A method for constructing a supply chain event graph based on causality
Through the causal relationship-based supply chain matter map construction method, the problem of difficulty in determining management quality caused by the complexity of supply chain events is solved. Through the extraction and representation of event causal relationships, the improvement of supply chain management quality and event development direction are achieved. Trust judgment.
Patent Information
- Application Number
- CN202111076868.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-14
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-09-14
AI Technical Summary
The existing supply chain management methods are difficult to effectively construct and manage complex event causal relationships, which makes it difficult to determine and improve the quality of supply chain management.
The supply chain event map construction method based on causal relationship is adopted, and the supply chain events and their causal logical evolution relationship are extracted and represented through six processing processes: text preprocessing, causal extraction, event classification, event representation, event fusion and event storage. Use Neo4j and Mongodb databases to store core events and relationships.
Help judge the development direction and outcome possibility of supply chain events, analyze the problems of low management quality, provide improvement measures, improve supply chain management efficiency and benefits, enhance the credibility of the causal logic evolution relationship of events, and optimize database storage and retrieval.
Smart Images

Figure CN114239828B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the construction of causal relationships of supply chain events, and in particular to a method for constructing a supply chain event graph based on causal relationships. Background Art
[0002] Supply chain management is a new management model. This open management model has become the focus of many multinational companies. The research on supply chain management methods has been continuously deepened. These methods not only grasp the core competitiveness of enterprises, products and markets, but also use external resources of enterprises to quickly respond to market demand. However, due to some characteristics of the supply chain itself: user demand characteristics, dynamic characteristics, complex characteristics, cooperative characteristics, information characteristics, virtual characteristics, etc., it is difficult to determine the common management quality of enterprises at all points of the supply chain.
[0003] Events are one of the core concepts of human society, and people's social activities are often event-driven. The evolutionary laws and patterns of events that occur one after another in time and causal logic are very valuable knowledge. However, the current research objects of knowledge graphs and knowledge bases such as semantic networks are not events. The current method of constructing a graph of events is still in the exploratory stage and is immature. The general definition of a graph of events is a directed cyclic graph. The complexity of the evolutionary laws of events in the real world determines that we must adopt this complex graph structure. Summary of the invention
[0004] In order to overcome the shortcomings of the current existing technology, the present invention provides a method for constructing a supply chain event graph based on causal relationships, which extracts the subject of each event and the causal logical evolution relationship between them from the complete text data through six processing processes: text preprocessing, causal extraction, event classification, event representation, event fusion and event storage. The specific steps of the method are as follows:
[0005] A method for constructing a supply chain event graph based on causality, the steps of the construction method are as follows:
[0006] S1: text preprocessing;
[0007] S2: Causal event analysis;
[0008] S3: Event classification;
[0009] S4: event extraction;
[0010] S5: event fusion;
[0011] S6: Event storage,
[0012] In step S1, the supply chain-related text data is segmented into long sentences and references are resolved, and dependency syntactic analysis is performed on the basis of retaining causal relationships to extract the triple information of the long sentences as candidate sentences.
[0013] Step S2 identifies the causal relationship of the supply chain related text data based on the causal knowledge base, and generates a set of cause event sentences, a set of result event sentences and a set of associated words.
[0014] In step S3, the causal event sentence set extracted from a text is taken as a whole, and the category of the event is determined based on the TextCNN algorithm. A total of four categories are defined, namely real events, predicted events, regular events, and non-events.
[0015] After removing non-event data in step S4, the core elements of real events, predicted events and regular events are extracted based on the BiLSTM+CRF algorithm that introduces character+word+part of speech+dependency syntactic features, and the three main elements of subject, action and object of supply chain-related events are defined.
[0016] In step S5, word2vec similarity calculation and LDA clustering are performed on the cause and result events of supply chain related events respectively, and the transition probability of the event is calculated by combining the causal event category with the Markov chain algorithm as the event relationship weight value.
[0017] S6 uses the Neo4j graph database to store the core elements and event relationships of supply chain related events, and uses the MongoDB database to store event IDs and event source information.
