Supplier relationship management method and system combined with big data analysis

Abstract

The application provides a supplier relationship management method and system combined with big data analysis, first constructs a conversation text semantic correlation network based on historical conversation text big data, then mines a supplier interaction intention set based on the conversation text semantic correlation network, then calls a natural language processing model to jointly analyze the intention set and the network, predicts a supplier relationship state, then locates an optimization improvement node according to the prediction result, and finally generates a supplier relationship dynamic management and control mechanism based on the optimization improvement node and the supplier interaction intention set and applies it to a cooperation process, so that the continuous maintenance and adjustment of the supplier relationship in the field of intelligent manufacturing production line robots can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data, in particular to a supplier relationship management method and system combined with big data analysis. BACKGROUND

[0002] The cooperation relationship between enterprises and suppliers is crucial for the stable operation and efficient output of production lines. At present, enterprises mainly rely on manual simple sorting and analysis of communication records with suppliers when managing supplier relationships. These communication records contain a large amount of conversation text information, but the traditional processing method is only to superficially classify and count the text, which is difficult to deeply mine the rich information hidden behind the text. For example, for the cooperation needs, potential problems, risk prompts and other contents expressed by the supplier in the conversation, manual analysis often cannot comprehensively and accurately capture and interpret. Moreover, the existing method lacks effective use of semantic association in conversation text, and cannot grasp the interaction logic and relationship context between suppliers and enterprises as a whole. This makes it difficult for enterprises to have scientific and systematic basis when evaluating the status of supplier relationships, to predict potential relationship changes in advance, and to quickly locate key factors and make targeted optimization and improvement when facing problems in supplier relationships, thereby affecting the smooth development of intelligent manufacturing production line robot related businesses and the long-term stable development of enterprises. SUMMARY

[0003] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a supplier relationship management method combined with big data analysis, which comprises:

[0004] A conversation text semantic association network is constructed, which is based on historical conversation text big data of suppliers and enterprises, contains multiple conversation text units and semantic association edges between conversation text units, and the association strength of the semantic association edges is determined according to the semantic association closeness between the conversation text units;

[0005] Based on the conversation text semantic association network, the supplier interaction intention in the conversation text unit is mined to obtain a set of supplier interaction intentions;

[0006] A natural language processing model is called to jointly analyze the set of supplier interaction intentions and the conversation text semantic association network, to predict the relationship status between the supplier and the enterprise, and to obtain a supplier relationship status prediction result;

[0007] According to the supplier relationship status prediction result, an optimization and improvement node in the supplier relationship is located, which is a conversation text association node that affects the relationship stability degree or leads to a potential relationship change tendency;

[0008] Based on the optimization improvement node and the supplier interaction intent set, a supplier relationship dynamic management mechanism is generated, and the supplier relationship dynamic management mechanism is applied to a supplier cooperation process.

[0009] In still another aspect, the embodiment of the present application also provides a supplier relationship management system combined with big data analysis, comprising a processor, a machine readable storage medium, the machine readable storage medium is connected with the processor, the machine readable storage medium is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the machine readable storage medium to realize the above-mentioned method.

[0010] Based on the above aspects, the embodiment of the present application describes the semantic association of the conversation text units between the supplier and the enterprise by constructing a conversation text semantic association network, based on the historical conversation text big data, mines the supplier interaction intent based on the conversation text semantic association network, obtains a set containing multiple key intents, calls a natural language processing model to jointly analyze the interaction intent set and the semantic association network, predicts the relationship state between the supplier and the enterprise, locates the optimization improvement node according to the relationship state prediction result, can quickly lock the key factors affecting the relationship, avoids blind decision, finally generates the supplier relationship dynamic management mechanism based on the optimization improvement node and the interaction intent set, and applies it to the supplier cooperation process, so that the supplier relationship can be continuously maintained and adjusted according to the actual situation, and the stability and cooperation efficiency of the supplier relationship are effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is the execution flow diagram of the supplier relationship management method combined with big data analysis provided by the embodiment of the present application.

[0012] Figure 2 is the schematic diagram of the exemplary hardware and software components of the supplier relationship management system combined with big data analysis provided by the embodiment of the present application. DETAILED DESCRIPTION

[0013] The present application will be specifically described below in combination with the drawings of the specification, Figure 1 is the flow diagram of the supplier relationship management method combined with big data analysis provided by an embodiment of the present application, and the supplier relationship management method combined with big data analysis will be described in detail below.

[0014] Step S110: a conversation text semantic association network is constructed, the conversation text semantic association network is based on the historical conversation text big data of the supplier and the enterprise, contains multiple conversation text units and the semantic association edges between the conversation text units, and the association strength of the semantic association edges is determined according to the semantic association closeness between the conversation text units.

[0015] This embodiment takes the intelligent manufacturing production line robot supplier relationship management as the application scenario. The enterprise has a long-term cooperative relationship with the robot core component supplier, the whole machine assembly supplier and the after-sales maintenance supplier. The historical conversation text big data covers procurement negotiation records, technical parameter confirmation documents, quality problem communication emails, delivery progress coordination minutes and other contents. When constructing the semantic association network of conversation text, these historical texts need to be sorted and processed first. For example, all conversation texts of X year with a certain robot core component supplier are split according to time sequence and theme content to form multiple independent conversation text units. Each unit is developed around a specific business theme, such as "servo motor procurement price negotiation", "encoder technical parameter confirmation", "quarterly delivery volume adjustment negotiation" and the like.

[0016] After determining the conversation text unit, the theme content of each unit is analyzed, and the core words representing the theme are extracted, such as "servo motor", "procurement price", "technical parameter", "delivery volume" and the like, to form a theme representation word list. For any two conversation text units, the semantic association closeness is calculated by comparing the theme representation word list. When the semantic association closeness exceeds the preset threshold, a semantic association edge is established, and the association closeness value is taken as the association strength. For example, the "servo motor procurement price negotiation" unit and the "quarterly delivery volume adjustment negotiation" unit have semantic association edges because they both involve procurement transaction related themes. The association strength is determined according to the overlap degree and semantic relevance of the theme words of the two units.

[0017] Step S111: The historical conversation text big data of the supplier and the enterprise is divided into multiple independent conversation text units according to time sequence and theme coherence. Each conversation text unit corresponds to a complete theme interaction content.

[0018] When dividing the historical conversation text big data, first, sort according to time sequence, and then split according to theme coherence. Taking the conversation records of the enterprise and the robot whole machine assembly supplier as an example, the content from March 1 to March 5 in the conversation text of X year is developed around "new type collaborative robot assembly process confirmation", which is divided into a conversation text unit. The content from March 10 to March 15 is focused on "coordination of delivery time of the first batch of sample machines", which is divided into another conversation text unit. Each conversation text unit contains all the interaction content under the theme, including meeting minutes, correspondence, instant messaging records, etc., ensuring that the theme content in the unit is complete and coherent.

[0019] Step S112: From each conversation text unit, select the theme representation words that can uniquely identify the business theme of the unit to form a theme representation word list corresponding to each conversation text unit.

[0020] For each session text unit, core words are extracted by a semantic analysis tool. For example, in the "new collaborative robot assembly process confirmation" unit, the words "collaborative robot", "assembly process", "component tolerance", "assembly flow", "quality detection standard" and the like are extracted, which can accurately reflect the business theme of the unit. These words are sorted by importance to form a theme representation word list for the session text unit. The length of the list is determined according to the complexity of the theme to ensure that the theme can be fully and uniquely identified.

[0021] Step S113: Comparing the theme representation word lists of any two session text units, counting the number of overlapping words and the number of semantically associated words in the two theme representation word lists, and calculating the semantic association closeness between the two session text units.

[0022] Select any two session text units, such as the "new collaborative robot assembly process confirmation" unit and the "collaborative robot quality detection standard revision" unit, and obtain their theme representation word lists. Compare the words in the lists and count the number of identical words, such as "collaborative robot" and "quality detection standard", as the number of overlapping words. For non-overlapping words, such as "assembly process" and "detection flow", the semantic association degree of the two words is determined by a semantic association dictionary. If the association degree exceeds a predetermined threshold, it is counted as a semantically associated word.

[0023] Step S1131: Selecting any two session text units in the session text semantic association network, and obtaining the theme representation word lists corresponding to the two session text units, respectively, denoted as the first theme word list and the second theme word list.

[0024] Randomly select two units in the session text semantic association network, for example, the first unit is "robot servo motor procurement contract signing", and its first theme word list contains "servo motor", "procurement contract", "payment method", "warranty period", "breach of contract liability"; the second unit is "servo motor quality problem claim negotiation", and its second theme word list contains "servo motor", "quality problem", "claim amount", "warranty period", "repair scheme".

