A fact verification method and system based on global to local aggregation fission network

Abstract

The present application discloses a fact verification method and system based on a global-to-local aggregated fission network, the method comprising: parsing the acquired evidence statements into semantic-level relationship triplets and constructing a fully connected graph as evidence nodes; performing multi-hop entity-level reasoning, updating the initial hidden layer representation of the evidence nodes to obtain a first fission representation; performing neighborhood information aggregation operations on the first aggregate representation of each evidence node to obtain a second aggregate representation; selecting a preset number of evidence nodes with the highest value according to the second aggregate representation; calculating the prediction probability according to the selected evidence nodes and the representation information of the statement, and outputting the prediction label with the largest prediction probability as the verification result of the statement; the present invention obtains fine-grained clue representations through a global-to-local aggregated fission network, and adopts iterative operations to realize multi-hop reasoning and higher-order information aggregation, obtains sufficient logical relationships, and improves the efficiency and accuracy of fact semantic verification.

Description

Technical Field

[0001] The present application relates to the technical field of natural language reasoning and fact verification, and more specifically, to a fact verification system based on a global to local aggregated fission network. Background Art

[0002] With the rapid development of the Internet, data information has also ushered in an era of explosive growth. However, in addition to correct information, there is also a large amount of false information on the Internet. Therefore, how to judge the authenticity of information on the Internet has become an urgent problem to be solved. The fact verification task refers to giving a statement and requiring the system to extract relevant sentences from a large-scale text corpus as evidence, and use these evidences to verify and judge the authenticity of the statement. The system needs to give three types of judgments on the statement based on evidence: support (SUPPORTED), opposition (REFUTED) or insufficient information (NOT ENOUGH INFO).

[0003] In recent years, many traditional fact verification systems use natural language reasoning technology to mine the relationship between evidence and claims to output judgments. These existing fact verification systems usually use a three-level pipeline structure to complete the fact verification task: first, document extraction, then evidence extraction, and finally claim verification. For the document extraction step, there are three main ideas. The first is entity linking, which maps certain strings in the text to the corresponding entities in the knowledge base; the second is component analysis plus keyword matching; the third is component analysis plus online retrieval. Evidence extraction can be achieved through keyword matching or supervised classification, such as using the ESIM enhanced sequence reasoning model to calculate the similarity score between sentences. The claim verification step is also the main focus of most existing models. For example, the model proposed by Yoneda et al. infers the accuracy of each claim-evidence pair and outputs the final prediction by aggregating multiple prediction labels; the model proposed by Hanselowski et al. encodes each claim-evidence pair separately, and finally uses a pooling function to aggregate features for prediction. However, most existing traditional fact verification methods only use simple models to extract relevant information from evidence, without considering the connection between each piece of evidence, which also limits their performance to a certain extent.

[0004] Furthermore, by integrating multi-evidence reasoning into fact verification based on constructed evidence graphs, the academic community has proposed many graph-enhanced natural language reasoning models and achieved better performance. Among them, the GEAR model used BERT to encode evidence for the first time, and designed a graph neural network to perform information aggregation operations on the constructed evidence graph by treating each piece of evidence as a node. The DREAM model uses XLNet, and uses GCN and GAT to perform evidence reasoning on the established semantic-level graph. KGAT innovatively uses a kernel graph attention network to achieve information aggregation by unifying the edge kernel mechanism and the node kernel mechanism on the evidence graph.