[0018] Preferably, the specific steps of S1 text preprocessing are as follows:
[0019] S1.1: Segmentation of long sentences;
[0020] S1.2: Reference elimination;
[0021] S1.3: Dependency parsing,
[0022] S1.1 long sentence segmentation uses the long sentence segmentation punctuation marks and line breaks in the chapter-level supply chain-related text data as segmentation points to segment the text;
[0023] Among them, S1.2 reference resolution uses the dilated convolution algorithm to resolve the reference of demonstrative nouns in the text, and sets rules to resolve the reference of demonstrative pronouns;
[0024] Among them, S1.3 dependency syntactic analysis extracts fact triplets centered on the predicate based on the dependency features in the long sentence; retains the causal relationship, preliminarily extracts the core facts, and eliminates other redundant noise numbers in the long sentence.
[0025] Preferably, the specific steps of S2 causal event analysis are as follows:
[0026] S2.1: Establishment of causal knowledge base;
[0027] S2.2: Identification of causal relationships,
[0028] Among them, S2.1 causal knowledge base is established, and the causal knowledge base is composed of causal connection words. Causal connection words are words that can be used as causal connection components in Chinese.
[0029] Among them, S2.2 causal relationship identification, based on the causal knowledge base established in the previous step, extracts the cause event sentence set, the result event sentence set and the causal connective word set.
[0030] Preferably, the specific steps of S3 event classification are as follows:
[0031] S3.1: Event type definition;
[0032] S3.2: Imbalanced Class Handling;
[0033] S3.3: Use TextCNN algorithm to classify events.
[0034] Among them, S3.1 event type definition predefines four categories of events according to event characteristics and event authenticity: real events, predicted events, regular events, and non-events;
[0035] The S3.2 imbalanced category processing adopts the following two methods:
[0036] (1) Increase the loss function weight of the classifier model: fine-tune by setting multiple sets of different loss function weights, and compare and select the weight that obtains the best result;
[0037] (2) Data enhancement: By replacing some adjectives and nouns in the annotated event data with semantic associations and synonyms, the original small sample data is automatically generated in batches.
[0038] S3.3 uses the TextCNN algorithm to classify events and obtains the feature representation of n-grams in sentences through one-dimensional convolution.
[0039] Preferably, a real event is defined as an event in which both the cause and the result are events, and the result event has actually occurred; a predicted event is defined as an event in which both the cause and the result are events, and the result event is predicted to be likely to occur; a regular event is defined as an event in which both the cause and the result are events, and the regular event is inferred based on historical laws or expert opinions; a non-event is defined as an event in which one or neither the cause or the result is an event, or the cause and the result event do not constitute a causal relationship.
[0040] Preferably, the specific steps of extracting S4 event elements are as follows:
[0041] S4.1: Event element definition;
[0042] S4.2: character + word + part of speech + dependency feature fusion;
[0043] S4.3: Use BiLSTM+CRF to extract event elements.
[0044] Among them, S4.1 event element definition, event representation is to present the unstructured text containing event information in a structured form. According to the research object of the event graph, which is the predicate event and its relationship, considering the incompleteness of the supply chain event type, the supply chain related event elements are defined as the subject, action, and object.
[0045] Among them, S4.2 character + word + part of speech + dependency feature fusion extracts the character, word, part of speech, and dependency syntactic features of the sentence, and then concatenates and fuses them and passes them to the word embedding layer.
[0046] S4.3 uses BiLSTM+CRF to extract event elements. All scores predicted by the BiLSTM layer are fed to the CRF layer, and the CRF layer selects the label sequence with the highest prediction score as the best answer for extraction.
[0047] Preferably, the specific steps of S5 event fusion are as follows:
[0048] S5.1: Word2vec similarity calculation;
[0049] S5.2: LDA clustering;
[0050] S5.3: Markov Chain probability transfer,
[0051] Among them, S5.1 Word2vec similarity calculation uses word2vec to generate word vectors and calculate the similarity relationship between events.