[0025] Step S1132: Compare each word in the first theme word list with each word in the second theme word list one by one, and record the number of identical words as the number of overlapping words.

[0026] Compare "servo motor" in the first theme word list with "servo motor" in the second theme word list, and determine that they are the same word; "warranty period" exists in both lists, and is also determined to be the same word. The remaining words such as "procurement contract" and "quality problem", "payment method" and "claim amount" are not the same. Therefore, the number of overlapping words is 2.

[0027] Step S1133: For the non-overlapping words in the first topic vocabulary list, find the semantically similar words in the second topic vocabulary list, judge the semantic correlation degree between the words through the semantic correlation dictionary, and record the number of words with a semantic correlation degree higher than the preset semantic threshold as the number of semantically correlated words.

[0028] The non-overlapping words in the first topic vocabulary list are "purchase contract", "payment method" and "breach of contract". In the second topic vocabulary list, find semantically similar words. "Purchase contract" and "claim amount" both involve contract amount related content, and the semantic correlation degree between the two is higher than the preset threshold. "Breach of contract" and "repair scheme" both involve problem handling related content, and the semantic correlation degree is also higher than the preset threshold. "Payment method" has no semantically similar words in the second topic vocabulary list. Therefore, the number of semantically correlated words is 2.

[0029] Step S1134: Calculate the sum of the number of overlapping words and the number of semantically correlated words, and record it as the total number of correlated words.

[0030] The total number of correlated words is equal to the sum of the number of overlapping words and the number of semantically correlated words, i.e. 2 plus 2 equals 4.

[0031] Step S1135: Calculate the ratio of the total number of correlated words to the total number of words in the two topic representation word lists to obtain the initial correlation coefficient.

[0032] The first topic vocabulary list has 5 words, the second topic vocabulary list has 5 words, and the total number of words is 5 plus 5, which equals 10. The initial correlation coefficient is the ratio of the total number of correlated words 4 to the total number of words 10.

[0033] Step S1136: Analyze the time interval of the two conversation text units in the historical conversation text big data. If the time interval is within the preset time range, the initial correlation coefficient is positively corrected; if it exceeds the preset time range, it is negatively corrected.

[0034] Suppose the "robot servo motor purchase contract signing" unit occurs in January X, and the "servo motor quality problem claim negotiation" unit occurs in March X, the time interval is 2 months, and the preset time range is within 3 months, so the initial correlation coefficient is positively corrected, and the correction amplitude is determined according to the proportion of the time interval to the preset range.

[0035] Step S1137: Determine the modified correlation coefficient as the semantic correlation tightness between the two conversation text units.

[0036] The correlation coefficient obtained after the forward correction is the semantic correlation tightness between the two conversation text units, which reflects the comprehensive correlation of the two in terms of theme content and time correlation.

[0037] Step S114: When the semantic correlation tightness between the two conversation text units is higher than the preset correlation threshold, a semantic correlation edge is established between the two conversation text units.

[0038] The preset correlation threshold is set according to the correlation of historical conversation text. If the semantic correlation tightness calculation result of the two conversation text units is higher than the threshold, a semantic correlation edge is added in the network for the two units. For example, the semantic correlation tightness of the above-mentioned “robot servo motor procurement contract signing” unit and the “servo motor quality problem claim negotiation” unit is higher than the preset threshold, and therefore a semantic correlation edge is established between the two.

[0039] Step S115: The numerical value of the semantic correlation tightness is directly assigned as the correlation strength of the corresponding semantic correlation edge, ensuring that the correlation strength and the semantic correlation tightness correspond one-to-one.

[0040] The correlation strength of the semantic correlation edge directly uses the calculation value of the semantic correlation tightness. For example, the semantic correlation tightness value of the above-mentioned two units is 0.6, and the correlation strength of the corresponding semantic correlation edge is 0.6, so that the correlation strength can accurately reflect the semantic correlation degree between the two units.

[0041] Step S116: All conversation text units and established semantic correlation edges are integrated to form a conversation text semantic correlation network containing unit attributes, edge connection relationships and correlation strengths.

[0042] All divided conversation text units are taken as network nodes, each node containing unit number, theme content, time information and other attributes; the established semantic correlation edges are arranged according to the connection relationship, and each edge contains starting node, ending node and correlation strength information. Through network visualization tools, these nodes and edges are integrated to form a complete conversation text semantic correlation network, which intuitively displays the semantic correlation between the conversation text units.

[0043] Step S120: Based on the conversation text semantic correlation network, the supplier interaction intention in the conversation text unit is mined to obtain a supplier interaction intention set, which contains cooperation demand intention, problem feedback intention, risk prompt intention and collaborative optimization intention.

[0044] On the basis of the constructed conversation text semantic association network, the intention of each conversation text unit is analyzed. For example, from the "quarterly delivery volume adjustment negotiation" unit, by analyzing the expressions "hope to increase the delivery volume next month" and "can the production plan be adjusted", the cooperation demand intention of the supplier is identified; from the "servo motor quality problem communication" unit, according to the expressions "part of the motor operation noise exceeds the standard" and "there is a jam phenomenon", the problem feedback intention is identified.

[0045] The interaction intentions of all conversation text units are classified and summarized to form a set of supplier interaction intentions. For conversation text units with an association relationship, the association between their intentions is comprehensively analyzed, for example, the "quality problem feedback" unit and the "repair scheme negotiation" unit both belong to problem feedback related intentions, which are classified into one category in the set.

[0046] Step S121: Extract all conversation text units from the conversation text semantic association network, arrange them in chronological order to form an ordered conversation text unit sequence.

[0047] Traverse the conversation text semantic association network, extract all conversation text units, obtain the timestamp information of each unit, and arrange them in chronological order. For example, arrange the conversation text units of a supplier in X year in the order of January, February, March, …, to form an ordered sequence from the beginning of the year to the end of the year, which facilitates subsequent analysis of the trend of interaction intentions.

[0048] Step S122: Disassemble each conversation text unit in the conversation text unit sequence into multiple independent semantic words, remove the virtual words without actual semantics, and retain the effective semantic words with business meaning.

[0049] Disassemble each conversation text unit into words, for example, "hope to increase the delivery volume of collaborative robots next month to meet the production line expansion demand" is disassembled into "hope", "increase", "next month", "collaborative robot", "delivery volume", "meet", "production line", "expand", "demand", etc. Remove the virtual words "hope" and "to" and retain the effective semantic words "increase", "next month", "collaborative robot", "delivery volume", "production line", "expand", "demand", etc.

[0050] Step S123: Combine the effective semantic words of each conversation text unit to form a semantic combination fragment that can reflect the core meaning of the unit.

[0051] Combine the effective semantic words according to semantic logic, for example, the effective semantic words of the above unit are combined into "increase the delivery volume of collaborative robots next month to meet the production line expansion demand", which can accurately reflect the core meaning of the conversation text unit, facilitating subsequent matching with typical semantic expressions.

[0052] Step S124: Establish a supplier interaction intention comparison library, which includes typical semantic expressions corresponding to cooperation demand intention, problem feedback intention, risk prompt intention and collaborative optimization intention respectively.

[0053] In the supplier interaction intention comparison library, the typical semantic expressions of cooperation demand intention include "adjust delivery volume", "negotiate purchase price", "increase order quantity", etc.; the typical semantic expressions of problem feedback intention include "product operation exception", "parameter not up to standard", "delivery delay", etc.; the typical semantic expressions of risk prompt intention include "raw material shortage may affect supply", "insufficient production capacity to meet demand", etc.; the typical semantic expressions of collaborative optimization intention include "optimize production process", "improve product performance", "cost reduction scheme", etc. These typical semantic expressions are formed by collecting historical interaction cases and are updated regularly.

[0054] Step S125: Match the semantic combination fragments of each session text unit with the typical semantic expressions in the supplier interaction intention comparison library to determine the corresponding supplier interaction intention of each session text unit.

[0055] The semantic combination fragments are compared one by one with the typical semantic expressions in the comparison library, for example, "increase next month's collaborative robot delivery volume to meet the production line expansion demand" matches the typical expression "adjust delivery volume" in cooperation demand intention, so the corresponding supplier interaction intention of this session text unit is cooperation demand intention; "part of the servo motor operation noise is out of standard, and there is a jamming phenomenon" matches the typical expression "product operation exception" in problem feedback intention, and the corresponding intention is problem feedback intention.

