Cross-border supplier traceability evaluation method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of cross-border supplier traceability evaluation, and discloses a cross-border supplier traceability evaluation method and device, electronic equipment and a storage medium, and the method comprises the steps: achieving the dynamic and three-dimensional verification of a supplier compliance declaration through integrating multi-modal objective evidence data and constructing a space-time correlation evidence chain, and improving the verification efficiency of the supplier compliance declaration. Through the automatic process of semantic analysis and cross validation, the evaluation efficiency is improved, and the evaluation result and the traceability report are generated based on the preset compliance database, so that compliance judgment can intelligently adapt to regulatory differences of different markets, and the flexibility and reliability of the method are enhanced. The supplier compliance evaluation method has the advantages that an efficient, accurate and traceable supplier compliance evaluation tool is provided for enterprises, the supply chain ESG risk is greatly reduced, a sustainable and transparent supply chain system is constructed, the objectivity and accuracy of evaluation are greatly improved, and false declarations are effectively restrained.

Description

Methods, devices, electronic equipment and storage media for sourcing and evaluation of cross-border suppliers Technical Field

[0001] This invention relates to the field of cross-border supplier traceability assessment technology, and in particular to a cross-border supplier traceability assessment method, apparatus, electronic device and storage medium. Background Technology

[0002] In the context of increasingly complex global supply chains, environmental, social, and governance (ESG) compliance assessments and traceability of cross-border suppliers have become a critical aspect of corporate risk management. However, existing assessment processes heavily rely on static compliance statements (such as reports or certificates) provided unilaterally by suppliers, lacking efficient and objective verification methods and making it difficult to verify their authenticity. Secondly, verification data sources are isolated; objective evidence such as satellite remote sensing, IoT sensors, and third-party certifications are scattered across different systems, forming data silos that cannot be effectively correlated and integrated, making it difficult to construct a dynamic chain of evidence reflecting the full picture of suppliers. Summary of the Invention

[0003] Based on this, it is necessary to propose a method, device, electronic equipment and storage medium for cross-border supplier traceability assessment to address the existing problems in cross-border supplier traceability assessment.

[0004] A method for tracing and assessing cross-border suppliers includes: acquiring a compliance statement text and multimodal objective evidence data from a designated supplier; wherein the multimodal objective evidence data consists of evidence data from multiple different data sources; performing semantic parsing on the compliance statement text to generate a structured semantic vector; constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; generating a credibility assessment result and a compliance tracing report for the designated supplier based on a preset compliance database and the risk quantification index; and determining whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance tracing report.

[0005] Furthermore, the step of semantically parsing the compliance statement text to generate a structured semantic vector includes: encoding the multilingual compliance statement text using a cross-language pre-trained large language model to obtain an encoding result; identifying and extracting target entity information related to a preset dimension from the encoding result; and vectorizing the target entity information to form the structured semantic vector.

[0006] Furthermore, the step of constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data includes: performing time synchronization and spatial registration on each piece of evidence data in the multimodal objective evidence data to obtain each registered target evidence data; associating and integrating the registered target evidence data according to time series and geographic spatial location to obtain an evidence matrix as the spatiotemporal correlation evidence chain.

[0007] Furthermore, the step of cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index includes: using a contrastive learning algorithm to calculate the semantic consistency score between the structured semantic vector and the spatiotemporal correlation evidence chain; and generating the risk quantification index based on the semantic consistency score.

[0008] Furthermore, in the step of generating a credibility assessment result and a compliance tracing report for the designated supplier based on a preset compliance database and the risk quantification indicators, the step of generating the credibility assessment result includes: obtaining the target market region corresponding to the designated supplier; retrieving the corresponding regional ESG regulatory subset from the compliance database according to the target market region; setting the verification weight of each regulatory clause in the regional ESG regulatory subset using a preset reinforcement learning model based on the risk quantification indicators; performing a weighted assessment of the risk quantification indicators according to the verification weight of each regulatory clause to obtain a weighted assessment result; and generating the credibility assessment result based on the weighted assessment result.

[0009] Furthermore, in the step of generating a credibility assessment result and a compliance traceability report for the designated supplier based on a preset compliance database and the risk quantification indicators, the step of generating the compliance traceability report includes: obtaining the designated supplier's raw material procurement information, production batch data, and logistics transportation trajectory; associating the compliance statement text content and multimodal objective evidence data with the designated supplier's raw material procurement information, production batch data, and logistics transportation trajectory to obtain a visualized traceability map; and generating the compliance traceability report based on the visualized traceability map.

[0010] Furthermore, the multimodal objective evidence data includes at least two of the following: satellite remote sensing image data, time-series data from IoT sensors deployed in production facilities, and digital records from third-party certification platforms.

[0011] A cross-border supplier traceability assessment device includes: an acquisition module for acquiring a compliance statement text and multimodal objective evidence data of a designated supplier; wherein the multimodal objective evidence data consists of evidence data corresponding to multiple different data sources; a parsing module for semantically parsing the compliance statement text to generate a structured semantic vector; a construction module for constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; a verification module for cross-verifying the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; a generation module for generating a credibility assessment result and a compliance traceability report for the designated supplier based on a preset compliance database and the risk quantification index; and a judgment module for judging whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance traceability report.

[0012] An electronic device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the following steps: acquiring a compliance statement text from a designated supplier and multimodal objective evidence data; wherein the multimodal objective evidence data consists of evidence data corresponding to multiple different data sources; performing semantic parsing on the compliance statement text to generate a structured semantic vector; constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; generating a credibility assessment result and a compliance tracing report for the designated supplier based on a preset compliance database and the risk quantification index; and determining whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance tracing report.

[0013] A computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the following steps: acquiring a compliance statement text from a designated supplier and multimodal objective evidence data; wherein the multimodal objective evidence data comprises evidence data corresponding to multiple different data sources; performing semantic parsing on the compliance statement text to generate a structured semantic vector; constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; generating a credibility assessment result and a compliance tracing report for the designated supplier based on a preset compliance database and the risk quantification index; and determining whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance tracing report.

[0014] The beneficial effects of this invention are as follows: By integrating multimodal objective evidence data and constructing a spatiotemporally correlated evidence chain, dynamic and three-dimensional verification of supplier compliance statements is achieved. The automated process of semantic parsing and cross-validation improves assessment efficiency. Based on a pre-set compliance database, assessment results and traceability reports are generated, enabling compliance judgments to intelligently adapt to regulatory differences in different markets. This enhances the flexibility and reliability of the method, providing enterprises with an efficient, accurate, and traceable supplier compliance assessment tool. This significantly reduces supply chain compliance risks, builds a sustainable and transparent supply chain system, greatly improves the objectivity and accuracy of assessments, and effectively curbs false statements. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0016] Wherein: Figure 1 is an application environment diagram of the cross-border supplier traceability assessment method in one embodiment; Figure 2 is a flowchart of the cross-border supplier traceability assessment method in one embodiment; Figure 3 is a structural block diagram of the cross-border supplier traceability assessment device in one embodiment; Figure 4 is a structural block diagram of the electronic device in one embodiment. Detailed Implementation

[0017] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Figure 1 illustrates an application environment diagram for cross-border supplier traceability assessment in one embodiment. Referring to Figure 1, the cross-border supplier traceability assessment method is applied to a cross-border supplier traceability assessment system. This system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; the mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to obtain the compliance statement text of a specified supplier, and the server 120 is used to generate the credibility assessment result of the specified supplier.