[0005] Although the existing fact verification models have made significant progress, they still have the following shortcomings. The first shortcoming is the lack of fine-grained representation of evidence clues. Existing models either connect evidence sentences into a single string or simply treat each evidence-claim pair as a single sentence-level node. Since these models only aggregate information at the sentence level, they have great difficulty in achieving fine-grained reasoning, which requires the model to distinguish the subtle differences between true and false statements, but most existing models find it difficult to make such detailed distinctions between these key entities. The second shortcoming is the difficulty in obtaining sufficient logical relationships and jump-based claim verification. Since false claims on the Internet are often artificially fabricated, they may be reasonable at the semantic level, but they are not supported by facts at the logical level. Therefore, the model needs to obtain sufficient logical relationships through existing clues and perform multi-hop reasoning to make accurate judgments. Taking the statement "The Rodney King riots occurred in the most populous county in the United States" as an example, the evidence clues show that "The Rodney King riots occurred in Los Angeles County" and "Los Angeles County is the most populous county in the United States". The fact verification model must mine with "Los Angeles County" as the core and obtain the potential relationship between "Rodney King riots" and "the most populous county in the United States". Only through multi-hop reasoning can an accurate judgment be made. The third deficiency is the lack of a filtering mechanism for noise evidence. Since the evidence statements in the fact verification process are retrieved from a complex background corpus, noise will inevitably be introduced. Worse, these noises may be further amplified in subsequent conversion processing, affecting the performance of the fact verification system. Summary of the invention

[0006] In response to at least one defect or improvement need in the prior art, the present invention provides a fact verification method and system based on a global to local aggregate fission network, which realizes fact verification under complex logical relationships by extracting fine-grained evidence clues to represent information, with the aim of improving the accuracy of claim verification.

[0007] To achieve the above object, according to one aspect of the present invention, a fact verification method based on a global to local aggregate fission network is provided, the method comprising the following steps:

[0008] Obtain a statement and multiple evidence statements, and parse each of the evidence statements into at least one semantic-level relationship triple; construct a fully connected graph using each of the semantic-level relationship triples as an evidence node and generate an initial hidden layer representation of each evidence node;

[0009] Perform multi-hop entity-level object reasoning between each evidence node and its adjacent evidence nodes, update the initial hidden layer representation of the evidence node according to the initial hidden layer representation of the adjacent evidence nodes, obtain the first fission representation of each evidence node and convert it into a first aggregate representation;

[0010] Perform neighborhood information aggregation operation on the first aggregate representation of each evidence node to obtain a second aggregate representation;

[0011] Selecting a preset number of evidence nodes with the highest value from the fully connected graph according to the second aggregate representation;

[0012] The prediction probability is calculated according to the second aggregate representation of the selected evidence node and the initial hidden layer representation of the statement, and the prediction label with the largest prediction probability is output as the verification result of the statement.

[0013] Furthermore, the above fact verification method based on the global to local aggregate fission network also includes:

[0014] The second aggregation representation of each evidence node is converted into a second fission representation, and a neighborhood information aggregation operation is iteratively performed on each evidence node based on the second fission representation.

[0015] Furthermore, in the above-mentioned fact verification method based on the global to local aggregated fission network, the first fission representation is obtained in the following manner:

[0016] For each evidence node, respectively calculate the first vector similarity between the object representation in its initial hidden layer representation and the subject representation in the initial hidden layer representation of each adjacent evidence node;

[0017] Calculate the first attention weight corresponding to each adjacent evidence node according to the first vector similarity;

[0018] Calculate an attention vector representing the adjacent evidence node according to the first attention weight and the initial hidden layer representation of the adjacent evidence node;

[0019] The initial hidden layer representation of the evidence node is updated according to the attention vector and the object representation of the evidence node to obtain a first fission representation.

[0020] Furthermore, in the above fact verification method based on the global to local aggregate fission network, the first aggregate representation of each evidence node is respectively subjected to neighborhood information operation to obtain the second aggregate representation specifically as follows:

[0021] Calculating a second vector similarity between nodes based on the first aggregate representation of each evidence node and its adjacent evidence nodes;

[0022] Calculate the second attention weight corresponding to each adjacent evidence node according to the second vector similarity;

[0023] The first aggregate representations of adjacent evidence nodes are integrated according to the second attention weight to obtain a second aggregate representation of each evidence node.

[0024] Furthermore, the fact verification method based on the global to local aggregate fission network described above, wherein a preset number of evidence nodes with the highest value are selected from the fully connected graph according to the second aggregate representation, is specifically:

[0025] Calculate the semantic similarity between the initial latent representation of the statement and the second aggregated representation between each evidence node;

[0026] A preset number of evidence nodes with the highest semantic similarity are selected through pooling operation.

