3.1.1 Ternary information structure.

The triple evaluation structure is a data model widely used for representing information, consisting of three components: subject, predicate, and object. It is extensively applied in fields such as knowledge graphs and the semantic web. The triple evaluation structure is highly flexible and can be employed to represent various types of information. Different combinations of subjects, predicates, and objects can capture rich semantic relationships, making it suitable for diverse knowledge representation. In summary, the advantages of the triple evaluation structure lie in its flexibility, clear semantic representation, scalability, and effective expression of relationships, making it a key technology in the fields of knowledge representation and the semantic web.

Based on the CEKGRL model, business activity information is defined as G = (E,R,S), where E is the set of enterprise entities; R is the set of enterprise relationships; S ⊆ E × R × E denotes the set of triples, and the set of triples is represented by (h,r,t), h, r and t represents the head entity (enterprise name), relationship (activity information direction) and tail entity (business activity information), respectively. In addition, this paper uses c to denote the activity categories and defines structure-based and category-based entity representations, representing the entity representation learned from the triad and the entity representation obtained by introducing the category representation, respectively. The overall architecture of the CEKGRL model is shown in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Overall architecture of the CEKGRL model.

https://doi.org/10.1371/journal.pone.0304835.g002

Among them, the ellipse composed of diagonal circles represents the structure-based vector representation, the ellipse composed of lattice-like circles represents the category-based vector representation, the ellipse composed of solid circles represents the vector representation of relationships, and the ellipse composed of hollow circles represents the vector representation of entity categories, a represents the attention fraction, and in order to integrate the two representation types, the energy function is defined in this paper as shown in Eq (1).

(1)

E_ss = ∥h_s + r − t_s∥ is the energy functions obtained by using structure-based entity representations for head and tail entities; h_s and t_s are the structure-based representations for head and tail entities, respectively; and the hyperparameters β is used to adjust the importance of the category-based representations in the CEKGRL model. E_cc = ∥h_c + r − t_c∥ is the energy functions obtained by using category-based entity representations for head and tail entities. It should be noted that both the structure-based and category-based representations of entities use a unified relational representation r during the training process, ensuring that the vector representation space of both types can be unified by the same relational representation.

In the training process, the correlation between the entity category representation and the triadic relationship, i.e., the attention score, is first obtained through the attention mechanism, and then this attention score is used to weight and sum the category representation and use it as the category-based entity representation.

3.1.2 Associated information categories and activity representation.

Different categories of enterprise entities can portray entities from multiple perspectives, and the same entity will focus on its different categories of information under different relationships, which is manifested by the different semantic relevance between different categories of the same entity and different relationships. In order to effectively utilize the potential correlation between the relationship and entity categories in the triad, this paper calculates and obtains the similarity between them by the following attention mechanism.

The essence of the attention mechanism lies in the ability to learn to ignore irrelevant information and focus on the key information when comparing similarities. In this paper, we first construct a scaled dot product attention model, combining the CEKGRL model with relationships r as query vectors and categories c as both vectors key and value vectors. In the implementation process, in order to speed up the processing efficiency, the attention is calculated in the form of a matrix, and the representation vectors of multiple inter-enterprise relationships and their corresponding activity category representation vectors are stitched into a relationship matrix R and a category matrix C, respectively. Then, the weight matrices W^Q, W^K and W^V to be trained are introduced, and the matrix multiplication operations are performed on the weight matrix, the relationship matrix and the category matrix to obtain the matrices query, key and value corresponding to Q, K and V and the attention scores as shown in Eqs (2) and (3). (2) (3) where k is the vector dimension of entities and relationships, and d_k is the weight matrix dimension. The higher the attention score obtained by the above attention calculation model, the stronger the relevance of the category c to the relationship r, when the category is assigned a higher weight.

3.2.1 Contrasting entity generation.

In this paper, the entity ambiguity calculation is carried out for the example of business entity names and business transaction activities. Based on Stanford NER for named entity identification, the set of entity designations of the candidate comparison blocks is obtained and denoted as M = {m₁, m₂, m₃, …}. When performing entity ambiguity calculation, contextual activity information plays an important role as evidence for entity designation. Therefore, the Stanford NER tool is used to extract the activity context information D from the activity information of the removed entity designation set M, which constitutes the context set of the activity context information D, denoted as C = {c₁, c₂, c₃, …}.

The named entity designations in the set M are passed through other blocks to obtain the candidate comparison entity set, and the information extraction and formatting of the prepared candidate entity set are sorted out and filtered to generate the final candidate comparison entity set N_i = {m_i1, m_i2, m_i3, …}.

