An intelligent negation construction method and system combined with multi-dimensional dynamic business rules

Abstract

This invention provides an intelligent negative word construction method and system that combines multi-dimensional dynamic business rules, belonging to the field of data processing technology. Specifically, it includes: a generative language model, through an injected thought chain mechanism and attribute decoupling attention mask, outputs search term attribute vectors and product attribute vectors through inference links and customized masks; constructs a bipartite graph with search term attribute vectors and product attribute vectors as nodes; calculates the semantic distance between attribute nodes using a soft-aligned attention mechanism; establishes synonym alignment connections and conflicting attribute pairs; inputs the conflicting attribute pairs into a discriminative network optimized by adversarial contrastive learning; the discriminative network focuses on the conflicting dimension with the greatest semantic difference through cross-attention gating units; calculates the conflict probability of the conflicting attribute pairs; and makes a comprehensive decision by combining multi-dimensional dynamic business rule thresholds with the conflict probability, thereby improving the reliability of negative word recognition and processing.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to an intelligent negative word construction method and system that combines multi-dimensional dynamic business rules. Background Technology

[0002] In search advertising on e-commerce and news feed platforms, negative search term strategies are crucial for controlling advertising costs and improving conversion rates. This means preventing ads from being shown to users when negative search terms are present. Existing technical solutions have the following main drawbacks:

[0003] Rule-based judgments based on static thresholds are rigid: they typically use a globally uniform fixed threshold (e.g., more than 30 clicks without conversion) to determine whether to negate a search term. This method ignores the differences in traffic tolerance for products across different categories, sales seasons, and lifecycle stages, easily leading to the false negative of potential keywords or the missed detection of invalid keywords, making it difficult to achieve refined operations.

[0004] Weak semantic understanding: Traditional models often rely on simple text matching or general pre-trained models for coarse-grained classification, lacking the ability to analyze the deep semantic relationships between search terms and products. They struggle to handle the problem of "feature entanglement," cannot accurately identify semantically related attributes that are not literal matches (such as "deep red" and "red"), and also have difficulty capturing the seasonal or scenario-based demand conflicts implied behind words such as "breathable."

[0005] Therefore, there is an urgent need for an intelligent negative keyword construction method and system that combines multi-dimensional dynamic business rules. Summary of the Invention

[0006] To achieve the objectives of this invention, the following technical solution is adopted:

[0007] Specifically, this application provides a method for constructing intelligent negative keywords by combining multi-dimensional dynamic business rules, which includes:

[0008] S1: Input unstructured search term text and product title text, and use a generative language model fine-tuned by ontology knowledge enhancement instructions. The generative language model outputs search term attribute vectors and product attribute vectors through an injected thought chain mechanism and attribute decoupling attention mask, and through inference chain and customized mask.

[0009] S2: Construct a bipartite graph with search term attribute vectors and product attribute vectors as nodes, use a soft alignment attention mechanism to calculate the semantic distance between attribute nodes, and establish synonym alignment connections and conflicting attribute pairs;

[0010] S3: Input the conflict attribute pair into a discriminant network optimized by adversarial contrastive learning. The discriminant network focuses on the conflict dimension with the greatest semantic difference through cross-attention gating units and calculates the conflict probability of the conflict attribute pair.

[0011] S4: Combining multi-dimensional dynamic business rule thresholds with the conflict probability to make a comprehensive decision, thereby constructing or downgrading negative search terms. When it is determined that the search term is not a negative search term, the downgrading strategy for the search term is determined based on the historical downgrading data of the product's search terms. Based on the downgrading strategy, and in combination with the product's exposure data and retrieval data containing negative keywords, the conflict probability identification and processing strategy for the conflict attribute pairs of search terms is determined.

[0012] The beneficial effects of this invention are as follows:

[0013] Extremely high discrimination accuracy: By introducing Domain Adaptive Pre-training (DAPT) and Adversarial Data Augmentation, the model significantly outperforms the traditional BERT model in recognizing ambiguous boundaries such as "long sleeves vs. three-quarter sleeves" and "business vs. casual".

[0014] Highly Explainable AI: With the help of CoT thinking chain and explicit conflict enumeration, the system can clearly point out whether the reason for the negative word is "collar mismatch" or "scenario conflict", solving the problem that "black box" models cannot be trusted by operations personnel.

[0015] Ontology-Augmented Instruction Tuning: Introducing the Chain-of-Thought (CoT) mechanism, it not only extracts explicit attributes but also captures implicit needs in search terms through logical reasoning.

[0016] Dynamic Graph-Based Semantic Alignment: Abandoning traditional hard matching, it constructs a bipartite graph and utilizes a soft alignment attention mechanism to effectively solve the vocabulary gap between search terms and product titles.

[0017] Siamese Interaction & Adversarial Perception: Employing an improved Siamese network structure and cross-attention gating units, this approach focuses on core conflict dimensions and enhances the model's discriminative ability on fuzzy boundaries through adversarial data augmentation.

[0018] Multi-dimensional dynamic threshold engine: Combines factors such as season, category, and life cycle to build a dynamic threshold model, realizes personalized negative word triggering. The combination of dynamic rule engine and semantic model enables the system to handle obvious attribute errors and adapt to the traffic tolerance of different seasons and life cycles, truly realizing refined management of people and goods matching.

[0019] Furthermore, the output includes search term attribute vectors and product attribute vectors, specifically:

[0020] A thought chain mechanism is injected into the instruction of the generative language model, which utilizes the structured attributes of the search terms in the search term text.

[0021] A customized masking mechanism is introduced into the decoding layer of the Transformer in the generative language model, which forces the model to focus on the attribute vector of the structured attributes of the input text when extracting structured attributes of different dimensions, so as to obtain the attribute vector of the search term.

[0022] Furthermore, a bipartite graph is constructed, specifically including:

[0023] The two sets of structured attributes input are defined as the left and right node sets of the graph, respectively. The search term attribute set constitutes the left node set, and the product title attribute set constitutes the right node set. Each attribute node contains the attribute type and a pre-trained semantic vector representation. During initialization, a full connection is established between the left and right node sets, that is, each attribute node on the left is connected to each attribute node on the right with a pre-defined edge to be evaluated, forming a complete bipartite graph structure, thereby exhaustively enumerating all potential cross-domain semantic association possibilities.

[0024] Furthermore, the semantic distance between attribute nodes is calculated, specifically including:

[0025] For each edge to be evaluated established in step one, a quantitative evaluation of semantic association strength is performed. This is done through a lightweight soft alignment attention layer. This layer takes the attribute nodes generated in the previous step as input and calculates the semantic association score between each pair of attribute nodes through a parameterized similarity function to obtain the semantic distance of the attribute nodes.

[0026] Furthermore, synonym alignment links and conflict attribute pairs are established, specifically including:

[0027] A configurable dynamic alignment threshold is set to make decisions on the calculated association scores. For each edge to be evaluated, if its association score is greater than or equal to the threshold, a strong alignment edge is established between the attribute nodes corresponding to the edge to be evaluated, and the attribute nodes are marked as satisfied.

[0028] After traversal, all attribute nodes are checked: if an attribute node fails to establish a strong alignment edge with any product-side attribute node, it is determined that the requirement of the attribute node lacks sufficient semantic correspondence in the product-side attribute nodes, thus the attribute node is identified as a potential conflict point;

[0029] Each tagged potential conflict point will be combined with the product side node with the highest associated score (regardless of whether the score meets the standard) to form a conflict attribute pair.

