A visual analysis method for large-scale news source media bias identification

Abstract

This invention discloses a visual analysis method for identifying media bias in large-scale news sources, comprising the following steps: Step 1, constructing a data corpus based on the Global Open News Dataset (GDELT); Step 2, constructing time-series encoding and dimensionality reduction representation of reporting features based on related reports from large-scale media sources; Step 3, constructing hierarchical media cluster visualization results; Step 4, generating visualization results of media reporting differences on different event types based on representative media selected by the user; Step 5, constructing the event development context of related news reports by selected media on specific topics; Step 6, allowing users to interactively compare and analyze the continuous evolution pattern of media bias in the development path of specific events based on the event development context. This invention provides users with an intuitive interactive platform to explore media bias identification results, reduces the workload of users reading and understanding news reports from different media sources, improves research efficiency, and helps users gain a deeper understanding of media bias.

Description

Technical Field

[0001] This invention belongs to the field of visualization and human-computer interaction technology, and in particular relates to a visual analysis method for identifying media bias in large-scale news sources. Background Technology

[0002] Currently, media bias is a crucial and unavoidable problem in the news industry, significantly impacting public perception and understanding of key international public issues. To provide the public with accurate and truthful information, journalists must identify and understand media bias. Visualizing media reporting data helps users intuitively understand media bias and form their own objective understanding of public events.

[0003] The Weibo public opinion visualization analysis method proposed in Chinese invention patent application number CN202210201566.5 is as follows: S1: Determine the Weibo accounts to be crawled and obtain their Weibo data during the public opinion period; extract the required research fields and preprocess the obtained research corpus; S2: Use a Bayesian model to classify the Weibo data into topics, then use IDF feature extraction and LDA topic model to mine the classified corpus into text topics, extract Weibo keywords, and finally combine the dissemination attributes such as the number of reposts to obtain hot topics; S3: Extract Weibo comments on hot topics in each time period to obtain the final classification results and explore people's sentiment trends on hot topics; S4: Perform visual analysis of Weibo public opinion during the two public opinion periods.

[0004] Existing media bias visualization analysis technologies target a small number of media organizations, using keyword information to summarize media news reports and employing clustering techniques to compare different media organizations' perspectives on independent, specific events, thus studying the diversity of news articles. These technologies, to some extent, present the diversity of media coverage of specific events, allowing users to identify different media biases without repeatedly reading large amounts of original news text, thereby improving analytical efficiency. However, existing media bias visualization analysis technologies cannot analyze media bias across a large range of media organizations. Analyzing media bias across a wide range of media organizations can provide users with a more comprehensive and objective understanding of different media reporting tendencies, reducing the impact of media bias on user perception. Existing media bias visualization analysis technologies cannot support the analysis of large-scale news sources. Secondly, existing media bias visualization analysis technologies tend to analyze media bias based on independent, specific events, ignoring the continuity of biased reporting by media organizations on certain topics or events. For example, a specific event may lead to a chain reaction, and the media may follow up with reports on subsequent events, which may continue for hours, days, or weeks. As time goes on, new information and details may emerge, making the reporting more complete and detailed. However, existing methods only analyze media bias between media outlets based on related news reports of independent events, and therefore cannot support comprehensive media bias analysis. Summary of the Invention

[0005] The purpose of this invention is to provide a visual analysis method for identifying media bias in large-scale news sources that can overcome the aforementioned technical problems. This invention includes the following steps:

[0006] Step 1: Construct a data corpus based on the Global News Open Dataset (GDELT):

[0007] The corpus for the Global News Open Dataset (GDELT) is constructed in two parts: first, the construction of a global news reporting event database; and second, the construction of a news reporting corpus to support users in exploring and analyzing media biases under different news topics.

[0008] Step 2: Construct time-series encoding and dimensionality-reduction representation of report features based on relevant reports from a large number of media sources:

[0009] The construction of global media temporal coding and dimensionality reduction representation is based on the following three requirements: First, dynamism, meaning the model should be able to describe the dynamic changes in media reporting bias over different time periods; second, simplicity, meaning the relative distance between points in a two-dimensional plane should be used to describe the extent of media reporting bias; and third, timeliness, meaning the model should be able to quickly obtain a two-dimensional distribution map of global media bias based on the user's selected time period of interest, including the following two steps:

[0010] Step 2.1: Construct a global media reporting bias design space. Based on the GDELT open-source data, select several time-series attributes representing media reporting bias: the number of reported events, the number of times the reported events are mentioned, the number of articles mentioning the reported events, the number of links mentioning the reported events, the degree of impact of the reported events on the region, and the attitude towards the reported events. Use an autovariable time-series encoder model to model the complex joint distribution of media reports as a latent feature representation to maintain the continuity of news reporting over time. That is, set a set of n initial integrated media feature vectors X = {x1, x2, ..., x...} n}, where x i This refers to the feature vector of the i-th news media report. The variable-time encoder model uses the hidden state h of the previous model as a basis. t and the vector set X at the corresponding time step t+1 Calculate state h t+1 As shown in the following formula:

[0011]

[0012] in, and Let b be the model weight matrix. r Let h be a constant vector, and the final state h of the model be... end The linear layer is mapped to the mean μ and standard deviation σ, as shown in the following formula:

[0013]

[0014]

[0015] in, and Let b be the model weight matrix. m With b s Given a constant vector, reparameterization is performed based on the mean and standard deviation, and a latent vector z is sampled from the distribution defined by the mean and standard deviation. The model passes the sampled latent vector to obtain the initial state of the decoder, as shown in the following formula:

[0016]

[0017]

[0018] The decoder is updated using backpropagation as shown in the following formula:

[0019]

[0020] in, and Let b be the model weight matrix. z bc With b out It is a constant vector;

[0021] Finally, the loss function of the network is constructed for joint training, as shown in the following formula:

[0022]

[0023] By training a multi-attribute autovariable temporal encoder model, the design space of global media reporting bias within a certain time period can be quickly obtained to meet the dynamic and timeliness requirements of media bias design space construction.