[0019] As shown in Figure 2, in one embodiment, a cross-border supplier traceability assessment method is provided. This method can be applied to both terminals and servers; this embodiment uses terminal application as an example. The cross-border supplier traceability assessment method specifically includes the following steps: S1: Obtain the compliance statement text and multimodal objective evidence data of a designated supplier; wherein the multimodal objective evidence data consists of evidence data corresponding to multiple different data sources; S2: Perform semantic parsing on the compliance statement text to generate a structured semantic vector; S3: Based on the multimodal objective evidence data, construct a spatiotemporal correlation evidence chain related to the compliance statement text; S4: Cross-validate the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; S5: Based on a preset compliance database and the risk quantification index, generate a credibility assessment result and a compliance traceability report for the designated supplier; S6: Determine whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance traceability report.

[0020] As described in step S1 above, obtain the compliance statement text and multimodal objective evidence data from the designated supplier. This involves identifying the designated supplier to be evaluated (e.g., supplier ID, registration information, or contract number) and collecting the compliance statement text provided by the supplier based on the designated supplier identifier. This includes, but is not limited to, compliance statements, scanned copies of certificates, written commitments, reports, and publicly available statements on websites. The multimodal objective evidence data originates from multiple different data sources, such as satellite remote sensing imagery, drone aerial photographs, time-series data from IoT sensors deployed in factories or logistics (temperature, energy consumption, emissions, pH levels, etc.), digital certification records from third-party organizations (e.g., EcoVadis (corporate sustainability) scores, certification certificate numbers and validity periods), public databases (customs, registration of origin), logistics trajectory data, and social media or news reports. The acquisition process includes: scanning / OCR (Optical Character Recognition) of the statement text, language recognition, and metadata extraction; collecting external evidence through APIs (Application Programming Interfaces), data stream pipelines, scheduled capture, or subscription services; simultaneously adding timestamps, geographic coordinates, and verifying data source and integrity (such as digital signatures, source authentication, hash verification, etc.) to the evidence, and performing preliminary cleaning and formatting (unifying time format, coordinate system, and unit standards) for subsequent fusion and retrieval. This step also involves permissions and privacy controls, determining which data can be stored locally on the terminal or needs to be encrypted before being transmitted to the server for processing.

[0021] As described in step S2 above, the compliance statement text undergoes semantic parsing to generate structured semantic vectors. The goal of this step is to transform unstructured or weakly structured statement texts in multiple languages ​​and formats into machine-processable structured representations. First, text preprocessing is performed, including denoising (removing template headers and footers, and duplicate expressions), language detection, segmentation and sentence division, and post-OCR error correction for scanned documents. Next, entity recognition and relation extraction are performed, automatically identifying key entities and their attributes (numerical values, units, time ranges, related factories or batches) such as "carbon emissions," "renewable energy percentage," "zero wastewater discharge," and "halal production certification." Simultaneously, entity disambiguation and standardization are performed, mapping various expressions to predefined ontologies or vocabularies (e.g., transforming "carbon dioxide emissions" into the standard indicator "carbon emissions"). Subsequently, the extracted entity-attribute pairs and inter-sentence semantic relationships are encoded into vector representations (semantic vectors). Vectorization can employ methods such as pre-trained or fine-tuned language model embedding, sentence vectors, or graph embedding, and each vector is appended with a confidence score, time stamp, and source identifier. The generated structured semantic vectors support both numerical comparison and semantic retrieval and fuzzy matching, serving as the core input for subsequent comparison with objective evidence. The structured semantic vectors refer to: encoding the OCR-processed and pre-processed compliance declaration text using the cross-language pre-trained model XLM-RoBERTa; then extracting entity-attribute pairs through named entity recognition (transformer + conditional random field); and finally, vectorizing each entity-attribute pair into d-dimensional (e.g., 768-dimensional) floating-point vectors using CLS vectors or entity embeddings and performing L2 normalization. The vectors also include source confidence, timestamp, and unit encoding. These vectors are stored in a vector database for subsequent similarity retrieval.

[0022] As described in step S3 above, a spatiotemporal evidence chain related to the compliance statement text is constructed based on the multimodal objective evidence data. In this step, the system performs spatiotemporal alignment and semantic association on the collected multimodal evidence to form a traceable evidence chain. First, time synchronization (unifying time zones and correcting clock drift) and spatial registration (unifying satellite imagery, sensor locations, and logistics GPS coordinates to the same coordinate system) are performed on each data source, and sampling consistency and missing value imputation are performed on the sensor data. Then, based on the entities mentioned in the statement text (e.g., a production line, a batch of materials, or the coordinates of a factory), matching items are retrieved from the evidence set. Multimodal evidence (images, sensor curves, authentication records, transaction documents) within the same time window and spatial range are organized according to the event chain to form a temporally sequenced and traceable evidence matrix. The evidence matrix can be constructed using a spatiotemporal index structure (such as an R-tree, a spatial index structure) for storage and retrieval. This evidence chain also records metadata (source, acquisition time, acquisition method, confidence level, calibration information) to ensure auditability; and data fusion techniques (e.g., multi-sensor fusion, image-time correlation) are used to enhance the reliability and interpretability of the evidence. The spatiotemporal correlation evidence chain is composed of spatiotemporal units as basic units, where each spatiotemporal unit is an evidence set mapped by (time window × geographic grid). Orthorectification and georegistration are performed on satellite imagery (based on ground control points or digital elevation models), and NTP (Network Time Protocol) clock correction and resampling are performed on sensor data. All evidence is stored in the spatiotemporal evidence matrix or graph database according to the spatiotemporal index. Evidence fusion employs multi-source confidence weighting or Kalman filtering, and each piece of evidence records its source and calibration metadata for auditing purposes.