[0052] Among them, S5.2LDA clustering uses LDA to mine event topic models and calculate the similarity of event subjects.
[0053] Among them, S5.3 Markov Chain probability transfer uses the Markov Chain algorithm and combines the event category weight to calculate the event transfer probability, which represents the evolution weight of the event logical relationship.
[0054] Preferably, the specific steps of S6 event storage are as follows:
[0055] S6.1: Neo4j;
[0056] S6.2: Mongodb,
[0057] Among them, S6.1 uses Neo4j to store the core elements and event relationships of supply chain related events;
[0058] S6.2 uses the MongoDB database to store supply chain related event IDs and event source information.
[0059] Beneficial effects: The present invention provides a method for constructing a supply chain event graph based on causal relationships. By constructing a supply chain event graph that describes the causal relationship between supply chain events, the causal logic evolution rules and patterns between events are revealed, which is helpful to help judge the development direction and possibility of results of supply chain events, analyze the problem of low quality performance of supply chain management, find out the reasons, and conduct supply chain management. Help each node enterprise to formulate improvement measures, complete the improvement, and ultimately achieve the efficiency and benefits of supply chain management. The present invention not only uses word2vec and LDA algorithms to calculate the similarity and cluster analysis of events, but also sets weights according to the categories of supply chain causal events, and calculates the event transfer probability based on the Markov chain algorithm, which helps users judge the credibility of the causal logic evolution relationship of events. The present invention combines the neo4j graph database with the mongodb database, which can effectively share the database storage pressure, and at the same time, its characteristics can be used to efficiently retrieve the graph. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 Constructing a flow chart for the supply chain cause-and-effect graph of the present invention;
[0061] Figure 2 It is the principle diagram of TextCNN of the present invention;
[0062] Figure 3 This is the BiLSTM+CRF principle diagram of the present invention. DETAILED DESCRIPTION
[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0064] like Figure 1 As shown, an embodiment of the present invention provides a method for constructing a supply chain event graph based on causality, comprising the following steps:
[0065] S1: text preprocessing;
[0066] S2: Causal event analysis;
[0067] S3: Event classification;
[0068] S4: event extraction;
[0069] S5: event fusion;
[0070] S6: Event storage,
[0071] Step S1 divides the supply chain related text data into long sentences and resolves references, and performs dependency syntactic analysis on the basis of retaining causal relationships, and extracts triple information of long sentences as candidate sentences.
[0072] The specific steps of S1 text preprocessing are as follows:
[0073] S1.1: Sentence segmentation: using the sentence segmentation punctuation marks and line breaks in the supply chain-related text data at the chapter level as segmentation points to segment the text;
[0074] S1.2: Co-reference elimination: using the dilated convolution algorithm to resolve the reference of demonstrative nouns in the text, and setting rules to resolve the reference of demonstrative pronouns;
[0075] S1.3: Dependency syntactic analysis: extract fact triplets centered on the predicate based on the dependency features in the long sentence; retain causal relationships, preliminarily extract core facts, and eliminate other redundant noise numbers in the long sentence.
[0076] Step S2 identifies the causal relationship of the supply chain related text data according to the causal knowledge base, and generates a set of cause event sentences, a set of result event sentences and a set of associated words.
[0077] The specific steps of S2 causal event analysis are as follows:
[0078] S2.1: Establishment of causal knowledge base. The causal knowledge base consists of causal connective words. Causal connective words are words that can be used as causal connective components in Chinese. According to the order of expression of causal events, the following two rules are set (each is divided into 4 subcategories according to the part of speech of the connective word):
[0079] Rule 1: <Cause> comes first, <Result> comes second
[0080] (1) {cause event}, <conjunction> {result event}, example: {it is raining} <therefore / therefore / so that...> {bring an umbrella}.
[0081] (2) {cause event}, <verb> {result event}, example: {it rained} <result / cause / cause...> {bring an umbrella}.
[0082] (3) <Conjunction> {cause event}, {result event}, example: <due to / because of / based on…> {it is raining}, {bring an umbrella}.