[0030] Furthermore, the system can construct or de-rank negative search terms, specifically including:

[0031] Dynamic business rule thresholds are dynamically calculated based on seasonal factors, category factors, product lifecycle factors, and operational preferences. The final decision-making logic is as follows:

[0032] If the number of clicks for a search term exceeds the threshold of the dynamic business rule and there is no conversion, the negative condition of the dynamic business rule is met, and it is directly marked as a negative search term.

[0033] If the probability of conflict output by the discrimination network of the search term is greater than the dynamic rule threshold but not greater than the preset confidence threshold, it is determined to be a "strong attribute conflict" search term and an early rejection operation is performed.

[0034] If the conflict probability output by the discrimination network for the search term does not exceed the preset confidence threshold, then a weight reduction processing scheme is determined according to the weight reduction processing strategy.

[0035] Secondly, the present invention provides a computer system comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-described intelligent negative word construction method combining multi-dimensional dynamic business rules when running the computer program.

[0036] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0038] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0039] Figure 1 This is a flowchart of an intelligent negative keyword construction method that combines multi-dimensional dynamic business rules;

[0040] Figure 2 This is a flowchart of the method for outputting search term attribute vectors and product attribute vectors;

[0041] Figure 3 This is a flowchart illustrating the methods for establishing synonym alignment links and conflicting attribute pairs;

[0042] Figure 4 This is a flowchart of a method for calculating the conflict probability of the conflict attribute pair. Detailed Implementation

[0043] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0044] Example 1

[0045] like Figure 1 As shown, this application provides a method for constructing intelligent negative keywords by combining multi-dimensional dynamic business rules, specifically including:

[0046] S1: Input unstructured search term text and product title text, and use a generative language model fine-tuned by ontology knowledge enhancement instructions. The generative language model outputs search term attribute vectors and product attribute vectors through an injected thought chain mechanism and attribute decoupling attention mask, and through inference chain and customized mask.

[0047] Furthermore, such as Figure 2 As shown, the output includes search term attribute vectors and product attribute vectors, specifically:

[0048] A thought chain mechanism is injected into the instruction of the generative language model, which utilizes the structured attributes of the search terms in the search term text.

[0049] A customized masking mechanism is introduced into the decoding layer of the Transformer in the generative language model, which forces the model to focus on the attribute vector of the structured attributes of the input text when extracting structured attributes of different dimensions, so as to obtain the attribute vector of the search term.

[0050] Specifically, the above steps go beyond traditional entity recognition (NER) by employing an ontology-augmented instruction tuning strategy, which mainly includes:

[0051] CoT Reasoning Injection:

[0052] A chain-of-thought mechanism is injected into the fine-tuning instructions of the generative language model. Instead of directly outputting attributes, the model is trained to generate "reasoning paths" first.

[0053] Exaggerated Logic: Input "Summer breathable running tee" -> Internal model inference: "User is looking for sports scenario -> needs breathable material -> Season is Summer" -> Output structured attributes. This mechanism significantly improves the ability to capture implicit attributes.

[0054] Attribute-Decoupling Attention Mask:

[0055] A customized masking mechanism is introduced in the decoding layer of Transformer to force the model to focus on different regions of the input text when extracting attributes of different dimensions (such as color and material), thereby preventing feature entanglement between attributes.

[0056] 1. Thought Chain Hint Reasoning (CoT) is a technique that guides large language models to "think in steps." Traditional models are "black boxes": they take text as input and output results directly (such as attribute labels). CoT, however, requires the model to output the logical chain of reasoning that led to the final answer before outputting the final answer itself.

[0057] Model: Based on Flan-T5 (a generative language model with good instruction compliance capabilities), fine-tuned. Training instructions: No longer simply "extract attributes from this text," but instead: "Please analyze the following search terms, infer the user's potential needs step by step, and then output structured attributes. Your reasoning process is: [Here the model generates the reasoning chain] The final attribute is: [Here the model outputs the attribute]"

[0058] Workflow example:

[0059] Input: “Summer breathable running tee” Model internal / output inference chain: “running” indicates the user is in a sports / sports scene; “breathable” indicates the user has a specific material requirement (breathability). “Summer” directly indicates the seasonal attribute; “tee” usually refers to a T-shirt, belonging to the upper garment category. The final output structured attributes are: {Scene: Sports, Material: Breathable, Season: Summer, Category: T-shirt}

[0060] Capturing latent attributes: This is the most crucial point. For example, the attribute "sports scene" doesn't appear directly in the search term but is inferred from the word "running." Similarly, the material requirement of "breathable" implicitly addresses scenarios like "coolness" and "sweating." Without CoT, the model might only extract the literal meanings of "summer," "breathable," and "T-shirt," losing the most important information about the "sports" scene.

[0061] Improved interpretability: Operations staff can see the reasons why the model negates certain words. For example, the system can report: "This word is negated because the model infers that the user needs a 'sports scenario,' while your product is a 'casual shirt,' creating a scenario conflict." This greatly enhances people's trust in AI decision-making.

[0062] It upgrades the model from a simple "pattern recognizer" to a "semantic understander" with basic logical reasoning capabilities. This is crucial for understanding user search intent, as users' true needs are often not stated directly.

[0063] 2. Attribute decoupling attention masking is a constraint mechanism imposed on the decoder part of Transformer models (such as T5 and BERT). The "mask" is like an indicator map that forces the model to "look" only at specific parts of the input text at different time steps when processing the text, while ignoring other parts.

[0064] Problem to be solved: Feature entanglement is a classic challenge in NLP. For example, consider the sentence: "A red cotton shirt." The task involves extracting the "color" and "material" attributes separately. A model without decoupling might distribute its attention evenly across "red," "cotton," and "shirt" when extracting "color." This results in a "red" feature vector that is mixed with information from both "cotton" and "shirt." This leads to:

[0065] When encountering "red silk shirt", the model may be affected in judging "red" due to the mismatch of "material" features. The model has difficulty distinguishing between "attribute A mismatch" and "attribute B mismatch".

[0066] Customized Mask: At each step of the model decoding (attribute generation), a manually designed attention mask is introduced. When the model decoder generates the attribute value of "color", the mask will force its attention to focus mainly on color-related words in the input text (such as "red" and "dark blue"), and suppress its high attention to words such as "cotton" and "shirt".

[0067] When generating the "material" attribute in the next step, the mask is switched again, forcing the model to focus on words such as "cotton" and "silk". As a result, the "color feature vector" and "material feature vector" learned by the model are relatively independent and pure, with minimal interference between them.

[0068] Improved accuracy: In the subsequent conflict determination stage, since the attribute features have been decoupled, the system can more clearly and accurately determine which dimension of the attribute is conflicting (is it the color that is wrong? Or the material that is wrong?), avoiding a "one-size-fits-all" approach.

[0069] Achieving fine-grained analysis: This is the technical foundation for achieving "ultra-fine-grained" semantic analysis. Only by untangling the intertwined features can we independently and accurately identify and compare very specific attributes such as "sleeve length," "collar type," and "waist type."

