A method for representing and constructing conditional knowledge graphs

Abstract

The present invention provides a method for representing and constructing a conditional knowledge graph, which relates to the field of industrial manufacturing; first, a new conditional knowledge graph representation method, namely, nested triple representation, is proposed. Secondly, the present invention designs a new extraction strategy, which uses a text hierarchical parsing module to extract the semantic relationship between facts and conditions; and then uses a triple extraction module to realize the extraction of triples. In addition, the present invention designs corresponding physical storage and query schemes to manage conditional knowledge, which is convenient for downstream applications. The present invention solves the problem that traditional knowledge graphs lose important conditional constraint information, and the existing conditional knowledge graphs ignore the semantic relationship between fact triples and condition triples and the logical relationship between condition triples.

Description

Technical Field

[0001] The present invention relates to the field of industrial manufacturing, and in particular to a method for representing and constructing a conditional knowledge graph. Background Art

[0002] In recent years, knowledge graphs have gradually become the core technology driving the development of smart factories, and they play a vital role in various applications such as smart search, intelligent decision-making, and knowledge reasoning. We have observed that a large number of facts in text data in some fields are stated in the form of conditional sentences. We have counted the number of conditional sentences contained in literature in various fields, and the proportion is above 20%. A complete conditional sentence generally contains two clauses: a clause expressing the condition, called a conditional clause; and a main clause expressing the result, called a conclusion sentence. The validity of the conclusion fact depends entirely on whether the condition is established.

[0003] Unstructured data accounts for a large proportion of the massive and complex data involved in industrial applications. Extracting it into a knowledge graph is a key technology for efficient, intelligent and in-depth utilization. Based on the knowledge graph, it can assist in the intelligent repair of equipment and assist workers in solving problems during operation. However, there is a large amount of conditional information in the data that supports the validity of facts. The traditional knowledge graph represents knowledge as a semantic network that describes entities and their relationships. Its construction only relies on triple information extracted from the text, without distinguishing whether it is a conditional triple or a factual triple. It loses important conditional constraint information and is insufficient to fully and accurately formalize the conditional semantics in the conditional sentence, which limits the expressive power of the knowledge graph and indirectly affects the exploration of downstream tasks such as knowledge reasoning. In the scenario of enterprise equipment operation, maintenance and inspection, the accumulated experience and knowledge of old experts are connected in series with facts with conditional information to construct the corresponding conditional knowledge graph. When a certain type of problem occurs, operation and maintenance personnel can use an application to query relevant information on the conditional knowledge graph, including which phenomena occur at the same time or any one situation can indicate a certain fault, when a similar fault occurred, what caused it, what the solution is, who is the expert who successfully solved the problem, and even directly issue instructions to the equipment to automatically solve related problems.

[0004] Recent research has begun to try to extract conditional information from sentences. They jointly extract fact triples and condition triples from sentences. Although an effective strategy is proposed for the extraction of fact triples and condition triples, some problems are also ignored. First, the conditional constraint relationship between fact triples and condition triples is ignored. The conditional triples and fact triples contained in the sentence are only extracted and listed separately, but the mutual constraint relationship between the triples cannot be judged. Secondly, the parallel logical relationship between condition triples is also ignored. Whether the specific logical relationship between condition triples is "and" or "or", the lack of parallel logical relationship makes it impossible to clearly distinguish the valid conditions that guarantee the establishment of facts.

[0005] Regarding storage, relational databases and graph databases are currently used to store knowledge. When using relational databases to store knowledge, the relational model stores wide tables, resulting in a waste of space, huge overhead for multi-hop queries, and low query efficiency. Graph databases have certain technical advantages in both knowledge graph storage and query. Summary of the invention

[0006] In view of the deficiencies of the prior art, the present invention provides a method for representing and constructing a conditional knowledge graph;

[0007] A method for representing and constructing a conditional knowledge graph, the specific steps are as follows:

[0008] Step 1: Construct a conditional knowledge graph representation method that uses nested triples to formally express the rich hierarchical semantic associations between facts and conditional triples, and between conditional triples;

[0009] The head node and the tail node in the nested triples may be entities or triples;

[0010] That is, there are three types of semantic associations between facts and conditional triples, and between conditional triples, expressed in nested triples:

[0011] Entity, relationship between entities, entity;

[0012] Triples, relations between triples, triples;

[0013] triples, the relationship between a triple and an entity, an entity or an entity, the relationship between an entity and a triple, a triple;