[0083] (4) <Conjunction 1> {cause event}, <Conjunction 2> {result event}, e.g. <Because / Since / Since…> {it is raining}, <So / Therefore / Then…> {bring an umbrella}.
[0084] Rule 2: <Result> comes first, <Cause> comes second
[0085] (1) {result event}, <conjunction> {cause event}, example: {bring an umbrella} <because / due to / based on…> {it is raining}.
[0086] (2) {result event}, <preposition> {cause event}, example: {bring an umbrella} <for the sake of / because of / because of…> {it is raining}.
[0087] (3) {result event}, <compound word> {cause event}, example: {bring an umbrella} <the reason is / the reasons include…> {it rained}.
[0088] (4) <Conjunction 1> {result event}, <Conjunction 2> {cause event}, e.g. <so / therefore…> {bring an umbrella}, <because / due to…> {it is raining}.
[0089] S2.2: Causal relationship identification, based on the causal knowledge base established in the previous step, extract the cause event sentence set, result event sentence set and causal connective word set respectively.
[0090] Step S3 takes the set of causal event sentences extracted from a text as a whole, and determines the category of the event based on the TextCNN algorithm. A total of four categories are defined, namely real events, predicted events, regular events, and non-events.
[0091] The specific steps for S3 event classification are as follows:
[0092] S3.1: Definition of event types. Based on the characteristics of events and the authenticity of events, four categories of events are pre-defined: real events, predicted events, regular events, and non-events, among which:
[0093] (1) Real event: Both the cause and the result are events, and the result event has actually occurred;
[0094] (2) Predicted events: Both the cause and the effect are events, and the effect event is predicted to occur;
[0095] (3) Regular events: Both the cause and the result are events, and regular events are inferred based on historical laws or expert opinions;
[0096] (4) Non-event: One or neither of the cause or effect is an event, or the cause and effect events do not constitute a causal relationship.
[0097] S3.2: Imbalanced category processing. After annotating a large amount of text, it was found that the amount of data in the two categories of predicted events and regular events was very small. When generating the training set, a serious data category imbalance problem occurred. To this end, the following two methods were used to deal with it:
[0098] (1) Increase the loss function weight of the classifier model: fine-tune by setting multiple sets of different loss function weights, and compare and select the weight that can obtain the best results;
[0099] (2) Data enhancement: By replacing some adjectives and nouns in the annotated event data with semantic associations and synonyms, the original small sample data is automatically generated in batches.
[0100] S3.3: Using the TextCNN algorithm to classify events. For event classification, the core is to extract text features. Because the causal event texts extracted by S2 are mostly short texts, the present invention uses the TextCNN algorithm, which has a strong ability to extract shallow text features, to achieve event classification. TextCNN uses one-dimensional convolution to obtain the feature representation of n-grams in sentences, such as Figure 2 As shown in the figure, the principle diagram of the TextCNN algorithm.
[0101] (1) Embedding: TextCNN uses pre-trained word vectors as the embedding layer. For all the words in the dataset, since each word can be represented as a vector, we can get a word embedding matrix. Each row in the word embedding matrix is a word vector, which can be static, that is, fixed or non-static, that is, it can be updated according to back propagation.
[0102] (2) Convolution: The word vector matrix can be regarded as an image, and a convolutional neural network is used to extract features. Since the correlation between adjacent words in a sentence is always high, a one-dimensional convolution can be used. That is, the difference between text convolution and image convolution is that convolution is performed only in one direction (vertical) of the text sequence, and the width of the convolution kernel is fixed to the dimension d of the word vector. The height is a hyperparameter and can be set. Perform a convolution operation on each possible window of the sentence word to obtain a feature map.
[0103] (3) Polling: The sizes of feature maps obtained by convolution kernels of different sizes are also different, so a pooling function is used for each feature map to make their dimensions the same.
[0104] (4)FullConnection and Softmax: The last layer is a fully connected softmax layer, which outputs the probability of each category.