[0023] As described in step S4 above, the structured semantic vector is cross-validated with the spatiotemporal evidence chain to generate risk quantification indicators. First, based on the semantic vector, each claim in the statement is semantically matched with the corresponding facts in the evidence chain to calculate semantic similarity or consistency score; for numerical claims (such as emissions, energy consumption), numerical comparison and statistical testing are performed to calculate the error range and anomaly degree. Methods such as contrastive learning, similarity measurement, temporal window consistency testing, and spatial overlap assessment are used to comprehensively determine whether there are contradictions or anomalies; for image-based evidence, visual-text comparison is performed by combining image analysis and target detection (such as detecting chimney emissions, hot spots) with textual claims, and the results of different evidence dimensions are summarized through a preset or dynamically adjusted weight model to generate a local risk score for each claim, and further, an aggregation strategy is used to calculate the overall risk quantification indicators (such as compliance risk score, conflict confidence index). Risk quantification indicators include: a semantic consistency score calculated for each claim (based on the cosine similarity between the text vector and the evidence vector in the same spatiotemporal unit), numerical differences calculated as relative errors, and the overall risk score (normalized to 0-1) obtained by weighting the scores with confidence levels and then using weighted summation or Bayesian fusion. The system can set thresholds (e.g., ≥0.8 indicates high conflict) to trigger subsequent in-depth audits or manual review.

[0024] As described in step S5 above, based on the preset compliance database and the risk quantification indicators, a credibility assessment result and a compliance tracing report for the designated supplier are generated. This step maps and determines the aforementioned risk quantification output with regulatory / standard semantics. First, the system retrieves the corresponding regulatory clauses, compliance thresholds, and proof requirements for the target market or customer from the preset compliance database, and maps these regulatory clauses into computable compliance rules or judgment logic (including mandatory clauses, recommended clauses, and their priorities). Then, the risk quantification indicators are substituted into the rule engine or intelligent decision-making model (which can be rule-driven, machine learning, or reinforcement learning components), and each compliance clause is evaluated item by item according to regulatory priority and evidence weight, resulting in a compliance pass / fail / requirement of supplementary proof determination. Based on this, the system generates a structured credibility assessment result (including overall score, dimensional score, evidence support for each claim, and conflict details) and a visualized compliance tracing report. The report includes the spatiotemporal location of key evidence, evidence chain diagram, assessment conclusions, recommended rectification measures, and items requiring further manual verification. The report supports export, multi-level viewing, and retains audit logs, providing a basis for compliance decisions and subsequent legal or business actions.

[0025] As described in step S6 above, the system determines whether the designated supplier meets the supply requirements based on the credibility assessment results and the compliance traceability report. The system executes a judgment process and triggers corresponding actions based on the credibility assessment results and the compliance traceability report. Judgment rules can be fixed thresholds (e.g., overall score not lower than a certain value and all key clauses passed) or business-oriented decision trees (combining procurement strategies, risk tolerance, supply urgency, etc.). For suppliers with serious inconsistencies or high risks, the system can automatically mark them as unqualified, recommend suspending procurement, or trigger compliance blocking measures. For minor inconsistencies, a rectification notice can be generated with a rectification deadline set, and relevant items can be added to the subsequent review schedule. Human-machine collaborative review is also supported, meaning the system's judgment results can be submitted to compliance or procurement managers for review, and review opinions and final rulings can be recorded. The judgment results are written back to the supplier management system, updating supplier status, triggering contract clauses, influencing allocation strategies, and using the judgment and related evidence for feedback training of reinforcement learning models to continuously optimize the weights and thresholds of subsequent assessments. Furthermore, the entire decision-making process maintains a complete audit chain to meet compliance and legal traceability requirements, and can generate notifications and reports for relevant stakeholders or regulatory agencies.

[0026] In one embodiment, step S2, which performs semantic parsing on the compliance statement text to generate a structured semantic vector, includes: S201: encoding the multilingual compliance statement text using a cross-language pre-trained large language model to obtain an encoding result; S202: identifying and extracting target entity information related to a preset dimension from the encoding result; and S203: vectorizing the target entity information to form the structured semantic vector.

[0027] As described in step S201 above, a cross-language pre-trained large language model is used to encode the multilingual compliance declaration texts to obtain the encoding results. This maps compliance declaration texts from different languages ​​and formats to a unified high-dimensional representation space within the model, facilitating subsequent entity recognition and semantic comparison. In practice, the original text is first pre-processed, including OCR post-correction (for scanned documents), language detection and segmentation, noise filtering (removing headers and footers, and duplicate template text), and necessary format standardization (e.g., standardizing numbers, units, and dates). Then, a cross-language pre-trained large language model is used to perform word / sub-word segmentation on the processed text [such as BPE (Byte Pair Encoding) or SentencePiece (sentence fragment / sub-word segmenter)], and the resulting text is input into the model for encoding. This model typically possesses a multilingual vocabulary and positional encoding, self-attention mechanisms, and other structures, enabling it to generate context-related vector representations of different languages ​​within a single model. The encoding results can include word-by-word or sub-word hidden state vectors, sentence vectors, or paragraph-level vectors, as well as confidence metrics and attention weight information output by the model. To improve domain adaptability, the model can be fine-tuned or have domain adaptation layers (such as adapters or cue engineering) added on ESG (Environmental, Social, and Governance) domain corpora to enhance sensitivity to industry terminology, units, and regulations. For quality control, the encoding results need to undergo confidence assessment and anomaly detection (such as excessively short vectors or significantly biased attention distributions), and the source language and version information should be recorded for auditing purposes. The large language model is a Transformer encoder containing 12 self-attention modules with a hidden layer dimension of 768. After pre-training on a general corpus using XLM-RoBERTa, the model is fine-tuned on at least 10,000 industry ESG corpora using a named entity recognition task, outputting a CLS (classification token) vector as the structured semantic vector.

[0028] As described in step S202 above, target entity information related to the preset dimensions is identified and extracted from the encoding results. The aim is to identify entities and attributes in the declaration text related to the evaluation dimensions (such as carbon emissions, energy consumption, country of origin, certification status, etc.) based on the encoding results. First, an ontology or tag set (containing entity type, attribute items, units, and possible expression variations) is established according to the preset dimensions, and an annotated corpus or rule base is prepared to support the recognition task. Entity recognition can employ sequence labeling methods (such as a Transformer-based named entity recognition model + CRF decoding), or it can combine rule-based pattern matching and regular expressions to reinforce numerical or formatted information (such as dates, certificate numbers). During the recognition process, entity disambiguation and referential parsing are also required to merge different expressions of the same entity in the text into a unified identifier, and to perform unit normalization (e.g., converting "tons CO2e / year" and "kg CO2e / month" to a unified unit) and numerical range verification. For complex relationships, relation extraction must be performed to identify semantic relationships between entities (such as "this batch of raw materials comes from the Congo mining area"). The extracted target entity information should include entity type, standardized name, original text fragment, numerical value and unit, time range, confidence level, and location information (sentence or segment index). Abnormal or low-confidence items should be marked for subsequent manual review or in-depth audit.