[0056] Step S1251: Extract the typical semantic expressions corresponding to cooperation demand intention, problem feedback intention, risk prompt intention and collaborative optimization intention respectively from the supplier interaction intention comparison library to form four typical expression sets.

[0057] The typical semantic expressions of the four types of intention are extracted from the comparison library respectively, the cooperation demand intention typical expression set includes expressions such as "adjust delivery volume", "negotiate purchase price", "increase order quantity", etc.; the problem feedback intention typical expression set includes expressions such as "product operation exception", "parameter not up to standard", "delivery delay", etc.; the risk prompt intention typical expression set includes expressions such as "raw material shortage may affect supply", "insufficient production capacity to meet demand", etc.; the collaborative optimization intention typical expression set includes expressions such as "optimize production process", "improve product performance", "cost reduction scheme", etc.

[0058] Step S1252: The semantic combination fragment of each session text unit is respectively compared with each typical semantic expression in the four typical expression sets, and four similarity values are calculated, respectively corresponding to the four types of intentions.

[0059] The semantic combination fragment "increase the next month's collaborative robot delivery volume to meet the production line expansion demand" is compared with "adjust the delivery volume" in the cooperation demand intention typical expression set, and a similarity value is obtained. Compared with the typical expressions of problem feedback intention, risk prompt intention, and collaborative optimization intention, three other similarity values are obtained respectively, and the four values correspond to the matching degree of the four types of intentions.

[0060] Step S1253: Find the maximum value in the four similarity values, and determine the intention type corresponding to the maximum value as the candidate intention.

[0061] In the four similarity values, the value corresponding to the cooperation demand intention is the largest, so the candidate intention is the cooperation demand intention.

[0062] Step S1254: Check whether the similarity value corresponding to the candidate intention is higher than the preset matching threshold value. If it is higher than the preset matching threshold value, the candidate intention is directly determined as the supplier interaction intention corresponding to the session text unit.

[0063] The preset matching threshold value is set according to the historical matching accuracy. If the similarity value of the above candidate intention is higher than the threshold value, it is determined that the supplier interaction intention corresponding to the session text unit is the cooperation demand intention.

[0064] Step S1255: If the similarity value corresponding to the candidate intention is lower than the preset matching threshold value, the associated unit of the session text unit in the session text semantic association network is extracted, and the supplier interaction intention corresponding to the associated unit is obtained.

[0065] When the matching degree of the semantic combination fragment and the typical expression is low, for example, the similarity of a certain fragment "device operation parameter needs to be further confirmed" with each type of typical expression is lower than the threshold value. At this time, the associated unit of the session text unit in the network is found. Assuming that the associated unit is "collaborative robot technical parameter confirmation", the corresponding interaction intention is the cooperation demand intention.

[0066] Step S1256: Count the intention type with the highest frequency of occurrence in the associated unit, and take the intention type as the auxiliary reference intention.

[0067] If the session text unit has multiple associated units, respectively corresponding to the cooperation demand intention, the collaborative optimization intention, and the cooperation demand intention, the intention type with the highest frequency of occurrence is the cooperation demand intention, which is taken as the auxiliary reference intention.

[0068] Step S1257: The semantic consistency of the candidate intent and the auxiliary reference intent is determined by the natural language processing model. If the consistency is higher than the preset consistency threshold, the auxiliary reference intent is determined as the supplier interaction intent corresponding to the session text unit. If the consistency is lower than the preset consistency threshold, the disassembly logic of the semantic combination segment is rechecked, and the matching is performed again after modification until the corresponding supplier interaction intent is determined.

[0069] The semantic consistency of the candidate intent and the auxiliary reference intent is analyzed by the natural language processing model. If the consistency is higher, for example, both involve cooperation-related content, the auxiliary reference intent is determined as the interaction intent of the unit. If the consistency is lower, the semantic combination segment is re-disassembled, such as modifying "device operating parameters need to be further confirmed" to "collaborative robot operating parameters need to be further confirmed" and then matching again until the intent type is determined.

[0070] Step S126: Collect all the supplier interaction intents corresponding to the session text units, classify and summarize them by intent type, and form a supplier interaction intent set containing cooperation demand intent, problem feedback intent, risk prompt intent, and collaborative optimization intent.

[0071] All the interaction intents of the session text units are classified by type, the number and distribution of each intent are counted, and a supplier interaction intent set is formed. For example, there are 15 cooperation demand intents, 8 problem feedback intents, 5 risk prompt intents, and 10 collaborative optimization intents in the set, clearly showing the distribution of the supplier's interaction intents in different aspects.

[0072] Step S130: The natural language processing model is called to jointly analyze the supplier interaction intent set and the session text semantic association network, predict the relationship state between the supplier and the enterprise, and obtain a supplier relationship state prediction result. The supplier relationship state prediction result contains a relationship stability description and a potential relationship change tendency description.

[0073] The natural language processing model fine-tuned by the professional corpus in the field of supplier cooperation is called, and the structure data of the supplier interaction intent set and the semantic association network of the conversation text are input. The model analyzes the intent set, and the proportion of cooperation demand intent and collaborative optimization intent is counted. If the proportion is high, the relationship is relatively stable; the network structure data is analyzed, and the association strength between the conversation text units corresponding to the problem feedback intent and the risk prompt intent is observed. If the association strength is high and concentrated, there is a potential relationship change tendency. For example, the analysis result shows that the cooperation demand intent accounts for 60%, the collaborative optimization intent accounts for 20%, and the relationship stability degree is described as "the current cooperation relationship is stable overall, and the cooperation willingness of both parties is strong"; the problem feedback intent is concentrated in the product quality aspect, and the association strength is high, and the potential relationship change tendency is described as "attention should be paid to the influence of product quality problems accumulation on the cooperation relationship, and there is a relationship fluctuation risk".

[0074] Step S131: Importing professional corpus in the field of supplier cooperation to fine-tune the natural language processing model, so that the natural language processing model can recognize the business semantic expression in the supplier interaction scenario.

[0075] Collect professional corpus in the field of supplier cooperation, including procurement contract text, technical agreement document, quality objection processing record, delivery coordination email, etc. After removing duplicates and irrelevant content, the corpus is labeled, and the labeled content includes interaction scenario type (such as procurement negotiation, technical confirmation, quality problem handling, etc.) and core intent label (such as cooperation demand, problem feedback, etc.). The labeled corpus is divided into training corpus and validation corpus, a pre-trained natural language processing model is loaded, learning parameters are set, and training is performed. The model weight is adjusted through back propagation until the recognition accuracy of the interaction scenario type and the labeling accuracy of the core intent label reach the preset standard.

[0076] Step S1311: Collecting professional corpus in the field of supplier cooperation, which covers text content of cooperation negotiation conversation, product delivery communication, quality problem handling and contract clause negotiation scenarios.

[0077] In the intelligent manufacturing production line robot supplier relationship management scenario, cooperation negotiation conversation includes robot procurement price negotiation meeting record, technical parameter confirmation meeting minutes, etc.; product delivery communication covers robot parts delivery time confirmation email, whole machine assembly progress report document, delivery acceptance report, etc.; quality problem handling includes robot running fault feedback, maintenance scheme communication record, quality rectification notice, etc.; contract clause negotiation involves procurement contract supplementary agreement draft, payment condition change negotiation record, after-sales service clause confirmation text, etc. These professional corpus are collected through enterprise internal document management system, email server, meeting record database, etc. to ensure that key interaction scenarios in the whole process of supplier cooperation are covered.

[0078] Step S1312: Remove repeated text segments, garbled text segments, and text segments irrelevant to supplier cooperation from the professional corpus, and retain the effective professional corpus.

[0079] The collected professional corpus is preprocessed. Repeated text segments such as repeated contract clause segments and the same delivery notices sent multiple times are identified and deleted by a text duplication checking tool. Garbled text segments caused by format errors, such as character transposition and abnormal symbols in email content, are detected and removed by an encoding verification tool. Text irrelevant to supplier cooperation, such as internal employee daily work arrangements and communication records with other non-supplier enterprises, is manually screened and removed. For example, when processing quality problem handling related corpus, the same fault description records submitted repeatedly are deleted, the garbled parts in the maintenance log with disordered format are removed, and the clear and complete fault analysis report and solution communication content are retained to form an effective professional corpus set.

[0080] Step S1313: Label each text segment in the effective professional corpus to obtain a labeled effective professional corpus. The labeling content includes the supplier interaction scene type to which the text segment belongs and the corresponding core intent label.