[0105] After eliminating non-event type data in step S4, the core elements of real events, predicted events and regular events are extracted based on the BiLSTM+CRF algorithm that introduces character+word+part of speech+dependency syntactic features, and the three main elements of subject, action and object of supply chain-related events are mainly defined.
[0106] The specific steps of S4 event element extraction are as follows:
[0107] S4.1: Definition of event elements. Event representation is to present the unstructured text containing event information in a structured form. According to the research object of the event graph, which is the predicate event and its relationship, and considering the incompleteness of the supply chain event type, the supply chain related event elements are defined as the subject, action, and object. Among them:
[0108] 1) Subject: It refers to the main object of the event, the person or thing that is mainly described, usually represented by nouns, pronouns, numerals, gerunds, infinitives, etc.;
[0109] (2) Action: It refers to the action, state, characteristic or behavior of an event;
[0110] (3) Object: It refers to the object of the subject’s action. It follows a transitive verb or preposition. Objects include nouns, pronouns, gerunds, numerals, and infinitives.
[0111] S4.2: Character + word + part of speech + dependency feature fusion. In general element recognition tasks, characters or words are mostly used as features to enter the word embedding layer. However, according to Chinese expression habits and grammatical information, part of speech and dependency syntactic features can help locate event elements. Therefore, the characters, words, parts of speech, and dependency syntactic features of the sentence are extracted, and they are concatenated and fused before being passed to the word embedding layer to help learn more semantic feature information.
[0112] S4.3: Use BiLSTM+CRF to extract event elements. The basic idea of the bidirectional recurrent neural network (BiLSTM) is to propose that each training sequence has two LSTMs forward and backward, and both are connected to an output layer. This structure provides the output layer with complete past and future context information for each point in the input sequence. BiLSTM is followed by a softmax layer, which can output the probability of each label. However, the outputs of the softmax layer are independent of each other. Although BiLSTM has learned the context information, the outputs do not affect each other. It just selects a label with the maximum probability value at each step. This will lead to problems such as B-Subject followed by another B-Subject. CRF has a transfer feature, that is, it will consider the order between the output labels, so the BiLSTM+CRF method is used to extract event elements, and CRF is used as the output layer of BiLSTM.
[0113] like Figure 3 As shown in the figure, the schematic diagram of BiLSTM+CRF, where the output of the BiLSTM layer is the score of each label. For example, for w0, the output of the BiLSTM node is 0.9 (B-Subject), 0.1 (I-Subject), 0.07 (B-Action), 0.06 (I-Action), 0.12 (B-Object), 0.86 (I-Object) and 0.7 (O). These scores will be the input of the CRF layer. Therefore, all the scores predicted by the BiLSTM layer are fed to the CRF layer, and the CRF layer selects the label sequence with the highest prediction score as the best answer.
[0114] Step S5 performs word2vec similarity calculation and LDA clustering on the cause and result events of supply chain related events, and uses the Markov chain algorithm combined with the causal event category to calculate the transition probability of the event as the event relationship weight value.
[0115] The specific steps of S5 event fusion are as follows:
[0116] S5.1: Word2vec similarity calculation, word2vec is a word embedding tool open sourced by Google, which can generate word vectors, and the similarity between words can be well measured through word vectors. The models adopted by word2vec include the continuous bag of words model (CBOW) and the Skip-Gram model. It can be used to perform efficient training on large amounts of data to obtain word vectors. The present invention extracts the event representation of the core elements of the event, uses word2vec to generate word vectors, and calculates the similarity relationship between events;
[0117] S5.2: LDA clustering, Latent Dirichlet Allocation (LDA) is a topic model that mines potential topics from a given document. The present invention uses LDA to mine event topic models and calculate the similarity of event subjects;
[0118] S5.3: Markov Chain probability transfer, Markov chain (MC) is a random process of transition from one state to another in the state space. At each step of the Markov chain, the system can change from one state to another or maintain the current state according to the probability distribution. The change of state is called transfer, and the probability associated with different state changes is called transfer probability. The present invention uses the Markov Chain algorithm, combined with the event category weight (real event: 1, predicted event 0.5, regular event 0.7) to calculate the event transfer probability, which is used to represent the evolution weight of the logical relationship of the event.
