A social bias discovery method and device based on multi-agent evidence slot negotiation

Abstract

The application discloses a social prejudice discovery method and device based on multi-agent evidence slot negotiation, which comprises the following steps: using an analysis agent unit to sample a to-be-detected text according to a preset evidence slot from different angles to obtain a structured feature; calculating the sampling stability of each analysis agent unit according to the structured feature; determining a to-be-repaired slot according to the representation content of a core evidence slot in the structured feature of each analysis agent unit and repairing the to-be-repaired slot; calculating the evidence effectiveness score of each analysis agent unit based on the to-be-detected text, the repaired structured feature and the sampling stability; re-determining the prejudice judgment dominant label of each analysis agent unit according to the core evidence slot; confirming the representation content of each core evidence slot according to the evidence effectiveness score and the consistency of the representation content of the core evidence slot of each analysis agent unit; and determining a final classification label. The method can avoid error cascade amplification and falling into confirmation bias.

Description

Technical Field

[0001] This invention belongs to the field of social bias detection technology, and particularly relates to a method and apparatus for social bias detection based on multi-agent evidence slot negotiation. Background Technology

[0002] With the widespread application of Large Language Models (LLMs) in the field of natural language processing, using large models for social bias detection and text content security analysis has become an important research direction. However, expressions of social bias are often highly complex and covert, not only manifesting as explicit aggressive or derogatory words, but also frequently embedded in the text through implicit inferences, value presuppositions, group attributions, or complex syntactic structures containing "quotations" and "refutations".

[0003] Existing conventional social bias detection tasks mostly employ a "single-round inquiry" direct judgment paradigm. This approach easily leads to models over-reliance on surface lexical cues, such as the simultaneous appearance of group terms and negative evaluation terms, resulting in severe confirmation bias. When faced with implicit biases lacking obvious offensive vocabulary, or in complex contexts where the speaker is refuting a bias proposition, single-round direct judgments often fail to proactively initiate necessary intermediate analysis processes, easily generating a large number of false positives and false negatives. Furthermore, the judgment results exhibit black-box characteristics and lack interpretability.

[0004] To overcome the cognitive blind spots and logical flaws of the single-round direct judgment paradigm in complex contexts, and to improve the accuracy and transparency of bias detection, a technical approach based on multi-agent collaboration and structured reasoning has emerged. Introducing agents with different perspectives for task decomposition and collaboration has become a cutting-edge direction for improving the bias detection capabilities of large models. Currently, a representative leading method is the Structured Reasoning for Fairness: A Multi-Agent Approach to Bias Detection in Textual Data (SRF-MA).

[0005] SRF-MA is a representative structured multi-agent baseline method for handling social bias detection tasks. This method aims to address the lack of intermediate reasoning processes in traditional black-box models. Technically, it abandons the approach of outputting conclusions all at once, dividing the complex social bias judgment process into multiple sequentially executed sub-tasks and constructing a three-stage multi-agent pipeline: First, the CheckAgent performs an initial screening of the input text's attributes, focusing on distinguishing whether the text expresses objective facts or subjective opinions; then, for subjective texts judged by the upstream to potentially contain biased expressions, the system hands them over to the VerificationAgent for further in-depth evaluation of their specific degree of bias; finally, the ExplanationAgent, based on the analysis of the first two stages, generates the corresponding judgment criteria and a brief natural language explanation. This pipeline-based task decomposition strategy, to a certain extent, improves the systematic nature of the bias detection process and the interpretability of the judgment results.

[0006] However, the above method has the following problems:

[0007] (1) It is prone to irreversible error cascading amplification and lacks an effective error correction mechanism. Existing multi-agent baseline methods mostly adopt a sequential execution unidirectional architecture. Under this architecture, the local misjudgment of the upstream agent will be directly passed to the downstream agent as factual premise, resulting in the error being irreversibly cascaded and amplified. For example, if the upstream agent incorrectly classifies a certain implicit bias as an objective statement, the downstream bias evaluation will directly fail.

[0008] (2) It is prone to confirmation bias, which can obliterate minority insights and reasonable objections. In the existing sequential decomposition architecture, once the preceding task establishes a biased dominant tendency, the subsequent interpretation generation often blindly follows this perspective, inducing a confirmation bias unique to the model. This is especially true when facing dynamic adversarial texts that contain irony, quote others' opinions, or refute certain stereotypes. Summary of the Invention

[0009] To address the problems of cascading error amplification and confirmation bias in existing technologies, this invention provides a method and apparatus for discovering social bias based on multi-agent evidence slot negotiation.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] The first aspect of this invention provides a method for detecting social bias based on multi-agent evidence slot negotiation, comprising the following steps:

[0012] At least two analysis agent units are used to independently sample the text to be detected at least three times from different angles according to the preset evidence slots to obtain structured features that are related to bias determination and contain the corresponding representation content of the evidence slots. The preset evidence slots include core evidence slots and auxiliary evidence slots. The core slots in the core evidence slots include at least the target object range, the speaker's position, the potential bias mechanism, and whether there is an unbiased interpretation supported by the original text. The auxiliary slots in the auxiliary evidence slots include at least the target group, literal evidence fragments, brief explanations, and preliminary classification labels. The preliminary classification labels represent whether there is bias or no bias.

[0013] The bias judgment dominant label is determined based on the preliminary classification label of each analysis agent unit, and the sampling stability of each analysis agent unit is calculated based on the structured features.