[0024] Step 2.2: Construct a media dimensionality reduction representation. Keep the relative distance in the original media deviation design space unchanged, and reduce the high-dimensional media encoding in the original design space to a two-dimensional plane. This will make media nodes with similar reporting habits relatively close to each other in the two-dimensional plane, and make media nodes with different reporting habits relatively far apart.

[0025] Step 3, construct a hierarchical media clustering visual result:

[0026] Step 3.1: Based on the two-dimensional representation of media reporting bias obtained in Step 2, representative media at different levels are selected by using hierarchical clustering and considering the number of media reports. This allows users to better analyze media reporting bias between and within clusters, enabling users to compare different media bias cluster representations at different exploration levels and achieving hierarchical exploration of media reporting bias clusters and bias cluster location based on fuzzy matching of media names. Step 3.1: Constructing a hierarchical media bias projection. Based on the two-dimensional representation of media reporting bias in Step 2, a distance-based hierarchical clustering method is used to construct a hierarchical projection of global news media bias by dividing the minimum cluster set. The media nodes at each level represent media with similar reporting bias distances under the current level's clustering strength.

[0027] Step 3.2, Hierarchical exploration of media bias clusters: Combining the hierarchical characteristics of media bias projection, different levels of exploration granularity are set during user exploration. The user's mouse zoom determines the level of exploration, thereby displaying representative media nodes at different levels. By moving the mouse over each media node, the user can see the similarities and differences in the number of news events reported, the number of times the reported events are mentioned, the number of articles mentioning the reported events, the number of links mentioning the reported events, the degree of impact of the reported events on the region, and the attitudes towards the reported events.

[0028] Step 3.3, based on the fuzzy matching of media names, the report deviation cluster positioning provides a fuzzy search function for media that the user is interested in. The user enters the domain name of the media they are interested in, and the fuzzy matching gives the matching results. The user can click on the matching results to automatically focus on the cluster of media that the user is interested in, so as to facilitate the user's exploration and use.

[0029] Step 4: Generate visualizations of media coverage differences across different event types based on the representative media selected by the user.

[0030] The topic comparison view (in this invention, it refers to the visualization results of the differences in media reporting on different event types) uses the CAMEO event category coding information from the GDELT database to visualize the category distribution of selected media reports in a barcode tree visualization format, and reflects the hierarchical structure of event categories. It supports users to click to interactively explore the distribution of specific event categories of specific media and to compare media. The topic comparison view in this invention refers to the visualization results of the differences in media reporting on different event types.

[0031] Using information on event initiators and recipients from the GDELT database, the topic comparison view uses a matrix-style visualization to show the media attitudes of different media outlets when reporting on specific subjects as both initiators and recipients of events. The matrix visualization is linked to the event selection component, meaning that as the user selects a different event category, the event category in the matrix visualization will also change accordingly.

[0032] Step 5: Construct the event development timeline of relevant news reports from selected media outlets on a specific topic:

[0033] Step 5.1: Media bias on a specific topic is reflected in overall news reporting over a period of time, rather than isolated news articles. To analyze the media's consistent support or opposition to a particular viewpoint when reporting on a specific topic, users need complete contextual information about the media over a time span. Various types of information should be presented, including complex topological structures, event evolution, event summaries, and summary text information, so that users can analyze the evolution of their news reports and thus analyze media bias. A path-aware story construction algorithm is used to process the input news documents and to perform topic clustering by repeatedly removing the edges with the highest media centrality through the construction of a keyword co-occurrence graph. Based on the extracted topics, a story construction algorithm based on path awareness is used to construct a story framework and combined with visualization methods to allow users to further analyze media bias from the evolution of media reports.

[0034] Step 5.2, storyline construction is an iterative process. For each newly generated event node in each time period, the storyline construction algorithm comprehensively evaluates the compatibility of placing the event node in all positions on the storyline. The algorithm iteratively constructs the eventline based on four factors: time interval, node consistency, path consistency, and topic consistency. That is, for each incoming event E... new And the existing storyline T, through event E new Event nodes are continuously added to the existing storyline by placing them in appropriate positions. Two types of operations are used to put incoming events into the storyline: merging and connecting. First, all nodes on the storyline are traversed to evaluate whether new events and existing events in the storyline represent the same theme. This is achieved by calculating the cosine similarity between the TF-IDF vectors of two documents. When the similarity exceeds a certain threshold, the new event E is added. new With existing events E in the tree j Merge; otherwise, continue iterating and calculating the new event E. new With the existing events E in each network j The connection probability P between con The connection probability is calculated based on three factors: (1) the time interval between the two events. gap (2) Node consistency between two event nodes sim , and (3) when event E new E connected in the event timeline j Path consistency p in subsequent times coh The longer the time interval between two events, the lower their correlation becomes. The time interval is calculated as follows:

[0035]

[0036] in, It is event E j The timeline of the event It is event E new The time nodes of events are used to evaluate node consistency and path consistency by vectorizing events using TF-IDF and assessing the cosine similarity between event nodes and story context nodes, as well as adjacent nodes along the path. The node consistency calculation method is as follows:

[0037]

[0038] Among them, TF-IDF(d j ) and TF-IDF(d new ) are events d j and d new TF-IDF vectorization representation;

[0039] The path consistency is calculated as follows:

[0040]

[0041] Among them, L T,j,new Indicates when event E new E connected in the event timeline j Then, from the root node of T to node E new The path, |L T,j,new | indicates the length of the path;

[0042] Reduce the dimensionality of all node vectors along the path using SVD, and then combine them with the new event node E. new Calculate cosine similarity and evaluate the new event node E using cosine similarity. new The compatibility with the overall topic along the branching paths of the storyline is used to determine whether the placement of new event nodes in the storyline is appropriate to maintain topic consistency. Finally, the eventline is iteratively constructed based on four factors: time interval, node consistency, path consistency, and topic consistency.

[0043] Step 6: Users interact and compare the continuous evolution pattern of media bias in the specific event development path based on the event development context.

[0044] Based on the event development storyline constructed in step 5 and the data corpus constructed in step 1, the visualization results of the event development storyline are displayed, while also supporting interactive exploration by users:

[0045] Step 6.1, Display of the visualization results of the event development timeline:

[0046] Each node in the event development timeline represents an aggregated event; the size of the node in the event development timeline represents the magnitude of the selected media's reporting deviation on the current aggregated event; the thickness of the line connecting two aggregated nodes represents the closeness of the event's development and evolution relationship; the size of different sectors on a node represents the proportion of articles published by the selected media on the current aggregated event; the node layout of the event development timeline is based on the layout method of the original node link tree, and time information mapping the event development is added in the horizontal direction to represent the development and evolution trend of aggregated event nodes; the development and changes of different branches on the timeline represent the development and evolution of different sub-topics; users can understand the evolution pattern of the selected media's reporting deviation on a specific topic by comparing and analyzing the changes in node size, node sector size, and the thickness of the line connecting nodes on different branches.

[0047] Step 6.2, user analysis of media bias on a single specific event:

[0048] Users can hover their mouse over an aggregated event node in the event timeline to see the number of news reports about the event from their selected media outlets, the number of times the event was mentioned in those reports, the number of articles mentioning the event, the number of links linking to the event, the impact of the event on the region, and the differences in attitudes towards the event. Clicking on an aggregated event node displays representative articles from media outlets most similar to the user's chosen event. By reading these representative articles, users can understand the differences in media coverage of the same event, while the large-scale pre-trained language model GPT-3 generates summaries of the representative articles to summarize the current aggregated event. Clicking on the title of a representative article redirects to the original article page, allowing users to trace the origins of the news event.

[0049] Step 6.3: The user analyzes the continuous evolution pattern of media deviation along different paths:

[0050] When a user clicks on a specific aggregated event node, a media reporting event line visualization is displayed, showing the path from the root node of the event's development to the current aggregated event node. This visualization lists all news reports related to the aggregated event along the current path, chronologically arranged by the media's reporting timeline. Each news report is mapped to a square, with the side length representing the article's impact on the region and the color representing the article's sentiment. News reports belonging to the same aggregated event are interconnected, as are news articles from the same media outlet. This allows for differentiation between different aggregated events and analysis of the changing trends in the impact of articles published by different media outlets on the same event. The media reporting event line visualization supports article viewing and source tracing. By exploring the media reporting event line, users can analyze the development status of a branch of the event's development and view the changing trends in the impact of articles reported by different media outlets on the region, as well as the changes in sentiment towards the same event.

[0051] The method described in this invention has the following beneficial effects:

[0052] 1. The method described in this invention supports a comprehensive analysis and understanding of media bias in large-scale media from multiple perspectives: the exploration of a global overview of media bias in large-scale news reporting enables users to not only selectively choose media sets with significant differences, but also search for and match media sets of interest. Based on this, combined with statistical data from the data corpus, a barcode tree view with hierarchical structure attributes is created to compare the number of reports on event categories, more intuitively reflecting the differences in reporting by different media on different event categories. At the same time, the visualization results based on matrix form show the media attitudes of different media when reporting on different event categories and different countries. A path-aware story context construction algorithm is proposed, combined with visualization analysis methods, to allow users to analyze media bias from the evolution of media reports.

[0053] 2. The method described in this invention has the capability of large-scale media bias analysis: a large-scale news media corpus is constructed based on the global news media open-source dataset GDELT. The media is encoded and dimensionality reduced using an autovariable temporal encoder model according to the reporting trends of the media on several characteristic attributes. The media mapping in the two-dimensional space reflects the design space of media bias. By constructing a large-scale media hierarchical exploration structure and realizing the search and matching of specific media, the method has the capability of large-scale media bias analysis.

[0054] 3. The method described in this invention has the ability to discover media reporting biases in different event categories: The topic comparison view uses a barcode tree visualization to present the category distribution of selected media in reporting and reflects the hierarchical structure of event categories. It supports users to click to interactively explore the distribution of specific event categories in specific media and compare media. It also supports users to analyze the emotional bias of media reporting on different countries. Furthermore, the interactive exploration allows users to gain a deeper understanding of the global overview of large-scale media biases and the reporting habits of local media clusters.