[0027] Furthermore, in the above-mentioned fact verification method based on the global to local aggregate fission network, the calculation of the prediction probability according to the second aggregate representation of the selected evidence node and the declared initial hidden layer representation is specifically as follows;

[0028] Generate fine-grained label prediction probabilities based on the declared initial hidden layer representation and the second aggregated representation of the selected evidence nodes;

[0029] Calculating the probability of evidence selection based on the semantic similarity between the initial latent representation of the statement and the second aggregate representation of the selected evidence node;

[0030] The prediction probability is calculated by the fine-grained label prediction probability and the evidence selection probability.

[0031] Furthermore, in the above-mentioned fact verification method based on the global to local aggregate fission network, the prediction probability is the sum of the products of the fine-grained label prediction probability of each selected evidence node and the evidence selection probability.

[0032] Furthermore, in the above-mentioned fact verification method based on the global to local aggregate fission network, the semantic-level relationship triples include the subject, relationship and object of the evidence node; the initial hidden layer representation of each evidence node includes the subject representation, relationship representation and object representation.

[0033] Furthermore, in the above-mentioned fact verification method based on global to local aggregate fission network, the predicted label is any one of support, opposition and insufficient information.

[0034] According to another aspect of the present invention, there is provided a fact verification system based on a global-to-local aggregate fission network, comprising:

[0035] A global semantic encoder is provided to obtain a statement and a plurality of evidence sentences, and parse each of the evidence sentences into at least one semantic-level relationship triple; a fully connected graph is constructed using each of the semantic-level relationship triples as an evidence node, and an initial hidden layer representation of each evidence node is generated respectively;

[0036] A local fission reasoning layer performs multi-hop entity-level object reasoning between each evidence node and its adjacent evidence nodes, updates the initial hidden layer representation of the evidence node according to the initial hidden layer representation of the adjacent evidence nodes, obtains the first fission representation of each evidence node and converts it into a first aggregate representation;

[0037] The global evidence aggregation layer performs neighborhood information aggregation on the first aggregate representation of each evidence node to obtain the second aggregate representation;

[0038] A pooling layer, selecting a preset number of evidence nodes with the highest value from the fully connected graph according to the second aggregate representation;

[0039] The multi-layer perceptron calculates the prediction probability according to the second aggregate representation of the selected evidence node and the initial hidden layer representation of the statement, and outputs the prediction label with the maximum prediction probability as the verification result of the statement.

[0040] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0041] (1) The present invention provides a fact verification method and system based on a global-to-local aggregate fission network. It performs multi-hop entity-level object reasoning between each evidence node and its adjacent evidence nodes, calculates the first fission representation of the target evidence node, and obtains fine-grained clue representation, solving the fine-grained reasoning problem. It selects a preset number of evidence nodes with the highest value, filters evidence nodes with lower value, and verifies fact semantics, thereby improving the efficiency and accuracy of fact semantic verification.

[0042] (2) The present invention provides a fact verification system based on a global-to-local aggregated fission network. By iteratively performing multi-hop entity-level object reasoning and neighborhood information aggregation operations in a global-to-local aggregated fission network model, multi-hop reasoning and higher-order information aggregation can be achieved to obtain sufficient logical relationships for more accurate fact verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 A flow chart of a fact verification method based on a global to local aggregate fission network provided in this embodiment;

[0045] Figure 2 This is a working mechanism diagram of multi-hop entity-level object reasoning in this embodiment;

[0046] Figure 3 This is a working mechanism diagram of the neighborhood information aggregation operation in this embodiment;

[0047] Figure 4 A flowchart of the global to local aggregation fission algorithm provided in this embodiment;

[0048] Figure 5 A structural diagram of a fact verification system based on a global to local aggregate fission network provided in this embodiment;

[0049] Figure 6 This is a graph of the attention value visualization results provided for this embodiment. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0051] The terms "first", "second", "third", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0052] In addition, in order to avoid obscuring the understanding of the present invention by the skilled person, well-known or widely used techniques, elements, structures and processes may not be described or shown in detail. Although the drawings represent exemplary embodiments of the present invention, the drawings are not necessarily drawn to scale, and certain features may be enlarged or omitted to better illustrate and explain the present invention.