[0070] These two technologies are not isolated; they work closely together in the patented "generative semantic decoupling module" to form a powerful semantic parsing engine: CoT is responsible for "logical reasoning" and "implicit mining": like humans, it first understands the context and intent, and infers which attributes need to be paid attention to (including explicit and implicit ones). The attribute decoupling attention mask is responsible for "precise extraction" and "feature purification": when extracting each inferred attribute, it acts like a precision instrument, ensuring that each extracted attribute feature is independent and free from interference.

[0071] The ultimate goal is to transform a vague, unstructured search term (such as "breathable running T-shirt for summer") into a high-dimensional, sparse, feature-decoupled, and implicitly information-rich structured attribute vector. This high-quality vector serves as the perfect input for subsequent precise semantic alignment and conflict determination.

[0072] S2: Construct a bipartite graph with search term attribute vectors and product attribute vectors as nodes, use a soft alignment attention mechanism to calculate the semantic distance between attribute nodes, and establish synonym alignment connections and conflicting attribute pairs;

[0073] Furthermore, a bipartite graph is constructed, specifically including:

[0074] The two sets of structured attributes input are defined as the left and right node sets of the graph, respectively. The search term attribute set constitutes the left node set, and the product title attribute set constitutes the right node set. Each attribute node contains the attribute type and a pre-trained semantic vector representation. During initialization, a full connection is established between the left and right node sets, that is, each attribute node on the left is connected to each attribute node on the right with a pre-defined edge to be evaluated, forming a complete bipartite graph structure, thereby exhaustively enumerating all potential cross-domain semantic association possibilities.

[0075] Furthermore, the semantic distance between attribute nodes is calculated, specifically including:

[0076] For each edge to be evaluated established in step one, a quantitative evaluation of semantic association strength is performed. This is done through a lightweight soft alignment attention layer. This layer takes the attribute nodes generated in the previous step as input and calculates the semantic association score between each pair of attribute nodes through a parameterized similarity function to obtain the semantic distance of the attribute nodes.

[0077] Furthermore, such as Figure 3 As shown, establishing synonym alignment links and conflict attribute pairs specifically includes:

[0078] A configurable dynamic alignment threshold is set to make decisions on the calculated association scores. For each edge to be evaluated, if its association score is greater than or equal to the threshold, a strong alignment edge is established between the attribute nodes corresponding to the edge to be evaluated, and the attribute nodes are marked as satisfied.

[0079] After traversal, all attribute nodes are checked: if an attribute node fails to establish a strong alignment edge with any product-side attribute node, it is determined that the requirement of the attribute node lacks sufficient semantic correspondence in the product-side attribute nodes, thus the attribute node is identified as a potential conflict point;

[0080] Each tagged potential conflict point will be combined with the product side node with the highest associated score (regardless of whether the score meets the standard) to form a conflict attribute pair.

[0081] Specifically, in the middle layer of the model: the Cross-Domain Feature Alignment Module, this is the key "bridge" connecting generation and discrimination in this algorithm. In order to solve the huge gap in expression habits between search terms (QueryDomain) and product titles (Product Domain), we designed a dynamic graph alignment mechanism.

[0082] The Dynamic Semantic Alignment Graph constructs a bipartite graph from search term attributes and product title attributes.

[0083] Soft-Alignment Attention:

[0084] Traditional methods only perform hard matching (string matching). This module introduces a lightweight attention layer to calculate the semantic distance between attribute values. Even if the literal meanings are different (e.g., ST: "Crimson" vs PT: "Red"), the model can recognize their proximity in high-dimensional space through pre-trained semantic vector projection, thus establishing "synonymous alignment" edges instead of discarding them directly. Only nodes that cannot be aligned and are semantically mutually exclusive are marked as potential conflict points.

[0085] Discriminative Conflict Perception Module:

[0086] This module is based on the BERT architecture but has undergone deep structural modifications. It transforms into a contrast learning mechanism where users (search terms) and merchants (product titles) describe the same thing using different languages. Hard matching (strings perfectly identical) will fail: Search term: "deep red dress" (attribute: color: Crimson), Product title: "bright red fitted maxi dress" (attribute: color: Red), Hard matching result: Crimson ≠ Red → marked as "color conflict" → misjudgment! This can lead to a good product being incorrectly rejected. The goal is to establish an intelligent "translation" or "bridge" mechanism to identify that Crimson and Red are essentially the same, rather than conflicting.

[0087] In one possible specific embodiment, the collaborative work steps specifically include:

[0088] Assuming we have obtained the following from the previous module: Search term attribute set (ST): {Color: Crimson, Category: Dress, Length: Long}, Product attribute set (PT): {Color: Red, Category: Long Dress, Material: Silk};

[0089] Step 1: Construct a property bipartite graph;

[0090] Left node set (L): All attribute values ​​of the search term. [Crimson, dress, long], Right node set (R): All attribute values ​​of the product. [Red, long dress, silk], Initial edge set (E): Fully connected. That is, each node in the left set has a potential edge to every node in the right set. The initial weights of these edges are unknown. Edge examples: (Crimson, Red), (Crimson, long dress), (Crimson, silk), (dress, Red), ...

[0091] At this point, the significance of the graph is to explore all possible semantic relationships between each demand point of the search term and each selling point of the product.

[0092] Step 2: Soft alignment attention calculation (calculate edge weights);

[0093] This is the most crucial technical step. For each potential edge (ST_Node_i, PT_Node_j) in the graph, perform the following operations: Vectorization: Using a pre-trained encoder (such as BERT), convert the text nodes Crimson and Red into high-dimensional semantic vectors V_crimson and V_red, respectively. This encoder has been trained on a large corpus, and it knows that V_crimson and V_red are very close in the vector space.

[0094] Calculating semantic similarity (soft-aligned attention score): A lightweight attention layer (usually a feedforward network followed by similarity calculation, such as cosine similarity) is used to calculate the association score between two vectors. The formula (conceptually): Attention_Score(i, j) = F( V_i, V_j ), calculation examples: F(V_crimson, V_red) → score 0.95 (highly similar semantically), F(V_dress, V_long skirt) → score 0.88 (semantically similar, but "dress" and "long skirt" are hierarchical, not completely equivalent), F(V_long style, V_silk) → score 0.05 (completely unrelated semantically), F(V_long style, V_long skirt) → score 0.75 ("long style" is an attribute, "long skirt" is a product name, but semantically related).

[0095] Step 3: Establish alignment edges and identify conflict points, setting an alignment threshold (e.g., threshold_alignment = 0.8). Traverse all edges: If Attention_Score(i, j) > threshold_alignment, then establish a strong alignment edge between nodes i and j. This means the system considers these two nodes to express the same or highly compatible semantics.

[0096] Results: (Crimson, Red) Strong alignment edges are established. (Dress, Long Skirt) Medium-strength alignment edges may be established. If Attention_Score(i, j) < threshold_alignment, then no strong alignment edges are established. Identify "unaligned and mutually exclusive" conflicting nodes: Rule: For a search term node ST_i, if it cannot find any PT_j with which it can establish a strong alignment edge in the entire set of right nodes, then the node is marked as a "potential conflicting point".

[0097] Analysis Example: Crimson: Red was found as a strong alignment partner → Aligned, safe. Dress: May establish a moderate alignment edge with long skirt (assuming score 0.88 > 0.8) → Aligned, safe. Long: It scores very low with silk (0.05), is irrelevant to Red (0.1), and its semantic relevance to long skirt (0.75) may not reach the strong alignment threshold (0.8). Therefore, no clear strong alignment attribute was found for long skirt on the product side.