[0055] 4. The method described in this invention has the ability to analyze the continuity of reporting deviations for certain topics or events: It uses an autovariable time-series encoder model to model the complex joint distribution of media reports in order to maintain the continuity of news media reports over time, enabling users to identify and analyze the persistent deviations in news media reports. This invention is the first in the technical field to use a path-aware storyline construction algorithm, which constructs a storyline based on the media set selected by the user. The overall development of the storyline represents the development pattern of the news topic, and different sub-branches represent the development and evolution of different sub-topics. Combined with visualization methods, it allows users to further analyze media deviations from the evolution of media reports. It supports interactive exploration of different paths on the path-aware storyline. News articles on the media report event line support the functions of viewing and tracing the source of articles. It allows users to explore the media report event line, analyze the event development status of a certain branch on the event development line, and view the changing trend of the impact of articles reported by different media on the region and the changes in the emotional attitude towards the same event.

[0056] 5. The method described in this invention has the ability to analyze media bias on independent specific events: it constructs a path-aware storyline, where nodes refer to the media reports on the same event selected by the user, which can be regarded as an aggregated event node. The size of the node represents the reporting bias of different media on the same event, supporting interactive exploration by users on specific aggregated events.

[0057] 6. The method described in this invention constructs a global overview of media biases existing among a large number of media outlets, supports users to interactively select representative media of interest, and generates a visual overview of the number of reports and the sentiment of the reports based on the representative media selected by the user, in order to assist users in identifying and analyzing biased media. In addition, a graph-based clustering method is used to summarize related news reports into events, and an overview of the development of media bias events is constructed. This not only supports users in exploring media bias in the reporting of individual events, but also supports exploring the continuous evolution pattern of media reporting bias, thereby improving the effectiveness of understanding media bias and identifying the underlying causes. Attached Figure Description

[0058] Figure 1 This is a flowchart of the method described in this invention;

[0059] Figure 2 This is a schematic diagram illustrating the overall effect of the media reporting model of the method described in this invention;

[0060] Figure 3 This is a schematic diagram illustrating the event development storyline effect of the method described in this invention. Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0062] This invention includes the following steps:

[0063] Step 1: Construct a data corpus based on the Global News Open Dataset (GDELT):

[0064] The corpus for the Global News Open Dataset (GDELT) is constructed in two parts: first, the construction of a global news reporting event database; and second, the construction of a news reporting corpus to support users in exploring and analyzing media biases under different news topics.

[0065] Step 2: Construct time-series encoding and dimensionality-reduction representation of report features based on relevant reports from a large number of media sources:

[0066] The construction of global media temporal coding and dimensionality reduction representation is based on the following three requirements: First, dynamism, meaning the model should be able to describe the dynamic changes in media reporting bias over different time periods; second, simplicity, meaning the relative distance between points in a two-dimensional plane should be used to describe the extent of media reporting bias; and third, timeliness, meaning the model should be able to quickly obtain a two-dimensional distribution map of global media bias based on the user's selected time period of interest, including the following two steps:

[0067] Step 2.1: Construct a global media reporting bias design space. Based on the GDELT open-source data, select several time-series attributes representing media reporting bias: the number of reported events, the number of times the reported events are mentioned, the number of articles mentioning the reported events, the number of links mentioning the reported events, the degree of impact of the reported events on the region, and the attitude towards the reported events. Use an autovariable time-series encoder model to model the complex joint distribution of media reports as a latent feature representation to maintain the continuity of news reporting over time. That is, set a set of n initial integrated media feature vectors X = {x1, x2, ..., x...} n}, where x i This refers to the feature vector of the i-th news media report. The variable-time encoder model uses the hidden state h of the previous model as a basis. t and the vector set X of the corresponding time step t+1 Calculate state h t+1 As shown in the following formula:

[0068]

[0069] in, and Let b be the model weight matrix. r Let h be a constant vector, and the final state h of the model be... end The linear layer is mapped to the mean μ and standard deviation σ, as shown in the following formula:

[0070]

[0071]

[0072] in, and Let b be the model weight matrix. m With b s Given a constant vector, reparameterization is performed based on the mean and standard deviation, and a latent vector z is sampled from the distribution defined by the mean and standard deviation. The model passes the sampled latent vector to obtain the initial state of the decoder, as shown in the following formula:

[0073]

[0074]

[0075] The decoder is updated using backpropagation as shown in the following formula:

[0076]

[0077] in, and Let b be the model weight matrix. z b c With b out It is a constant vector;

[0078] Finally, the loss function of the network is constructed for joint training, as shown in the following formula:

[0079]

[0080] By training a multi-attribute autovariable temporal encoder model, the design space of global media reporting bias within a certain time period can be quickly obtained to meet the dynamic and timeliness requirements of media bias design space construction.

[0081] Step 2.2: Construct a media dimensionality reduction representation. Keep the relative distance in the original media deviation design space unchanged, and reduce the high-dimensional media encoding in the original design space to a two-dimensional plane. This will make media nodes with similar reporting habits relatively close to each other in the two-dimensional plane, and make media nodes with different reporting habits relatively far apart.

