CN116186331B - Graph interpretation method and system - Google Patents
Graph interpretation method and system Download PDFInfo
- Publication number
- CN116186331B CN116186331B CN202310464984.8A CN202310464984A CN116186331B CN 116186331 B CN116186331 B CN 116186331B CN 202310464984 A CN202310464984 A CN 202310464984A CN 116186331 B CN116186331 B CN 116186331B
- Authority
- CN
- China
- Prior art keywords
- chart
- interpretation
- training
- text
- business
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Machine Translation (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The application discloses a chart interpretation method and system, which are used for solving the technical problem that the matching performance of chart interpretation and business scenes is low. Wherein, a chart interpretation scheme comprises the following steps: obtaining a chart object; according to the chart object, determining a business scene, element data and a chart type of the corresponding chart object in a database; determining a target parameter analysis function corresponding to the service scene according to the service scene; according to the element data and the chart type, a target parameter analysis function corresponding to the service scene is adopted, and a target parameter is obtained through calculation; inputting the element data, the chart type and the target parameters into the pre-training interpretation model to obtain a chart interpretation text. And the pre-training interpretation model selects an interpretation template with better sense of language to generate a chart interpretation text according to the business scene of the corresponding chart object, so that the interpretation pertinence of the chart interpretation text is improved, and the matching property of the chart interpretation and the business scene is further improved.
Description
Technical Field
The present disclosure relates to the field of data processing technologies, and in particular, to a chart interpretation method and system.
Background
Along with the way that more and more enterprises carry out digital transformation, the digital enterprises introduce a data management concept in daily informatization management, and the service data is visually displayed in a chart form, so that the data characteristics are found, and the data analysis is promoted.
In the prior art, the characteristics of service data are intuitively displayed to a user through various visual chart forms, and the user is matched with corresponding chart abstracts or prompts and other interpretation texts to help the client to better understand the service gist.
In implementing the prior art, the inventors found that:
at present, only a fixed rule template is used for applying a generating mode of the interpretation text, and the generated text language is dead. When facing to a changeable business scene, the interpretation pertinence of the interpretation text generated by the fixed rule template is weaker.
Therefore, a new chart interpretation scheme is needed to solve the technical problem of low matching between chart interpretation and business scenario.
Disclosure of Invention
The embodiment of the application provides a novel chart interpretation scheme which is used for solving the technical problem that the matching performance of chart interpretation and business scenes is low.
Specifically, a chart interpretation method includes the following steps:
obtaining a chart object;
according to the chart object, determining a business scene, element data and a chart type of the corresponding chart object in a database;
determining a target parameter analysis function of the corresponding business scene according to the business scene of the corresponding chart object;
according to the element data of the corresponding chart object and the chart type of the corresponding chart object, a target parameter analysis function of the corresponding business scene is adopted, and a target parameter is obtained through calculation;
inputting element data corresponding to the chart object, the chart type corresponding to the chart object and the target parameter into the pre-training interpretation model to obtain a chart interpretation text.
Further, the training process of the pre-training interpretation model includes:
inputting training chart objects formed by training service scenes, training element data and training chart types into an original interpretation model to obtain a first chart interpretation text;
updating data corresponding to the training service scene in the first chart interpretation text, and generating a chart interpretation text set with elements at least including a second chart interpretation text and a third chart interpretation text;
performing cross-dimensional interference evaluation on the elements in the chart interpretation text set respectively to generate element scoring values with corresponding relations with the elements in the chart interpretation text set;
determining the element with the highest element grading value as an optimized graph interpretation text;
and taking the training chart and the optimizing chart interpretation text as rewarding results, and updating the original interpretation model into a pre-training interpretation model by adopting a near-end strategy optimization algorithm.
Further, the business scenario of the corresponding chart object comprises at least one of supervision report, enterprise operation analysis, marketing analysis, business report and leading cockpit.
Further, the chart type of the corresponding chart object includes at least one of a detail table, a cross table, a grouping table, a fixed table, a bar chart, a current situation chart, a pie chart, a multidimensional pie chart, a dashboard, a duty ratio chart, a radar chart, a bubble chart, an index card, a word cloud chart, a pyramid chart and a rendering map.
Further, the chart interpretation method further includes the steps of:
and playing the chart interpretation text through a voice synthesis tool.
The embodiment of the application also provides a chart interpretation system.
Specifically, a chart interpretation system includes:
the acquisition module is used for acquiring the chart object;
the interpretation module is used for determining a business scene, element data and a chart type of the corresponding chart object in the database according to the chart object; the target parameter analysis function is used for determining a target parameter analysis function corresponding to the business scene according to the business scene corresponding to the chart object; the method is also used for calculating and obtaining target parameters by adopting a target parameter analysis function of the corresponding business scene according to the element data of the corresponding chart object and the chart type of the corresponding chart object; and the method is also used for inputting element data corresponding to the chart object, the chart type corresponding to the chart object and the target parameter into the pre-training interpretation model to obtain a chart interpretation text.