[0029] As described in step S203 above, the target entity information is vectorized to form the structured semantic vector. The identified and standardized target entity information is converted into a computable and comparable vector representation, facilitating similarity calculation and rule matching with multimodal evidence. The vectorization method can be implemented hierarchically: for textual entities (such as authentication names and locations), entity embedding or graph embedding methods using pre-trained language models can be used to generate semantic vectors; for numerical attributes (such as emissions and energy consumption), a numerical encoding strategy is used, normalizing the original numerical values ​​and concatenating them with unit codes and time window vectors; for time and geographic information, time series embedding and geographic coordinate encoding (e.g., latitude and longitude encoding or geohashing mapping) are used respectively. Furthermore, the entity's contextual information (e.g., source document ID, paragraph context vector, extracted confidence) can be encoded as additional dimensions to form a structured semantic vector rich in metadata. After vector construction, dimensionality reduction and standardization processing (such as principal component analysis or L2 normalization) are typically performed, and an index (such as a vector database or vector index) is established to accelerate similarity retrieval. To support cross-modal comparison, semantic consistency of the vector space should be ensured (e.g., by using contrastive learning or co-embedding training to create a comparable metric space between text entity vectors and image / sensor feature vectors). Finally, the generated structured semantic vectors should be recorded with version number, generation time, and the model and parameters used for vectorization to ensure traceability and reproducibility.

[0030] In one embodiment, step S3, which constructs a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data, includes: S301: performing time synchronization and spatial registration on each piece of evidence data in the multimodal objective evidence data to obtain each registered target evidence data; S302: associating and integrating the registered target evidence data according to time series and geographic location to obtain an evidence matrix as the spatiotemporal correlation evidence chain.

[0031] As described in step S301 above, time synchronization and spatial registration are performed on each piece of evidence data in the multimodal objective evidence data to obtain each registered target evidence data. The aim is to unify evidence data from different sources, sampling frequencies, and coordinate systems to the same time reference and spatial coordinate system, so as to facilitate consistent spatiotemporal comparison and fusion. Time synchronization first standardizes the timestamps of each data source, including time zone normalization (e.g., converting all to Coordinated Universal Time), time format standardization, and correction for sensor local clock drift (using NTP calibration records or time deviation estimation based on reference events). It also resamples or interpolates (e.g., linear interpolation, spline interpolation, or Kalman filtering) to obtain a consistent time resolution for sampling inconsistencies. At the same time, it annotates data with missing values ​​or anomalous jumps and performs imputation or removal strategies. Spatial registration includes unifying various types of geographic information to the same geographic reference system, performing projection transformation and accuracy verification on the original coordinates from GPS / sensor points and logistics trajectories, and performing orthorectification, distortion correction, and georegistration (using ground control points or digital elevation models for orthorectification and image registration) on satellite or aerial imagery, and mapping its pixels to geographic coordinates. For indoor or localized scenarios, coordinate transformation and pairing are performed on the sensor coordinate system (e.g., mapping relative coordinates to factory plane coordinates). During registration, metadata for each piece of evidence is recorded and output: original source, acquisition time and corrected timestamp, registered latitude / longitude / grid cell identifier, accuracy estimate and confidence level, transformation parameters and correction log, and integrity verification information (e.g., hash value, signature). The final output, "registered evidence data for each target," includes both synchronized and registered numerical or image data, as well as the aforementioned detailed auditable metadata, laying the foundation for subsequent spatiotemporal correlation and evidence fusion.

[0032] As described in step S302 above, the registered target evidence data are associated and integrated according to time series and geospatial location to obtain an evidence matrix as the spatiotemporal associated evidence chain. Through spatiotemporal connection and data fusion mechanisms, the registered evidence is organized into a searchable and computable evidence matrix or tensor structure to compare each claim in the statement text. First, the spatiotemporal association rules are defined: the time window length (fixed window or dynamic window, such as based on event duration), spatial buffer or grid granularity (such as binning by administrative region / latitude and longitude grid / geohash) is determined, and each piece of evidence is mapped to the corresponding spatiotemporal unit accordingly. Then, the spatiotemporal connection operation is performed to aggregate multimodal evidence (image slices, sensor time series, certificate records, transport documents, etc.) that are located in the same time window and spatially overlap or in the buffer into the same record. The aggregation process employs a multi-source fusion strategy: for numerical data, statistical summarization (mean, median, variance) and time-series feature extraction are used; for images or videos, event labels and confidence scores are extracted using object detection / semantic segmentation; and for structured records, key fields are extracted and uniformly encoded. To support subsequent comparison and tracing, each unit of the evidence matrix includes: a spatiotemporal index, a list of aggregated evidence sub-items, confidence weights for each piece of evidence, the fused metric and its uncertainty estimate, and can also generate a graph-based index (with events or entities as nodes and evidence as edges) to represent causal or temporal relationships between evidence. To improve retrieval efficiency and scalability, a spatiotemporal index structure (such as R-tree, Quad-tree, Geohash, or spatiotemporal partitioning table) and vector index are used to index the evidence matrix and store it in a scalable database or graph database. At the same time, complete audit logs and data verification information are retained. The final evidence matrix is ​​the structured representation of the spatiotemporal related evidence chain, which can provide a multimodal evidence set that is accurately corresponding in time and location for the cross-validation of semantic vectors, and supports subsequent conflict detection, risk quantification, and manual review.

[0033] In one embodiment, step S4, which involves cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index, includes: S401: using a contrastive learning algorithm to calculate the semantic consistency score between the structured semantic vector and the spatiotemporal correlation evidence chain; S402: generating the risk quantification index based on the semantic consistency score.

[0034] As described in step S401 above, a contrastive learning algorithm is used to calculate the semantic consistency score between the structured semantic vector and the spatiotemporal related evidence chain. The goal is to quantify the degree of semantic and factual consistency between each semantic claim in the statement and the corresponding spatiotemporal evidence set. The generated structured semantic vector (including entity, attribute, time, and location encodings, etc.) and the evidence matrix of the obtained spatiotemporal related evidence chain (each unit is vectorized or contains multimodal features that can be mapped to vectors) are input. First, the multimodal evidence features are mapped to an embedding space shared with the semantic vector. This is typically achieved through joint embedding training or cross-modal contrastive learning (e.g., using a dual-tower structure, momentum encoder, or CLIP (Contrastive Language-Image)). Pre-training uses a contrastive learning paradigm similar to the contrastive language-image pre-training model (SimCLR, a contrastive learning framework for visual representation) to ensure that text vectors and image / temporal features are comparable in the metric space. During training, a pairwise / negative sampling strategy is used to construct positive and negative sample pairs. Contrastive loss (such as InfoNCE, a commonly used contrastive learning loss function, or contrastive loss with a temperature coefficient) is used to optimize the model, maximizing the similarity of true matching pairs and minimizing the similarity of non-matching pairs. During online computation, for each claim, evidence units that overlap with the claim in time and space are retrieved from the evidence matrix. The similarity between vectors is calculated (a combination of cosine similarity, normalized dot product, or inverse distance metric can be used), and the relative error of numerical fields and temporal consistency score (e.g., error is |observation-claim| / baseline value), as well as evidence confidence, are considered. The system considers multiple dimensions (source reliability, sensor accuracy, and image detection confidence) and integrates them into a single semantic consistency score. This integration can be achieved using weighted averaging, weighted geometric averaging, or a learning-based fusion network. Simultaneously, it provides uncertainty estimates for missing evidence or low-confidence terms, generates structured semantic vectors from a large language model, and maps multimodal evidence (satellite image features extracted and projected using ResNet50 (50 layers of residual network), sensor temporal sequences summarized and projected using Transformer / TCN, and third-party records projected using MLP (Multilayer Perceptron)) into a unified evidence vector. Offline, a multi-tower projection head is trained using contrastive learning to construct a common embedding space and build an index. For each candidate piece of evidence, three components are calculated: semantic similarity, temporal similarity, and spatial similarity. These are then weighted and summed according to preset weights to obtain the semantic consistency score. To improve interpretability, the output should also include a decomposition of the score (e.g., text matching score, image support score, numerical consistency score, and temporal / spatial overlap), and record the parameters used for calculation and the model version for auditing purposes. The contrastive learning algorithm uses a dual-tower architecture and the InfoNCE loss function. During training, it employs in-batch negative sample and hard negative sample mining, with a temperature parameter τ of 0.07 and a training batch size of no less than 128.