[0014] Based on the representation content of the core evidence slots in the structured features of each analysis agent unit, the slots to be repaired are determined and repaired to obtain the repaired structured features of each analysis agent unit.

[0015] The evidentiary value score of each analysis proxy unit is calculated based on the text to be detected, the repaired structured features, and the sampling stability.

[0016] The dominant label for bias judgment of each analytical agent unit is re-determined based on the core evidence slot;

[0017] Based on the evidence validity score and the consistency of the representation content of the core evidence slots of each analysis agency unit, the representation content of each core evidence slot is confirmed.

[0018] The final classification label is determined based on the content of the dominant label, the evidentiary value score, and the preset rules, which are redefined by each analysis agent unit to determine the bias judgment.

[0019] A second aspect of the present invention provides a social bias detection device based on multi-agent evidence slot negotiation, comprising a memory and a controller connected in sequence, wherein the memory stores a computer program, and the controller is used to read the computer program and execute the social bias detection method based on multi-agent evidence slot negotiation described in the first aspect.

[0020] The beneficial effects of this invention are:

[0021] 1. This invention improves robustness in complex texts by enabling judgments from different perspectives to cross-validate each other through parallel independent reasoning and global coordinated repair of multiple analysis agent units, thus avoiding irreversible error cascading.

[0022] 2. This invention objectively audits the reasoning process through multiple indicators, making the final result not only more accurate and stable, but also more interpretable and traceable.

[0023] 3. By introducing a counter-evidence proxy and an evidence slot for whether there is an unbiased interpretation supported by the original text, the present invention enables the system to proactively examine whether there is an unbiased interpretation of the text, such as quotations, rebuttals, irony, or reasonable descriptions, thereby reducing the risk of false alarms and avoiding falling into confirmation errors. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of a social bias detection method based on multi-agent evidence slot negotiation proposed in this invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0029] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0030] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0031] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0032] The first aspect of this invention discloses a method for detecting social bias based on multi-agent evidence slot negotiation. This method is executed by a device, which can be a cloud server, a local server, a laptop, a mobile phone, or other smart devices. Specifically, as... Figure 1 As shown, this method includes steps S1 to S6.

[0033] Step S1: Use at least two analysis agent units to independently sample the text to be detected at least three times from different angles according to the preset evidence slots to obtain structured features that are related to the bias judgment and contain the representation content corresponding to the evidence slots. The preset evidence slots include core evidence slots and auxiliary evidence slots.

[0034] Specifically, the analytical proxy unit includes at least two of the following: textual proxy, discourse practice proxy, social practice proxy, and counter-evidence proxy. Preferably, all four analytical proxy units are used simultaneously.

[0035] Among them, the text proxy is used to analyze the bias of the test text from the perspective of literal expression; the discourse practice proxy is used to analyze the bias of the test text from the perspective of the speaker's expression mode; the social practice proxy is used to analyze the bias of the test text from the perspective of underlying social significance; and the counter-evidence proxy is used to analyze the bias of the test text from the perspective of finding that the input text does not constitute bias, so as to find the strongest reasonable and compliant explanation that the input text does not constitute bias.

[0036] The preset evidence slots include core evidence slots and auxiliary evidence slots.

[0037] Among them, the core evidence slots include at least four core slots: target_scope, speaker's stance, potential bias mechanism, and whether there is an unbiased interpretation supported by the original text.

[0038] The target scope and speaker stance are the upstream slots, while the potential bias mechanism and whether there is an unbiased interpretation supported by the original text are the downstream slots.

[0039] The target audience includes groups and individuals. 1. None (no explicit object) and 2. Unclear (object scope uncertain).

[0040] The speaker's position is that of a quoting ,refute ,support Unclear, etc.

[0041] Potential bias mechanisms include stereotypes, denigration, exclusion, and ambiguity.

[0042] That is, the output of each agency unit regarding the core evidence slot. for:

[0043]

[0044] The auxiliary evidence slots include at least four slots: target_group, evidence_spans, short_rationale, and provisional_label. (Preliminary classification label) The labels are binary, including biased or unbiased. The `target_group` parameter identifies the specific social group mentioned in the text, i.e., who the group is, providing the model with explicit socio-semantic objects in the expression of bias.

[0045] 1 indicates bias, and 0 indicates no bias; It is a natural number greater than or equal to 1 and less than or equal to K, where K is the total number of independent samples; A natural number greater than or equal to 1 and less than or equal to I, where I is the number of analysis agent units.

[0046] Example:

[0047] The text to be detected is:

[0048] I am not saying all immigrants are criminals, but they always bring trouble and should be kept out of our neighborhoods..

[0049] The text was manually annotated and found to contain social bias. The text mentions the social group "immigrants," includes the generalized expression "they always bring trouble," and the exclusionary expression "should be kept out of our neighborhoods." Although the sentence begins with a superficial denial, "I am not saying all immigrants are criminals," the subsequent text uses a contrastive structure to introduce negative judgments and exclusionary claims against immigrants. Therefore, this sample can be used to verify the method's ability to identify social biases within complex pragmatic structures.

[0050] The above four proxy units are used to independently sample the text to be detected at least three times according to the above preset evidence slots. The random seed is fixed at 42 and the stability fusion parameter is 0.5. In the evidence validity score, the overall stability weight is 0.3, the evidence effectiveness weight is 0.3, the cross-proxy consistency weight is 0.4, and the ambiguity threshold is 0.10.

