that the credibility result is that the inference result of the preset language model on the target statement is unreliable.

[0115] In some embodiments, the apparatus includes: a construction unit, configured to: divide at least one sample statement into topics to obtain at least one set of topic statements; perform truth / false labeling on at least one topic statement based on the statement truthfulness of at least one topic statement in a first set of topic statements to obtain at least one statement truth / false label corresponding to at least one topic statement, wherein the first set of topic statements is any set of at least one set of topic statements; perform semantic polarity labeling on at least one topic statement based on the semantic polarity of at least one topic statement to obtain at least one semantic polarity label corresponding to at least one topic statement, wherein semantic polarity is used to indicate whether at least one topic statement is an affirmative statement or a negative statement; input at least one topic statement into a preset language model for inference to obtain at least one sample activation vector corresponding to at least one topic statement; and construct a set of sample statements based on at least one sample activation vector, at least one statement truth / false label, and at least one semantic polarity label.

[0116] In some embodiments, the apparatus includes: a fitting unit, configured to determine a first sample activation vector mean of at least one sample activation vector corresponding to a first set of topic statements; construct a first linear combination relationship for each activation vector in the at least one sample activation vector, including an initial statement truth direction vector and an initial semantic polarity direction vector, based on the first sample activation vector mean, at least one statement truth label and at least one semantic polarity label; and perform fitting processing on the at least one first linear combination relationship based on a preset fitting algorithm to obtain a truth direction vector and a polarity direction vector.

[0117] In some embodiments, the training unit is configured to: center at least one sample activation vector based on the mean of the first sample activation vector to obtain at least one updated sample activation vector; project at least one updated sample activation vector based on the truth direction vector and semantic polarity direction vector to obtain at least one sample feature vector corresponding to at least one updated sample activation vector; normalize at least one sample feature vector based on the number of lexical units corresponding to each topic statement in at least one topic statement to obtain at least one updated sample feature vector; and train the initial true / false classifier by using at least one updated sample feature vector as input to the initial true / false classifier and at least one statement true / false label as supervision signal to obtain the true / false classifier.

[0118] The methods and apparatus provided in the embodiments of this application have been described above. To implement the functions of the methods provided in the embodiments of this application, the electronic device may include a hardware structure and software modules, and may implement the above functions in the form of a hardware structure, software modules, or a hardware structure plus software modules. One of the above functions may be executed in the form of a hardware structure, software modules, or a hardware structure plus software modules.

[0119] Figure 4 This is a block diagram illustrating an electronic device 400 for implementing the above-described model credibility assessment method, according to an exemplary embodiment. For example, the electronic device 400 may be a mobile phone, computer, messaging device, game console, tablet device, medical device, fitness equipment, personal digital assistant, etc.

[0120] Reference Figure 4 The electronic device 400 may include a communication interface 401, capable of interacting with other devices; a processor 402, connected to the communication interface 401 to interact with other devices, used to execute the methods provided by one or more of the above-mentioned technical solutions when running a computer program; and a memory 403, on which the computer program is stored. Specifically, the specific processing procedure of the processor 402 can refer to the model reliability evaluation method described in the above embodiments of this application.

[0121] Of course, in practical applications, the various components in electronic device 400 are coupled together through bus system 404. It can be understood that bus system 404 is used to realize the connection and communication between these components. In addition to a data bus, bus system 404 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, in... Figure 4 The general designated all buses as Bus System 404.

[0122] The memory 403 in this embodiment is used to store various types of data to support the operation of the electronic device 400. Examples of such data include any computer program used to operate on the electronic device 400.

[0123] The methods disclosed in the above embodiments of this application can be applied to processor 402, or implemented by processor 402. Processor 402 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in processor 402 or by instructions in the form of software. The processor 402 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 402 can implement or execute the various methods, steps, and logic block diagrams of the application in the embodiments of this application. A general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the method applied in the embodiments of this application can be directly manifested as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in memory 403. Processor 402 reads the information in memory 403 and combines its hardware to complete the steps of the aforementioned method.

[0124] In an exemplary embodiment, the electronic device 400 may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0125] Embodiments of this application also propose a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the model credibility evaluation method described in the above embodiments of this application.

[0126] Embodiments of this application also propose a computer program product, including a computer program that is executed by a processor using the model credibility evaluation method described in the above embodiments of this application.

[0127] Embodiments of this application also propose a chip including one or more interface circuits and one or more processors; the interface circuits are used to receive signals from the memory of an electronic device and send signals to the processors, the signals including computer instructions stored in the memory, and when the processor executes the computer instructions, it causes the electronic device to perform the model credibility evaluation method described in the above embodiments of this application.

[0128] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0129] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0130] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0131] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processing module, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (control method), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic device, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0132] It should be understood that various parts of the embodiments of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0133] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0134] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc.

[0135] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for evaluating the credibility of a model, characterized in that, The method includes: The activation vector generated by the preset language model when reasoning about the target statement is obtained. The activation vector is used to represent the internal cognitive result formed by the preset language model after completing the semantic understanding of the target statement. Based on the pre-constructed truth direction vector and semantic polarity direction vector of the statement, the activation vector is projected to obtain the feature vector of the activation vector in the truth direction and semantic polarity direction of the statement. The feature vector is used to characterize the truth determination result of the target statement and the semantic sentence type of the target statement. Based on the feature vectors and the true / false classifier, the reasoning credibility of the preset language model is analyzed to obtain the credibility evaluation result. The true / false classifier is pre-trained based on the constructed sample sentence set.