[0035] As described in step S402 above, the risk quantification index is generated based on the semantic consistency score. The purpose is to convert the semantic consistency assessment results into a risk quantification index that can be used for decision-making, indicating the credibility of the claims and potential compliance risks. The consistency score of each claim, along with its uncertainty estimate, evidence confidence level, and spatiotemporal coverage, are input as input. First, the consistency score is normalized and calibrated (e.g., mapped to the 0-1 range and subjected to temperature scaling, or probabilistic calibration using a calibration set). Then, according to a preset risk mapping rule or a learning-based scoring model, the consistency score is converted into a local risk score (e.g., using 1-similarity as a risk measure, or outputting risk probability through a trained regression model). Next, an aggregation strategy is used to merge the various local risk scores into a comprehensive risk quantification index at the supplier level. This can be achieved using weighted summation (weights based on regulatory priority, clause importance, or evidence source reliability), hierarchical merging (aggregating by dimension first and then synthesizing), or Bayesian fusion methods to integrate uncertainty information. The generated risk quantification metrics typically include an overall risk score (numerical scale), detailed risk breakdowns by dimension (environmental / social / governance dimensions), a ranking of top risk factors, and trigger threshold indicators (e.g., a conflict value > 0.8 triggers a deep audit). Furthermore, this step should provide interpretable output for manual review: listing key inconsistencies leading to high risk, relevant evidence links and temporal-spatial connections, and recommended actions (re-certification, audit, order suspension, etc.). Finally, the risk quantification metrics are stored along with audit logs, model versions, and input evidence to support subsequent model retraining, compliance tracing, and legal justification.

[0036] In one embodiment, step S5, which generates a credibility assessment result and a compliance traceability report for the designated supplier based on a preset compliance database and the risk quantification indicators, includes the following steps for generating the credibility assessment result: S501: Obtaining the target market region corresponding to the designated supplier; S502: Retrieving the corresponding regional ESG regulatory subset from the compliance database based on the target market region; S503: Setting the verification weight of each regulatory clause in the regional ESG regulatory subset using a preset reinforcement learning model based on the risk quantification indicators; S504: Performing a weighted evaluation on the risk quantification indicators based on the verification weights of each regulatory clause to obtain a weighted evaluation result; S505: Generating the credibility assessment result based on the weighted evaluation result.

[0037] As described in step S501 above, the target market region corresponding to the designated supplier is obtained, i.e., the target market field (e.g., country, region, customs territory, or specific trading market) is extracted from supplier files, purchase contracts, order information, or user input. If the field is empty, it can be inferred and verified through counterparty information, delivery destination, invoice header, or logistics destination. The step includes a data verification process: standardizing the extracted region names, handling composite regions (e.g., cases where both EU and member state legislation apply), and recording time windows (some regulations are only applicable during specific effective periods). Furthermore, this step also needs to consider cross-regional application scenarios (e.g., products sold to multiple countries), thus supporting the generation of a multi-target region list and priority labeling (primary market, secondary market). If necessary, geographic information services or a regulatory application judgment module can be invoked for further confirmation, such as determining applicable regulations based on HS codes, product categories, or trade terms. The output includes the final confirmed set of target market region IDs, priority ranking, judgment criteria, and audit logs.

[0038] As described in step S502 above, the corresponding regional ESG regulatory subset is retrieved from the compliance database based on the target market region. Specifically, based on the identified target market, the corresponding regulatory clause set is retrieved and extracted from the pre-built or updated compliance database to form a computable compliance rule base. This is implemented by first matching the target market region with the regulatory metadata in the compliance database (regulatory type, effective date, scope of application, clause priority, etc.), considering the hierarchical relationship of regulations (international conventions → regional regulations → national laws → industry standards) and conflict resolution strategies (if a conflict exists, the applicable clause is selected according to priority or the principle of applicable law). Then, the retrieved regulatory clauses are machine-readable, that is, the textual clauses are converted into rule-based judgment logic or mapped to comparable thresholds (e.g., "products must not contain X substance > Y ppm" is mapped to a numerical threshold), and metadata is established for each clause (risk level, mandatory / guidance, evidence requirements, and penalties). This step also handles the timeliness of the regulations (if the regulations are under revision, the version is indicated) and regional adaptability (if the regulations have other provisions for specific product categories, conditions are added).

[0039] As described in step S503 above, based on the risk quantification indicators, a preset reinforcement learning model is used to set the verification weights of each regulatory clause in the regional ESG regulatory subset. Intelligent methods are used to assign verification weights to each applicable regulatory clause to reflect the importance of each clause to compliance risk assessment in the current situation. In implementation, the risk quantification indicators (including consistency scores between various statements and corresponding evidence, uncertainty measures, evidence confidence, etc.) are used as input state information for the reinforcement learning model. The model's action space is defined as assigning or adjusting weights (continuous or discrete values) to each regulatory clause. The reward function can be set as the accuracy of posterior compliance judgment, historical audit feedback (cost of misjudgments discovered after manual review), and business decision-making costs (such as business losses caused by misjudgments). The reinforcement learning model is trained offline using labeled historical cases (historical audit conclusions, subsequent compliance results, and penalty records serve as monitoring signals), and continuously fine-tuned online by comparing with real audit results (using policy gradient or Q-learning methods). To ensure interpretability and enforceability of regulations, the model embeds a rule constraint layer in its action selection (e.g., minimum weight threshold for mandatory clauses, and the requirement that certain clause weights cannot be zero), and outputs the recommended weight and confidence interval for each clause. Furthermore, the model records the decision-making basis (input state features, recent update history) for auditing and backtracking. The reinforcement learning model employs the PPO algorithm. The state vector includes risk quantification indicators and regulatory embeddings, the action space is the regulatory clause weight vector, and the reward function is a linear combination of the difference between audit accuracy improvement and misjudgment cost.