3.2.2 Contextual information association graph model construction.

In this paper, the contextual structure in an enterprise entity is represented as an entity-related graph model G = (V, E), where V denotes the vertex set and E denotes the edge set. The construction of the entity-related graph is divided into two steps: vertex set construction and edge set construction.

(1) Vertex set construction. Vertex set V: the set D of entity contextual activities that appear with a given enterprise activity, where c_i denotes the ith contextual information corresponding to this activity, V = {c_i |∀i, c_i ∈ C}.

Each vertex in the graph is assigned a Confidence Measure (CM), which indicates the importance of the node without considering other contextual information, and is calculated as shown in Eq (4). (4) Where ResultScore(c_i) is the matching score obtained based on Google Knowledge Graph, and the higher the matching score, the more accurate the contextual information represents.

(2) Side set construction. The edges of the graph model in this paper consist of the path correlations of the contextual information corresponding to this enterprise activity.

In this paper, we propose a new method to determine the path association degree between two messages, i.e., the two-way shortest path determination method, considering the hyperlink structure between messages as a directed graph, remembering that the path from message A link to message B is the forward shortest path, and the path from entity B link to entity A is the backward shortest path, and the shortest path between two nodes is calculated as shown in Eq (5). (5) Where FShortPath denotes the forward shortest path length and BShortPath denotes the backward shortest path length. In order to better describe the path correlation between two nodes, the route length is converted into path correlation and calculated as shown in Eq (6).

(6)

From this formula, the smaller the path length between two nodes, the greater the path association between the two nodes. According to the theory of "six degrees of separation", when exploring the shortest path length of two nodes, the upper limit of the shortest path length is set to six, that is, if the path length between two nodes exceeds six, the two nodes are considered to have no path association.

3.2.3 Inter-block similarity calculation.

In this paper, the semantic similarity is calculated using the name of the business entity and the specific content of the business transaction activities, using SimText(A,B) to denote the semantic similarity of the blocks represented by block A and block B, respectively, and vectorizing the semantic description information represented by block A and block B as A = {m₁₁,m₁₂,m₁₃,…} and B = {m₂₁,m₂₂,m₂₃,…}, respectively, using cosine similarity calculated as shown in Eq (7): (7)

The normalization process is shown in Eq (8): (8)

Finally, the total consistency of the blockchain is calculated using the average measure of the similarity of these blocks, and the average similarity is calculated as shown in Eq (9): (9) where N is the total number of blocks to be compared, i is the blocks other than the blocks to be compared, and SimText(A,i) is the similarity measure between the blocks to be compared and the other blocks.

3.4.1 Source-based information trustworthiness characterization.

(1) Credibility of information sources. In this paper, we assume that the higher the click-through rate of information sources, the higher the credibility. The click through rate is considered an important basis for measuring the credibility of information sources. Firstly, the click through rate reflects the level of interest and recognition of users towards the content of the information source. Users are more likely to click on information they consider useful, authentic, and trustworthy. Therefore, information sources with high click through rates often imply user trust in the source, thus becoming one of the indicators for evaluating credibility. Secondly, click through rate is a direct feedback of user behavior. The behavior of users clicking on a certain information source indicates that they are interested in the content of that source and believe that the content is valuable to them. This positive user behavior feedback is seen as an affirmation of the credibility of the information source. Due to the limitations of certain information sources (e.g., forums, blogs) for access restrictions and their own audience limitations, their click-through rates as a whole will be lower than those of official information sources (e.g., news sites, official websites), so the adverse effects of the above should be smoothed out. The information source credibility model is shown in Eq (10): (10) Where, Site_R(i) is the credibility of a certain information source, hits denotes the click rate, i denotes the columns related to business activities in the information source, m denotes the number of all columns in the information source, and n is the number of a certain type of information source. The model soothes out the heat by calculating the ratio of column clicks to the whole information source, reducing the rise in clicks on the business activity column due to interest in other information in a source.

(2) Credibility of information pages. To a certain extent, the web page where the information is located also characterizes the credibility of the information. In this paper, we propose a method to characterize the trustworthiness of a page based on the information of the relevant links on the page, i.e., the accessibility of the links in the page and the availability of the reached page to judge the trustworthiness of an information source. The specific model is shown in Eq (11): (11) where Page_R(i_A,B,C) denotes the credibility of the information page, A denotes the set of reachable links among the relevant links, B denotes the set of unreachable links among the relevant links, and C denotes the set of unavailable pages among the pages reached by the links. Reachability means that the link can link and open the specified page normally, and availability means that the new page linked to has sufficient information and relevance.