[0014] Step 2: Based on the conditional knowledge graph representation method in step 1, the hierarchical parsing module is used to perform hierarchical parsing on the text structure of the conditional statement to obtain a set of text units and the hierarchical semantic relationships between the text units;

[0015] A RST recursive hierarchical parse tree is constructed based on the conditional sentence to parse out the relationship between the condition and the fact, between the condition and the condition, and between the fact and the fact in the conditional sentence. At the same time, the conditional sentence is converted into multiple text units and used as the input of the triple extraction task in step 3; the text units input to step 3 are conditional units and fact units;

[0016] The conditional sentence includes a conditional clause and a concluding sentence;

[0017] Step 2.1: Given a conditional statement C: If c, f., where "If" can be replaced by other conditional introductory words; c represents a logical combination description of several conditions, and f represents a logical combination description of several facts;

[0018] Step 2.2: Perform text-level semantic analysis on the conditional statement C: If c, f., and construct a Rhetorical Structure Theory RST recursive hierarchical parsing tree, where the lowest leaf node is a set of text units {c1, ..., cm, f1, ..., fn, s1, ..., sr};

[0019] Step 2.3: Classify a set of text units {c1, ..., cm, f1, ..., fn, s1, ..., sr} into three types, ct(t∈{1, ..., m}) is a conditional unit, fj(j∈{1, ..., n} is a factual unit, and sk(k∈{1, ..., r}) is a supplementary unit;

[0020] Among them, the conditional unit and factual unit are simple sentences with complete structures and only contain a subject-predicate structure, and the supplementary unit is in the form of a noun phrase structure;

[0021] Step 3: Use the triple extraction module to extract triples from the text units parsed in step 2, namely, conditional units and factual units, convert the text units into triples, and substitute the extracted triples into the leaf nodes of the hierarchical parse tree; finally, the conditional knowledge graph representation method in step 1 represented by the hierarchical parse tree is obtained;

[0022] Step 3.1: Use the triple extraction module to extract triples from the text units parsed in step 2, i.e., conditional units and factual units, and substitute the extracted triples into the leaf node positions of the hierarchical parse tree to obtain a tree-structured conditional graph representation model; obtain the conditional knowledge graph representation method in step 1 represented by a hierarchical parse tree;

[0023] Step 3.2: Use the non-leaf nodes in the hierarchical parse tree as relationship edges, and recursively use the non-leaf nodes as relationship edges to connect the nested structures on both sides from bottom to top until recursively reaching the root node, and convert the hierarchical parse tree obtained in step 3.1 into a nested triple conditional knowledge graph representation method;

[0024] Step 4: Based on the conditional knowledge graph representation method of step 1 and the hierarchical parse tree obtained in step 3, design the corresponding physical storage and query scheme of the hierarchical parse tree to manage the conditional knowledge graph;

[0025] For storage, the tree structure of the hierarchical parsing tree obtained in step 3.1 is regarded as a directed acyclic graph, and the conditional knowledge graph is stored in the graph database; the triple of the leaf node of the hierarchical parsing tree is abstracted into a triple node, and the child nodes of the triple node are the head and tail entities in the triple; according to the parsing tree structure obtained in step 3, storage in the graph database is realized;

[0026] For query, start from the root node of the hierarchical parse tree and traverse the entire hierarchical parse tree from top to bottom. When backtracking, recursively use non-leaf nodes as relationship edges layer by layer to connect the nested structure of child nodes of non-leaf nodes until backtracking to the root node. In this way, the conditional knowledge graph representation method in step 1 can be obtained again; the conditional sentence is used as the root node of the hierarchical parse tree; for the convenience of query, the corresponding root node is saved in the entity during storage, so that when querying a certain entity, all conditional sentences associated with the node can be queried based on the root node it has.