[0051] The structured features of the text proxy output are as follows:

[0052]

[0053] The format of the structured features output by other agent units is similar to that described above.

[0054] The text agent primarily analyzes sentences based on their literal meaning. In this test case, the text agent identified "immigrants" as the evaluated social group in all three samples and considered "they always bring trouble" to be an inappropriate generalization targeting this group. Specifically, the structured outputs of the first and third samples identified the potential bias mechanism as inappropriate generalization or stereotype, while the structured output of the second sample identified it as exclusion. However, all three samples initially identified the classification label as biased. The system selected the first sample as the representative output, with the target object being the social group, the speaker's stance being approval, the potential bias mechanism being inappropriate generalization or stereotype, the existence of an unbiased interpretation supported by the original text being false, and the literal evidence fragment being "they always bring trouble." The brief interpretation is that this fragment makes a generalized negative judgment about the immigrant group.

[0055] The discourse practice agent analyzes sentences from an expressive perspective. In this test case, the agent noticed that although there was a denial at the beginning of the sentence, the phrases "they always bring trouble" and "should be kept out of our neighborhoods" following the transition word "but" constituted the speaker's final judgment. Therefore, the agent output the target audience as a social group, the speaker's stance as approval, the potential bias mechanism as exclusion, the existence of an unbiased interpretation supported by the original text as false, the preliminary classification label as biased, and the literal evidence fragments as "but they always bring trouble" and "should be kept out of our neighborhoods." The brief interpretation is that although there is denial at the beginning of the sentence, the content after "but" reflects the speaker's final exclusionary claim. The agent's three samplings were completely consistent.

[0056] The social practice agent primarily analyzes sentences from a social meaning perspective. In this test case, the social practice agent argues that the text not only generalizes immigrant groups as troublemakers but also proposes their exclusion from the community. Therefore, the structured output from all three samples is: target audience is the social group, speaker's stance is approval, potential bias mechanism is exclusion, existence of unbiased interpretation supported by the original text is false, preliminary classification label is biased, literal evidence fragments are "they always bring trouble" and "should be kept out of our neighborhoods," and the brief interpretation is that the text assigns negative social meaning to immigrant groups and makes exclusionary claims.

[0057] The counter-evidence agent is used to proactively seek the strongest unbiased interpretation. In this test case, the counter-evidence agent noticed that the phrase "I am not saying all immigrants are criminals" at the beginning of the sentence might provide an unbiased interpretation, resulting in discrepancies across three samplings. In the first sampling, the structured output showed an unclear speaker stance, an unclear potential bias mechanism, and an unclear existence of an unbiased interpretation supported by the original text, leading to an initial classification label of unbiased. In the second sampling, the speaker stance label remained unbiased. In the third sampling, however, the system identified a subsequent exclusionary expression and classified the initial classification label as biased. Based on the majority label, the system selected the structured output from the first sampling as the representative output. Its target audience was a social group, with an unclear speaker stance, an unclear potential bias mechanism, and an unclear existence of an unbiased interpretation supported by the original text. The literal evidence fragment was "I am not saying all immigrants are criminals," and the brief interpretation was that the initial denial provided a weak unbiased interpretation.

[0058] Step S2: Determine the dominant label for bias judgment based on the preliminary classification labels of each analysis agent unit, and calculate the sampling stability of each analysis agent unit based on the structured features. Specifically, this step includes steps S21 to S23.

[0059] Step S21: Based on the voting mechanism, obtain the bias judgment dominant label of each analysis agent unit according to the preliminary classification label obtained by each independent sampling.

[0060] No. The first analysis agent unit Preliminary classification labels generated from independent sampling of the text to be detected. The first one was determined through a voting mechanism. Bias judgment dominant label of each analysis agent unit ,in,

[0061] ,

[0062] In the formula, This is the mode function, used to find the value that appears most frequently in the input set. Indicates possible binary classification bias labels;

[0063] Step S22: Calculate the label stability based on the consistency between the preliminary classification label obtained by each analysis agent unit through independent sampling and the dominant label of the bias determination.

[0064] Specifically, calculate the first Tag stability of individual analysis agent units for:

[0065] ,

[0066] In the formula, For indicator functions, when The value is 1 if the condition is true, and 0 otherwise.

[0067] Step S23: Select a set of labels that are consistent with the dominant label for bias determination and the corresponding preliminary classification label, and select the corresponding structured features from the set of labels as the benchmark sample for the analysis agent unit.

[0068] benchmark sample The determination is based on the consistency between the dominant bias judgment label and the corresponding preliminary classification label. Specifically, for each analysis agent unit, a set of labels whose dominant bias judgment label and corresponding preliminary classification label are consistent is first selected. In random seed From the set of labels under constraints A label is extracted with a moderate probability. This label serves as the index of the representative sample, and its corresponding structured features serve as the benchmark sample. ,in, This represents a uniform random sampling function.

[0069] Step S24: Calculate the core slot consistency of the corresponding analysis agent unit in the core evidence slot for the structured features corresponding to the benchmark sample and the label set.

[0070] benchmark sample In the core slot The representation content is The slot consistency of each analysis agent unit in the core slot is calculated as follows:

[0071] First, based on the consistency between the core slot representation content of each analysis agent unit in the tag set and the core slot representation content of the corresponding benchmark sample, the single-slot conditional consistency is calculated. ,in,

[0072] .

[0073] Then, the core slot consistency is calculated based on the single slot condition consistency. ,

[0074] .