[0053] In view of the shortcomings of existing fact verification models, such as the lack of fine-grained evidence clue representation, the difficulty in obtaining sufficient logical relationships and jump-point-based statement verification, and the lack of a filtering mechanism for noise evidence, this application proposes a fact verification solution based on a global-to-local aggregate fission network to complete the fact verification task. In this application, the challenging fine-grained reasoning problem is solved by parsing the evidence statements into three-level reasoning clues and modeling the clues as functions that map from subject entities to object entities. The experimental results of this application on the FEVER dataset achieved a test score of 76.2%, achieving quite excellent performance.

[0054] On the one hand, this embodiment provides a fact verification method based on a global to local aggregate fission network. Figure 1 For a flowchart of a fact verification method based on a global to local aggregate fission network provided in this embodiment, please refer to Figure 1 , the fact verification method comprises the following steps:

[0055] (1) Obtaining a statement and multiple evidence statements, parsing each of the evidence statements into at least one semantic-level relationship triple; constructing a fully connected graph using each of the semantic-level relationship triples as an evidence node and generating an initial hidden layer representation of each evidence node;

[0056] Specifically, get the given declaration c and l Evidence statements, from l n semantic-level relationship triplets are extracted from the evidence sentences, and the semantic-level relationship triplets are used as evidence nodes to construct the evidence graph G. The evidence sentences are parsed into three-level reasoning clues, and the clues are modeled as a function mapping from the subject entity to the object entity. The set of evidence nodes can be expressed as formula (1):

[0057] E={e 1 ,…,e i ,…,e n}(1)

[0058] Among them, 1≤ i ≤n, where n is a natural number greater than 1.

[0059] In a specific example, the semantic-level relationship triplets in the set E of evidence nodes are obtained by applying an existing semantic role labeling (SRL) tool on l evidence sentences, and each evidence sentence can be parsed into multiple semantic-level relationship triplets.

[0060] The semantic level relationship triples include the subject, relationship and object of the evidence node. i , can be expressed as a relation triple (subject, relation, object), see formula (2),

[0061] ei = (s i , r i , o i )(2)

[0062] Among them, s i is the subject of the evidence node, r i Indicates the relationship between the subject and the object in the evidence node, o i An object that is an evidence node.

[0063] The n evidence nodes in the evidence node set E are connected by edges, and finally a fully connected graph (evidence graph) G is constructed. Based on the fully connected graph G, the prediction probability is obtained by performing the reasoning process on the semantic level relationship triples (graph nodes), and finally the predicted label with the highest probability is returned as the claim verification result.

[0064] In this embodiment, the semantic role labeling SRL tool connects each semantic-level relationship triple as "[CLS] subject [SEP] relationship [SEP] object", where the [CLS] mark is at the beginning of the sentence and the [SEP] mark is used to separate the input sentences. "[CLS] subject [SEP] relationship [SEP] object" is input into the BERT model to obtain the initial hidden layer representation of all evidence nodes and claim nodes, and output the initial hidden layer representation of the evidence nodes. The initial hidden layer representation of each evidence node includes the subject representation, the relationship representation and the object representation.

[0065] For evidence nodes , The initial hidden layer representation of can be expressed as formula (3),

[0066] (3)

[0067] Similarly, the calculation method of the initial hidden layer representation of the declared node is shown in formula (4),

[0068] (4)

[0069] (2) performing multi-hop entity-level object reasoning between each evidence node and its adjacent evidence nodes, updating the initial hidden layer representation of the evidence node according to the initial hidden layer representation of the adjacent evidence nodes, obtaining the first fission representation of each evidence node and converting it into the first aggregate representation;

[0070] Specifically, for each evidence node, the first vector similarity between the object representation in its initial hidden layer representation and the subject representation in the initial hidden layer representation of each adjacent evidence node is calculated; the first attention weight corresponding to each adjacent evidence node is calculated according to the first vector similarity; the attention vector representing the adjacent evidence node is calculated according to the first attention weight and the initial hidden layer representation of the adjacent evidence node; the initial hidden layer representation of the evidence node is updated according to the attention vector and the object representation of the evidence node to obtain the first fission representation.