[0081] For each effective professional corpus text segment, the interaction scene type is determined based on the content. For example, the "robot welding precision fault analysis record" is labeled as a quality problem handling scene, and the "robot annual procurement volume price negotiation conference minutes" is labeled as a cooperation negotiation conversation scene. At the same time, the core intent label is labeled according to the core meaning of the text segment, such as "email requesting to speed up the delivery progress of robot core components" labeled as a cooperation demand intent label, "report reflecting robot control system compatibility problem" labeled as a problem feedback intent label, "letter warning of rising prices of raw materials for parts" labeled as a risk warning intent label, and "proposal to optimize robot maintenance cycle" labeled as a collaborative optimization intent label. During the labeling process, for text segments with ambiguous expressions, the labeling content is confirmed through cross-department collaboration meetings to ensure the accuracy of the labeling.

[0082] Step S1314: Divide the labeled effective professional corpus into fine-tuning training corpus and fine-tuning verification corpus according to a preset proportion.

[0083] According to the preset proportion, such as the proportion of seven to three, the annotated effective professional corpus is divided. Randomly select the corresponding proportion of text fragments from the corpus corresponding to each interaction scene type and core intent label to form the fine-tuning training corpus and the fine-tuning verification corpus. For example, in the quality problem processing scene corpus, seven of the fault feedback forms and maintenance scheme communication records are selected as the fine-tuning training corpus, and the remaining three are selected as the fine-tuning verification corpus; in the cooperation negotiation conversation scene corpus, the price negotiation records, technical parameter confirmation records, etc. are also divided by proportion to ensure that the fine-tuning training corpus and the fine-tuning verification corpus are consistent in scene type and intent label distribution, avoiding the influence of uneven data distribution on the model fine-tuning effect.

[0084] Step S1315: Load the pre-trained natural language processing model, set the learning parameters of the natural language processing model fine-tuning, including the learning step, the training batch size, and the iteration termination condition.

[0085] Load the natural language processing model pre-trained based on the general corpus, which has basic text semantic understanding ability. According to the characteristics of the intelligent manufacturing production line robot supplier cooperation corpus, set the learning parameters: set the learning step to a value suitable for professional field corpus to ensure the stability of model parameter adjustment; the training batch size is determined according to the corpus data volume and computing resource situation to balance the training efficiency and model convergence effect; the iteration termination condition is set to stop training when the recognition accuracy of the fine-tuning verification corpus remains stable for multiple rounds and is not less than the preset standard, or terminate when the maximum iteration number is reached. For example, set the maximum iteration number to a preset number of rounds, if the scene recognition accuracy and intent label annotation accuracy of the verification corpus exceed the preset threshold for three consecutive rounds during training, terminate the iteration in advance.

[0086] Step S1316: Input the fine-tuning training corpus into the pre-trained natural language processing model, with the annotated interaction scene type and core intent label as the training target, and adjust the internal weight parameters of the natural language processing model through back propagation.

[0087] Input the divided fine-tuning training corpus into the pre-trained natural language processing model in batches, the model extracts features and analyzes semantics for each text fragment, and outputs the predicted interaction scene type and core intent label. Compare the predicted results with the annotated target values, calculate the loss value, and adjust the internal weight parameters of the model from the output layer to the input layer layer by layer through the back propagation algorithm, to strengthen the recognition ability of professional terms and business logic in the supplier cooperation scene. For example, for the training corpus "robot servo motor operation noise exceeds standard problem rectification communication record", the model continuously optimizes the recognition weight of "quality problem processing" scene type and "problem feedback" intent label through learning, and improves the judgment accuracy of similar texts.

[0088] Step S1317: After each round of training, input the fine-tuning verification corpus into the natural language processing model to test the recognition accuracy of the natural language processing model for the interactive scene type and the labeling accuracy of the core intent label.

[0089] After each round of training, input the fine-tuning verification corpus into the natural language processing model in the current training state, and the model predicts the scene type and intent label for the text segments in the verification corpus. The proportion of the number of correctly predicted scene types to the total number of scene types in the verification corpus is calculated to obtain the interactive scene type recognition accuracy. The proportion of the number of correctly predicted core intent labels to the total number of intent labels in the verification corpus is calculated to obtain the core intent label labeling accuracy. For example, after a certain round of training, 100 pieces of verification corpus related to quality problem processing are input, the model correctly identifies the scene type of 92 pieces, and the core intent label labeling is correct for 90 pieces. The scene recognition accuracy of this round is the corresponding proportion, and the intent labeling accuracy is the corresponding proportion.

[0090] Step S1318: If the recognition accuracy and labeling accuracy of the natural language processing model both meet the preset performance standard, stop fine-tuning; if the preset performance standard is not met, adjust the learning parameters and continue training until the performance of the natural language processing model meets the preset standard.

[0091] The preset performance standard is set according to actual business needs, for example, the interactive scene type recognition accuracy is not less than a preset value, and the core intent label labeling accuracy is not less than a preset value. When the verification result after a certain round of training shows that both accuracies meet or exceed the preset standard, stop fine-tuning and save the current model parameters as the fine-tuned natural language processing model; if the standard is not met, analyze the loss value trend, adjust the learning step or training batch size, for example, reduce the learning step when the loss value decreases slowly, and adjust the batch size when the model fluctuates greatly, and then continue the next round of training. Repeat this process until the model performance meets the preset standard to ensure that the model can accurately recognize the business semantic expression in the interactive scene of the intelligent manufacturing production line robot supplier.

[0092] Step S132: Integrate each supplier interaction intent in the supplier interaction intent set with the corresponding conversation text unit association information to form intent association data readable by the natural language processing model.

[0093] The identification information, timestamp, topic representation word list and other associated information of each intent in the set of supplier interaction intent are integrated with the intent type. For example, the "cooperation demand intent - increase in robot procurement quantity" is associated with the corresponding session text unit ID, interaction time, and topic words "procurement quantity, production capacity, delivery period" and other information, and is structured and processed according to the format required by the natural language processing model, and is converted into intent associated data containing text sequence, intent label, and context features, wherein the text sequence is the core content of the session text unit, the intent label is the corresponding interaction intent type, and the context features include time information and topic word features, ensuring that the model can read and understand the association between the intent and the text unit.

[0094] Step S133: Extracting the connection number of the session text unit, the association strength distribution of the semantic association edge, and the unit topic association density from the semantic association network of the session text, forming network structure data readable by the natural language processing model.

[0095] The connection number of the session text unit refers to the total number of semantic association edges established by each unit with other units in the semantic association network. For example, a session text unit about robot technology upgrade has an association with 5 other units, and the connection number is 5. The association strength distribution of the semantic association edge refers to the distribution of the association strength values of all associated edges of the unit, such as the proportion of strength values in different intervals. The unit topic association density is calculated by the overlap degree of the topic representation words of the unit and the topic representation words of the associated units. These data are converted into vector form, with the connection number as a one-dimensional feature, the association strength distribution divided into multiple dimensional features according to the interval, and the topic association density as a one-dimensional feature, collectively forming network structure data that adapts to the input format of the natural language processing model.

[0096] Step S134: Inputting the intent associated data and the network structure data into the fine-tuned natural language processing model at the same time, and performing semantic deep analysis of the intent associated data by the natural language processing model to identify the association logic between different supplier interaction intents.

[0097] After the fine-tuned natural language processing model receives the intent associated data, it converts the text sequence into a semantic vector through an embedding layer, extracts features through multiple Transformer encoder layers, and captures the context semantic relationship in the text. For example, analyzing the intent associated data of "cooperation demand intent - increase in procurement quantity" and "cooperative optimization intent - increase in production capacity", it is identified that there is a causal association logic between the two, i.e. the demand for increasing procurement quantity prompts the supplier to propose a cooperative optimization suggestion to increase production capacity; analyzing the associated data of "problem feedback intent - high robot failure rate" and "risk prompt intent - unstable quality of parts", it is identified that there is a progressive association logic between problem feedback and risk prompt.

[0098] Step S135: Topological analysis of network structure data is performed by the natural language processing model to determine the semantic association stability and association change trend between the conversation text units.

[0099] The model analyzes the number of connections, association strength distribution, and topic association density in the network structure data. If the number of connections of a conversation text unit is stable, the association strength distribution is concentrated, and the topic association density is high, it is determined that the semantic association stability is strong. If the number of connections fluctuates greatly, the association strength distribution is scattered, and the topic association density decreases, it is identified that the association change trend is weakening. For example, analyzing the network structure data of conversation text units about robot after-sales service, if the number of connections with other units remains at a high level, the association strength is concentrated in the high value interval, and the topic association density is stable, it is determined that the semantic association stability of the unit is strong. If the number of connections gradually decreases, and the association strength shifts to the low value interval, it is identified that the association change trend is weakening.