[0040] As described in step S504 above, the risk quantification indicators are weighted according to the verification weights of each regulatory clause to obtain a weighted evaluation result. The clause-level weights are then combined with the risk quantification indicators corresponding to the declaration to calculate a weighted evaluation result reflecting the combined impact of regulatory priority and evidence consistency. Specific practices typically include: first, standardizing and calibrating the risk quantification indicators to ensure comparable scales across different dimensions (numerical error, semantic inconsistency, evidence coverage); then, determining the mapping relationship between each regulatory clause and the risk indicators, for example, if a clause focuses on "hazardous substance content," it is mapped to the relevant numerical error indicator in the declaration; then, a weighted aggregation method is used to calculate the weighted risk score at the clause level (e.g., clause risk = clause weight × corresponding risk indicator value). If a clause is associated with multiple risk indicators, they are combined based on mapping weights or a multiple regression model. Finally, the clause risks of all clauses are aggregated according to regulatory priority or regulatory hierarchy rules (e.g., hierarchical weighting or minimum / maximum rules) to obtain the supplier's weighted evaluation result in the target market. During implementation, the impact of missing or low-confidence evidence needs to be addressed (through uncertainty propagation, confidence level correction, or requiring supplementary evidence), and the interpretability output of the calculation process needs to be recorded (ranking of the contribution of each clause, sensitivity analysis). The final weighted assessment result can be a continuous compliance risk score, a graded label, or a multi-dimensional compliance vector, providing direct input for credibility assessment.

[0041] As described in step S505 above, the credibility assessment result is generated based on the weighted assessment result. According to preset mapping rules or a trained discrimination model, the weighted compliance risk score is mapped to a credibility level (e.g., "high credibility / medium credibility / low credibility" or a 0-100 scoring scale). In conjunction with uncertainty estimation, it is selected whether certain judgments should be marked as "requires supplementary documentation" or "requires manual auditing." The credibility assessment result should include an overall score, detailed scores according to regulatory clauses or ESG dimensions, a list of key inconsistencies leading to low credibility, corresponding evidence links and time-space orientations, and recommended remedial measures (e.g., requiring supplementary documentation, initiating supplier audits, suspending transactions, or a compliance correction plan). This step also generates a visual summary (e.g., radar chart, evidence chain timeline) and stores all assessment outputs, input evidence, and audit logs in a traceable database for subsequent retrieval and legal evidence collection.

[0042] In one embodiment, step S5, which generates a credibility assessment result and a compliance traceability report for the designated supplier based on a preset compliance database and the risk quantification indicators, includes the following steps for generating the compliance traceability report: S511: Obtaining the raw material procurement information, production batch data, and logistics transportation trajectory of the designated supplier; S512: Associating the compliance statement text content and multimodal objective evidence data with the raw material procurement information, production batch data, and logistics transportation trajectory of the designated supplier to obtain a visualized traceability map; S513: Generating the compliance traceability report based on the visualized traceability map.

[0043] As described in step S511 above, the raw material procurement information, production batch data, and logistics transportation trajectory of the designated supplier are obtained. This aims to collect the structured supply chain data required for building end-to-end traceability, so that claims and objective evidence can be accurately correlated along the material flow and timeline. Specifically, this includes extracting data items from Enterprise Resource Planning (ERP), Procurement Management System (MES), Warehouse Management System (WMS), logistics tracking platform, and third-party shipping / carrier APIs, such as purchase order number, supplier ID, raw material category and batch number, warehousing / outbound time, production batch number, production line / factory identifier, production / inspection timestamp, quality inspection report reference, shipping order number, transportation origin and destination latitude and longitude and time series, vehicle / voyage information, and receipt records, etc. The data acquisition process requires field mapping and standardization: a unified coding system (e.g., SKU / HS codes), unit conversion, time format and time zone alignment, and verification of key identifiers such as batch numbers and order numbers (format, checksum, source signature). Metadata (data source, retrieval time, data integrity check hash, access permissions, and privacy tags) must also be recorded. If external suppliers or cross-enterprise data sharing exist, authorization must be obtained through contractual interfaces, data sharing protocols, or intermediate trust layers (e.g., API gateways, encrypted channels), and necessary anonymization or encrypted storage must be completed. The output structure is a standardized, traceable procurement-production-transportation dataset that can be used for subsequent graph construction.

[0044] As described in step S512 above, the compliance statement text and multimodal objective evidence data are associated with the designated supplier's raw material procurement information, production batch data, and logistics transportation trajectory to obtain a visualized traceability map. Entity parsing and multi-source linking technologies are used to establish one-to-one or many-to-many relationships between the statement, evidence, and supply chain events in time and space, generating a traceability map that can be displayed and analyzed. The implementation method includes: based on the structured semantic vector and spatiotemporal evidence matrix obtained in previous steps, identifying entities involved in the statement (such as raw material name, batch number, production location, certificate number) and matching them with corresponding fields in the procurement / production / logistics data; the matching strategy includes precise identifier matching (batch number, order number), rule and regular expression matching (contract number, certificate format), and fuzzy matching based on similarity / score (name variants, transcription differences, events within approximate time windows); simultaneously, time windows and geographical buffer constraints are used to improve matching accuracy (e.g., ±24 hours centered on shipment time, location buffer 500 meters). Multimodal evidence (images, sensor time series, certificate records) is attached as attributes to graph nodes or edges, annotating the evidence type, confidence level, timestamp, and source. The graph is represented using a graph database (such as Neo4j or JanusGraph) or a knowledge graph framework. Nodes represent raw material batches, production batches, shipping events, certificates, etc., while edges represent the "supply-production-transportation-verification" relationship and carry metrics (carbon footprint, energy consumption). To enhance credibility, key evidence can be written to an immutable evidence storage layer (such as blockchain or signature logs) and referenced in the graph using the evidence storage hash. The visualized traceability graph extracts key entities and attributes from standardized procurement, production batches, logistics trajectories, compliance statements, and multimodal evidence. Based on spatiotemporal registration (Geohash / latitude and longitude grid + time window), nodes (batch, factory, transportation, certificate, evidence fragment) and edges (procurement → production → shipment → verification) are constructed in the graph database. Nodes / edges are accompanied by confidence level, carbon footprint, and evidence storage hash. The association process employs precise ID matching, rule-based matching, and fuzzy vector similarity matching, while conflict candidates are retained for manual confirmation. The front end presents the data synchronously in a map, timeline, and topology view, supporting node clicks to view evidence snapshots, filtering by time / region / confidence level, risk highlighting, and path tracing. Results can be exported as an interactive URL, PDF, or JSON, and audit logs and access controls are recorded to ensure traceability and compliance.