[0071] Figure 2 This is a diagram of the working mechanism of multi-hop entity-level object reasoning in this embodiment. Different from the existing reasoning model that uses the entire evidence sentence as a node, this embodiment performs entity-level object reasoning. Figure 2 , for the evidence node , calculate the evidence node The representation of the object (tail entity) and its adjacent evidence nodes in the initial hidden layer representation The similarity of the first vector of the subject (head entity) representation in the initial hidden layer representation can be expressed as formula (5):

[0072] (5)

[0073] in, An object representation representing an evidence node, Represents the main representation of adjacent evidence nodes, 1≤ j ≤n.

[0074] Then the evidence node is obtained by normalization calculation The first attention weight is expressed as formula (6):

[0075] (6)

[0076] in, Represents evidence node Its adjacent evidence nodes The first vector similarity of k ≤n.

[0077] Based on the evidence nodes The first attention weight of and the initial hidden layer representation of the adjacent evidence nodes can be used to calculate the attention vector representing the adjacent evidence nodes, which is expressed as formula (7):

[0078] (7)

[0079] Furthermore, based on the above attention vector and the object representation of the evidence node, the evidence node is updated The fission state of is characterized by the first fission, which is expressed as formula (8):

[0080] (8)

[0081] in, and are all trainable weights, Represents the target evidence node object representation.

[0082] Updated evidence node The first fission vector representation of is expressed as formula (9),

[0083] (9)

[0084] Finally, in order to facilitate the subsequent global information sharing, an aggregation operation is performed on each evidence node, and the first fission representation is converted into the first aggregation representation as shown in formula (10):

[0085] (10)

[0086] in, is a trainable matrix, Characterizes the first fission.

[0087] (3) Perform neighborhood information aggregation on the first aggregate representation of each evidence node to obtain a second aggregate representation;

[0088] Specifically, the second vector similarity between nodes is calculated based on the first aggregate representation of each evidence node and its adjacent evidence nodes; the second attention weight corresponding to each adjacent evidence node is calculated respectively based on the second vector similarity; the first aggregate representation of each adjacent evidence node is integrated according to the second attention weight to obtain the second aggregate representation of each evidence node.

[0089] Figure 3 For a working mechanism diagram of the neighborhood information aggregation operation in this embodiment, please refer to Figure 3 , for the evidence node , calculate the evidence nodes in the aggregate state With adjacent evidence nodes The second vector similarity in can be expressed as formula (11),

[0090] (11)

[0091] in, Represents evidence node The first polymer characterization of Represents adjacent evidence nodes First polymer characterization of .

[0092] Then the evidence node is obtained by normalization calculation The second attention weight is expressed as formula (12):

[0093] (12)

[0094] After integrating the first aggregate representation of the surrounding n evidence nodes (including itself), according to the evidence node The second attention vector is calculated to obtain the second aggregate representation expressed as formula (13):

[0095] (13)

[0096] At the same time, in order to perform the global aggregation operation in subsequent iterations, a fission operation needs to be performed to convert the second aggregation representation into the second fission representation expressed as formula (14):

[0097] (14)

[0098] in is a trainable matrix, It is the second aggregate representation of the evidence node.

[0099] The normalization calculation may use a softmax function, and of course other functions may also be used, and this embodiment does not make any specific limitation.

[0100] Figure 4 For the workflow diagram of the global to local aggregation fission algorithm provided in this embodiment, please refer to Figure 4 ,By iteratively performing multi-hop entity object reasoning and neighborhood information aggregation, multi-hop reasoning and higher-order information aggregation can be achieved to obtain sufficient logical relationships for more accurate fact verification.

[0101] (4) selecting a preset number of evidence nodes with the highest value from the fully connected graph according to the second aggregate representation;

[0102] Through the pooling operation, a noise filtering mechanism is used to select the evidence nodes, filter out the evidence nodes with insufficient evidence, and select a preset number of evidence nodes with the highest value from the fully connected graph. Specifically, the semantic similarity between the initial hidden layer representation of the statement and the second aggregate representation between each evidence node is first calculated; then a preset number of evidence nodes with the highest semantic similarity are selected through the pooling operation.