[0119] S6 uses Neo4j graph database to store the core elements and event relationships of supply chain related events, and uses MongoDB database to store event IDs and event source information.
[0120] The specific steps of S6 event storage are as follows:
[0121] S6.1: Neo4j, Neo4j is a high-performance, NOSQL graph database that stores structured data on the network rather than in tables. The present invention uses Neo4j to store the core elements and event relationships of supply chain related events.
[0122] S6.2: Mongodb, MongoDB is a database based on distributed file storage. The biggest feature of Mongo is that it supports a very powerful query language. Its syntax is somewhat similar to an object-oriented query language. It can achieve almost all functions similar to single-table queries in relational databases, and it also supports indexing of data. Use the Mongodb database to store supply chain-related event IDs and event source information.
[0123] First, this method optimizes the problem of event classification definition by combining practical applications, and proposes the definitions of real events, predicted events, regular events and non-events, which optimizes its application value and is also the basis for effectively carrying out the next step of event extraction and event probability transfer calculation. Secondly, in the problem of event representation, the character + word + part of speech + dependency features are introduced, and the core elements of supply chain related events are extracted based on the BiLSTM + CRF algorithm. This method has good scalability and supports event extraction in various fields, and has strong compatibility. Then, in the event fusion method, Word2vec and LDA algorithms are used to calculate the similarity and topic clustering of events respectively, so as to align similar events, and Markov Chain is used to calculate the event transfer probability to represent the evolution weight of the logical relationship of supply chain related events. Finally, in the event storage method, Neo4j graph database and Mongodb database are used to store the core of supply chain related events and the causal relationship and event source between events respectively, and the two databases are linked by event id, which not only optimizes the storage structure of events, but also expands the source text of events in the graph, and improves the credibility of the supply chain causal graph.
[0124] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprising a reference structure" do not exclude the existence of other identical elements in the process, method, article or device including the elements.
[0125] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for constructing a supply chain event graph based on causality, the steps of the construction method are as follows: S1: text preprocessing; S2: Causal event analysis; S3: Event classification; S4: event extraction; S5: event fusion; S6: Event storage, The step S1 divides the supply chain related text data into long sentences, resolves references, and performs dependency syntactic analysis on the basis of retaining causal relationships, extracting triple information of long sentences as candidate sentences. The step S2 identifies the causal relationship of the supply chain related text data according to the causal knowledge base, and generates a set of cause event sentences, a set of result event sentences and a set of associated words. The step S3 takes the causal event sentence set extracted from a text as a whole, and determines the category of the event based on the TextCNN algorithm. A total of four categories are defined, namely real events, predicted events, regular events and non-events. After removing the non-event data in step S4, the core elements of real events, predicted events and regular events are extracted based on the BiLSTM+CRF algorithm that introduces the features of character+word+part of speech+dependency syntax, and the three main elements of subject, action and object of supply chain related events are mainly defined. The step S5 performs word2vec similarity calculation and LDA clustering on the cause and result events of supply chain related events respectively, and uses the Markov chain algorithm combined with the causal event category to calculate the transition probability of the event as the event relationship weight value. The S6 uses the Neo4j graph database to store the core elements and event relationships of supply chain related events, and uses the mongodb database to store event IDs and event source information.
2. A method for constructing a supply chain event graph based on causal relationships according to claim 1, characterized in that: The specific steps of S1 text preprocessing are as follows: S1.1: Segmentation of long sentences; S1.2: Reference elimination; S1.3: Dependency parsing, The S1.1 long sentence segmentation uses the long sentence segmentation punctuation marks and line breaks in the chapter-level supply chain-related text data as segmentation points to segment the text; The S1.2 reference resolution uses a dilated convolution algorithm to resolve references to demonstrative nouns in the text, and sets rules to resolve references to demonstrative pronouns; The S1.3 dependency syntactic analysis extracts fact triplets centered on the predicate according to the dependency features in the long sentence; retains the causal relationship, preliminarily extracts the core facts, and eliminates other redundant noise numbers in the long sentence.