[0082] Step 3, construct a hierarchical media clustering visual result:

[0083] Step 3.1: Based on the two-dimensional representation of media reporting bias obtained in Step 2, representative media at different levels are selected by using hierarchical clustering and considering the number of media reports. This allows users to better analyze media reporting bias between and within clusters, enabling users to compare different media bias cluster representations at different exploration levels and achieving hierarchical exploration of media reporting bias clusters and bias cluster location based on fuzzy matching of media names. Step 3.1: Constructing a hierarchical media bias projection. Based on the two-dimensional representation of media reporting bias in Step 2, a distance-based hierarchical clustering method is used to construct a hierarchical projection of global news media bias by dividing the minimum cluster set. The media nodes at each level represent media with similar reporting bias distances under the current level's clustering strength.

[0084] Step 3.2, Hierarchical exploration of media bias clusters: Combining the hierarchical characteristics of media bias projection, different levels of exploration granularity are set during user exploration. The user's mouse zoom determines the level of exploration, thereby displaying representative media nodes at different levels. By moving the mouse over each media node, the user can see the similarities and differences in the number of news events reported, the number of times the reported events are mentioned, the number of articles mentioning the reported events, the number of links mentioning the reported events, the degree of impact of the reported events on the region, and the attitudes towards the reported events.

[0085] Step 3.3, based on the fuzzy matching of media names, the report deviation cluster positioning provides a fuzzy search function for media that the user is interested in. The user enters the domain name of the media they are interested in, and the fuzzy matching gives the matching results. The user can click on the matching results to automatically focus on the cluster of media that the user is interested in, so as to facilitate the user's exploration and use.

[0086] Step 4: Generate visualizations of media coverage differences across different event types based on the representative media selected by the user.

[0087] The topic comparison view (in this invention, it refers to the visualization results of the differences in media reporting on different event types) uses the CAMEO event category coding information from the GDELT database to visualize the category distribution of selected media reports in a barcode tree visualization format, and reflects the hierarchical structure of event categories. It supports users to click to interactively explore the distribution of specific event categories of specific media and to compare media. The topic comparison view in this invention refers to the visualization results of the differences in media reporting on different event types.

[0088] Using information on event initiators and recipients from the GDELT database, the topic comparison view uses a matrix-style visualization to show the media attitudes of different media outlets when reporting on specific subjects as both initiators and recipients of events. The matrix visualization is linked to the event selection component, meaning that as the user selects a different event category, the event category in the matrix visualization will also change accordingly.

[0089] Step 5: Construct the event development timeline of relevant news reports from selected media outlets on a specific topic:

[0090] Step 5.1: Media bias on a specific topic is reflected in overall news reporting over a period of time, rather than isolated news articles. To analyze the media's consistent support or opposition to a particular viewpoint when reporting on a specific topic, users need complete contextual information about the media over a time span. Various types of information should be presented, including complex topological structures, event evolution, event summaries, and summary text information, so that users can analyze the evolution of their news reports and thus analyze media bias. A path-aware story construction algorithm is used to process the input news documents and to perform topic clustering by repeatedly removing the edges with the highest media centrality through the construction of a keyword co-occurrence graph. Based on the extracted topics, a story construction algorithm based on path awareness is used to construct a story framework and combined with visualization methods to allow users to further analyze media bias from the evolution of media reports.

[0091] Step 5.2, storyline construction is an iterative process. For each newly generated event node in each time period, the storyline construction algorithm comprehensively evaluates the compatibility of placing the event node in all positions on the storyline. The algorithm iteratively constructs the eventline based on four factors: time interval, node consistency, path consistency, and topic consistency. That is, for each incoming event E... new And the existing storyline T, through event E new Event nodes are continuously added to the existing storyline by placing them in appropriate positions. Two types of operations are used to put incoming events into the storyline: merging and connecting. First, all nodes on the storyline are traversed to evaluate whether new events and existing events in the storyline represent the same theme. This is achieved by calculating the cosine similarity between the TF-IDF vectors of two documents. When the similarity exceeds a certain threshold, the new event E is added. new With existing events E in the tree j Merge; otherwise, continue iterating and calculating the new event E. new With the existing events E in each network j The connection probability P between con The connection probability is calculated based on three factors: (1) the time interval between the two events. gap(2) Node consistency between two event nodes sim , and (3) when event E new E connected in the event context j Path consistency p in the later stages coh The longer the time interval between two events, the lower their correlation becomes. The time interval is calculated as follows:

[0092]

[0093] in, It is event E j The timeline of the event It is event E new The time nodes of events are used to evaluate node consistency and path consistency by vectorizing events using TF-IDF and assessing the cosine similarity between event nodes and story context nodes, as well as adjacent nodes along the path. The node consistency calculation method is as follows:

[0094]

[0095] Among them, TF-IDF(d j ) and TF-IDF(d new ) are events d j and d new TF-IDF vectorization representation;

[0096] The path consistency is calculated as follows:

[0097]

[0098] Among them, L T,j,new Indicates when event E new E connected in the event context j Then, from the root node of T to node E nww The path, |L T,j,new | indicates the length of the path;

[0099] Reduce the dimensionality of all node vectors along the path using SVD, and then combine them with the new event node E. new Calculate cosine similarity and evaluate the new event node E using cosine similarity. new The compatibility with the overall topic along the branching paths of the storyline is used to determine whether the placement of new event nodes in the storyline is appropriate to maintain topic consistency. Finally, the eventline is iteratively constructed based on four factors: time interval, node consistency, path consistency, and topic consistency.