[0098] Conflict Detection: The user clearly desires a long garment (they want it to be long), but none of the product attribute sets {Red, Long Skirt, Silk} explicitly and strongly correlates with "this is a long product" ("Long Skirt" implies length, but the model considers it not to meet a strong alignment standard). Therefore, "long" is marked as a "potential conflict point."

[0099] Step 4: Output to the next stage, output the aligned attribute pairs: perform feature fusion or association marking on the node pairs (Crimson, Red) and (dress, long skirt) that have established strong alignment edges, as the "matched" part.

[0100] Output conflict attribute pairs: Combine the conflict points on the search term side (e.g., long style) with its most relevant product attribute (e.g., long skirt, even if it does not meet the criteria) or an "empty" attribute to form a (long style, NULL) pair, or directly pass the feature vector of long style as an isolated point to the next stage "discriminative conflict perception module". The task of this module is to finally determine whether the (long style, long skirt) pair is an "acceptable compatibility relationship" or an "unacceptable conflict relationship".

[0101] It bridges the lexical gap: no longer string matching, but semantic space matching. Crimson = Red, SUV ≈ off-road vehicle. Refined processing: by controlling the strictness of alignment through thresholds, it can flexibly handle complex relationships such as synonyms and hyponyms. Efficient filtering: the bipartite graph model and attention mechanism can quickly filter out truly suspicious points (i.e., unaligned nodes) from all possible attribute combinations, greatly reducing the computational burden of subsequent complex discrimination models.

[0102] This lays the foundation for discriminant analysis: it ensures that the final conflict discriminator receives truly ambiguous and difficult-to-determine cases (such as "long" vs. "long skirt"), rather than simple cases that are misclassified by hard matching (such as "deep red" vs. "red"). This allows the entire system to remain efficient while focusing its accuracy on the most challenging problems.

[0103] S3: Input the conflict attribute pair into a discriminant network optimized by adversarial contrastive learning. The discriminant network focuses on the conflict dimension with the greatest semantic difference through cross-attention gating units and calculates the conflict probability of the conflict attribute pair.

[0104] Furthermore, such as Figure 4 As shown, calculating the conflict probability of the conflict attribute pair specifically includes:

[0105] Step 1: Twin coding and feature extraction:

[0106] The input conflict attribute pairs are split into two parts: search term side attribute features and product side attribute features. Both are input into an encoder network with the same structure and shared parameters. Each encoder is modified based on a deep pre-trained model in the discriminant network to independently encode the input text or structured attributes into a high-dimensional, context-aware semantic feature vector.

[0107] Step 2: Cross-attention interaction and focus:

[0108] A cross-attention gating unit is introduced on top of the dual-tower encoder. The cross-attention gating unit receives the semantic feature vectors output by the dual towers and calculates the bidirectional attention weights between the semantic feature vectors. It automatically identifies and amplifies the dimension with the most significant semantic difference between the two semantic feature vectors and constructs fused features based on the bidirectional attention weights and the semantic feature vectors.

[0109] Step 3: Conflict detection and probability output

[0110] The fused features after cross-attention interaction are fed into a multilayer perceptron classifier head, which performs two tasks: first, outputs an overall conflict probability to quantify the degree of mismatch between the conflicting attribute pairs; second, performs multi-label classification to determine which predefined fine-grained conflict types the specific conflicting attribute pairs belong to.

[0111] Specifically, the Siamese Interaction Attention Network (Siamese Interaction Attention Network) architecture design abandons the simple [SEP] concatenation of inputs and instead uses a Siamese network structure to process aligned attribute pairs.

[0112] Tower A: Encodes search term attribute features, vst; Tower B: Encodes product attribute features, vpt.

[0113] Cross-Attention Gating Unit: At the top of the twin towers, a cross-attention mechanism is introduced to calculate Attention(vst, vpt). This gating unit can automatically "focus" on the dimensions with the greatest semantic differences (i.e., conflict points) while suppressing noise from irrelevant dimensions (such as the same gender or the same category).

[0114] In one possible embodiment, specifically exemplified by the search term (Query): "Summer long-sleeved breathable business shirt", and the product title (Product): "Men's pure cotton short-sleeved slim-fit professional shirt".

[0115] Input processed by the pre-processing module (entering this module): Aligned attribute pairs (from the soft alignment module): (Business, Occupation) - Semantically highly similar, aligned, non-conflicting. (Shirt, Shirt) - Synonyms, aligned, non-conflicting. (Breathable, Cotton) - "Cotton" has some breathability, which may be judged as weakly related or pending by the soft alignment module. We assume it enters this module as a conflict candidate pair.

[0116] Conflict candidate pairs (from the soft alignment module): Candidate pair A: (long sleeve, short sleeve) - explicit conflict, Candidate pair B: (summer, short sleeve) - potential scenario conflict (long sleeves are needed for sun protection in summer vs. the product is short sleeve), Candidate pair C: (breathable, pure cotton) - ambiguous judgment ("breathable" is an explicit requirement, "pure cotton" is a material but its breathability is not top-notch).

[0117] Module workflow (with case study):

[0118] Phase 1: Forward reasoning and decision-making, Step 1: Twin coding and feature extraction, the model's dual-tower encoders process candidate pairs respectively.

[0119] For candidate pair A (long sleeve, short sleeve): Tower A receives "long sleeve" and outputs a feature vector V_long, which encodes semantics such as "covering most of the arm", "warmth", "sun protection", and "formal". Tower B receives "short sleeve" and outputs a feature vector V_short, which encodes semantics such as "covering the forearm", "cool", "summer", and "casual".

[0120] Step 2: Cross-attention interaction and focusing. The cross-attention gating unit calculates the attention between V_long and V_short. The model will automatically discover that in the "sleeve length" dimension, the signal of "covering most of the arm" of V_long and "covering the forearm" of V_short form a strong contrast, and the attention weights surge in this dimension.

[0121] In terms of the "seasonal" association dimension, V_long's "sunscreen" and V_short's "summer" are somewhat related but also contradictory, thus gaining moderate attention.

[0122] Attention was suppressed on dimensions that were similar or unrelated, such as "gender" (implicit) and "category". As a result, the model focused on the core difference of "sleeve length" and filtered out other noise.

[0123] Step 3: Conflict Detection and Probability Output: The fused features are fed into the classification head. Result: Conflict probability: 0.99 (extremely high);

[0124] Conflict type: [SLEEVE_LENGTH_CONFLICT, SEASONAL_SCENARIO_CONFLICT]. The model not only determined the direct conflict in sleeve length, but also inferred the possible sun protection scenarios from the demand for "long sleeves in summer", which conflicted with the scenario of "short sleeves".

[0125] For candidate pair C (breathable, pure cotton): This is a hard example, and the cross-attention mechanism requires in-depth analysis: "breathable" is a functional requirement, while "pure cotton" is a material. Pure cotton is naturally breathable, but not optimal (e.g., linen and some synthetic fibers may be more breathable). The model needs to combine a lot of e-commerce knowledge to determine: in the context of "business shirts," are users' requirements for "breathability" so stringent that "pure cotton" cannot meet them, or is "pure cotton" sufficient to meet general needs?