3. According to the method for constructing a supply chain event graph based on causal relationships as described in claim 1, it is characterized in that: The specific steps of S2 causal event analysis are as follows: S2.1: Establishment of causal knowledge base; S2.2: Identification of causal relationships, The S2.1 causal knowledge base is established. The causal knowledge base is composed of a causal connection word library. The causal connection words are words that can be used as causal connection components in Chinese. The causal relationship identification in S2.2 is based on the causal knowledge base established in the previous step, and extracts the cause event sentence set, the result event sentence set and the causal connective word set respectively.
4. According to the method for constructing a supply chain event graph based on causal relationships as described in claim 1, it is characterized in that: The specific steps of S3 event classification are as follows: S3.1: Event type definition; S3.2: Imbalanced class handling; S3.3: Use TextCNN algorithm to classify events. The event type definition in S3.1 predefines four categories of events according to the event characteristics and the authenticity of the event: real events, predicted events, regular events, and non-events; The S3.2 imbalanced category processing adopts the following two methods: (1) Increase the loss function weight of the classifier model: fine-tune by setting multiple sets of different loss function weights, and compare and select the weight that obtains the best result; (2) Data enhancement: By replacing some adjectives and nouns in the annotated event data with semantic associations and synonyms, the original small sample data is automatically generated in batches. S3.3 uses the TextCNN algorithm to classify events and obtains the feature representation of n-grams in sentences through one-dimensional convolution.
5. A method for constructing a supply chain event graph based on causal relationships according to claim 4, characterized in that: The real event is defined as an event in which both the cause and the result are events, and the result event has actually occurred; the predicted event is defined as an event in which both the cause and the result are events, and the result event is predicted to be likely to occur; the regular event is defined as an event in which both the cause and the result are events, and the regular event is inferred based on historical laws or expert opinions; the non-event is defined as an event in which one or neither of the cause or the result is an event, or the cause and the result event do not constitute a causal relationship.
6. According to the method for constructing a supply chain event graph based on causal relationships as described in claim 1, it is characterized in that: The specific steps of extracting the S4 event elements are as follows: S4.1: Event element definition; S4.2: character + word + part of speech + dependency feature fusion; S4.3: Use BiLSTM+CRF to extract event elements. The definition of event elements in S4.1, event representation is to present the unstructured text containing event information in a structured form. According to the research object of the event graph, which is the predicate event and its relationship, and considering the incompleteness of the supply chain event type, the supply chain related event elements are defined as the three major elements of subject, action and object; The S4.2 character + word + part of speech + dependency feature fusion extracts the character, word, part of speech, and dependency syntactic features of the sentence, and then concatenates and fuses them and passes them into the word embedding layer. The S4.3 uses BiLSTM+CRF to extract event elements. All scores predicted by the BiLSTM layer are fed to the CRF layer, and the CRF layer selects the label sequence with the highest prediction score as the best answer for extraction.
7. A method for constructing a supply chain event graph based on causal relationships according to claim 1, characterized in that: The specific steps of the S5 event fusion are as follows: S5.1: Word2vec similarity calculation; S5.2: LDA clustering; S5.3: Markov Chain probability transfer, The S5.1 Word2vec similarity calculation uses word2vec to generate word vectors and calculate the similarity relationship between events. The S5.2 LDA clustering uses LDA to mine event topic models and calculate event subject similarity. The S5.3 Markov Chain probability transfer uses the Markov Chain algorithm and combines the event category weight to calculate the event transfer probability, which represents the evolution weight of the event logical relationship.
8. A method for constructing a supply chain event graph based on causal relationships according to claim 1, characterized in that: The specific steps of S6 event storage are as follows: S6.1: Neo4j; S6.2: Mongodb, Said S6.1 uses Neo4j to store the core elements and event relationships of supply chain related events; The S6.2 uses the MongoDB database to store supply chain related event IDs and event source information.
Citation Information
Cited By
Remediation generation for situation event graphs
US20250077851A1