[0100] Step 6: Users interact and compare the continuous evolution pattern of media bias in the specific event development path based on the event development context.

[0101] Based on the event development storyline constructed in step 5 and the data corpus constructed in step 1, the visualization results of the event development storyline are displayed, while also supporting interactive exploration by users:

[0102] Step 6.1, Display of the visualization results of the event development timeline:

[0103] Each node in the event development timeline represents an aggregated event; the size of the node in the event development timeline represents the magnitude of the selected media's reporting deviation on the current aggregated event; the thickness of the line connecting two aggregated nodes represents the closeness of the event's development and evolution relationship; the size of different sectors on a node represents the proportion of articles published by the selected media on the current aggregated event; the node layout of the event development timeline is based on the layout method of the original node link tree, and time information mapping the event development is added in the horizontal direction to represent the development and evolution trend of aggregated event nodes; the development and changes of different branches on the timeline represent the development and evolution of different sub-topics; users can understand the evolution pattern of the selected media's reporting deviation on a specific topic by comparing and analyzing the changes in node size, node sector size, and the thickness of the line connecting nodes on different branches.

[0104] Step 6.2, user analysis of media bias on a single specific event:

[0105] Users can hover their mouse over an aggregated event node in the event timeline to see the number of news reports about the event from their selected media outlets, the number of times the event was mentioned in those reports, the number of articles mentioning the event, the number of links linking to the event, the impact of the event on the region, and the differences in attitudes towards the event. Clicking on an aggregated event node displays representative articles from media outlets most similar to the user's chosen event. By reading these representative articles, users can understand the differences in media coverage of the same event, while the large-scale pre-trained language model GPT-3 generates summaries of the representative articles to summarize the current aggregated event. Clicking on the title of a representative article redirects to the original article page, allowing users to trace the origins of the news event.

[0106] Step 6.3: The user analyzes the continuous evolution pattern of media deviation along different paths:

[0107] When a user clicks on a specific aggregated event node, a media reporting event line visualization is displayed, showing the path from the root node of the event's development to the current aggregated event node. This visualization lists all news reports related to the aggregated event along the current path, chronologically arranged by the media's reporting timeline. Each news report is mapped to a square, with the side length representing the article's impact on the region and the color representing the article's sentiment. News reports belonging to the same aggregated event are interconnected, as are news articles from the same media outlet. This allows for differentiation between different aggregated events and analysis of the changing trends in the impact of articles published by different media outlets on the same event. The media reporting event line visualization supports article viewing and source tracing. By exploring the media reporting event line, users can analyze the development status of a branch of the event's development and view the changing trends in the impact of articles reported by different media outlets on the region, as well as the changes in sentiment towards the same event.

[0108] Figure 1 This document illustrates the interactive flowchart between various parts of the method described in this invention, including corpus construction, media report feature representation, news media cluster visualization, event category analysis, storyline construction, and storyline visualization. A large-scale news media corpus is constructed based on the global news media open-source dataset GDELT. Media representations are built based on the constructed corpus and the reporting characteristics of different news media for specific events. A hierarchical media cluster visualization is constructed based on the dimensionality reduction representation of the constructed media report features. Users can explore and discover news media that may have media bias based on the visualization results, and generate visualization results of the reporting differences of news media on different event types based on the selected media. Based on the visualization results of the reporting differences on different event types, users can select the relevant event types they want to explore. Based on the user's selection, the system automatically constructs the news event development path of relevant news reports under specific topics and provides users with visualization results of the news event development path, enabling users to compare and analyze the continuous evolution pattern of media bias in the development path of specific events based on the visualization results.

[0109] Figure 2 and Figure 3 The diagram illustrates the technical effects of the system implementation in a specific embodiment of the present invention.

[0110] Specifically Figure 2The overview view of media reporting patterns is shown: the closer the media are, the more similar their reporting patterns are; otherwise, the greater the difference in their reporting patterns, the more similar their media reporting patterns are. The clustering results in the overview view reflect the media sets with different reporting habits. Users can view the media distribution overview at different levels by zooming in and out with the mouse. In addition, users can locate the clustering position of media by searching for specific media.

[0111] Figure 3 The presentation shows an event development storyline view: In the story tree view, each pie chart represents a specific event node, which includes one or more news articles from one or more media outlets. The degree of deviation of the selected media's news coverage of a specific event is represented by the radius of the event node, and the width of the story tree's edge represents the connection probability between events. When the mouse hovers over an event node, the path from the event node to the root node is highlighted in black. Users can click on event nodes to select paths and focus on analyzing the different stances of the media on a specific event within a specific path. Based on the event development storyline, users can not only analyze media bias on a single specific event, but also analyze the continuous evolution pattern of media bias on different paths.