[0126] Judgment result: Conflict probability: 0.65 (medium confidence, fuzzy boundary), Conflict type: [MATERIAL_FUNCTION_CONFLICT]

[0127] Explanation: The product's material, "pure cotton," may not perfectly match the user's functional requirement of "breathability," but it is not an absolute conflict. The system may combine this result with lower-weighted criteria or have it reviewed by operations personnel.

[0128] S4: Combining multi-dimensional dynamic business rule thresholds with the conflict probability to make a comprehensive decision, thereby constructing or downgrading negative search terms. When it is determined that the search term is not a negative search term, the downgrading strategy for the search term is determined based on the historical downgrading data of the product's search terms. Based on the downgrading strategy, and in combination with the product's exposure data and retrieval data containing negative keywords, the conflict probability identification and processing strategy for the conflict attribute pairs of search terms is determined.

[0129] Furthermore, the specific steps for training and updating the discriminant network are as follows:

[0130] Step 1: Domain Adaptive Pre-training:

[0131] Based on the weights of the general language model, secondary pre-training is performed using a large-scale domain-specific corpus;

[0132] Step 2: Phased training under the course learning strategy:

[0133] The secondary training process of the model follows a course learning strategy from easy to difficult, using samples with clear labels for training, and then gradually introducing samples containing complex semantic relationships such as synonyms and hyponyms.

[0134] Step 3: Apply adversarial data augmentation:

[0135] During the secondary training process, an adversarial generative network is introduced. This adversarial generative network continuously attempts to create "adversarial samples" that can deceive the current discriminative model by performing targeted semantic perturbations on the original training samples.

[0136] Step 4: Optimize the conflict feature enhancement loss function:

[0137] During training, a composite loss function is used to optimize the model. This function mainly consists of three parts: first, the focus loss, which reduces the weight of samples with a conflict probability greater than a preset conflict probability threshold, forcing the model to concentrate on overcoming difficult-to-distinguish boundary samples; second, the contrast margin loss, which aims to explicitly increase the distance between attribute pairs marked as conflicting in the feature space, while simultaneously narrowing the distance between compatible attribute pairs; and third, the regularization loss, which controls model complexity and prevents overfitting. By balancing and optimizing this composite objective, the feature space ultimately learned by the model possesses extremely strong class discrimination capabilities.

[0138] Specifically, to ensure the model possesses extremely high practical capabilities, the following training paradigm was adopted:

[0139] 5.1 Adversarial Data Augmentation:

[0140] A generative adversarial network (GAN) generator was trained specifically to create "adversarial examples." The generator's operation logic involves deliberately tweaking a word in the search term (e.g., changing "long sleeves" to "three-quarter sleeves") in an attempt to deceive the discriminative model. In this game of interaction with the generator, the discriminative model is forced to learn extremely subtle semantic boundaries, thereby significantly improving its sensitivity to attribute conflicts.

[0141] 5.2 Curriculum Learning Strategies:

[0142] Model training does not involve randomly feeding data; instead, it follows the principles of human learning.

[0143] Phase 1 (Easy): Learn about obvious conflicts (e.g., men's shoes vs. women's shoes).

[0144] Phase Two (Medium): Learning the relationship between synonyms and hyponyms (e.g., Crimson is a type of Red).

[0145] Phase Three (Hard): Learning extremely blurred boundaries and scene conflicts (e.g., business style vs. casual style, short sleeves vs. three-quarter sleeves).

[0146] This strategy ensures faster model convergence and stronger generalization ability on long-tailed complex cases.

[0147] 5.3 Domain-Adaptive Pre-training (DAPT):

[0148] Before using general BERT / T5 weights, a second pre-training was performed on an e-commerce corpus of tens of millions of words. This enabled the model to naturally understand the meaning of e-commerce specific terms such as "3 / 4 sleeve", "poly", and "tote", rather than treating them as ordinary general text.

[0149] Step 1: Domain-Adaptive Pre-training. Before training begins, the model has already learned from a corpus of tens of millions of e-commerce terms. Therefore, it can deeply understand that "long sleeves" is not only a single word, but is also associated with concepts such as "sunscreen," "air-conditioned room," and "business formal wear." The inherent properties of "pure cotton" include "moisture absorption," "breathability," and "wrinkle-resistant," which are often used as selling points in product descriptions. This allows the model to generate feature vectors rich in domain knowledge in the first stage of encoding.

[0150] Step 2: Course learning and training. Phase 1 (Easy): The model first learns to judge candidate pairs A (long sleeve, short sleeve) that are absolutely conflicting, such as (men's shoes, women's shoes) and (dress, jacket). Phase 2 (Medium): The model learns to handle synonym alignment such as (dark red, red) and relationships that require certain domain knowledge reasoning, such as (breathable, pure cotton). Phase 3 (Hard): The model finally conquers complex conflicts involving scenario and implicit need reasoning, such as candidate pair B, such as (business style, casual style) and (summer long sleeve, short sleeve).

[0151] Step 3: Adversarial Data Augmentation. During training, the adversarial generator (GAN) will generate things such as changing the search term "long-sleeved shirt" to "three-quarter sleeve shirt" and changing the product "pure cotton fabric" to "cotton fabric". In order not to be deceived, the discrimination model must learn to recognize that "long sleeve" and "three-quarter sleeve" are both sleeve lengths, but they are different and mutually exclusive specifications (should be judged as conflict), and "pure cotton" and "cotton" are synonymous expressions (should be judged as compatible).

[0152] It is through countless such "games" that the model has gained a keen ability to judge the absolute difference between "long sleeve / short sleeve" in candidate pair A and the subtle difference between "breathable / pure cotton" in candidate pair C.

[0153] Step 4: Conflict-Aware Loss Function. To address the imbalance between positive and negative samples, we designed a composite loss function: L total = λ1LFocal + λ2Lcontrastive + λ3LRegularization.

[0154] Focal Loss focuses on "hard examples," which are hidden cases that appear related but actually conflict in attributes (e.g., Deep V-neck vs V-neck). Contrastive Margin Loss forces an increase in the Euclidean distance between conflicting attribute pairs in the feature space, while simultaneously reducing the distance between compatible attribute pairs. Specifically, the conflict feature enhancement loss function is optimized as follows: for candidate pair A (easily distinguishable examples), Focal Loss reduces its loss weight to prevent the model from focusing too much on these learned, simple cases; for candidate pair C (hard examples), Focal Loss increases its loss weight, forcing the model to concentrate on optimizing the judgment of such ambiguous boundaries.

[0155] The contrastive interval loss performs the following operations: it pushes the feature vector of the conflicting pair (long sleeve, short sleeve) further apart in space, pulls the feature vector of the compatible pair (business, professional) closer together, and for (breathable, pure cotton), it learns to place it in a fuzzy boundary region between "obvious conflict" and "obvious compatibility".

[0156] It should be noted that the specialized corpus includes e-commerce titles, comments, and specifications.

[0157] Furthermore, the dynamic business rules include a seasonal factor, used to dynamically adjust the tolerance for peak and off-peak seasons; a category factor, which calculates the conversion benchmark for different categories based on historical data; a life cycle factor, which is lenient during the exploration period of new products and strict during the maturity period; and an operational preference factor.

[0158] Specifically, step one: multi-dimensional business factor collection and quantification. When a "search term-product" matching instance to be judged is received, the system first collects and quantifies factors of four core dimensions from the business database in real time.

[0159] Determine and quantify the seasonal factor (S_factor): Data source: current system date, historical sales seasonal curve of the product category; Logic: determine whether the current sales season is the off-season, average season, or peak season for the product category.