[0103] When t iterations have been completed, the evidence nodes with fewer clues will be discarded, and a parameter k will be given to select the k most valuable evidence nodes. Its formal definition is as follows (15):

[0104] (15)

[0105] in, Represents a collection of evidence nodes The set of evidence nodes obtained after evidence selection, and .

[0106] Evidence Node The value of is defined as the semantic similarity between the initial hidden representation of the key vector of statement c and the second aggregate representation between each evidence node, expressed as formula (16),

[0107] (16)

[0108] in, To declare the key vector of c, Evidence Node The updated second aggregate representation, is a trainable matrix.

[0109] (5) Calculate the prediction probability based on the second aggregate representation of the selected evidence node and the initial hidden layer representation of the statement, and output the prediction label with the largest prediction probability as the verification result of the statement. The prediction label is any one of support, opposition, and insufficient information.

[0110] The fact verification task aims to verify the accuracy of a statement based on multiple evidence sentences and return a predicted label; predicted label ∈ {support, oppose, insufficient information}.

[0111] Specifically, a fine-grained label prediction probability is generated based on the initial latent representation of the statement and the second aggregate representation of the selected evidence node; the evidence selection probability is calculated based on the semantic similarity between the initial latent representation of the statement and the second aggregate representation of the selected evidence node; the prediction probability is calculated through the fine-grained label prediction probability and the evidence selection probability.

[0112] This example follows the standard graph label prediction setting and divides the prediction into two components: the probability of evidence selection P(e i | G) and fine-grained label prediction probability P(y| e i , G).

[0113] The fine-grained label prediction probability is generated based on the declared initial hidden layer representation and the second aggregate representation of the selected evidence node, which can be expressed as formula (17):

[0114] (17)

[0115] in, Evidence Node The second polymer characterization, An object representation of a declaration node.

[0116] The probability of evidence selection is calculated based on the semantic similarity between the second aggregate representation of the selected evidence node and the initial hidden representation of statement c, and is expressed as formula (18):

[0117] (18)

[0118] The sum of the product of the fine-grained label prediction probability of each selected evidence node and the evidence selection probability is the final prediction probability, which can be expressed as formula (19):

[0119] (19)

[0120] The objective loss function of model training is to minimize the cross entropy, which is expressed as formula (20):

[0121] (20)

[0122] in, is the true predicted label.

[0123] On the other hand, this embodiment also provides a fact verification system based on a global to local aggregate fission network. Figure 5 For a structural diagram of a fact verification system based on a global to local aggregate fission network provided in this embodiment, please refer to Figure 5 , the fact verification system includes:

[0124] A global semantic encoder (GSE) obtains a statement and multiple evidence sentences, and parses each of the evidence sentences into at least one semantic-level relationship triple; constructs a fully connected graph using each of the semantic-level relationship triples as an evidence node and generates an initial hidden layer representation of each evidence node;

[0125] In this embodiment, the global semantic encoder is used to obtain the initial representation of the statement c and all evidence nodes in the fully connected graph G; in the global semantic encoder, semantic relationship triplets are first extracted from the evidence sentences as evidence nodes through the semantic role labeling (SRL) tool, and each triple is connected as "[CLS] subject [SEP] relation [SEP] object" (the [CLS] mark is at the beginning of the sentence, and the [SEP] mark is used to separate the input sentences) as the input of BERT, and then the initial hidden layer representation of all nodes is obtained through BERT.

[0126] A local fission reasoning layer (LFR) performs multi-hop entity-level object reasoning between each evidence node and its adjacent evidence nodes, updates the initial hidden layer representation of the evidence node according to the initial hidden layer representation of the adjacent evidence nodes, obtains the first fission representation of each evidence node and converts it into the first aggregate representation;

[0127] In order to achieve fine-grained claim verification, it is necessary to update the object information in the triple node. This embodiment uses a local fission reasoning layer to perform 2-hop entity-level reasoning between the target node and its connected nodes, and finally obtains the updated object representation. Specifically, the local fission reasoning layer takes the output representation of the global semantic encoder as input, is responsible for performing 2-hop object-level reasoning on fine-grained claim verification, and outputs updated node representations, which will serve as input to the global evidence aggregation layer after aggregation operations.