[0075] In the above formula, Representative tag set The total number of valid samples; The total number of slots representing core evidence slots; For the first The content representing the slot in the second sampling.

[0076] Step S25: Perform a weighted fusion calculation on the tag stability and core slot consistency to obtain the sampling stability of each analysis agent unit.

[0077] Specifically, a weighted fusion of tag stability and slot consistency is performed to obtain the sampling stability of each analysis agent unit. ,

[0078] ,

[0079] in, The balance coefficient is preferably set to 0.5.

[0080] This step yields the sampling stability of multiple analysis agent units. For the example in step 1, the method used in this step is employed to calculate:

[0081] Tag stability of text proxies The value is 1.0, the core slot consistency is 0.9167, and the sampling stability is 0.9583.

[0082] The three samplings of the discourse practice agent were completely consistent, and the sampling stability was 1.0.

[0083] The three samplings conducted by the social practice agent were completely consistent, and the sampling stability was 1.0.

[0084] Tag stability of counter-evidence proxy The value is 0.6667, the core slot consistency is 0.75, and the sampling stability is 0.7083.

[0085] Step S3: Determine the slots to be repaired based on the representation content of the core evidence slots in the structured features of each analysis agent unit, and repair the slots to be repaired to obtain the repaired structured features of each analysis agent unit.

[0086] Specifically, this step is implemented based on a coordinator, which repairs the slot to be repaired based on steps S31 to S32.

[0087] Step S31: Determine the core evidence slots to be repaired based on the consistency of the representation content of the corresponding core evidence slots in the structured features of each analysis agent unit.

[0088] In this step, if the representation content of a core evidence slot of an analysis agent unit is different from the representation content of the corresponding core evidence slot of other analysis agent units, then the core evidence slot is regarded as a slot to be repaired.

[0089] Step S32: Repair the core evidence slots to be repaired based on the repair package consisting of candidate values, literal evidence fragments, and brief explanations, which are composed of the representation content of the core evidence slots to be repaired in each analysis agent unit.

[0090] If all the core evidence slots to be repaired are upstream slots or all of them are upstream slots, a single-round repair mechanism will be adopted.

[0091] If the core evidence slot to be repaired includes both upstream slots and upstream sub-slots, preferably, this step employs a two-round repair mechanism, that is, repairing the core evidence slot to be repaired is based on the principle of prioritizing the repair of upstream slots. In this case, the specific step is as follows:

[0092] During the first round of repair, the upstream slots are re-examined based on the literal evidence fragments and brief explanations of each analysis agent unit, and the representation content of the core evidence slots to be repaired in the upstream slots is determined from the candidate values.

[0093] During the second round of repair, the downstream slots are re-examined based on the literal evidence fragments and brief explanations of each analysis agent unit, and the representation content of the core evidence slots to be repaired in the downstream slots is determined from the candidate values.

[0094] This step repairs the representation content of the core evidence slots based on the representation content, literal evidence fragments, and brief explanations of each analysis agent unit. For the example in step 2, the coordinator first compares the differences among the analysis agent units in the core evidence slots. It finds that all four analysis agent units believe the text ultimately points to the social group "immigrants," therefore the target scope does not need repair. However, there are conflicts in the three core slots: speaker's stance, potential bias mechanisms, and the existence of an unbiased interpretation supported by the original text, requiring partial controlled re-examination.

[0095] In the first round of repair, the system prioritizes repairing upstream slots, i.e., the speaker's stance. The coordinator constructs a repair package containing candidate values, evidence fragments, and brief explanations. In this round, the candidate values ​​are "acceptance" and "ambiguity." Candidates for an "acceptance" speaker stance come from textual agents, discourse practice agents, and social practice agents. The literal evidence fragments include "they always bring trouble," "but they always bring trouble," and "should be kept out of our neighborhoods," with the corresponding brief explanation being that the content following the transition word "but" reflects the speaker's final acceptance of a negative judgment and exclusionary claim. Candidates for an "ambiguous" speaker stance come from counter-evidence agents, with the literal evidence fragment "I am not saying all immigrants are criminals," and the corresponding brief explanation being that the initial denial provides weak counter-evidence.

[0096] Upon receiving the aforementioned repair package, each analysis agent unit only re-examines the speaker's stance; it cannot modify other slots or generate new evidence. After comparison, the system determines that initial denial is insufficient to offset the generalization and exclusionary expressions in the subsequent transitional structure. Therefore, the speaker's stance in all four analysis agent units is repaired to acceptance. The output of the first round of repair is: the target audience for all four agents remains social groups, and the speaker's stance is acceptance in all cases.

[0097] In the second round of repair, the system addressed downstream slots, namely bias mechanisms and unbiased interpretation support. For bias mechanisms, the candidate value provided by the text agent was inappropriate generalization or stereotype, with the literal evidence fragment "they always bring trouble," which could be interpreted as "always" representing a generalized judgment about immigrant groups. The candidate value provided by the discourse practice agent and the social practice agent was exclusion, with the literal evidence fragment "should be kept out of our neighborhoods," which could be interpreted as the expression excluding immigrant groups from the community. The candidate value provided by the counter-evidence agent was ambiguity, with the literal evidence fragment "I am not saying all immigrants are criminals," which could be interpreted as initial denial weakening the certainty of a direct bias judgment.