[0112] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A visual analysis method for identifying media bias in large-scale news sources, characterized in that, Includes the following steps: Step 1: Construct a data corpus based on the Global News Open Dataset (GDELT). The construction of the data corpus based on the Global News Open Dataset (GDELT) includes two parts: first, the construction of a global news reporting event database; and second, the construction of a news reporting corpus to support users in exploring and analyzing media biases under different news topics. Step 2: Construct time-series coding and dimensionality reduction representation of reporting features based on reports from a large number of media sources. A variational autoencoder is used to construct global media time-series coding and dimensionality reduction representation based on the following three requirements: First, dynamism, meaning the model should be able to describe the dynamic changes in media reporting deviations over different time periods; second, simplicity, meaning the relative distance between points in a two-dimensional plane is used to describe the distance of different media reporting deviations; and third, timeliness, meaning the model should be able to quickly obtain a two-dimensional distribution map of global media deviations based on the user's selected time period of interest. Step 3: Construct a hierarchical media cluster visualization result. Based on the two-dimensional representation of media reporting deviation obtained in Step 2, representative media at different levels are selected by hierarchical clustering method and considering the number of media reports. This allows users to better analyze the media reporting deviation between and within clusters, enabling users to compare different media deviation cluster representations at different exploration levels and realize hierarchical exploration of media reporting deviation clusters and deviation cluster positioning based on fuzzy matching of media names. Step 4: Generate a visualization of the differences in media reporting on different event types based on the representative media selected by the user. In the topic comparison view, the event classification component uses the CAMEO event category coding information from the GDELT database to visualize the category distribution of the selected media in terms of reporting using a barcode tree visualization format, and reflects the hierarchical structure of event categories. It supports users to click to interactively explore the distribution of specific event categories of specific media and to compare media. In this invention, the topic comparison view refers to the generated visualization of the differences in media reporting on different event types. Using information on event initiators and recipients from the GDELT database, the topic comparison view uses a matrix-style visualization to show the media attitudes of different media outlets when reporting on specific subjects as both initiators and recipients of events. The matrix visualization is linked to the event selection component, meaning that as the user selects a different event category, the event category in the matrix visualization will also change accordingly. Step 5: Construct the event development timeline of relevant news reports from selected media outlets on a specific topic: Step 5.1: Media bias on a specific topic is reflected in overall news reporting over a period of time, rather than isolated news articles. To analyze the media's consistent tendency to support or oppose a specific viewpoint when reporting on a particular topic, users need complete contextual information about the media over a time span, presenting various types of information, including complex topological structures, event evolution, event summaries, and summary text information, so that users can analyze the evolution of their news reports and thus analyze media bias. A path-aware story construction algorithm is used to process the input news documents and to perform topic clustering by repeatedly removing the edges with the highest media centrality through the construction of a keyword co-occurrence graph. Based on the extracted topics, a story construction algorithm based on path awareness is used to construct a story framework, and combined with visualization methods, users can further analyze media bias from the evolution of media reports. Step 5.2, storyline construction is an iterative process. For each newly generated event node in each time period, the storyline construction algorithm comprehensively evaluates the compatibility of placing the event node in all positions on the storyline. The algorithm iteratively constructs the eventline based on four factors: time interval, node consistency, path consistency, and topic consistency. That is, for each incoming event E... new And the existing storyline T, through event E new Event nodes are continuously added to the existing storyline by placing them in appropriate positions. Two types of operations are used to put incoming events into the storyline: merging and connecting. First, all nodes on the storyline are traversed to evaluate whether new events and existing events in the storyline represent the same theme. This is achieved by calculating the cosine similarity between the TF-IDF vectors of two documents. When the similarity exceeds a certain threshold, the new event E is added. new With existing events E in the tree j Merge; otherwise, continue iterating and calculating the new event E. new With the existing events E in each network j The connection probability P between con The connection probability is calculated based on three factors: (1) the time interval between the two events. gap (2) Node consistency between two event nodes sim , and (3) when event E new E connected in the event context j Path consistency p in the later stages coh The longer the time interval between two events, the lower their correlation becomes. The time interval is calculated as follows: in, It is event E j The timeline of the event It is event E new The time nodes of events are used to evaluate node consistency and path consistency by vectorizing events using TF-IDF and assessing the cosine similarity between event nodes and story context nodes, as well as adjacent nodes along the path. The node consistency calculation method is as follows: Among them, TF-IDF(d j ) and TF-IDF(d new ) are events d j and d new TF-IDF vectorization representation; The path consistency is calculated as follows: Among them, L T，j，new Indicates when event E new E connected in the event timeline j Then, from the root node of T to node E new The path, |L T,j,new | indicates the length of the path; Reduce the dimensionality of all node vectors along the path using SVD, and then combine them with the new event node E. new Calculate cosine similarity and evaluate the new event node E using cosine similarity. new The compatibility with the overall topic along the branching paths of the storyline is used to determine whether the placement of new event nodes in the storyline is appropriate to maintain topic consistency. Finally, the eventline is iteratively constructed based on four factors: time interval, node consistency, path consistency, and topic consistency. Step 6: Users interact and compare the continuous evolution pattern of media deviation in the specific event development path based on the event development context. Based on the event development story context constructed in Step 5 and the data corpus constructed in Step 1, the visualization results of the event development story context are displayed, while supporting the user's interactive exploration process.