[0100] Step S136: Combine the semantic deep analysis result and the topological analysis result, and calculate the proportion of cooperation demand intention and collaborative optimization intention in all intentions to generate a relationship stability description.

[0101] Calculate the proportion of the total number of cooperation demand intention and collaborative optimization intention in the supplier interaction intention set to the total number of intentions. If the proportion is high, and the semantic deep analysis shows that there is positive association logic between the two and other intentions, and the topological analysis shows that the semantic association stability of the corresponding conversation text unit is strong, then the relationship stability description is "the overall stability of the supplier relationship is stable, the cooperation willingness is positive, and the collaborative optimization atmosphere is good". If the proportion is medium, some association logic fluctuates, and the association stability is general, then the description is "the supplier relationship is basically stable, and there is certain cooperation optimization space". If the proportion is low, the association logic is loose and the association stability is weak, then the description is "the stability of the supplier relationship is insufficient, and the cooperation willingness needs to be improved".

[0102] Step S137: Analyze the centralized occurrence scene and association unit distribution of problem feedback intention and risk prompt intention to generate a potential relationship change trend description.

[0103] The interaction scenarios in which the problem feedback intention and the risk prompt intention are concentrated, such as quality problem processing scenarios, delivery delay communication scenarios, and the like; the distribution of the semantic units corresponding to the conversation text units of the intentions in the semantic association network is analyzed, if the units are concentrated in a certain specific theme unit (such as a robot core component quality related unit), the potential relationship change tendency is identified. For example, if the problem feedback intention and the risk prompt intention are concentrated in the quality problem processing scenario, and the associated units are mostly robot core component quality related units, the potential relationship change tendency is described as “affected by the core component quality problem, the supplier relationship has a potential tendency to develop in the direction of tension”; if they are concentrated in the delivery communication scenario, and the associated units are mostly logistics delay related units, it is described as “affected by the delivery progress, the supplier relationship has a potential tendency to decrease the cooperation efficiency”.

[0104] Step S138: integrating the relationship stability degree description and the potential relationship change tendency description to obtain a supplier relationship state prediction result.

[0105] The relationship stability degree description and the potential relationship change tendency description are integrated in a logical order to form a complete supplier relationship state prediction result. For example, the integrated result is “the supplier relationship is stable overall, the cooperation willingness is positive, and the collaborative optimization atmosphere is good; but affected by the core component quality problem, there is a potential tendency to develop in the direction of tension, and the quality improvement progress needs to be focused on”. The supplier relationship state prediction result presents the stable condition and possible change trend of the current supplier relationship.

[0106] Step S140: locating an optimization improvement node in the supplier relationship according to the supplier relationship state prediction result, the optimization improvement node being a conversation text association node that affects the relationship stability degree or leads to the potential relationship change tendency.

[0107] Based on the relationship stability degree description and the potential relationship change tendency description in the supplier relationship state prediction result, the corresponding conversation text units and associated units in the semantic association network of the conversation text are found, which are the key nodes that may affect the supplier relationship state. For example, according to the stability degree description of “positive cooperation willingness”, the corresponding cooperation negotiation conversation unit is found; according to the change tendency description of “core component quality problem leading to potential tension tendency”, the quality problem feedback unit and the associated component detection unit are found, which are used as candidate optimization improvement nodes.

[0108] Step S141: extracting key semantic expressions related to relationship stability from the relationship stability degree description of the supplier relationship state prediction result, the key semantic expressions being conversation text segments directly reflecting cooperation stability.

[0109] In the relationship stability degree description, text segments that explicitly reflect the cooperation state are screened out, such as "high proportion of cooperation demand intention", "consensus on collaborative optimization measures", "delivery on time rate meets the agreement" and other key semantic expressions. These expressions come from the core content in the conversation text unit, such as "both parties reach an agreement on the annual procurement plan of the robot" and "the supplier completes the delivery of customized parts on time", which directly reflects the stability of the supplier relationship.

[0110] Step S142: From the potential relationship change tendency description of the supplier relationship state prediction result, the triggering semantic expression that triggers the relationship change is extracted, which is the conversation text segment that may cause relationship fluctuation.

[0111] In the potential relationship change tendency description, text segments related to relationship changes are extracted, such as "core component failure rate exceeds standard", "delivery delay times increase", "payment condition dispute unresolved" and other triggering semantic expressions. These expressions correspond to the problem feedback content in the conversation text unit, such as "frequent operation failure of robot servo motor" and "three times of third quarter component delivery delay", which are key factors that may trigger supplier relationship fluctuation.

[0112] Step S143: Find the conversation text unit containing the key semantic expression in the conversation text semantic association network, and mark it as a stable impact unit.

[0113] Through text retrieval tools, conversation text units containing key semantic expressions are matched in the conversation text semantic association network, such as finding conversation units containing "agreement on procurement plan" for cooperation negotiation, and containing "collaborative optimization measures landing" for technical communication, etc. These units are marked as stable impact units. These units have positive semantic association with other units in the semantic association network, and play a supporting role in the stability of the supplier relationship.

[0114] Step S144: Find the conversation text unit containing the triggering semantic expression in the conversation text semantic association network, and mark it as a change triggering unit.

[0115] Similarly, through text retrieval, conversation text units containing triggering semantic expressions are located in the conversation text semantic association network, such as quality feedback units containing "frequent failures", and logistics communication units containing "delivery delay", etc. These units are marked as change triggering units. These units may be connected to multiple problem association units in the network, and are the source nodes that trigger relationship changes.

[0116] Step S145: Find all conversation text units that have semantic association edges with the stable impact unit, and mark them as stable association units.

[0117] All semantic association edges of the stable influence unit in the conversation text semantic association network are traversed to find other conversation text units connected thereto, such as the "capacity planning confirmation" unit, the "price clause refinement" unit and the like associated with the "purchase plan agreement" unit, and the units are marked as stable association units. The stable association units and the stable influence units jointly constitute the association network supporting the stable relationship.

[0118] Step S146: All conversation text units having a semantic association edge with the change trigger unit are found and marked as change association units.

[0119] The semantic association edges of the change trigger unit are traversed to find the associated conversation text units, such as the "part detection report" unit, the "repair scheme negotiation" unit, the "quality rectification requirement" unit and the like associated with the "frequent failure" unit, and the units are marked as change association units. The change association units and the change trigger unit form an association chain that may cause the relationship to fluctuate.

[0120] Step S147: The sum of the association strengths of the stable influence units, the stable association units, the change trigger units and the change association units in the conversation text semantic association network is counted, and units having a sum of association strengths higher than a preset strength threshold are screened out.

[0121] The association strength values of all semantic association edges of each marked unit are obtained, and the sum of all association strength values of the same unit is obtained as the sum of association strengths. The preset strength threshold is determined according to the average sum of association strengths of all units in the network, for example, 1.5 times the average sum of association strengths is taken as the threshold. The sum of association strengths of each unit is compared with the threshold, and units having a higher sum of association strengths are screened out, which have stronger influence in the network.

[0122] Step S1471: All semantic association edges corresponding to each stable influence unit, stable association unit, change trigger unit and change association unit in the conversation text semantic association network are obtained.

[0123] The semantic association edge list of each marked unit is extracted by the network traversal tool, and the target unit connected by each edge and the corresponding association strength value are determined. For example, the semantic association edges of the stable influence unit "purchase plan agreement" include the edge connected to the "capacity planning confirmation" unit and the edge connected to the "price clause refinement" unit, and each edge has a corresponding association strength value.

[0124] Step S1472: The association strength value of each semantic association edge is extracted, and the sum of the association strength values of all semantic association edges of each unit is summed to obtain the sum of association strengths of each unit.

[0125] For each marking unit, the association strength values of all its semantic association edges are added up, for example, the association strength values of the three association edges of a stable influence unit are and, and the sum is the addition result of the three; the association strength values of the two association edges of a change trigger unit are and, and the sum is the addition result of the two, and thus the association strength sum of each unit is obtained.

[0126] Step S1473: A preset strength threshold of the association strength sum is set, which is determined according to the average association strength sum of all units in the conversation text semantic association network.

[0127] The average value of the association strength sum of all units in the conversation text semantic association network is calculated, and the preset strength threshold is set as a certain multiple, such as 1.2 times or 1.5 times, of the average value, and the specific multiple is adjusted according to the network size and unit association density. For example, if the average association strength sum of all units in the network is, the preset strength threshold is set as 1.5 times of the value.