[0098] Regarding the support for unbiased interpretation, the text agent, discourse practice agent, and social practice agent all support false, meaning that the current text lacks a sufficiently valid strong unbiased interpretation path. When the system believes that the text lacks a sufficiently valid unbiased interpretation, this slot is set to false. The literal evidence fragments are "they always bring trouble," "but they always bring trouble," and "should be kept out of our neighborhoods." The brief interpretation is that the subsequent transitional content has already given a negative generalization and exclusionary claim, so the unbiased interpretation cannot be fully supported. The support from the counter-evidence agent is unclear, and its evidence is still the initial denial fragment. The interpretation provides some alternative interpretation space for this fragment.

[0099] After the second round of partial review, the system obtained the final structured results for the four analysis agent units, namely:

[0100] The structured features corresponding to the text agent include a target group as the target object, an endorsement from the speaker, inappropriate generalization or stereotypes as the potential bias mechanism, the existence of an unbiased interpretation supported by the original text as false, and the literal evidence fragment as "they always bring trouble," which can be interpreted as a generalized negative judgment about immigrant groups.

[0101] In the structured features corresponding to the discourse practice agency, the target object is a social group, the speaker's position is approval, the potential bias mechanism is exclusion, the existence of an unbiased interpretation supported by the original text is false, and the literal evidence fragments are "but they always bring trouble" and "should be kept out of our neighborhoods". The brief interpretation is that the content after "but" reflects the exclusionary claim that the current speaker ultimately approves of.

[0102] In the structured features corresponding to social practice agency, the target object is a social group, the speaker's position is approval, the potential bias mechanism is exclusion, the existence of an unbiased interpretation supported by the original text is false, and the literal evidence fragments are "they always bring trouble" and "should be kept out of our neighborhoods". In short, the text assigns negative social meaning to the immigrant group and proposes exclusion.

[0103] After repair, the structured features of the counter-evidence proxy include a target group as a social group, a speaker's stance of approval, a potential bias mechanism of exclusion, the existence of an unbiased interpretation supported by the original text being false, and the literal evidence fragment being "should be kept out of our neighborhoods". The brief interpretation is that, in the complete context, the initial denial is insufficient to support an unbiased interpretation.

[0104] Instead of simply removing the objections from the counter-evidence attorney, this stage incorporates their evidence into a remediation package along with other attorneys' candidate arguments, allowing each attorney to re-compare evidence and interpretations within a controlled framework. Under conditions of complete evidence, the counter-evidence attorney's original unbiased interpretation based on initial denial was deemed insufficient, thus completing the partial remediation.

[0105] Step S4: Calculate the evidentiary strength score of each analysis proxy unit based on the text to be detected, the repaired structured features, and the sampling stability.

[0106] This step is performed by the auditor, specifically including steps S41 to S43.

[0107] Step S41: Based on the matching degree between the text to be detected and the literal evidence fragments in the repaired structured features, the evidence validity rate is obtained. .

[0108] This step calculates the validity of evidence based on strict string matching verification. ,in,

[0109] ,

[0110] In the formula, To analyze the exact number of substrings in proxy unit i that are completely identical to the text to be detected, To extract the total number of evidence fragments, the denominator is used to avoid the risk of division by zero.

[0111] Step S42: Calculate the consistency between the core slots of the structured features after repair of each analysis agent unit and the corresponding core slots of other analysis agent units to obtain the cross-agent consistency.

[0112] Cross-agent consistency for:

[0113] ,

[0114] In the formula, M represents the total number of analysis agent units; and These represent the analysis agent units respectively. With analysis agent unit In the core slot of the repaired structured features The content represented by it.

[0115] Step S43: Based on the sampling stability The evidentiary validity score of each analytical agency unit is obtained by calculating the evidentiary validity rate and cross-agency consistency.

[0116] In the analysis agent unit Evidence validity for:

[0117] ,

[0118] in , and These represent the weighting coefficients of sampling stability, evidence validity, and cross-agent consistency in the total score, respectively, and are preferably set as follows: , , .

[0119] For the example in step 3, this step first checks whether the literal evidence fragments output by each analysis agent unit are exact substrings in the text to be detected. In this test case, the literal evidence fragments "they always bring trouble" from the text agent, "but they always bring trouble" and "should be kept out of our neighborhoods" from the discourse practice agent, "they always bring trouble" and "should be kept out of our neighborhoods" from the social practice agent, and "should be kept out of our neighborhoods" from the counter-evidence agent all match precisely in the original text. Therefore, the evidence effectiveness of all four agents is... It is 1.0.

[0120] The four analytical agent units are completely consistent in terms of target scope, speaker stance, and the existence of unbiased interpretations supported by the original text, differing only in their potential bias mechanisms. The text agent's potential bias mechanism is inappropriate generalization or stereotype, while the other three agents' potential bias mechanism is exclusion. Therefore, after processing using this method, the text agent has 9 slot matches and 12 slot comparison items when compared with other agents, with a cross-agent consistency of 0.75; the discourse practice agent, social practice agent, and refutation agent each have 11 slot matches and 12 slot comparison items when compared with other agents, with a cross-agent consistency of 0.9167 for each.

[0121] Next, the evidentiary strength score for each analysis agency unit is calculated based on sampling stability, evidentiary effectiveness, and cross-agency consistency. In this embodiment, , , .

[0122] The sampling stability of the text proxy is 0.9583, the evidence validity rate is 1.0, and the cross-proxy consistency is 0.75. Therefore, its evidentiary validity is 0.8875.

[0123] The sampling stability of the discourse practice agency is 1.0, the evidence validity rate is 1.0, and the cross-agency consistency is 0.9167. Therefore, its evidentiary validity score is 0.9667.