2. The method according to claim 1, characterized in that, The projection processing of the activation vector based on the pre-constructed statement truth direction vector and semantic polarity direction vector to obtain the feature vectors of the activation vector in the statement truth direction and semantic polarity direction includes: Based on at least one topic statement set in the sample statement set, the mean value of the target sample activation vector corresponding to the activation vector is determined, wherein the mean value of the target sample activation vector is the mean value of the sample activation vector of at least one sample activation vector in the topic statement set corresponding to the target statement; Based on the mean of the target sample activation vector, the activation vector is centered to obtain the updated activation vector corresponding to the target statement. Based on the truth direction vector of the statement, the update activation vector is projected to obtain the truth direction projection value of the update activation vector along the truth direction of the statement. Based on the semantic polarity direction vector, the updated activation vector is projected to obtain the polarity direction projection value of the updated activation vector along the semantic polarity direction; Based on the number of lexical units corresponding to the target statement, the truth direction projection value is normalized to obtain an updated truth direction projection value. The updated truth direction projection value and the polarity direction projection value are then integrated to obtain the feature vector.

3. The method according to claim 1, characterized in that, The step of analyzing the reasoning credibility of the preset language model based on the feature vector and the true / false classifier to obtain the credibility evaluation result includes: The feature vector is input into the true / false classifier to generate a confidence probability value corresponding to the preset language model; If the credibility probability value is greater than or equal to a preset probability threshold, the credibility result is determined to be a credible inference result of the preset language model for the target statement; If the credibility probability value is less than the preset probability threshold, the credibility result is determined to be that the inference result of the preset language model on the target statement is unreliable.

4. The method according to claim 1, characterized in that, The method includes: Divide at least one sample statement into topics to obtain at least one set of topic statements; Based on the authenticity of at least one topic statement in the first topic statement set, the at least one topic statement is marked as true or false to obtain at least one statement true or false label corresponding to the at least one topic statement. The first topic statement set is any topic statement set in the at least one topic statement set. Based on the semantic polarity of the at least one topic statement, semantic polarity annotation is performed on the at least one topic statement to obtain at least one semantic polarity label corresponding to the at least one topic statement. The semantic polarity is used to indicate whether the at least one topic statement is an affirmative statement or a negative statement. The at least one topic statement is input into the preset language model for inference to obtain at least one sample activation vector corresponding to the at least one topic statement; The sample statement set is constructed based on the at least one sample activation vector, the at least one statement truth / false label, and the at least one semantic polarity label.

5. The method according to claim 4, characterized in that, The method involves projecting the activation vector based on pre-constructed statement truth direction vectors and semantic polarity direction vectors to obtain feature vectors of the activation vector in the statement truth direction and semantic polarity direction. Prior to this, the method includes: Determine the mean value of the first sample activation vector of at least one sample activation vector corresponding to the first topic statement set; Based on the mean of the first sample activation vector, the at least one statement truth value label, and the at least one semantic polarity label, a first linear combination relationship is constructed for each activation vector in the at least one sample activation vector, which includes an initial statement truth value direction vector and an initial semantic polarity direction vector; Based on a preset fitting algorithm, at least one of the first linear combination relationships is fitted to obtain the true value direction vector and the polarity direction vector.

6. The method according to claim 5, characterized in that, The method for analyzing the reasoning credibility of the preset language model based on the feature vector and the true / false classifier to obtain a credibility evaluation result includes: Based on the mean of the first sample activation vector, the at least one sample activation vector is centered to obtain at least one updated sample activation vector. Based on the truth value direction vector of the statement and the semantic polarity direction vector, the at least one updated sample activation vector is projected to obtain at least one sample feature vector corresponding to the at least one updated sample activation vector. Based on the number of lexical units corresponding to each topic statement in the at least one topic statement, the at least one sample feature vector is normalized to obtain at least one updated sample feature vector. The initial true / false classifier is trained by using the at least one updated sample feature vector as input to the initial true / false classifier and the at least one statement true / false label as supervision signal to the initial true / false classifier.

7. A model credibility assessment device, characterized in that, The device includes: The acquisition unit is used to acquire the activation vector generated by the preset language model when reasoning about the target statement. The activation vector is used to characterize the internal cognitive result formed by the preset language model after completing the semantic understanding of the target statement. The projection unit is used to project the activation vector based on the pre-constructed statement truth direction vector and semantic polarity direction vector to obtain the feature vector of the activation vector in the statement truth direction and semantic polarity direction. The feature vector is used to characterize the truth determination result of the target statement and the semantic sentence type of the target statement. The analysis unit is used to analyze the reasoning credibility of the preset language model based on the feature vector and the true / false classifier to obtain a credibility evaluation result. The true / false classifier is pre-trained based on a constructed set of sample sentences.

8. An electronic device, characterized in that, include: The processor and the memory used to store computer programs that can run on the processor. When the processor is used to run the computer program, it performs the method according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images