[0128] Step S1474: The association strength sum of each unit is compared with the preset strength threshold, and units with an association strength sum higher than the preset strength threshold are screened out to form a target unit set.

[0129] For each stable influence unit, stable association unit, change trigger unit and change association unit, its association strength sum value is extracted one by one and compared with the preset strength threshold. If the association strength sum of a unit exceeds the preset strength threshold, the unit is included in the target unit set. For example, in the conversation text semantic association network of the robot core component supplier, the association strength sum of a certain stable influence unit is, and the preset strength threshold is 1.5 times of the value, so the unit is included in the target unit set because its association strength sum is higher than the threshold; the association strength sum of another change trigger unit does not reach the threshold, so it is not included temporarily.

[0130] Step S1475: The subject representation words of each unit in the target unit set are analyzed to determine whether the subject of the unit is directly related to supplier relationship maintenance.

[0131] Extract the theme representation word list of each unit in the target unit set, and judge the theme relevance through the theme matching rule. The theme matching rule contains core theme words related to supplier relationship maintenance, such as "quality assurance", "delivery cycle", "price negotiation", "technical support", "contract performance" and the like. If the theme representation word list of the unit contains at least one core theme word, it is judged that the theme of the unit is directly related to supplier relationship maintenance; if it does not contain any core theme word, it is judged as theme irrelevant. For example, the theme representation word of a certain unit is "robot reducer life test" and "quality detection standard", because it contains the core theme word "quality detection standard", it is determined that the theme is relevant; the theme representation word of another unit is "workshop equipment layout planning", which does not contain the core theme word, and is determined to be theme irrelevant.

[0132] Step S1476: Keep the units whose themes are directly related to supplier relationship maintenance, and remove the units whose themes are irrelevant, to obtain a preliminary screening unit set.

[0133] According to the theme relevance judgment result, the target unit set is screened. The units determined to be theme related are kept to form a preliminary screening unit set; the units determined to be theme irrelevant are removed from the target unit set. For example, in the target unit set, after theme analysis, the units containing theme words such as "quality assurance agreement", "delivery delay handling" and "technical parameter adjustment" are kept, and the units involving irrelevant themes such as "workshop environment temperature control" and "logistics transportation route planning" are removed, to form a preliminary screening unit set.

[0134] Step S1477: Check the supplier interaction intention corresponding to each unit in the preliminary screening unit set. If the intention type is problem feedback intention or risk prompt intention, further verify whether the association strength sum of the unit is continuously higher than the preset strength threshold.

[0135] Iterate through each unit in the preliminary screening unit set, and query the intention type corresponding to it in the supplier interaction intention set. For the units whose intention type is problem feedback intention or risk prompt intention, the association strength sum record in different time periods is called to check whether it is higher than the preset strength threshold in continuous multiple periods. For example, the intention type corresponding to a certain unit is problem feedback intention, and the feedback content is "robot servo motor failure rate exceeds the standard". The association strength sum record of the unit in the past three months is called, and if the association strength sum of each month is higher than the preset threshold, the verification is passed; if the association strength sum of a certain month is lower than the threshold, the stability of the unit needs to be re-evaluated.

[0136] Step S1478: The units that pass the verification are finally confirmed, kept in the preliminary screening unit set, and a final optimized improvement node set is formed.

[0137] After the verification of the sum of the correlation strength, the units that pass the verification are determined as the final optimization improvement nodes and are retained in the preliminary screening unit set. The subject relevance and correlation strength data of the units that fail the verification are checked again. If the units still do not meet the requirements, the units are removed. The final optimization improvement node set contains all the units that are subject-related, meet the sum of the correlation strength, and are stable. Each node is marked with the corresponding influence type (stable influence type or change trigger type) and the associated intent type. For example, the "quality detection standard negotiation" unit is determined as a stable influence type optimization improvement node, and the "servo motor failure rate feedback" unit is a change trigger type optimization improvement node, which together form the optimization improvement node set.

[0138] Step S150: Based on the optimization improvement nodes and the supplier interaction intent set, a supplier relationship dynamic management mechanism is generated. The supplier relationship dynamic management mechanism is applied to the supplier cooperation process to realize the continuous maintenance and adjustment of the supplier relationship.

[0139] Based on the optimization improvement node set and the supplier interaction intent set, corresponding management measures are developed for different types of optimization improvement nodes. The execution subject, execution cycle, and effect evaluation standard are specified. The management measures are embedded in the supplier cooperation process and dynamically optimized to form a complete supplier relationship dynamic management mechanism. For example, relationship strengthening measures are developed for stable influence type nodes, and risk prevention and control measures are developed for change trigger type nodes. By executing these measures in the cooperation process according to the cycle, stable factors and risk factors in the supplier relationship are timely addressed to realize the continuous maintenance and adjustment of the relationship.

[0140] For example, step S151: For each optimization improvement node, the supplier interaction intent corresponding to the optimization improvement node is queried from the supplier interaction intent set to determine the influence property of the optimization improvement node.

[0141] Each node in the optimization improvement node set is analyzed one by one. Through the association relationship between the node and the supplier interaction intent set, the supplier interaction intent type corresponding to each node is found. If the intent type corresponding to the node is mainly cooperation demand intent or collaborative optimization intent, the influence property of the node is determined as stable promotion property. If the intent type corresponding to the node is mainly problem feedback intent or risk prompt intent, the influence property of the node is determined as risk triggering property. For example, the "technology parameter collaborative optimization" node corresponds to the collaborative optimization intent, and the influence property is stable promotion property. The "raw material quality fluctuation feedback" node corresponds to the problem feedback intent, and the influence property is risk triggering property.

[0142] Step S152: If the influence property of the optimization and improvement node is the stable promotion property, a relationship strengthening control measure is constructed, and the relationship strengthening control measure includes increasing the frequency of communication of the optimization and improvement node corresponding to the theme, determining the response process of the cooperation demand under the theme, and establishing a tracking mechanism for collaborative optimization.

[0143] For the optimization and improvement node with the stable promotion property, specific relationship strengthening measures are designed around the corresponding theme. Increasing the frequency of communication can be set to increase the frequency of communication of the theme video conference once a month or the frequency of communication of the on-site technical exchange once a quarter; the cooperation demand response process needs to clearly define the responsibility departments and time limit requirements of each link of demand receiving, evaluation, feedback and execution; the collaborative optimization tracking mechanism needs to establish a tracking account to record the execution progress, stage results and content that need to be adjusted of the optimization measures. For example, for the node of "robot assembly process collaborative optimization", the process optimization special communication is increased once a month, the demand response needs to be completed within two working days for preliminary evaluation, and the optimization measure execution tracking table is established to update the execution situation every week.

[0144] Step S153: If the influence property of the optimization and improvement node is the risk triggering property, a risk prevention and control control measure is constructed, and the risk prevention and control control measure includes establishing a rapid feedback channel of the optimization and improvement node corresponding to the problem, constructing a risk prompt response plan, and setting a progress monitoring node of problem solving.

[0145] For the optimization and improvement node with the risk triggering property, risk prevention and control measures are developed to reduce potential risk impact. The rapid feedback channel can set up dedicated interface personnel, emergency contact email and real-time communication group to ensure real-time transmission of problem information; the response plan needs to clearly define the risk level division standard, the disposal process of different risk levels and the responsible personnel; the progress monitoring node needs to set key check points in the problem solving process, such as problem confirmation, scheme development, measure implementation and effect verification, etc. Each node sets a time limit for completion. For example, for the node of "robot core component delivery delay risk prompt", a full-time delivery coordinator is set up, an emergency communication group for delivery risk is established, a response plan including delay warning, alternative scheme activation and loss assessment is developed, and three progress monitoring nodes are set at half of the delivery cycle, before the expected delay occurs and after the problem is solved.

[0146] Step S154: For each relationship strengthening control measure and risk prevention and control control measure, an execution subject is specified, and the execution subject is a specific department or post in the enterprise responsible for supplier relationship maintenance.

[0147] According to the business attributes and responsibility division of the control measures, the execution subject of each measure is determined. The measures involving technical exchanges in the relationship strengthening control measures can be executed by the technical research and development department, and the measures involving cooperation demand response can be executed by the procurement department; the measures involving problem feedback in the risk prevention and control control measures can be executed by the quality management department, and the measures involving preplan implementation can be executed by the supply chain management department. For example, the measure of “increasing the communication frequency of process synergy optimization” is executed by the process engineer position of the technical research and development department; the measure of “delivery delay risk response preplan implementation” is executed by the supplier management specialist position of the supply chain management department.