[0027] Beneficial effects of the present invention: First, the present invention proposes a new conditional knowledge graph representation method - nested triple representation. This representation method can not only represent entity-level semantic relationships, but also represent triple-level hierarchical semantic associations. This representation is scalable and supports reasoning, and can more accurately formalize the rich semantic information of complex conditional sentences. Secondly, the present invention designs a new extraction strategy, which uses a text hierarchical parsing module to convert complex sentences into semantic hierarchical representations, so that the deep semantic relationship between facts and conditions can be extracted; a triple extraction module is used to realize the extraction of fact triples and condition triples. In addition, according to the logical representation model of the conditional knowledge graph proposed by the present invention, a corresponding physical storage scheme is designed to store and manage conditional knowledge, which is convenient for downstream applications. By using the conditional knowledge graph representation and construction method of the present invention, more structured facts and conditional triples can be extracted from massive unstructured data, thereby constructing a knowledge graph containing rich conditional knowledge, indirectly providing important data support for applications such as intelligent question answering and decision-making systems, and providing more powerful support and guarantee for major industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A flow chart of a conditional knowledge graph representation and construction method according to an embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of an example of a conditional knowledge graph representation model according to an embodiment of the present invention;

[0030] Figure 3 It is a schematic diagram of an example of an extraction model of fact and condition triples and hierarchical relations between triples according to an embodiment of the present invention;

[0031] Figure 4 Schematic diagram of a triple extraction module model according to an embodiment of the present invention;

[0032] Figure 5 A schematic diagram of an example of conditional knowledge storage according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments;

[0034] A method for representing and constructing conditional knowledge graphs, such as Figure 1 As shown, the specific steps are as follows:

[0035] Step 1: Construct a conditional knowledge graph representation method that uses nested triples to formally express the rich hierarchical semantic associations between facts and conditional triples, and between conditional triples;

[0036] The head node and the tail node in the nested triples may be entities or triples;

[0037] That is, there are three types of semantic associations between facts and conditional triples, and between conditional triples, expressed in nested triples:

[0038] Entity, relationship between entities, entity;

[0039] Triples, relations between triples, triples;

[0040] triples, the relationship between a triple and an entity, an entity or an entity, the relationship between an entity and a triple, a triple;

[0041] The first type is defined as a normal triple, and the latter two types are defined as nested triples. Figure 2 As shown, any type of triple can be further used as the head node or tail node of other triples. Taking the conditional sentence "The fuse will blow, if the load current exceeds its rated current or passes through no appliance in the circuit" as an example, according to the representation model proposed in this implementation, the following representation can be obtained:

[0042] ((The fuse,will blow,null),if,((the load current,exceeds,its ratedcurrent),or,((the load current,passes through,no appliance),in,thecircuit))).

[0043] Overall, this is a nested triple, in which the conditional subordinate conjunction "if" is used as a relationship, the head node belongs to the first type of fact triple, and the tail node is a conditional triple that ensures the validity of the fact triple, which belongs to the second type. For the tail node, its head node belongs to the first type of ordinary triple, and the tail node is a nested triple of the third type. They are connected by the parallel conjunction "or", which means that any one of the two conditional triples can make the fact triple valid. It can be seen that the nested triple logic representation model proposed in this embodiment can effectively represent the complex semantic information of conditional sentences.

[0044] Triples are the basic units of knowledge representation in traditional knowledge graph construction. The structure is concise and clear, with powerful knowledge expression capabilities. It can depict the associations between entities in the real world and is easier to be understood by humans and computers. However, due to the rich semantic associations between triples, such as conditional relationships, logical relationships, etc., it exceeds the expression capabilities of the original triples. However, for entity-level semantic relationships, the triple structure is still simple and effective. Therefore, we strive to propose a new conditional knowledge graph representation method without abandoning the advantages of triple representation - nested triple representation. Different from traditional knowledge representation, the head node and tail node in the nested triple can not only be an entity, but also a triple. In other words, a triple can be a member of another triple, thus forming a nested representation. This representation method can not only represent entity-level semantic relationships, but also triple-level semantic relationships, and can more clearly formalize the hierarchical semantic information of conditional sentences.

[0045] Step 2: Based on the conditional knowledge graph representation method in step 1, the hierarchical parsing module is used to perform hierarchical parsing on the text structure of the conditional sentence to obtain a set of text units and the hierarchical semantic relationships between the text units;

[0046] According to the conditional sentence, a RST recursive hierarchical parse tree is constructed to parse the relationship between the condition and the fact, between the condition and the condition, and between the fact and the fact in the conditional sentence. At the same time, the conditional sentence is converted into multiple text units and used as the input of the triple extraction task in step 3; the text units input to step 3 are conditional units and fact units;

[0047] A complete conditional sentence generally consists of two clauses: a clause expressing the condition, called a conditional clause; and a main clause expressing the result, called a concluding clause.