2. The visual analysis method for identifying media bias in large-scale news sources according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Construct a global media reporting bias design space. Based on the GDELT open-source data, select several time-series attributes representing media reporting bias: the number of reported events, the number of times the reported event is mentioned, the number of articles mentioning the reported event, the number of links mentioning the reported event, the degree of impact of the reported event on the region, and the attitude towards the reported event. Use a variational autoencoder to model the complex joint distribution of media reports as a latent feature representation to maintain the continuity of news reporting over time. That is, set a set of n initial integrated media feature vectors X = {x1, x2, ..., x...} n }, where x i This refers to the feature vector of the i-th news media report. The variable-time encoder model uses the hidden state h of the previous model as a basis. t and the vector set X at the corresponding time step t+1 Calculate state h t+1 As shown in the following formula: in, and Let b be the model weight matrix. r Let h be a constant vector, and the final state h of the model be... end The linear layer is mapped to the mean μ and standard deviation σ, as shown in the following formula: in, and Let b be the model weight matrix. m With b s Given a constant vector, reparameterization is performed based on the mean and standard deviation, and a latent vector z is sampled from the distribution defined by the mean and standard deviation. The model passes the sampled latent vector to obtain the initial state of the decoder, as shown in the following formula: The decoder is updated using backpropagation as shown in the following formula: in and Let b be the model weight matrix. z b c With b out It is a constant vector; Finally, the loss function of the network is constructed for joint training as shown in the following formula: By training a multi-attribute autovariable temporal encoder model, the design space of global media reporting bias within a certain time period can be quickly obtained to meet the dynamic and timeliness requirements of media bias design space construction. Step 2.2: Construct a media dimensionality reduction representation. Keep the relative distances in the original media deviation design space unchanged, and reduce the high-dimensional media encoding in the original design space to a two-dimensional plane so that media nodes with similar reporting habits are relatively close to each other in the two-dimensional plane, and media nodes with different reporting habits are relatively far apart.

3. The visual analysis method for identifying media bias in large-scale news sources according to claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: Based on the two-dimensional representation of media reporting bias obtained in Step 2, representative media at different levels are selected by using hierarchical clustering and considering the number of media reports. This allows users to better analyze media reporting bias between and within clusters, enabling users to compare different media bias cluster representations at different exploration levels and achieving hierarchical exploration of media reporting bias clusters and bias cluster location based on fuzzy matching of media names. Step 3.1: Constructing a hierarchical media bias projection. Based on the two-dimensional representation of media reporting bias in Step 2, a distance-based hierarchical clustering method is used to construct a hierarchical projection of global news media bias by dividing the minimum cluster set. The media nodes at each level represent media with similar reporting bias distances under the current level's clustering strength. Step 3.2, Hierarchical exploration of media bias clusters: Combining the hierarchical characteristics of media bias projection, different levels of exploration granularity are set during user exploration. The user's mouse zoom determines the level of exploration, thereby displaying representative media nodes at different levels. By moving the mouse over each media node, the user can see the similarities and differences in the number of news events reported, the number of times the reported events are mentioned, the number of articles mentioning the reported events, the number of links mentioning the reported events, the degree of impact of the reported events on the region, and the attitudes towards the reported events. Step 3.3, based on fuzzy matching of media names, provides a fuzzy search function for media that the user is interested in. The user enters the domain name of the media they are interested in, and the fuzzy matching function provides matching results. The user can click on the matching results to automatically focus on the cluster of media that the user is interested in, so as to facilitate the user's exploration and use.

4. The visual analysis method for identifying media bias in large-scale news sources according to claim 1, characterized in that, Step 6 includes the following steps: Step 6.1, Display of the visualization results of the event development timeline; Each node in the event development timeline represents an aggregated event; the size of the node in the event development timeline represents the magnitude of the deviation in the selected media's reporting on the current aggregated event; the thickness of the line connecting two aggregated nodes represents the closeness of the event's development and evolution; the size of different sectors on the node represents the percentage of articles published by the selected media on the current aggregated event; The node layout of the event development path is based on the layout method of the original node link tree. The time information of the event development is added in the horizontal direction to represent the development and evolution trend of the aggregated event nodes. The development and changes of different branches in the path represent the development and evolution of different sub-topics. By comparing and analyzing the changes in node size, node sector size and the thickness of the connection between nodes on different branches, users can understand the evolution pattern of the selected media's reporting deviation on a specific topic as a whole. Step 6.2: The user analyzes media bias on a single specific event; Users can hover their mouse over an aggregated event node in the event timeline to see the number of news reports about the event from their selected media outlets, the number of times the event was mentioned in those reports, the number of articles mentioning the event, the number of links linking to the event, the impact of the event on the region, and the differences in attitudes towards the event. Clicking on an aggregated event node displays representative articles from media outlets most similar to the user's chosen event. By reading these representative articles, users can understand the differences in media coverage of the same event, while the large-scale pre-trained language model GPT-3 generates summaries of the representative articles to summarize the current aggregated event. Clicking on the title of a representative article redirects to the original article page, allowing users to trace the origins of the news event. Step 6.3: The user analyzes the continuous evolution pattern of media deviation on different paths; When a user clicks on a specific aggregated event node, a media reporting event line visualization is displayed, showing the path from the root node of the event's development to the current aggregated event node. This visualization lists all news reports related to the aggregated event along the current path, chronologically arranged by the media's reporting timeline. Each news report is mapped to a square, with the side length representing the article's impact on the region and the color representing the article's sentiment. News reports belonging to the same aggregated event are interconnected, as are news articles from the same media outlet, distinguishing different aggregated events and analyzing the changing trends in the impact of articles published by different media outlets on the same event. The media reporting event line visualization supports article viewing and source tracing. By exploring the media reporting event line, users can analyze the development status of a branch of the event's development and view the changing trends in the impact of articles reported by different media outlets on the region, as well as the changes in sentiment towards the same event.

Images