[0124] Social practice agency also received 0.9667.

[0125] The sampling stability of the counter-evidence representation is 0.7083, the evidence validity rate is 1.0, and the cross-representation consistency is 0.9167. Therefore, its evidentiary value is 0.8792.

[0126] Step S5: Based on the core evidence slots, redetermine the dominant labels for each analysis agent unit's bias judgment.

[0127] This step involves calling a rule validation function based on the repaired structured features. Dominant label for bias judgment Perform corrections.

[0128] Specifically, rule validation function The rules are as follows:

[0129] When the target audience is an individual or there is no specific target, the dominant label of bias judgment should be forcibly corrected to unbiased.

[0130] When the speaker's position is either quoting or refuting and there is an unbiased interpretation supported by the original text, the dominant bias label is forcibly corrected to unbiased.

[0131] When the target audience is a group, the speaker's stance is one of acceptance bias, the potential bias mechanism belongs to the pre-defined set of bias mechanisms, and there is no unbiased interpretation supported by the original text, the dominant label for bias judgment is forcibly corrected to bias. That is:

[0132]

[0133] Representing the individual, The representation has no clear object. Characteristic citation; Rebuttal by representation; Characterization? The characteristic is, Characterizes recognition bias, For a set of pre-set bias mechanisms; To represent other situations.

[0134] When the target object scope is unclear, the bias judgment dominant label will not be directly corrected, but will continue to participate in subsequent rules or score judgments.

[0135] For the example in step 4, the four corrected analysis agent units all satisfy the following conditions: the target object is a social group, the speaker's stance is approval, the potential bias mechanism belongs to the bias mechanism set, and the existence of an unbiased interpretation supported by the original text is not true. Therefore, the local labels of all four agents are recalculated as having bias.

[0136] Step S6: Based on the evidence validity score and the consistency of the representation content of the core evidence slots in each analysis agency unit, confirm the representation content of each core evidence slot. Specifically, this step includes steps 61 to 64.

[0137] Step 61: Calculate the results for each analytical agency unit based on the evidence strength score and the consistency of the representation content of the core evidence slots of each analytical agency unit. Each core evidence slot Representation content support .

[0138] in, .

[0139] Step 62: Based on all the representation content of each core evidence slot, sum up the support of each representation content to obtain the support weight of each representation content of that core evidence slot.

[0140] Step 63: For each representation of each core evidence slot, calculate the relative dominance of the difference between the highest and second-highest support weights. .

[0141] ,

[0142] In the formula, To prevent extremely small positive numbers from being divided by zero.

[0143] Step 64: In response to the relative dominance being greater than or equal to a preset threshold, the representation content corresponding to the highest support weight is used as the representation content of the core evidence slot for all analysis agent units. .

[0144] When the relative dominance is less than a preset threshold, Set to unclear.

[0145] In this step, the preset threshold can be set to any value between 0.05 and 0.15.

[0146] Based on the output of the example in step 5, this step performs global slot aggregation and weights the slot values ​​according to the comprehensive evidentiary strength score of each analysis agent unit.

[0147] For the core slots of the target object scope, all four analysis agent units support social groups. Therefore, the global target object scope is social groups, with a total weight of 3.7001.

[0148] For the core slot of speaker stance, all four analysis agent units support approval, therefore the global speaker stance is approval, with a total weight of 3.7001.

[0149] In the potential bias mechanism slots, if the text proxy supports inappropriate generalization or stereotypes, the weight of representing inappropriate generalization or stereotypes is 0.8875; if the discourse practice proxy, social practice proxy, and refutation proxy support exclusion, each representing content has a weight of 0.9125, then the total weight of representing content as exclusion is 2.8126. Since the total weight of exclusion 2.8126 is much greater than the total weight of inappropriate generalization or stereotypes 0.8875, the global potential bias mechanism aggregates into exclusion.

[0150] Does the original text support an unbiased interpretation slot? All four proxies support false, therefore the global unbiased interpretation support is false, and the total weight is 3.7001.

[0151] In response, the output structured features indicate that the target audience is a social group, the speaker's stance is approval, the potential bias mechanism is exclusion, and the existence of an unbiased interpretation supported by the original text is false.

[0152] Step 7: Determine the dominant label based on the biases redefined for each analytical agent unit. The content of the representation and the evidentiary effect The final category label is determined based on preset rules. Specifically, this step includes steps 71 and 72.

[0153] Step 71: Determine the dominant label based on the biases redefined for each analytical agent unit. The content of the representation and the evidentiary effect Calculate the weighted support scores for biased and unbiased individuals.

[0154] Among them, the weighted support score for bias for:

[0155] ,

[0156] Unbiased weighted support score for:

[0157] .

[0158] Based on the example in step 6, due to the dominant label of the four analysis agent units... All the content represented was biased, so the weighted support score for bias was the sum of the combined evidence strength scores of the four agents, which is 3.7001; the weighted support score for non-biasedness was 0.

[0159] Step 712: Determine the final classification label based on preset rules and weighted support scores.

[0160] The default rules here are:

[0161] When the target audience is an individual or there is no specific target, the classification label represents the content as unbiased.

[0162] When the speaker's position is either a quotation or a rebuttal and there is an unbiased interpretation supported by the original text, the category label represents the content as unbiased.

[0163] When the target object is a group, the speaker's position is supportive, the potential bias mechanism belongs to the bias in the set of pre-set bias mechanisms, and there is no unbiased interpretation supported by the original text, the content represented by the classification label is biased.