[0045] As described in step S513 above, the compliance traceability report is generated based on the visualized traceability map. This step integrates the structured information, evidence chain, and compliance judgment results in the map and presents them as a human-readable compliance traceability report, facilitating compliance review, procurement decisions, and legal evidence collection. The report generation first selects the display dimensions according to the assessment purpose and audience (such as procurement, compliance, and law): overall compliance summary, evidence support by raw materials / production batches / transportation segments, key inconsistencies, timeline view, geographical distribution, and carbon footprint summary, etc. Interactive diagrams are rendered using map visualization components (nodes / edges are colored according to confidence level or risk level), and evidence snapshots (satellite image slices, sensor curves, certificate scans) and clickable evidence details are embedded. The compliance judgment section will reference the credibility assessment results generated in S5, listing the applicable regulatory clauses, corresponding evidence mappings, judgment basis, and recommended actions (re-certification, in-depth audit, suspension of cooperation) item by item. The report includes metadata and audit chain (input data source, processing model version, timestamp, evidence hash), and provides exportable formats (PDF, JSON, visualization URL) and API access interfaces, supporting evidence-level download and chain auditing.

[0046] In one embodiment, the multimodal objective evidence data includes at least two of the following: satellite remote sensing image data, time-series data from IoT sensors deployed in production facilities, and digital records from a third-party certification platform.

[0047] Satellite remote sensing imagery refers to surface observation images acquired through sensors such as optical, multispectral, thermal infrared, and radar. It is used to monitor visual evidence related to compliance claims, such as thermal anomalies in factory areas, chimney emissions, land use / vegetation changes, and water anomalies. Acquisition methods may include commercial high-resolution satellite image subscriptions, open data crawling, or on-demand task scheduling. Image preprocessing typically includes radiometric correction, atmospheric correction, geometric correction, and orthorectification to eliminate sensor and atmospheric effects and accurately register pixels to the geographic coordinate system. Cloud obstruction detection, noise filtering, and time-series stitching are then performed for long-term dynamic monitoring. Image analysis may employ target detection, heatmap generation, change detection, and spectral index calculation, outputting event annotations and evidence slices with time, latitude, longitude, and confidence levels. To ensure auditability, the data pipeline must record the original image source, acquisition time, resolution, processing steps and parameters, and generate tamper-proof hash evidence for high-value evidence; when performing cross-modal comparisons, it must also provide a pixel-to-geographic entity mapping so that the image evidence can be accurately matched with the specific factory or pipeline mentioned in the statement.

[0048] IoT sensor time-series data includes high-order time series recorded by sensors installed in production lines, wastewater treatment plants, chimneys, boiler rooms, or storage facilities, such as energy consumption (current / electricity), temperature, pressure, emission concentration, wastewater pH, and chemical oxygen demand. Data acquisition is performed via industrial communication protocols or edge gateways, with sampling frequencies ranging from seconds to days. Preprocessing requires clock synchronization, missing value interpolation, noise reduction filtering (e.g., low-pass filtering, wavelet denoising), outlier detection, and calibration coefficient application (sensor calibration curves, drift compensation). The data also includes the preservation of the original and corrected time series and sensor metadata (installation location, range, calibration time, accuracy). Time-series data facilitates behavioral pattern recognition (e.g., production load curves), sudden event detection (peak emissions), energy consumption and emission measurement estimation, and can be correlated with production batch information for batch-level energy consumption / emission attribution. Data security and integrity protection are paramount, requiring the use of signatures, encryption, and access control, and the recording of sensor certificates and calibration records to support subsequent admissibility and legal proof.

[0049] The digital records of third-party certification platforms include structured or semi-structured data such as compliance certificates, scoring reports, audit conclusions, certificate numbers, and validity periods issued by authoritative institutions or assessment platforms. These records can be obtained through platform API crawling, uploading electronic certificate archives, or verification links provided by suppliers. These records must be verified for authenticity (verifying certificate numbers and issuing authorities, digital signature verification, or cross-verification through the platform), and mapped to machine-readable fields (certificate type, issuance date, validity period, scope, scoring dimensions, and specific non-compliance items). During integration, the timeliness of certificates (expiration / revocation markings), certificate coverage (whether it covers a specific production line or batch), and differences in third-party assessment methodologies (differences in scoring criteria) must be addressed. These digital records provide highly credible certification evidence, but they also need to be cross-verified with real-time data to avoid misleading judgments with forged or outdated information. Therefore, in the chain of evidence, third-party records typically carry higher weight as supporting evidence, and their source, validity, and correspondence with the actual operating time window are clearly stated in the report.

[0050] Referring to Figure 3, the present invention also provides a cross-border supplier traceability assessment device, the device comprising: an acquisition module 902, used to acquire a compliance statement text and multimodal objective evidence data of a designated supplier; wherein the multimodal objective evidence data are evidence data corresponding to multiple different data sources; a parsing module 904, used to perform semantic parsing on the compliance statement text to generate a structured semantic vector; a construction module 906, used to construct a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; a verification module 908, used to cross-verify the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; a generation module 910, used to generate a credibility assessment result and a compliance traceability report for the designated supplier based on a preset compliance database and the risk quantification index; and a judgment module 912, used to judge whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance traceability report.

[0051] In one embodiment, the parsing module 904 includes: an encoding submodule, used to encode the multilingual compliance statement text using a cross-language pre-trained large language model to obtain an encoding result; an extraction submodule, used to identify and extract target entity information related to a preset dimension from the encoding result; and a vectorization submodule, used to vectorize the target entity information to form the structured semantic vector.

[0052] In one embodiment, the construction module 906 includes: a registration submodule, used to perform time synchronization and spatial registration on each piece of evidence data in the multimodal objective evidence data to obtain each registered target evidence data; and an association submodule, used to associate and integrate the registered target evidence data according to time sequence and geographic location to obtain an evidence matrix as the spatiotemporal associated evidence chain.

[0053] In one embodiment, the verification module 908 includes: a calculation submodule, used to calculate the semantic consistency score between the structured semantic vector and the spatiotemporal correlation evidence chain using a contrastive learning algorithm; and a risk quantification index generation submodule, used to generate the risk quantification index based on the semantic consistency score.

[0054] In one embodiment, the generation module 910 includes: a target market area acquisition submodule, used to acquire the target market area corresponding to the specified supplier; a calling submodule, used to call the corresponding regional ESG regulatory subset from the compliance database according to the target market area; a verification weight setting submodule, used to set the verification weight of each regulatory clause in the regional ESG regulatory subset based on the risk quantification indicator and using a preset reinforcement learning model; a weighted evaluation submodule, used to perform a weighted evaluation on the risk quantification indicator according to the verification weight of each regulatory clause to obtain a weighted evaluation result; and a credibility evaluation result generation submodule, used to generate the credibility evaluation result based on the weighted evaluation result.

[0055] In one embodiment, the generation module 910 includes: a raw material procurement information acquisition submodule, used to acquire the raw material procurement information, production batch data, and logistics transportation trajectory of the designated supplier; a visual traceability map acquisition submodule, used to associate the compliance statement text content and multimodal objective evidence data with the raw material procurement information, production batch data, and logistics transportation trajectory of the designated supplier to obtain a visual traceability map; and a compliance traceability report generation submodule, used to generate the compliance traceability report based on the visual traceability map.