[0160] Quantification example:

[0161] Peak season (e.g., winter down jackets): S_factor = 1.2 (increased tolerance, allowing more exploratory clicks), off-season: S_factor = 1.0 (baseline tolerance), off-season (e.g., summer down jackets): S_factor = 0.8 (decreased tolerance, stricter requirements).

[0162] Identify and quantify the category factor (C_factor): Data source: historical conversion rate data and average cost-per-click data of the category to which the product belongs; Logic: the threshold can be relaxed for categories with high conversion and low decision cost; the threshold needs to be tightened for categories with low conversion and high decision cost.

[0163] Quantification example:

[0164] High-frequency, low-priced product categories (such as socks and snacks): C_factor = 1.3 (fast conversion, low cost per click, more lenient); low-frequency, high-priced product categories (such as jewelry and large appliances): C_factor = 0.7 (slow conversion, high value per click, extremely strict).

[0165] Identify and quantify the lifecycle factor (L_factor): Data source: product listing time, historical sales growth curve, operational tags; Logic: new products need exposure and exploration, while older products need efficient harvesting.

[0166] Quantitative examples: New product phase (listed for <30 days): L_factor = 1.5 (vigorous exploration, tolerance for invalid clicks), growth phase: L_factor = 1.0 (benchmark), maturity / decline phase: L_factor = 0.6 (meticulous cultivation, strict prevention of waste).

[0167] Determine the operational preference coefficient (P_user): Data source: Global or campaign-level settings in the system backend by advertisers or operations personnel. Logic: Allows manual intervention to adjust the overall risk preference of the system.

[0168] Quantitative examples: Aggressive (pursuing market share): P_user = 1.2 (willing to take more risk), Balanced: P_user = 1.0 (default), Conservative (pursuing absolute ROI): P_user = 0.8 (risk averse).

[0169] Step 2: Dynamic Threshold Calculation

[0170] The quantified factors are then substituted into a preset calculation formula to generate a personalized threshold for the current judgment instance.

[0171] Formula: Threshold_dynamic =Base × S_factor × C_factor × L_factor ×P_user

[0172] Base (basic threshold): An initial value set based on global experience, such as "negative if there is no conversion after 15 clicks".

[0173] Calculation process: The four factors adjust the base threshold in a multiplicative manner. Tightening any factor (value < 1) will lower the final threshold, making the system more likely to trigger rejection; conversely, relaxing any factor (value > 1) will raise the threshold, making the system more lenient.

[0174] Step 3: Threshold Output and Application

[0175] The calculated Threshold_dynamic is output to the "Decision Fusion Execution Module" in real time. The output is a specific numerical value, such as the dynamic click threshold = 25 times.

[0176] Application logic: This threshold will serve as the benchmark for subsequent "hard rule judgments". The system will compare the cumulative number of clicks or ad spend for the current search term under this product. If it exceeds this dynamic threshold and there is still no conversion, it will be marked as a negative candidate based on business rules, regardless of the semantic model's judgment.

[0177] Furthermore, the system can construct or de-rank negative search terms, specifically including:

[0178] Dynamic business rule thresholds are dynamically calculated based on seasonal factors, category factors, product lifecycle factors, and operational preferences. The final decision-making logic is as follows:

[0179] If the number of clicks for a search term exceeds the threshold of the dynamic business rule and there is no conversion, the negative condition of the dynamic business rule is met, and it is directly marked as a negative search term.

[0180] If the probability of conflict output by the discrimination network of the search term is greater than the dynamic rule threshold but not greater than the preset confidence threshold, it is determined to be a "strong attribute conflict" search term and an early rejection operation is performed.

[0181] If the conflict probability output by the discrimination network for the search term does not exceed the preset confidence threshold, then a weight reduction processing scheme is determined according to the weight reduction processing strategy.

[0182] The system combines the outputs of the rule layer and the model layer to make the final decision: Hard rule judgment: If the dynamic threshold condition is met (e.g., clicks > Threshold_dynamic and no conversions), it is marked as pending rejection.

[0183] Model-assisted judgment (prediction): If the rule layer is not triggered, but the HG-SCPN model determines that there is a high confidence conflict (e.g., Score>0.95 and type is SLEEVE_LENGTH_CONFLICT), it is marked as "strong attribute conflict".

[0184] Implementation strategy: Keywords with "strong attribute conflicts" can be negated or downgraded in advance, regardless of the click threshold, to achieve zero click waste.

[0185] Specifically, the method for determining the weight reduction processing scheme is as follows:

[0186] Input: A search term to be processed and its associated products. The system has calculated the "conflict probability (P_conflict)" of the search term relative to the product (e.g., 0.8), as well as its clicks and conversions. Define the match probability: P_match = 1 - P_conflict. P_match represents the match degree between the search term and the product. The higher the P_conflict, the lower the P_match. Obtain historical demotion data: The system retrieves a list of all search terms that have historically been demoted for this product from the database, i.e., the "demotion search term library". Record the total number (N_down) and the "demotion weight value (W_history)" for each term.

[0187] Based on the historical de-ranking data of the search terms for the product, the search terms for the product that have been de-ranked are identified and used as de-ranked search terms.

[0188] Based on the click data of the search term, the click volume of the search term is determined;

[0189] Based on the composition data of the search terms to be downgraded and the number of clicks on the search terms, a downgrade processing scheme for the search terms is determined.

[0190] For example, the product is a newly listed "Women's Autumn Lightweight Wool Sweater". The search term to be processed is "Winter Thick Sweater". The current data is as follows: the number of clicks brought to this product by this search term is 120, and the number of conversions brought to this product by this search term is 1.

[0191] The conflict probability (P_conflict) determined by the HG-SCPN semantic model is 0.85 (the model determines strong conflicts related to "season" and "thickness"). Therefore, the matching probability (P_match) = 1 - 0.85 = 0.15.

[0192] Historical demotion data for this product: Total number of search terms with historical demotion (N_down): 8; Details of historical demotion terms: Among them, 3 terms have a demotion weight value W_history>0.7 (belonging to high-weight conflict terms), and the remaining 5 terms have lower weights.

[0193] It is understood that in the above steps, the conversion rate of the search term is determined by the ratio of the conversion volume to the click volume. When the conversion rate of the search term is greater than the preset conversion rate threshold, it is determined that the search term does not need to be downgraded.

[0194] Specific example: Initial conversion rate screening, calculating conversion rate: CVR = Conversions / Clicks = 1 / 120 ≈ 0.83%, judgment: 0.83% < 2% (T_cvr). The conversion rate does not meet the standard and cannot be exempted. Proceed to the next step.

[0195] Additionally, it should be noted that when the conversion rate of the search term is not greater than the preset conversion rate threshold, it is then determined whether the number of the downgraded search terms is less than the preset downgraded search term number threshold. If not, the search term is downgraded using a preset strategy; if so, proceed to the next step.

[0196] Specifically, the product's historical ranking penalty status is checked, and the data obtained is as follows: the total number of historically penalized keywords for this product, N_down = 8, and the judgment is: 8 < 10(T_num_down). The number of historically penalized keywords has not reached the risk threshold. At this time, the probability of a search statement containing multiple penalized keywords is small, and the preset strategy is not triggered, proceeding to the next step.