[0048] In order to capture the deep semantic relationship between facts and conditions, inspired by the Rhetorical Structure Theory (RST) in linguistics, a semantic hierarchical parse tree is constructed to fragment and structure the conditional sentences. The entire tree structure can clearly express the hierarchical relationship structure of the conditional sentences, and naturally obtain the semantic association between the facts and the local text of the condition.

[0049] Step 2.1: Given a conditional statement C: If c, f., where "If" can be replaced by other conditional introductory words; c represents a logical combination description of several conditions, and f represents a logical combination description of several facts;

[0050] Step 2.2: Perform text-level semantic analysis on the conditional statement C: If c, f., and construct a Rhetorical Structure Theory RST recursive hierarchical parsing tree, where the lowest leaf node is a set of text units {c1, ..., cm, f1, ..., fn, s1, ..., sr};

[0051] Step 2.3: Classify a set of text units {c1, ..., cm, f1, ..., fn, s1, ..., sr} into three types, ci (i ∈ {1, ..., m}) is a conditional unit, fj (j ∈ {1, ..., n}) is a factual unit, and sk (k ∈ {1, ..., r}) is a supplementary unit;

[0052] Among them, the conditional unit and factual unit are simple sentences with complete structures and only contain a subject-predicate structure, and the supplementary unit is in the form of a noun phrase structure;

[0053] In this embodiment, the supplementary unit is generally a short noun phrase, which can be used as the entity component of the triple. The fact unit and the condition unit are simple sentences and are the main input objects of the triple extraction model. Due to the simplification of the input, only a single triple can be extracted from each fact unit or condition unit, so the extraction model does not need to consider the problem that the sentence may contain overlapping triples or multiple triples.

[0054] This implementation uses a set of transformation rules designed based on grammar drive to perform pattern matching in order to produce more systematic and fine-grained output. We conducted an in-depth linguistic analysis of the conditional sentences to be processed. From a linguistic perspective, we mainly designed grammar transformation patterns for several language scenarios such as conditional adverbial clauses, parallel sentences, parallel phrases (including parallel verbs, parallel nouns, parallel adjectives) and prepositional phrases. The specific Tregex transformation pattern is shown in Table 1.

[0055] Table 1 Tregex conversion mode

[0056]

[0057]

[0058] This implementation uses the entire conditional sentence as the root node of the parse tree, and recursively applies the Tregex conversion pattern from top to bottom to simplify the sentence until a set of basic text units that no pattern can match is obtained, including conditional units, factual units, and supplementary units. In order to more clearly express the deep semantic relationship between the three types of text units, this implementation retains the prompt keywords with logical attributes in the conditional sentence (such as if / for / during / …) as the intermediate nodes of the hierarchical parse tree to connect the associated text units.

[0059] Step 3: Use the triple extraction module to extract triples from the text units parsed in step 2, namely, conditional units and factual units, convert the text units into triples, and substitute the extracted triples into the leaf nodes of the hierarchical parse tree; finally, the conditional knowledge graph representation method in step 1 represented by the hierarchical parse tree is obtained;

[0060] Step 3.1: Use the triple extraction module to extract triples from the text units parsed in step 2, i.e., conditional units and factual units, and substitute the extracted triples into the leaf node positions of the hierarchical parse tree to obtain a tree-structured conditional graph representation model; obtain the conditional knowledge graph representation method in step 1 represented by a hierarchical parse tree;

[0061] Step 3.2: Use the non-leaf nodes in the hierarchical parse tree as relationship edges, and recursively use the non-leaf nodes as relationship edges to connect the nested structures on both sides from bottom to top until recursively reaching the root node, and convert the hierarchical parse tree obtained in step 3.1 into a nested triple conditional knowledge graph representation method; This shows that the representation method in step 1 and the hierarchical parse tree in step 3.1 can be converted to each other, that is, the conditional knowledge graph representation method in step 1 is realized by using the hierarchical parse tree;

[0062] This implementation method designs a multi-input single-output model, such as Figure 4As shown in the figure, when the input is a fact unit, the extracted triple is the fact triple; when the input is a condition unit, the extracted triple is the condition triple. The encoder-decoder model is adopted, in which the encoder encoder tries two neural networks based on the bidirectional encoder representation technology Bidirectional Encoder Representations from Transformers (BERT) and the bidirectional long short-term memory network Bi-LSTM, and the decoder decoder uses the long short-term memory network LSTM for decoding. The output vector obtained after decoding will enter the feed forward neural network Feed Forward neural Network (FFN):

[0063] d i =Decoder(Encoder(w i ))