[0056] In one embodiment, the multimodal objective evidence data includes at least two of the following: satellite remote sensing image data, time-series data from IoT sensors deployed in production facilities, and digital records from a third-party certification platform.

[0057] Figure 4 illustrates the internal structure of an electronic device in one embodiment. This electronic device can be a terminal or a server, specifically a computer device. As shown in Figure 4, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a cross-border supplier traceability assessment method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement a cross-border supplier traceability assessment method. Those skilled in the art will understand that the structure shown in Figure 4 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. Specific electronic devices may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0058] In one embodiment, an electronic device is provided, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the following steps: acquiring a compliance statement text from a designated supplier and multimodal objective evidence data; wherein the multimodal objective evidence data consists of evidence data corresponding to multiple different data sources; performing semantic parsing on the compliance statement text to generate a structured semantic vector; constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; generating a credibility assessment result and a compliance tracing report for the designated supplier based on a preset compliance database and the risk quantification index; and determining whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance tracing report.

[0059] By integrating multimodal objective evidence data and constructing a spatiotemporally correlated evidence chain, dynamic and comprehensive verification of supplier compliance statements is achieved. The automated process of semantic parsing and cross-validation improves assessment efficiency. Based on a pre-set compliance database, assessment results and traceability reports are generated, enabling compliance judgments to intelligently adapt to regulatory differences in different markets. This enhances the flexibility and reliability of the methodology, providing enterprises with an efficient, accurate, and traceable supplier compliance assessment tool. This significantly reduces supply chain compliance risks, builds a sustainable and transparent supply chain system, greatly improves the objectivity and accuracy of assessments, and effectively curbs false statements.

[0060] In one embodiment, a computer-readable storage medium is provided, storing a computer program that, when executed by a processor, causes the processor to perform the following steps: acquiring a compliance statement text and multimodal objective evidence data from a designated supplier; wherein the multimodal objective evidence data comprises evidence data corresponding to multiple different data sources; performing semantic parsing on the compliance statement text to generate a structured semantic vector; constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; generating a credibility assessment result and a compliance tracing report for the designated supplier based on a preset compliance database and the risk quantification index; and determining whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance tracing report.

[0061] By integrating multimodal objective evidence data and constructing a spatiotemporally correlated evidence chain, dynamic and comprehensive verification of supplier compliance statements is achieved. The automated process of semantic parsing and cross-validation improves assessment efficiency. Based on a pre-set compliance database, assessment results and traceability reports are generated, enabling compliance judgments to intelligently adapt to regulatory differences in different markets. This enhances the flexibility and reliability of the methodology, providing enterprises with an efficient, accurate, and traceable supplier compliance assessment tool. This significantly reduces supply chain compliance risks, builds a sustainable and transparent supply chain system, greatly improves the objectivity and accuracy of assessments, and effectively curbs false statements.

[0062] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0063] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0064] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for sourcing and evaluating cross-border suppliers, characterized in that, The method includes: acquiring a compliance statement text from a designated supplier and multimodal objective evidence data; wherein the multimodal objective evidence data consists of evidence data corresponding to multiple different data sources; performing semantic parsing on the compliance statement text to generate a structured semantic vector; constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; generating a credibility assessment result and a compliance tracing report for the designated supplier based on a preset compliance database and the risk quantification index; and determining whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance tracing report.

2. The cross-border supplier traceability assessment method according to claim 1, characterized in that, The step of semantically parsing the compliance statement text to generate a structured semantic vector includes: encoding the multilingual compliance statement text using a cross-language pre-trained large language model to obtain an encoding result; identifying and extracting target entity information related to a preset dimension from the encoding result; and vectorizing the target entity information to form the structured semantic vector.

3. The cross-border supplier traceability assessment method according to claim 1, characterized in that, The step of constructing a spatiotemporal evidence chain related to the compliance statement text based on the multimodal objective evidence data includes: performing time synchronization and spatial registration on each piece of evidence data in the multimodal objective evidence data to obtain each registered target evidence data; and associating and integrating the registered target evidence data according to time series and geographic location to obtain an evidence matrix as the spatiotemporal evidence chain.

4. The cross-border supplier traceability assessment method according to claim 1, characterized in that, The step of cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index includes: using a contrastive learning algorithm to calculate the semantic consistency score between the structured semantic vector and the spatiotemporal correlation evidence chain; and generating the risk quantification index based on the semantic consistency score.

5. The cross-border supplier traceability assessment method according to claim 1, characterized in that, The step of generating the credibility assessment result and compliance traceability report for the designated supplier based on the preset compliance database and the risk quantification indicators includes the following steps: obtaining the target market region corresponding to the designated supplier; retrieving the corresponding regional ESG regulatory subset from the compliance database according to the target market region; setting the verification weight of each regulatory clause in the regional ESG regulatory subset using a preset reinforcement learning model based on the risk quantification indicators; performing a weighted evaluation on the risk quantification indicators according to the verification weight of each regulatory clause to obtain a weighted evaluation result; and generating the credibility assessment result based on the weighted evaluation result.

6. The cross-border supplier traceability assessment method according to claim 1, characterized in that, The step of generating a credibility assessment result and a compliance traceability report for the designated supplier based on a preset compliance database and the risk quantification indicators includes the following steps: obtaining the designated supplier's raw material procurement information, production batch data, and logistics transportation trajectory; associating the compliance statement text content and multimodal objective evidence data with the designated supplier's raw material procurement information, production batch data, and logistics transportation trajectory to obtain a visualized traceability map; and generating the compliance traceability report based on the visualized traceability map.

7. The cross-border supplier traceability assessment method according to claim 1, characterized in that, The multimodal objective evidence data includes at least two of the following: satellite remote sensing image data, time-series data from IoT sensors deployed in production facilities, and digital records from third-party certification platforms.

8. A cross-border supplier traceability and assessment device, characterized in that, The apparatus includes: an acquisition module for acquiring a compliance statement text and multimodal objective evidence data from a designated supplier; wherein the multimodal objective evidence data consists of evidence data from multiple different data sources; a parsing module for semantically parsing the compliance statement text to generate a structured semantic vector; a construction module for constructing a spatiotemporal correlation evidence chain related to the compliance statement text based on the multimodal objective evidence data; a verification module for cross-validating the structured semantic vector with the spatiotemporal correlation evidence chain to generate a risk quantification index; a generation module for generating a credibility assessment result and a compliance tracing report for the designated supplier based on a preset compliance database and the risk quantification index; and a judgment module for judging whether the designated supplier meets the supply requirements based on the credibility assessment result and the compliance tracing report.

9. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the cross-border supplier sourcing assessment method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the cross-border supplier traceability assessment method as described in any one of claims 1 to 7.

Images