[0197] Based on the weight values ​​of different de-weighted search terms, determine the de-weighted search terms whose weight values ​​are greater than a preset weight threshold. Determine whether the number of de-weighted search terms whose weight values ​​are greater than the preset weight threshold is greater than a preset search term number threshold. If so, use a preset strategy to de-weight the search terms. If so, proceed to the next step.

[0198] Specifically, the key conflict density check involves obtaining data: among historically downgraded words, the number of high-weight words (W_history>0.7) N_high = 3, and the judgment is: 3>2(T_num_high). The number of high-weight conflict words exceeds the threshold.

[0199] Decision: Trigger preset strategy (Strategy A). Although the total number is not large, the high density of conflicting keywords indicates that this product may attract mismatched traffic through keywords with lower ranking, requiring stricter management. Redirect to Strategy A to calculate the weight.

[0200] Determine whether the number of clicks on the search term is less than a preset click threshold (e.g., 10). If so, determine that the search term does not need to be downgraded. If not, use the second preset strategy to downgrade the search term.

[0201] Understandably, a preset strategy is used to reduce the ranking of search terms, specifically including:

[0202] The correction factor is determined based on the number of the downgraded search terms, and the difference between 1 and the conflict probability is taken as the matching probability.

[0203] The product of the correction factor and the matching probability is used as the weight value of the search term.

[0204] Specifically, for example, α = 1 / sqrt(1+N_down)=1 / sqrt(1+8) =1 / sqrt(9)=1 / 3≈0.333, calculate the final weight (W_final): W_final = α *P_match=0.333*0.15=0.05.

[0205] It should be noted that N_down represents the number of search terms that have been demoted in search ranking.

[0206] It should be noted that the correction factor is related to the number of the de-weighted search terms; the more de-weighted search terms there are, the smaller the correction factor will be.

[0207] Specifically, the second preset strategy is to directly use the matching probability as the weight value of the search term.

[0208] Furthermore, the method for determining the conflict probability identification and processing strategy for conflict attribute pairs of search terms is as follows:

[0209] This embodiment illustrates how to dynamically determine "whether to initiate deep semantic conflict analysis (i.e., call the HG-SCPN model)" based on the product's historical exposure, negative keyword search volume, and the performance of the current search term. Essentially, this is a "pre-filter" that balances model computing resources, business benefits, and risk control.

[0210] Based on the exposure data of the product, determine the exposure volume of the product on different dates;

[0211] Based on the search data containing negative keywords for the product, determine the search volume containing negative keywords for the product on different dates, and use it as the negative keyword search volume;

[0212] Based on the aforementioned demotion strategy, the product's exposure on different dates, and the search volume of negative keywords, an identification and processing strategy is used to determine the conflict probability of conflict attribute pairs of search terms.

[0213] It should be noted that the exposure volume is determined based on the number of users who push the product on different dates.

[0214] It should be noted that the negative keyword retrieval volume is determined based on the number of users whose search text contains the negative keyword.

[0215] In a possible specific embodiment, suppose we are evaluating a search term "urban casual sneakers" under the product "men's waterproof hiking shoes". The product's historical exposure data is as follows: in the past 7 days, the product has been exposed to an average of 5,000 users per day. The product's historical negative keyword retrieval data is as follows: in the past 7 days, users have searched and triggered an average of 100 searches per day containing the negative keywords set for the product.

[0216] Current search term data: Click count: 8 times, dynamic business rule threshold (calculated by the dynamic rule module): 50 clicks.

[0217] It should be noted that the identification and processing strategy for determining the conflict probability of conflict attribute pairs of search terms is based on the aforementioned demotion processing strategy, the exposure volume of the product on different dates, and the search volume of negative keywords.

[0218] Case 1: If the number of clicks on the search term is greater than the preset ratio of the dynamic business rule threshold, where the preset ratio is less than 1, then it is determined that the number of clicks on the search term is high, and the conflict probability identification processing of the conflict attribute pairs of the search term is directly performed.

[0219] Specifically, to determine whether it falls under "Case 1 - Extremely High Click Count", calculate: η × dynamic threshold = 0.8 × 50 = 40 times. Decision: Current click count is 8 times < 40 times; therefore, it does not fall under Case 1. Proceed to subsequent case assessments.

[0220] Case 2: If the number of clicks on the search term is not greater than the preset dynamic business rule threshold, and the average exposure of the product on different dates is not less than the preset exposure threshold, then it is determined that no conflict probability identification processing of the conflict attribute pairs of the search term is required.

[0221] Specifically, to determine if it falls under "Situation 2 - Sufficient Product Exposure", the criteria are: daily average product exposure is 5000 times < 10000 times (preset exposure threshold). Decision: It does not fall under Situation 2 (exposure is insufficient). Proceed to the judgment of Situation 3 / 4 / 5.

[0222] Case 3: If the average exposure of the product on different dates is less than the preset exposure threshold, obtain the ratio of the negative keyword search volume to the exposure volume. If the ratio of the negative keyword search volume to the exposure volume is greater than the preset threshold, it is determined that no conflict probability identification processing of the conflict attribute pairs of the search terms is required.

[0223] Specifically, to determine if it falls under "Situation 3 - High Negative Keyword Rate", calculate the historical negative keyword rate: Daily average negative keyword search volume / Daily average exposure volume = 100 / 5000 = 2%. Judgment: 2% < 5% (preset negative keyword rate threshold). Decision: Not belonging to Situation 3 (overall negative keyword management of the product is good). Proceed to the judgment of Situation 4 / 5.

[0224] Case 4: If the ratio of the negative keyword search volume to the exposure volume is not greater than a preset threshold, and if the click volume of the search term is less than the preset click volume threshold, then no conflict probability identification processing of the conflict attribute pair of the search term will be performed.

[0225] Case 5: If the number of clicks on the search term is not greater than the preset ratio of the dynamic business rule threshold, and is not less than the preset click volume threshold, then the conflict probability of the conflict attribute pair of the search term is identified and processed.

[0226] Step 4: Determine whether it belongs to "Case 4" or "Case 5", and determine the relationship between the number of clicks and the preset click threshold: Current search term clicks = 8 times, preset click threshold = 10 times, 8 times < 10 times, decision: belongs to Case 4. At this time, even if deep semantic conflict analysis is performed, it will not be downgraded. Since the downgrade strategy requires more than 10 clicks, the final decision is: do not start deep semantic conflict identification processing (i.e. call the computationally intensive HG-SCPN model for analysis) for the search term "urban casual sneakers".

[0227] This "method for determining conflict probability identification and processing strategies" is essentially an "intelligent computing resource scheduler" and a "business analysis trigger".

[0228] It avoids computational waste: it prevents the expensive HG-SCPN model from analyzing every irrelevant search term. It implements risk-tiered response: it handles urgent risks (high clicks) immediately and sorts non-urgent risks by value. It incorporates product lifecycle management: it adopts different analysis strategies for popular products with sufficient exposure and products plagued by problems.

[0229] It ensures the timeliness and reliability of the analysis: only when the amount of data is sufficient (the number of clicks meets the standard) is it in line with the dimensionality reduction strategy, and the benefits of semantic analysis are higher. Therefore, deep semantic analysis is necessary.

[0230] Example 2

[0231] Secondly, the present invention provides a computer system comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-described intelligent negative word construction method combining multi-dimensional dynamic business rules when running the computer program.