(3) Credibility of the information publisher. The credibility of the information publisher is mainly focused on the credibility judgment of the publisher, only for the credibility of the person to judge whether the information is credible, and then gradually converted to credible information users and untrustworthy information users after a certain knowledge reserve with the deepening of people’s understanding of information and the acquisition of knowledge.

In this paper, we consider the network users who can receive information as a whole, and the total number of users is denoted as N. Initially, the users who receive data are considered as unknowns ignorant; as the information spreads on the network, these users can be further divided into two categories: those who consider the information credible believed and those who consider the information untrustworthy unbelieved. Based on this, this paper defines a benchmark for measuring the credibility of the information itself, which is used to discern the probability that the publisher of the information releases the information, the probability of determining the truthfulness of the information when it is published.

Therefore, based on the above considerations, the user states in the network can be divided into three categories:

1. Ignorant. This state refers to the state in the network when a piece of information has not yet been disseminated; or, when the information is published, users who do not know the truth or falsity of the information and have not yet made a judgment.
2. Believed. After having a certain knowledge base, when seeing the information, make a judgment that the information released by the publisher is credible.
3. Unbelieved. After having a certain knowledge base, when seeing the information, make a judgment that the information released by the publisher is not credible.

The schematic diagram of state transition between network nodes during information dissemination is shown in Fig 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Inter-node state transition.

https://doi.org/10.1371/journal.pone.0304835.g003

Suppose a publisher publishes a message publicly, then the user who sees the message for the first time is regarded as an unknowing person, and the probability that the message is seen by other users in the unit time starting from moment t is α; in contrast, the probability that the user thinks the message is untrustworthy is β.

With the passage of time and the accumulation of their own learned knowledge and awareness, the users’ cognitive state of information will change. Therefore, assuming that the probability that a trusted user thinks the information is untrustworthy is γ per unit time starting from moment t, then it is shown that the trusted information user converts to an untrustworthy information user; at the same time, the probability that an untrustworthy information user thinks the information is trustworthy is δ, then the untrustworthy information user converts to a trusted information user. Therefore, based on the above information content-based interaction rules, a system of equations is established as shown in Eq (12). (12) Where: I(t), B(t) and U(t) denote the proportion of the total number of users at the time of the unknowns, trusted information users, and untrusted information users, respectively; I(0), B(0) and U(0) denote the value of the number of users in each state in the initial state, and all are non-negative numbers.

In the above interaction rule, assume that the user’s trustworthiness of the message publisher is Publish_R. When other users see the message, they can directly judge whether the message is trustworthy or not by their intuitive impression of the publisher of the message, then the average trustworthiness probability of the publisher is calculated as shown in Eq (13).

(13)

The probability of implausibility is shown in Eq (14): (14)

(4) The fusion of credibility characterization results of the source. The source-based credibility is calculated as shown in Eq (15).

(15)

3.4.2 Information trustworthiness characterization based on evaluation features.

Information credibility is assessed by calculating the extrinsic representations of the comments, which are first defined as follows.

Invalid comments: Comments that do not contain emotional words or are not related to the credibility of the content of the message.

Explicit comments: Comments that contain words that can clearly characterize the tendency of credibility.

Implicit comments: Comments that contain emotional words but do not have the tendency to explicitly characterize credibility words.

Explicit and implicit comments have different contributions to the evaluation of information credibility based on comments. These two types of comments reveal different attitudes and perspectives of users when facing information. By comprehensively analyzing them, the credibility of information can be more comprehensively evaluated. Explicit comments usually contain clear attitudes, opinions, or feedback from users. This direct expression helps to understand the user’s true view of the information, thereby providing direct clues about the credibility of the information. And Explicit comments often contain more facts, data, or specific information. This clear statement makes it easier to evaluate the accuracy of information and helps confirm whether the information is based on verifiable content. Implicit comments may involve emotional vocabulary, emotional color, or suggestive language, revealing the user’s emotional feedback. This is of great significance for evaluating the user experience and impact of information. Implicit comments can provide additional context or information for explicit comments. Sometimes users may not express themselves clearly in their words, but through the tone and emotions in implicit comments, they can gain a deeper understanding of their views. These implicit viewpoints help to gain a more comprehensive understanding of the user’s overall perception of information, providing more dimensions for credibility assessment.