[0128] The global evidence aggregation layer (GEA) performs neighborhood information aggregation on the first aggregate representation of each evidence node to obtain the second aggregate representation;

[0129] In this embodiment, the global evidence aggregation layer realizes evidence information sharing by performing a first-order neighborhood information integration operation. The global evidence aggregation layer is introduced to realize information sharing and multi-evidence joint reasoning.

[0130] A pooling layer, selecting a preset number of evidence nodes with the highest value from the fully connected graph according to the second aggregate representation;

[0131] In order to overcome the deficiency of the existing method in that there is no noise evidence filtering mechanism, this embodiment adopts a graph pooling layer to perform node selection and noise filtering; the pooling layer filters the existing noise data through the graph pooling operation, and selects the probability of Select valuable evidence nodes.

[0132] A multi-layer perceptron (MLP) calculates the prediction probability according to the second aggregate representation of the selected evidence node and the initial hidden layer representation of the statement, and outputs the prediction label with the maximum prediction probability as the verification result of the statement.

[0133] It should be noted that the local fission reasoning layer and the global evidence aggregation layer can perform two-hop reasoning and first-order neighborhood aggregation respectively. By iteratively executing local fission reasoning and global evidence aggregation, multi-hop reasoning and higher-order information aggregation can be achieved to obtain sufficient logical relationships for more accurate fact verification.

[0134] In a specific embodiment, Table 1 is an example of a statement and evidence, please refer to Table 1:

[0135] Table 1. Examples of statements and evidence

[0136]

[0137] Table 2 shows examples of claims and evidence clues. By using the semantic role labeling (SRL) tool to extract semantic-level relationship triples from evidence sentences, we can obtain the claim clues and evidence clues shown in Table 2:

[0138] Table 2: Examples of statements and evidence clues

[0139]

[0140] Figure 6 The attention value visualization result diagram provided in this embodiment is as follows: Figure 6 As shown, the first four columns represent the second attention weights of nodes 0-3 (corresponding to the four evidence triplets in Table 2), and the last column represents the value of the evidence node.

[0141] from Figure 4 It can be seen that by aggregating the surrounding evidence node information, evidence node 1 has the highest value (0.982), and its value mainly comes from the relevant information of aggregated evidence node 2 and evidence node 3, while evidence node 2 (0.125) has a lower semantic value and will be filtered by the pooling layer. Therefore, it can be further inferred that the 2-hop reasoning chain is from node 1 to node 3, that is, the evidence (The Rodney King riots, occurred in, Los Angeles County) and (Los Angeles County, is, the most populous county in the USA) infer that the statement is true, and the output prediction label is "support".

[0142] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0143] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A fact verification method based on global to local aggregate fission network, characterized in that: The method comprises the following steps: Obtaining a statement and a plurality of evidence statements, parsing each of the evidence statements into at least one semantic level relationship triple; Constructing a fully connected graph using each of the semantic-level relationship triples as evidence nodes and generating an initial hidden layer representation of each evidence node; Perform multi-hop entity-level object reasoning between each evidence node and its adjacent evidence nodes, update the initial hidden layer representation of the evidence node according to the initial hidden layer representation of the adjacent evidence nodes, obtain the first fission representation of each evidence node and convert it into a first aggregate representation; the first fission representation is obtained in the following way: for each evidence node, respectively calculate the first vector similarity between the object representation in its initial hidden layer representation and the subject representation in the initial hidden layer representation of each adjacent evidence node; Calculate the first attention weight corresponding to each adjacent evidence node according to the first vector similarity; Calculating an attention vector representing an adjacent evidence node according to the first attention weight and the initial hidden layer representation of the adjacent evidence node; updating the initial hidden layer representation of the evidence node according to the attention vector and the object representation of the evidence node to obtain a first fission representation; Perform neighborhood information aggregation operation on the first aggregate representation of each evidence node to obtain a second aggregate representation; Selecting a preset number of evidence nodes with the highest value from the fully connected graph according to the second aggregate representation; The prediction probability is calculated according to the second aggregate representation of the selected evidence node and the initial hidden layer representation of the statement, and the prediction label with the largest prediction probability is output as the verification result of the statement.