[0064] x i =FFN(d i )

[0065] Among them, w i refers to the i-th word in the conditional sentence, d t Represents the output vector of the word obtained by decoding the i-th word, and then uses the softmax function to generate each word w t The corresponding label probability distribution:

[0066]

[0067] Among them, L represents the set of all labels. We use seven labels to annotate the head entity, relation, and tail entity in the triple, in the form of "A0 / P / A1-B / I" and "O" labels. "A0", "P", and "A1" respectively indicate that the current word belongs to the head entity, relation, and tail entity. "B" represents the head of an entity or relation, "I" represents the part of an entity or relation other than the head, and "O" represents a non-entity or relation.

[0068] Step 4: Based on the conditional knowledge graph representation method of step 1 and the hierarchical parse tree obtained in step 3, design the corresponding physical storage and query scheme of the hierarchical parse tree to manage the conditional knowledge graph;

[0069] For storage, the tree structure of the hierarchical parsing tree obtained in step 3.1 is regarded as a directed acyclic graph, and the conditional knowledge graph is stored in the graph database; the triple of the leaf node of the hierarchical parsing tree is abstracted into a triple node, and the child nodes of the triple node are the head and tail entities in the triple; according to the parsing tree structure obtained in step 3, storage in the graph database is realized;

[0070] For query, start from the root node of the hierarchical parse tree and traverse the entire hierarchical parse tree from top to bottom. When backtracking, recursively use non-leaf nodes as relationship edges layer by layer to connect the nested structure of child nodes of non-leaf nodes until backtracking to the root node. In this way, the conditional knowledge graph representation method in step 1 can be obtained again; the conditional sentence is used as the root node of the hierarchical parse tree; for the convenience of query, the corresponding root node is saved in the entity during storage, so that when querying a certain entity, all conditional sentences associated with the node can be queried based on the root node it has.

[0071] According to the conditional graph representation method and the tree structure of nested triples proposed in this embodiment, this embodiment uses a graph database to store and manage conditional knowledge. The advantage of this method is that it greatly reduces the storage space, while maintaining the graph structure properties of the knowledge graph itself, and improving query efficiency.

[0072] Because the node of the nested triple can be either an entity or a triple. In a graph database, a node can only represent an entity of a triple, and it is impossible to directly store the conditional knowledge representation proposed in this paper. This implementation method constructs a virtual node "tuple" to represent Figure 3 The ordinary triples corresponding to the leaf nodes in the tree are connected to the head node and the tail node of the ordinary triples. The virtual node can be further used as the head node or the tail node in the nested triples. For the prompt keywords with logical attributes that represent the hierarchical relationship between fact triples and conditional triples and conditional triples, this implementation method directly uses them as an independent node to connect the associated ordinary / nested triples. According to the above method, the knowledge storage in the graph database can be realized based on the parse tree structure of the conditional sentence obtained in step 3, such as Figure 5 shown.

[0073] In terms of query, this implementation uses a graph database to store the massive knowledge in the conditional knowledge graph. The graph database provides users with an interface to access data through a formal query language. This implementation uses the inverted index idea to store several root node ids {rootId1, rootId2, ...} of the tree structure corresponding to the conditional sentence in each leaf node stored in the tree storage structure. When querying a node, all structural relationships associated with the node can be queried based on the root node it has.

[0074] In the existing conditional knowledge graph representation and construction methods, although the importance of conditions is taken into account, the joint extraction of condition triples and fact triples is achieved, and a knowledge graph containing facts and conditions supporting the validity of facts is constructed, the extraction method can only extract fact triples and condition triples separately, and cannot obtain the corresponding constraint relationship between fact triples and condition triples and the logical relationship between condition triples. In view of the shortcomings of the prior art, the present invention provides a new conditional knowledge graph representation and construction method, which uses a text structure hierarchical parsing module to capture the hierarchical semantic association of conditional statements, and uses an open information extraction method based on sequence annotation to extract facts and condition triples. And based on the graph database, a corresponding knowledge storage strategy is designed to realize the mapping of knowledge logical representation to physical storage model to ensure the portability of knowledge.