[0232] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0233] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0234] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for constructing intelligent negative keywords by combining multi-dimensional dynamic business rules, characterized in that, Specifically, it includes: S1 takes unstructured search term text and product title text as input, and uses a generative language model fine-tuned by ontology knowledge enhancement instructions. The generative language model outputs search term attribute vectors and product attribute vectors through an injected thought chain mechanism and attribute decoupling attention mask, and through inference chain and customized mask. S2 constructs a bipartite graph with search term attribute vectors and product attribute vectors as nodes, and uses a soft alignment attention mechanism to calculate the semantic distance between attribute nodes, establishing synonym alignment connections and conflicting attribute pairs; S3 inputs the conflict attribute pair into a discriminant network optimized by adversarial contrastive learning. The discriminant network focuses on the conflict dimension with the greatest semantic difference through a cross-attention gating unit and calculates the conflict probability of the conflict attribute pair. By combining multidimensional dynamic business rule thresholds with the conflict probability, a comprehensive decision is made to construct or reduce the ranking of negative search terms. When it is determined that the search term is not a negative search term, the ranking reduction strategy for the search term is determined based on the historical ranking reduction data of the product's search terms. The method for determining the conflict probability identification and processing strategy for the conflict attribute pairs of the search terms is as follows: Based on the exposure data of the product, determine the exposure volume of the product on different dates; Based on the search data containing negative keywords for the product, determine the search volume containing negative keywords for the product on different dates, and use it as the negative keyword search volume; Based on the weight reduction processing strategy, the exposure volume of the product on different dates, and the negative keyword retrieval volume, an identification and processing strategy is used to determine the conflict probability of conflict attribute pairs of search terms. The strategy for identifying the conflict probability of the conflict attribute pairs of the search terms is to dynamically decide whether to initiate a strategy for deep semantic conflict analysis of a certain search term, wherein the certain search term is a search term that has not undergone deep semantic conflict analysis. If not started, there is no need to perform deep semantic conflict analysis on a search term; The demotion strategy includes using a preset strategy to demotion the search term, not performing demotion, and using a second preset strategy to demotion the search term.

2. The intelligent negative keyword construction method combining multi-dimensional dynamic business rules as described in claim 1, characterized in that, Output the search term attribute vector and the product attribute vector, specifically including: A thought chain mechanism is injected into the instruction of the generative language model, which utilizes the structured attributes of the search terms in the search term text. A customized masking mechanism is introduced into the decoding layer of the Transformer in the generative language model, which forces the model to focus on the attribute vector of the structured attributes of the input text when extracting structured attributes of different dimensions, so as to obtain the attribute vector of the search term.

3. The intelligent negative keyword construction method combining multi-dimensional dynamic business rules as described in claim 1, characterized in that, Constructing a bipartite graph specifically includes: The two sets of structured attributes input are defined as the left and right node sets of the graph, respectively. The search term attribute set constitutes the left node set, and the product title attribute set constitutes the right node set. Each attribute node contains the attribute type and a pre-trained semantic vector representation. During initialization, a full connection is established between the left and right node sets, that is, each attribute node on the left is connected to each attribute node on the right with a pre-defined edge to be evaluated, forming a complete bipartite graph structure, thereby exhaustively enumerating all potential cross-domain semantic association possibilities.

4. The intelligent negative keyword construction method combining multi-dimensional dynamic business rules as described in claim 3, characterized in that, Calculate the semantic distance between attribute nodes, specifically including: For each edge to be evaluated, a quantitative assessment of semantic association strength is performed. This is done through a lightweight soft alignment attention layer, which takes the attribute nodes generated in the previous step as input and calculates the semantic association score between each pair of attribute nodes through a parameterized similarity function to obtain the semantic distance of the attribute nodes.

5. The intelligent negative keyword construction method combining multi-dimensional dynamic business rules as described in claim 3, characterized in that, Establishing synonym alignment links and conflicting attribute pairs specifically includes: A configurable dynamic alignment threshold is set to make decisions on the calculated association scores. For each edge to be evaluated, if its association score is greater than or equal to the threshold, a strong alignment edge is established between the attribute nodes corresponding to the edge to be evaluated, and the attribute nodes are marked as satisfied. After traversal, all attribute nodes are checked: if an attribute node fails to establish a strong alignment edge with any product-side attribute node, it is determined that the requirement of the attribute node lacks sufficient semantic correspondence in the product-side attribute nodes, thus the attribute node is identified as a potential conflict point; Each tagged potential conflict point will be combined with the product side node with the highest associated score to form a conflict attribute pair.

6. The intelligent negative keyword construction method combining multi-dimensional dynamic business rules as described in claim 1, characterized in that, Calculating the conflict probability of the conflict attribute pairs specifically includes: Step 1: Twin coding and feature extraction: The input conflict attribute pairs are split into two parts: search term side attribute features and product side attribute features. Both are input into an encoder network with the same structure and shared parameters. Each encoder is modified based on a deep pre-trained model in the discriminant network to independently encode the input text or structured attributes into a high-dimensional, context-aware semantic feature vector. Step 2: Cross-attention interaction and focus: A cross-attention gating unit is introduced on top of the dual-tower encoder. The cross-attention gating unit receives the semantic feature vectors output by the dual towers and calculates the bidirectional attention weights between the semantic feature vectors. It automatically identifies and amplifies the dimension with the most significant semantic difference between the two semantic feature vectors and constructs fused features based on the bidirectional attention weights and the semantic feature vectors. Step 3: Conflict detection and probability output; The fused features after cross-attention interaction are fed into a multilayer perceptron classifier head, which performs two tasks: first, outputs an overall conflict probability to quantify the degree of mismatch between the conflicting attribute pairs; second, performs multi-label classification to determine which predefined fine-grained conflict types the specific conflicting attribute pairs belong to.

7. The intelligent negative keyword construction method combining multi-dimensional dynamic business rules as described in claim 1, characterized in that, To construct or reduce the ranking of negative search terms, specifically including: Dynamic business rule thresholds are dynamically calculated based on seasonal factors, category factors, product lifecycle factors, and operational preferences. The final decision-making logic is as follows: If the number of clicks for a search term exceeds the threshold of the dynamic business rule and there is no conversion, the negative condition of the dynamic business rule is met, and it is directly marked as a negative search term. If the probability of conflict output by the discrimination network of the search term is greater than the dynamic rule threshold but not greater than the preset confidence threshold, it is determined to be a search term with strong attribute conflict and an early rejection operation is performed. If the conflict probability output by the discrimination network for the search term does not exceed the preset confidence threshold, then a weight reduction processing scheme is determined according to the weight reduction processing strategy.

8. The intelligent negative keyword construction method combining multi-dimensional dynamic business rules as described in claim 1, characterized in that, The method for determining the weight reduction processing scheme is as follows: Based on the historical demotion data of the search terms for the product, the search terms that are subject to demotion are determined and used as demotion search terms; Based on the click data of the search term, the click volume of the search term is determined; Based on the composition data of the search terms to be downgraded and the number of clicks on the search terms, a downgrade processing scheme for the search terms is determined.

9. A computer system, comprising: A memory and processor connected by communication, and a computer program stored on the memory and capable of running on the processor, characterized in that, when the processor runs the computer program, it executes a method for constructing intelligent negative words in combination with multidimensional dynamic business rules as described in any one of claims 1-8.

Images