[0148] Step S155: Set an execution cycle for each control measure, which is determined according to the associated strength of the optimization and improvement node corresponding to the interaction intent type of the measure.

[0149] For control measures of stable promotion nature, if the corresponding intent is a long-term synergy optimization intent and the node association strength is high, the execution cycle can be set to monthly or quarterly; if the association strength is low, the execution cycle can be set to semi-annual. For control measures of risk triggering nature, if the corresponding intent is an urgent problem feedback intent and the node association strength is high, the execution cycle can be set to weekly or monthly; if it is a general risk prompt intent, the execution cycle can be set to quarterly. For example, the measure of “technical parameter synergy optimization tracking” has a high association strength and is a long-term synergy intent, so the execution cycle is set to once a month; the measure of “raw material quality risk investigation” has a medium association strength, so the execution cycle is set to once a quarter.

[0150] Step S156: Divide all control measures, execution subjects and execution cycles according to the links of the supplier cooperation process to form a control execution list by link.

[0151] The supplier cooperation process includes pre-investigation, contract signing, production delivery, quality acceptance, after-sales service and renewal evaluation, etc. According to the business scene of each control measure, it is classified into the corresponding process link. For example, the measure of “cooperation demand response process execution” is classified into the contract signing link; the measure of “delivery progress monitoring” is classified into the production delivery link; the measure of “quality problem rapid feedback” is classified into the quality acceptance link. The control execution list of each link includes all control measures, corresponding execution subjects and execution cycles in the link, and is arranged in the order of the process to form a structured list.

[0152] Step S157: Construct the effect evaluation standard of the control measures, which includes the change of the proportion of positive intents in the corresponding conversation text unit after the execution of the measures and the reduction ratio of risk intents.

[0153] The effect evaluation criteria need to quantify the implementation effect of the control measures. The change in the proportion of positive intentions refers to the difference in the proportion of the number of cooperation demand intentions and collaborative optimization intentions in the corresponding conversation text unit after the implementation of the measures to the total number of intentions before the implementation. The risk intention reduction ratio refers to the reduction ratio of the number of problem feedback intentions and risk prompt intentions after the implementation of the measures compared with the number before the implementation. For example, the effect evaluation criteria of a certain relationship strengthening measure is set to improve the proportion of positive intentions by a certain percentage, and the criteria of a certain risk prevention and control measure is set to reduce the proportion of risk intentions by a certain percentage.

[0154] Step S158: Embed the control implementation list of each link into the corresponding link of the supplier cooperation process, and trigger the implementation of the control measures according to the implementation period during the process execution.

[0155] The control implementation list is associated and configured with the supplier cooperation process through the process management system. When the process advances to a certain link, the system automatically prompts the corresponding control measures and implementation requirements of the link. After receiving the prompt, the execution subject carries out the implementation of the control measures according to the implementation period, and records the implementation in the system. For example, when the cooperation process enters the production delivery link, the system automatically triggers the implementation prompt of the "delivery progress monitoring" measure. The supplier management specialist carries out monthly monitoring and records the monitoring results.

[0156] Step S159: Collect conversation text data after the implementation of the control measures regularly, analyze the changes of intention types in the conversation text data, judge the control effect by comparing with the effect evaluation criteria, and adjust the execution subject, implementation period or specific content of the control measures according to the control effect, update the control implementation list of each link, and realize the iterative optimization of the supplier relationship dynamic control mechanism.

[0157] The conversation text data after the implementation of the control measures is collected according to a fixed period (such as every quarter), the intention type identification and statistics of the data are carried out, the change in the proportion of positive intentions and the reduction ratio of risk intentions are calculated. The statistical results are compared with the effect evaluation criteria. If the criteria are met, it is determined that the control effect is good, and the existing measures are maintained. If the criteria are not met, the reasons are analyzed, the execution subject is adjusted (such as replacing a more professional department to execute), the implementation period is shortened or lengthened (such as changing the quarterly implementation to monthly implementation), or the measure content is modified (such as increasing the communication frequency). The updated control implementation list is re-embedded into the cooperation process, and the iterative optimization of the control mechanism is completed. For example, the risk intention reduction ratio of a certain risk prevention and control measure does not meet the standard after the implementation, and it is found that the implementation period is too long, so the implementation period is adjusted from quarterly to monthly, and the control implementation list is updated.

[0158] Figure 2An exemplary hardware and software components of the vendor relationship management system 100 that can implement the idea of the present application in connection with big data analysis are shown in the schematic diagram. For example, the processor 120 can be used in the vendor relationship management system 100 in connection with big data analysis and for performing the functions in the present application.

[0159] For example, the vendor relationship management system 100 in connection with big data analysis can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. Exemplarily, the vendor relationship management system 100 in connection with big data analysis can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof. The methods of the present application can be implemented according to the program instructions. The vendor relationship management system 100 in connection with big data analysis also includes an I / O interface 150 between the computer and other input / output devices.

[0160] In addition, the present application also provides a readable storage medium, in which computer executable instructions are preset, and when a processor executes the computer executable instructions, the vendor relationship management method in connection with big data analysis is implemented.

[0161] It should be noted that, in order to simplify the description of the present application and to help the understanding of one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes incorporated into one embodiment, drawing, or description thereof.

Claims

1. A method of supplier relationship management in conjunction with big data analysis, characterized by, The method comprises: constructing a conversation text semantic association network, which is based on historical conversation text big data of a supplier and an enterprise, contains a plurality of conversation text units and semantic association edges between the conversation text units, and the association strength of the semantic association edges is determined according to the semantic association closeness between the conversation text units; based on the conversation text semantic association network, mining a supplier interaction intention in the conversation text unit to obtain a supplier interaction intention set; calling a natural language processing model to jointly analyze the supplier interaction intention set and the conversation text semantic association network, predicting a relationship state between the supplier and the enterprise to obtain a supplier relationship state prediction result; locating an optimization improvement node in the supplier relationship according to the supplier relationship state prediction result, the optimization improvement node being a conversation text association unit that affects the relationship stability degree or leads to a potential relationship change tendency; based on the optimization improvement node and the supplier interaction intention set, generating a supplier relationship dynamic management and control mechanism, and applying the supplier relationship dynamic management and control mechanism to a supplier cooperation process.

2. The supplier relationship management method in connection with big data analysis according to claim 1, characterized in that, The method comprises: dividing historical conversation text big data of a supplier and an enterprise into a plurality of independent conversation text units according to time sequence and theme coherence, each conversation text unit corresponding to a complete theme interaction content; selecting theme representation words capable of uniquely identifying the business theme of each conversation text unit from the conversation text unit to form a theme representation word list corresponding to each conversation text unit; comparing the theme representation word lists of any two conversation text units, counting the number of coincident words and the number of semantic association words in the two theme representation word lists, and calculating the semantic association closeness between the two conversation text units; when the semantic association closeness between the two conversation text units is higher than a preset association threshold, establishing a semantic association edge between the two conversation text units; directly assigning the numerical value of the semantic association closeness to the association strength of the corresponding semantic association edge, so as to ensure that the association strength corresponds to the semantic association closeness one-to-one; integrating all conversation text units and established semantic association edges to form a conversation text semantic association network containing unit attributes, edge connection relationships and association strengths. 3.The supplier relationship management method in connection with big data analysis according to claim 2, wherein, The method comprises: selecting any two conversation text units in the conversation text semantic association network, and obtaining the theme representation word lists corresponding to the two conversation text units, denoted as a first theme word list and a second theme word list; comparing each word in the first theme word list with each word in the second theme word list one by one, recording the number of completely same words as the number of coincident words; For the words not coinciding in the first subject word list, find the words with similar semantics in the second subject word list, judge the semantic correlation degree between the words through the semantic correlation dictionary, record the number of words with a semantic correlation degree higher than a preset semantic threshold as the number of semantic correlation words; Calculate the sum of the number of coinciding words and the number of semantic correlation words, denoted as the total number of correlation words; Calculate the ratio of the total number of correlation words to the total number of words in the two subject word lists to obtain an initial correlation coefficient; Analyze the time interval of the two conversation text units in the historical conversation text big data. If the time interval is within a preset time range, the initial correlation coefficient is positively corrected. If the time interval exceeds the preset time range, the initial correlation coefficient is negatively corrected. Determine the corrected correlation coefficient as the semantic correlation closeness between the two conversation text units. 4.The supplier relationship management method in connection with big data analysis of claim 1, wherein, Based on the conversation text semantic correlation network, the supplier interaction intention in the conversation text unit is mined to obtain a supplier interaction intention set, including: Extract all conversation text units from the conversation text semantic correlation network, arrange them in chronological order to form an ordered conversation text unit sequence; Disassemble each conversation text unit in the conversation text unit sequence into multiple independent semantic words, remove the virtual words without actual semantics, and retain the effective semantic words with business meaning; Perform semantic combination on the effective semantic words of each conversation text unit to form a semantic combination segment that can reflect the core meaning of the unit; Establish a supplier interaction intention reference library, which contains typical semantic expressions corresponding to cooperation demand intention, problem feedback intention, risk prompt intention, and collaborative optimization intention; Match the semantic combination segment of each conversation text unit with the typical semantic expressions in the supplier interaction intention reference library to determine the corresponding supplier interaction intention of each conversation text unit; Collect all the supplier interaction intentions corresponding to the conversation text units, classify and summarize them by intention type to form a supplier interaction intention set containing cooperation demand intention, problem feedback intention, risk prompt intention, and collaborative optimization intention.