[0164] When the core slot representation content does not fall into any of the above three categories, the representation content of the category label is determined based on the highest weighted support score.

[0165] Right now:

[0166]

[0167] In the above formula, , , , The system uses four core slot representations obtained after step S6: target_scope, speaker stance, potential bias mechanism, and whether a benign interpretation supported by the original text exists. Only when the key slot information is ambiguous and cannot trigger a strong rule, the system makes a final decision based on the globally weighted score.

[0168] The system executes strong logical constraints. Since the global slots satisfy the conditions of "target object scope is a social group, speaker's position is approval, bias mechanism is exclusion, and unbiased interpretation support is not true", the system triggers the bias judgment rule and outputs the final label as "social bias exists".

[0169] In summary, the final output of this embodiment is as follows: the text to be detected is determined to contain social bias, with a reliability level of high; the global target object is the social group, the global speaker's stance is approval, the global potential bias mechanism is exclusion, and the existence of an unbiased interpretation supported by the original text is false; the literal evidence fragments are "they always bring trouble" and "should be kept out of our neighborhoods"; the brief interpretation is that the current speaker ultimately approves of the negative generalization and exclusionary claims against immigrant groups, thus constituting an expression of social bias.

[0170] The social bias detection method based on multi-agent evidence slot negotiation in this embodiment can be widely applied to texts in various fields, such as news reports, blogs, social media conversations, literary works, large model interactive logs, and corporate promotional copy, to accurately detect expressions of social bias such as gender discrimination, racial bias, occupational stereotypes, regional exclusion, age bias, and appearance derogatory remarks. Social bias detection technology has been applied in various industrial sectors. For example, in online content moderation, it detects potentially antagonistic biased statements in massive amounts of user posts, comments, and forum dialogues, accurately distinguishing between malicious stereotype propagation and legitimate viewpoint quotations / criticisms. This allows platform moderators to promptly intercept harmful information and significantly reduce false positives. In artificial intelligence and large-scale model security governance, platforms can use the bias detection technology to analyze and act as a "safety barrier" for the content generated by intelligent customer service and large-scale model chatbots, enabling real-time monitoring and blocking of potentially veiled toxic output from large models. In corporate human resources and internal compliance, companies can use social bias detection technology to monitor the compliance of job descriptions (JDs), internal communications, and interview evaluations, promptly identifying potential exclusionary and discriminatory risks. For instance, it can be used in intelligent recruitment systems of large enterprises to promptly detect implicit gender or age discrimination in recruitment materials, preventing serious public relations crises and legal disputes.

[0171] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting social bias based on multi-agent evidence slot negotiation, characterized in that, Includes the following steps: At least two analysis agent units are used to independently sample the text to be detected at least three times from different angles according to the preset evidence slots to obtain structured features that are related to bias determination and contain the corresponding representation content of the evidence slots. The preset evidence slots include core evidence slots and auxiliary evidence slots. The core slots in the core evidence slots include at least the target object range, the speaker's position, the potential bias mechanism, and whether there is an unbiased interpretation supported by the original text. The auxiliary slots in the auxiliary evidence slots include at least the target group, literal evidence fragments, brief explanations, and preliminary classification labels. The preliminary classification labels represent whether there is bias or no bias. The bias judgment dominant label is determined based on the preliminary classification label of each analysis agent unit, and the sampling stability of each analysis agent unit is calculated based on the structured features. Based on the representation content of the core evidence slots in the structured features of each analysis agent unit, the slots to be repaired are determined and repaired to obtain the repaired structured features of each analysis agent unit. The evidentiary value score of each analysis proxy unit is calculated based on the text to be detected, the repaired structured features, and the sampling stability. The dominant label for bias judgment of each analytical agent unit is re-determined based on the core evidence slot; Based on the evidence validity score and the consistency of the representation content of the core evidence slots of each analysis agency unit, the representation content of each core evidence slot is confirmed. The final classification label is determined based on the content of the dominant label, the evidentiary value score, and the preset rules, which are redefined by each analysis agent unit to determine the bias judgment.

2. The method for detecting social bias based on multi-agent evidence slot negotiation according to claim 1, characterized in that, The analytical proxy unit includes at least two of the following: text proxy, discourse practice proxy, social practice proxy, and counter-evidence proxy. The text proxy is used to perform bias analysis on the text to be detected from a literal perspective. The discourse practice agent is used to analyze biases in the detected text from the perspective of the speaker's expression. The aforementioned social practice proxy is used to analyze biases in the text being tested from the perspective of underlying social significance. The counter-evidence proxy is used to find angles from which the input text does not constitute bias in the bias analysis of the text to be tested.

3. The method for detecting social bias based on multi-agent evidence slot negotiation according to claim 1, characterized in that, The step of calculating the sampling stability of each analysis agent unit based on the structured features includes: Based on the voting mechanism, the bias judgment dominant label of each analysis agent unit is obtained from the preliminary classification labels obtained by each independent sampling. The label stability is calculated based on the consistency between the preliminary classification label obtained by each analysis agent unit in each independent sampling and the dominant label of the bias judgment; A set of labels that are consistent with the dominant label for bias judgment and the corresponding preliminary classification label is selected, and a label is selected from the set of labels and its corresponding structured features are used as the benchmark sample for the analysis agent unit. The consistency of the core slot in the core evidence slot is calculated for the structured features corresponding to the benchmark sample and the label set. The sampling stability of each analysis agent unit is obtained by weighted fusion calculation of the tag stability and core slot consistency.