2. The fact verification method based on global to local aggregate fission network according to claim 1, characterized in that: Also includes: The second aggregation representation of each evidence node is converted into a second fission representation, and a neighborhood information aggregation operation is iteratively performed on each evidence node based on the second fission representation.

3. The fact verification method based on global to local aggregate fission network according to claim 1 is characterized in that: The first aggregate representation of each evidence node is subjected to neighborhood information operation to obtain the second aggregate representation, which is specifically: Calculating a second vector similarity between nodes based on the first aggregate representation of each evidence node and its adjacent evidence nodes; Calculate the second attention weight corresponding to each adjacent evidence node according to the second vector similarity; The first aggregate representations of adjacent evidence nodes are integrated according to the second attention weight to obtain a second aggregate representation of each evidence node.

4. The fact verification method based on global to local aggregate fission network according to claim 1, characterized in that: The step of selecting a preset number of evidence nodes with the highest value from the fully connected graph according to the second aggregate representation is specifically: Calculate the semantic similarity between the initial latent representation of the statement and the second aggregated representation between each evidence node; A preset number of evidence nodes with the highest semantic similarity are selected through pooling operation.

5. The fact verification method based on global to local aggregate fission network according to claim 1 or 4, characterized in that: The calculation of the prediction probability according to the second aggregate representation of the selected evidence node and the declared initial hidden layer representation is specifically as follows; Generate fine-grained label prediction probabilities based on the declared initial hidden layer representation and the second aggregated representation of the selected evidence nodes; Calculating the probability of evidence selection based on the semantic similarity between the initial latent representation of the statement and the second aggregate representation of the selected evidence node; The prediction probability is calculated by the fine-grained label prediction probability and the evidence selection probability.

6. The fact verification method based on global to local aggregate fission network according to claim 5 is characterized in that: The prediction probability is the sum of the products of the fine-grained label prediction probability and the evidence selection probability of each selected evidence node.

7. The fact verification method based on global to local aggregate fission network according to claim 1 is characterized in that: The semantic-level relationship triples include the subject, relationship and object of the evidence node; the initial hidden layer representation of each evidence node includes the subject representation, relationship representation and object representation.

8. The fact verification method based on global to local aggregate fission network according to claim 1, characterized in that: The predicted label is any one of support, opposition, and insufficient information.

9. A fact verification system based on global to local aggregate fission network, characterized in that: include: A global semantic encoder, which obtains a statement and a plurality of evidence sentences, and parses each of the evidence sentences into at least one semantic level relation triple; Constructing a fully connected graph using each of the semantic-level relationship triples as evidence nodes and generating an initial hidden layer representation of each evidence node; The local fission reasoning layer performs multi-hop entity-level object reasoning between each evidence node and its adjacent evidence nodes, updates the initial hidden layer representation of the evidence node according to the initial hidden layer representation of the adjacent evidence nodes, obtains the first fission representation of each evidence node and converts it into a first aggregate representation; the first fission representation is obtained in the following way: for each evidence node, the first vector similarity between the object representation in its initial hidden layer representation and the subject representation in the initial hidden layer representation of each adjacent evidence node is calculated respectively; Calculate the first attention weight corresponding to each adjacent evidence node according to the first vector similarity; Calculating an attention vector representing an adjacent evidence node according to the first attention weight and the initial hidden layer representation of the adjacent evidence node; updating the initial hidden layer representation of the evidence node according to the attention vector and the object representation of the evidence node to obtain a first fission representation; The global evidence aggregation layer performs neighborhood information aggregation on the first aggregate representation of each evidence node to obtain the second aggregate representation; A pooling layer, selecting a preset number of evidence nodes with the highest value from the fully connected graph according to the second aggregate representation; The multi-layer perceptron calculates the prediction probability according to the second aggregate representation of the selected evidence node and the initial hidden layer representation of the statement, and outputs the prediction label with the maximum prediction probability as the verification result of the statement.

Images