5. The supplier relationship management method in connection with big data analysis according to claim 4, characterized in that, The matching of the semantic combination segment of each conversation text unit with the typical semantic expressions in the supplier interaction intention reference library to determine the corresponding supplier interaction intention of each conversation text unit includes: Extract the typical semantic expressions corresponding to cooperation demand intention, problem feedback intention, risk prompt intention, and collaborative optimization intention from the supplier interaction intention reference library to form a set of four types of typical expressions; Compare the semantic combination segment of each conversation text unit with each typical semantic expression in the set of four types of typical expressions for similarity to calculate four similarity values, each corresponding to one of the four types of intention; Find the maximum value among the four similarity values to determine the type of intention corresponding to the maximum value as the candidate intention; Check whether the similarity value corresponding to the candidate intention is higher than a preset matching threshold. If it is higher than the preset matching threshold, the candidate intention is directly determined as the supplier interaction intention corresponding to the conversation text unit. If the similarity value corresponding to the candidate intent is lower than the preset matching threshold, an associated unit of the session text unit in the session text semantic association network is extracted, and a supplier interaction intent corresponding to the associated unit is obtained; An intent type with the highest occurrence frequency in the associated unit is counted, and the intent type is taken as an auxiliary reference intent; The semantic consistency of the candidate intent and the auxiliary reference intent is judged by the natural language processing model, if the consistency is higher than a preset consistency threshold, the auxiliary reference intent is determined as the supplier interaction intent corresponding to the session text unit, if the consistency is lower than the preset consistency threshold, the disassembly logic of the semantic combination fragment is rechecked, and the matching is performed again after correction, until the corresponding supplier interaction intent is determined.

6. The supplier relationship management method in connection with big data analysis according to claim 1, characterized by, The natural language processing model is fine-tuned by importing professional corpus in the supplier cooperation field, so that the natural language processing model can recognize business semantic expressions in the supplier interaction scene; Each supplier interaction intent in the supplier interaction intent set is integrated with the corresponding session text unit association information to form intent association data readable by the natural language processing model; The connection number of the session text unit, the association strength distribution of the semantic association edge, and the unit theme association density are extracted from the session text semantic association network to form network structure data readable by the natural language processing model; The intent association data and the network structure data are input into the fine-tuned natural language processing model at the same time, the natural language processing model performs semantic deep analysis on the intent association data, and identifies the association logic between different supplier interaction intents; The natural language processing model performs topological analysis on the network structure data, and judges the semantic association stability and association change trend between session text units; The proportion of cooperation demand intent and collaborative optimization intent in all intents is counted in combination with the semantic deep analysis result and the topological analysis result, and a relationship stability degree description is generated; The centralized occurrence scene and the associated unit distribution of problem feedback intent and risk prompt intent are analyzed to generate a potential relationship change tendency description; The relationship stability degree description and the potential relationship change tendency description are integrated to obtain the supplier relationship state prediction result. The natural language processing model is fine-tuned by importing professional corpus in the supplier cooperation field, so that the natural language processing model can recognize business semantic expressions in the supplier interaction scene, including:

7. The supplier relationship management method in connection with big data analysis according to claim 6, characterized by, Professional corpus in the supplier cooperation field is collected, and the professional corpus covers text content in cooperation negotiation session, product delivery communication, quality problem handling and contract clause negotiation scene; Repeated text fragments, garbled text fragments and text fragments irrelevant to supplier cooperation in the professional corpus are removed, and effective professional corpus is retained; ​ annotating each text segment in the effective professional corpus to obtain an annotated effective professional corpus, the annotation content including a supplier interaction scene type to which the text segment belongs and a corresponding core intent label; dividing the annotated effective professional corpus into fine-tuning training corpus and fine-tuning verification corpus according to a preset proportion; loading the pre-trained natural language processing model, setting learning parameters for fine-tuning of the natural language processing model, including learning step, training batch size and iteration termination condition; inputting the fine-tuning training corpus into the pre-trained natural language processing model, taking the annotated interaction scene type and core intent label as the training target, and adjusting the internal weight parameters of the natural language processing model through back propagation; after completing each round of training, inputting the fine-tuning verification corpus into the natural language processing model to test the recognition accuracy of the natural language processing model for the interaction scene type and the annotation accuracy of the core intent label; if the recognition accuracy and annotation accuracy of the natural language processing model both reach the preset performance standard, stop fine-tuning; if the preset performance standard is not reached, adjust the learning parameters and continue training until the performance of the natural language processing model meets the preset standard.

8. The supplier relationship management method in connection with big data analysis according to claim 1, characterized by, The optimization improvement node in the supplier relationship is located according to the supplier relationship state prediction result, including: extracting a key semantic expression related to relationship stability from the relationship stability description in the supplier relationship state prediction result, the key semantic expression being a conversation text segment directly reflecting cooperation stability; extracting a trigger semantic expression causing relationship change from the potential relationship change tendency description in the supplier relationship state prediction result, the trigger semantic expression being a conversation text segment that may cause relationship fluctuation; finding a conversation text unit containing the key semantic expression in the conversation text semantic association network and marking it as a stable influence unit; finding a conversation text unit containing the trigger semantic expression in the conversation text semantic association network and marking it as a change trigger unit; finding all conversation text units having a semantic association edge with the stable influence unit and marking them as stable association units; finding all conversation text units having a semantic association edge with the change trigger unit and marking them as change association units; calculating the association strength sum of each unit in the stable influence unit, stable association unit, change trigger unit and change association unit in the conversation text semantic association network, and comparing the association strength sum of each unit with a preset strength threshold to screen out units with an association strength sum higher than the preset strength threshold; determining the screened units as optimization improvement nodes in the supplier relationship, each optimization improvement node corresponding to a determined influence type, the influence type being a stable influence type or a change trigger type. 9.The supplier relationship management method in connection with big data analysis of claim 8, wherein, The association strength sum of each unit in the stable influence unit, stable association unit, change trigger unit and change association unit in the conversation text semantic association network is calculated, and the association strength sum of each unit is compared with a preset strength threshold to screen out units with an association strength sum higher than the preset strength threshold, including: Obtaining all semantic association edges corresponding to each stable influence unit, stable association unit, change trigger unit and change association unit in the conversation text semantic association network; Extracting the association strength value of each semantic association edge, summing the association strength values of all semantic association edges of each unit to obtain the association strength sum of each unit; Setting a preset strength threshold of the association strength sum, the preset strength threshold being determined according to the average association strength sum of all units in the conversation text semantic association network; Comparing the association strength sum of each unit with the preset strength threshold, screening out units with an association strength sum higher than the preset strength threshold to form a target unit set; Analyzing the subject representation words of each unit in the target unit set to determine whether the subject of the unit is directly related to supplier relationship maintenance; Retaining units with a subject directly related to supplier relationship maintenance and eliminating units with an irrelevant subject to obtain a preliminary screening unit set; Checking the supplier interaction intent corresponding to each unit in the preliminary screening unit set, and if the intent type is a problem feedback intent or a risk prompt intent, further verifying whether the association strength sum of the unit is continuously higher than the preset strength threshold; Finally confirming the units that pass the verification, retaining them in the preliminary screening unit set to form a final optimized improvement node set.

10. A supplier relationship management system in conjunction with big data analytics, characterized by, The supplier relationship management system combined with big data analysis includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to realize the supplier relationship management method combined with big data analysis in any one of claims 1-9.

Images