Claims

1. A method for representing and constructing a conditional knowledge graph, characterized in that: The specific steps are as follows: Step 1: Construct a conditional knowledge graph representation method that uses nested triples to formally express the rich hierarchical semantic associations between facts and conditional triples, and between conditional triples; Step 2: Based on the conditional knowledge graph representation method in step 1, the hierarchical parsing module is used to perform hierarchical parsing on the text structure of the conditional statement to obtain a set of text units and the hierarchical semantic relationships between the text units; A RST recursive hierarchical parse tree is constructed based on the conditional sentence to parse out the relationship between the condition and the fact, between the condition and the condition, and between the fact and the fact in the conditional sentence. At the same time, the conditional sentence is converted into multiple text units and used as the input of the triple extraction task in step 3; the text units input to step 3 are conditional units and fact units; Step 3: Use the triple extraction module to extract triples from the text units parsed in step 2, namely, conditional units and factual units, convert the text units into triples, and substitute the extracted triples into the leaf nodes of the hierarchical parse tree; finally, the conditional knowledge graph representation method in step 1 represented by the hierarchical parse tree is obtained; Step 3.1: Use the triple extraction module to extract triples from the text units parsed in step 2, i.e., conditional units and factual units, and substitute the extracted triples into the leaf node positions of the hierarchical parse tree to obtain a tree-structured conditional graph representation model; obtain the conditional knowledge graph representation method in step 1 represented by a hierarchical parse tree; Step 3.2: Use the non-leaf nodes in the hierarchical parse tree as relationship edges, and recursively use the non-leaf nodes as relationship edges to connect the nested structures on both sides from bottom to top until recursively reaching the root node, and convert the hierarchical parse tree obtained in step 3.1 into a nested triple conditional knowledge graph representation method; Step 4: Based on the conditional knowledge graph representation method of step 1 and the hierarchical parse tree obtained in step 3, design the corresponding physical storage and query scheme of the hierarchical parse tree to manage the conditional knowledge graph; The corresponding physical storage of the hierarchical parse tree is: For storage, the tree structure of the hierarchical parsing tree obtained in step 3.1 is regarded as a directed acyclic graph, and the conditional knowledge graph is stored in the graph database; the triple of the leaf node of the hierarchical parsing tree is abstracted into a triple node, and the child nodes of the triple node are the head and tail entities in the triple; according to the parsing tree structure obtained in step 3, storage in the graph database is realized; The corresponding query scheme of the hierarchical parse tree is: For query, start from the root node of the hierarchical parse tree and traverse the entire hierarchical parse tree from top to bottom. When backtracking, recursively use non-leaf nodes as relationship edges layer by layer to connect the nested structure of child nodes of non-leaf nodes until backtracking to the root node. Then, the conditional knowledge graph representation method in step 1 can be obtained again; the conditional sentence is used as the root node of the hierarchical parse tree; for the convenience of query, the corresponding root node is saved in the entity during storage. When querying an entity, all conditional sentences associated with the node are queried based on the root node it has.

2. A method for representing and constructing a conditional knowledge graph according to claim 1, characterized in that: The nested triples in step 1 are: The head node and tail node in a nested triple are entities or triples.

3. A method for representing and constructing a conditional knowledge graph according to claim 1, characterized in that: There are three types of rich hierarchical semantic associations between the expression facts and conditional triples, and between conditional triples, as described in step 1, which are: Entity, relationship between entities, entity; Triples, relations between triples, triples; Triple, relationship between triple and entity, entity or entity, relationship between entity and triple, triple.

4. A method for representing and constructing a conditional knowledge graph according to claim 1, characterized in that: The conditional statement of step 2 is: A conditional sentence consists of a conditional clause and a concluding clause.

5. A method for representing and constructing a conditional knowledge graph according to claim 1, characterized in that: The step 2 is: Step 2.1: Given a conditional statement C: If c, f., "If" can be replaced by other conditional introductory words; c represents a logical combination description of several conditions, and f represents a logical combination description of several facts; Step 2.2: Perform text-level semantic analysis on the conditional statement C: If c, f. and construct a Rhetorical Structure Theory RST recursive hierarchical parse tree, where the lowest leaf node is a set of text units {c1,…,cm,f1,…,fn,s1,…,sr}; Step 2.3: Classify a set of text units {c1,…,cm,f1,…,fn,s1,…,sr} into three types, ci(i∈{1,…,m}) is a conditional unit, fj(j∈{1,…,n}) is a factual unit, and sk(k∈{1,…,r}) is a supplementary unit; Among them, the conditional unit and the factual unit are simple sentences with complete structures and only contain a subject-predicate structure, and the supplementary unit is in the form of a noun phrase structure.

Images