1. Dispositional analysis of explicit commentary representations. In this paper, the sentiment intensity is adjusted by the modifier window strategy to control the representational value. If the current analyzed word is a sentiment word, the word is placed in the corresponding calculator based on the specific sentiment tendency, and the degree to which the current sentiment word is modified by the modifier words in the sliding window is taken out to obtain the value of the word’s contribution to the sentiment of the sentence. If the sliding window contains a negative word when the current sentiment word is analyzed, the sentiment tendency is reversed and the sliding window range is used to control the range of action of the negative word. Finally, the sentiment tendency is calculated for the whole comment tendency, which is calculated as shown in Eq (16). (16) where R_review_dominant denotes the value of a user’s comment on the credibility of the message, N denotes the number of sentences in the total comment, R denotes the score of credible representations in a sentence, D denotes the score of suspicious representations in a sentence, and L denotes the length of a sentence. When the single-sentence sentiment score lies in the (- 0.1, + 0.1) interval, the comment has a weak tendency to be characterized and is not considered.
2. Implicit commentary representation propensity analysis. The analysis of representational tendency for implicit comments is performed in two steps: first, the relevance of the comments to the message is calculated; second, the representational tendency value is calculated for the relevant implicit comments.
   1. Relevance of comments to information. In this paper, a probabilistic retrieval model is used to calculate the relevance of information and comments, i.e., comments and information are considered as retrieval problems. The set of information constitutes the set of documents when the comment C is assumed to be a query string and the information I is treated as a document. On the basis of the classical probabilistic retrieval model, it is smoothed as shown in Eq (17). (17) where I is the message, C is an entry in the set of comments of I, w is a word in C, p(C|I) is the probability of I generating C, and p(w|I) is the probability of w appearing in I. t_I is the set of topics of message I, t is a topic in t_I, λ is a parameter, p(w|t) is the probability of word w appearing in topic t, and p(t|I) is the probability of topic t appearing in message I. p(C|I) is calculated for all comments in message I. If the probability value is greater than a certain threshold, the relevant comment is determined to be relevant to the message.
   2. Related implicit commentary on representational tendencies. In addition to explicit emotion words, some implicit words also have attitudinal characteristics, and for implicit evaluation representation tendency values are calculated as shown in Eq (18). (18) where a denotes a paragraph, sentence or word, and R corresponds to a whole comment, paragraph or sentence, the sentiment score of R is denoted as w(R), p(a_i) denotes the position of the sentiment unit a_i in R, and count(a|R) denotes the total number of R containing a.
3. Comment-based fusion of credibility characterization results. The process of fusion is divided into two steps: first, the objects of the comments are obtained and classified; second, the results of the classified comments are fused in the way shown in Eq (19). (19) where R_review is the trustworthiness of the message based on the comment, Tar(R_i) is the characterization of the trustworthiness of the message by the object i being commented on, R_j is the characterization of the trustworthiness of a comment j on i, and R_k is the characterization of the trustworthiness of a direct comment on the message.

3.4.3 Information trustworthiness characterization based on content features.

Existing methods of content credibility calculation assess the content of a certain type of activity information with the help of external information and lack a direct and reliable evaluation algorithm for information from multiple sources. This paper proposes a method for characterizing information trustworthiness using data content features based on a study of enterprise activity information characteristics, which is based on two assumptions:

1. Feature words determine the professionalism of information, i.e. the more words characterizing enterprise activities in the information content, the more professional the information is.
2. The professionalism of the information determines the credibility of the information, i.e. the more professional the information is, the more credible it is.

Considering the uneven distribution of corporate activity information features in information from multiple sources, this paper uses a corporate activity feature dictionary to count the frequency of corporate activity feature words in various types of information. The credibility of the content is calculated as shown in Eq (20). (20) where word_f, word_s and word_t denote the number of common, characteristic and specialized words occurring in the message, respectively, α, β and λ denote the associated weights, and μ is the information type characteristic coefficient.

Finally, information trustworthiness in the blockchain is a combined measure of source-based trustworthiness, comment-based trustworthiness and content-based trustworthiness, as shown in Eq (21).

(21)

Among them α < γ < β, since the information source is the key factor to measure the credibility of information, such as official information is 100% credible, the credibility of the information source has the largest weight, the content credibility is the second, and the comment-based credibility has the smallest weight since the comment depends on the subjectivity of the commenter. Some situation may affect the method like different fields may have different professional standards and vocabulary usage habits. Some fields may place more emphasis on the use of professional terminology, while others may place more emphasis on common language. Therefore, the relationship between the number of feature words and professionalism may vary between different fields. Sometimes, the number of feature words does not always reflect the depth and detail level of information. Some content may require more professional terminology and detailed explanations, which may not necessarily lead to an increase in the number of feature words. Therefore, quantity cannot fully represent the professionalism and credibility of the text. But in most cases, these issues will not have a significant impact on the conclusions of the model output, so no additional discussion is needed.