4. The method for detecting social bias based on multi-agent evidence slot negotiation according to claim 3, characterized in that, The calculation of the core slot consistency of the corresponding analysis proxy unit in the core evidence slot for the structured features corresponding to the benchmark sample and the label set includes: Based on the consistency between the core slots of each analysis agent unit in the tag set and the core slot representation content of the corresponding benchmark sample, the single slot conditional consistency is calculated. Core slot consistency is calculated based on single slot condition consistency.

5. The method for detecting social bias based on multi-agent evidence slot negotiation according to claim 1, characterized in that, The process involves determining the slots to be repaired based on the representation content of the core evidence slots in the structured features of each analysis agent unit, and repairing the slots to be repaired to obtain the repaired structured features of each analysis agent unit, including: The core evidence slots to be repaired are determined based on the consistency of the representation content of the corresponding core evidence slots in the structured features of each analysis agent unit. The repair package, consisting of candidate values, literal evidence fragments, and brief explanations, based on the representation content of the core evidence slots to be repaired from each analysis agent unit, repairs the core evidence slots to be repaired. When a core evidence slot to be repaired contains both upstream and downstream slots, a two-round repair mechanism is used to repair the core evidence slots to be repaired. During the first round of repair, the upstream slots are re-examined based on the literal evidence fragments and brief explanations of each analysis agent unit. The representation content of the core evidence slots to be repaired in the upstream slots is determined from the candidate values. The upstream slots include the target object range and the speaker's position. During the second round of repair, the downstream slots are re-examined based on the literal evidence fragments and brief interpretations of each analysis agent unit. The representation content of the core evidence slots to be repaired in the downstream slots is determined from the candidate values. The downstream slots include potential bias mechanisms and whether there is an unbiased interpretation supported by the original text.

6. The method for detecting social bias based on multi-agent evidence slot negotiation according to claim 1, characterized in that: The evidentiary strength score of each analysis proxy unit is calculated based on the text to be detected, the repaired structured features, and sampling stability, including: The evidence validity rate is obtained based on the matching degree between the text to be detected and the literal evidence fragments in the repaired structured features. The consistency between the core slots of the structured features repaired by each analysis agent unit and the corresponding core slots of other analysis agent units is calculated to obtain the cross-agent consistency. The evidentiary validity score of each analysis agent unit is calculated based on the sampling stability, evidentiary validity rate, and cross-agent consistency.

7. The method for discovering social bias based on multi-agent evidence slot negotiation according to claim 1, characterized in that: The process of re-determining the dominant label for bias judgment of each analysis agent unit based on the core evidence slot includes: Based on the repaired structured features, the bias judgment dominant label is corrected by calling a rule validation function. The rules in the rule validation function are as follows: When the target audience is an individual or there is no specific target, the dominant label of bias judgment should be forcibly corrected to unbiased. When the speaker's position is either quoting or refuting and there is an unbiased interpretation supported by the original text, the dominant bias label is forcibly corrected to unbiased. When the target is a group, the speaker's stance is one of acceptance bias, the potential bias mechanism belongs to the set of pre-set bias mechanisms, and there is no unbiased interpretation supported by the original text, the dominant label of bias judgment will be forcibly corrected to bias.

8. The method for discovering social bias based on multi-agent evidence slot negotiation according to claim 1, characterized in that: Based on the consistency of the evidentiary value score and the representational content of the core evidentiary slots of each analysis agency unit, the representational content of each core evidentiary slot is determined, including... The results for each analytical agency unit are calculated based on the evidentiary value score and the consistency of the representation content of the core evidentiary slots of each analytical agency unit. Each core evidence slot represents the degree of support for the content; Based on all the representations in each core evidence slot, the support of each representation is summed to obtain the support weight of each representation in that core evidence slot. For each representation of each core evidence slot, calculate the relative dominance of the difference between the highest and second-highest support weights. In response to the relative advantage being greater than or equal to a preset threshold, the representation content corresponding to the highest support weight is used as the representation content of the core evidence slot for all analysis agent units.

9. The method for discovering social bias based on multi-agent evidence slot negotiation according to claim 1, characterized in that: The dominant label is determined based on the bias identified by each analysis agent unit. The representational content, evidentiary value score, and pre-set rules determine the final classification labels, including: Based on the bias determination of each analytical agent unit, the dominant label is determined. The weighted support scores for biased and unbiased elements are calculated based on the content of the representation and the evidentiary force. The final category label is determined based on preset rules and weighted support scores. The preset rules are as follows: When the target audience is an individual or there is no specific target, the classification label represents the content as unbiased. When the speaker's position is either a quotation or a rebuttal and there is an unbiased interpretation supported by the original text, the category label represents the content as unbiased. When the target object is a group, the speaker's position is supportive, the potential bias mechanism belongs to the bias in the set of pre-set bias mechanisms, and there is no unbiased interpretation supported by the original text, the content represented by the classification label is biased. When the core slot representation content does not fall into any of the above three categories, the representation content of the category label is determined based on the highest weighted support score.

10. A social bias detection device based on multi-agent evidence slot negotiation, comprising a memory and a controller connected in sequence, wherein the memory stores a computer program, characterized in that: The controller is used to read the computer program and execute the social bias discovery method based on multi-agent evidence slot negotiation as described in any one of claims 1-9.

Images