Further, the training process of the pre-training interpretation model includes:
inputting training chart objects formed by training service scenes, training element data and training chart types into an original interpretation model to obtain a first chart interpretation text;
updating data corresponding to the training service scene in the first chart interpretation text, and generating a chart interpretation text set with elements at least including a second chart interpretation text and a third chart interpretation text;
performing cross-dimensional interference evaluation on the elements in the chart interpretation text set respectively to generate element scoring values with corresponding relations with the elements in the chart interpretation text set;
determining the element with the highest element grading value as an optimized graph interpretation text;
and taking the training chart and the optimizing chart interpretation text as rewarding results, and updating the original interpretation model into a pre-training interpretation model by adopting a near-end strategy optimization algorithm.
Further, the business scenario of the corresponding chart object comprises at least one of supervision report, enterprise operation analysis, marketing analysis, business report and leading cockpit.
Further, the chart type of the corresponding chart object includes at least one of a detail table, a cross table, a grouping table, a fixed table, a bar chart, a current situation chart, a pie chart, a multidimensional pie chart, a dashboard, a duty ratio chart, a radar chart, a bubble chart, an index card, a word cloud chart, a pyramid chart and a rendering map.
Further, the chart interpretation system further includes:
and the playing module is used for playing the chart interpretation text through a voice synthesis tool.
The technical scheme provided by the embodiment of the application has at least the following beneficial effects:
and determining a target parameter analysis function corresponding to the business scene according to the business scene corresponding to the chart object, and improving reading knowing efficiency. And the pre-training interpretation model selects an interpretation template with better sense of language to generate a chart interpretation text according to the business scene of the corresponding chart object, so that the interpretation pertinence of the chart interpretation text is improved, and the matching property of the chart interpretation and the business scene is further improved.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application and do not constitute an undue limitation to the application. In the drawings:
FIG. 1 is a block flow diagram of a graph interpretation method according to an embodiment of the present application;
fig. 2 is a schematic structural diagram of a graph interpretation system according to an embodiment of the present application.
The reference numerals in the drawings are as follows:
100. chart interpretation system
11. Acquisition module
12. Interpretation module
13. And a playing module.
Detailed Description
For the purposes, technical solutions and advantages of the present application, the technical solutions of the present application will be clearly and completely described below with reference to specific embodiments of the present application and corresponding drawings. It will be apparent that the described embodiments are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are within the scope of the present disclosure.
Referring to fig. 1, in order to solve the technical problem of low matching between chart interpretation and service scene, the present application provides a chart interpretation method, which includes the following steps:
s110: a chart object is obtained.
S120: and determining the business scene, the element data and the chart type of the corresponding chart object in the database according to the chart object.
It will be appreciated that the chart object at least includes a statistical chart or table, and in a specific application scenario, the chart object may be represented as a chart billboard.
In a scene of displaying data characteristics facing to clients, a client selects a chart object, a server determines element data and chart types of the corresponding chart object in a database according to the chart object selected by the client, and builds and renders a chart billboard, and then visually displays the chart billboard at the client.
Different service scenes need to be displayed and have different interpretation texts, for example, the service scenes of enterprise operation are often provided with a detail table and an index card, and the newly added index type, the newly added index duty ratio, the newly added index reference value and the newly added index maximum value are displayed on a side without paying attention to various index trends. The business report is often matched with a grouping table and a current situation diagram, and the types and the trends of various indexes are displayed on the emphasis instead of paying attention to the reference value and the maximum value of the newly added index.
Therefore, the chart objects displayed by different service scenes are different, and the required acquisition, statistics and calculation target parameters are also different. In the prior art, a generation mode of the interpretation text is only applied by adopting a fixed rule template, so that the generated text language is dead. When facing to a changeable business scene, the interpretation pertinence of the interpretation text generated by the fixed rule template is weaker. Therefore, in the specific application scenario provided in the present application, the chart object has data information such as a service scenario, element data, a chart type, and the like, and these data information are stored in a database. The method and the device are specially recorded in the business scenes of the chart object and stored in the database so as to generate different interpretation texts according to different business scenes.
Specifically, the business scenario corresponding to the chart object includes at least one of supervision report, enterprise operation analysis, marketing analysis, business report and lead cockpit.
Further, in a preferred embodiment provided in the present application, the database further stores a business field corresponding to the chart object. The business field at least comprises at least one of finance, electric power, medical treatment, manufacturing and environmental protection. Therefore, the interpretation text content of the corresponding chart object can be ensured to be matched with the service scene by comprehensively considering the various aspects such as different service fields, service scenes, chart types and the like.
S130: and determining a target parameter analysis function of the corresponding business scene according to the business scene of the corresponding chart object.
S140: and according to the element data of the corresponding chart object and the chart type of the corresponding chart object, calculating to obtain a target parameter by adopting a target parameter analysis function of the corresponding business scene.
It can be appreciated that the target parameter profiling function has an association with a business scenario. In a specific application scenario, different traffic scenarios correspond to different objective parameter profiling functions. The target parameter profiling function defines target parameters that need to be collected, counted, calculated. For example, the target parameter profiling function may be represented as a SUM summation function for calculating an index SUM. The target parameter profiling function may be represented as an AVERAGE function for calculating an index AVERAGE. The target parameter profiling function may be represented as a MAX maximum function for collecting index maxima. The target parameter profiling function may be represented as a COUNT statistics function for counting each index type.
Of course, according to the service scenario, a plurality of target parameter parsing functions corresponding to the service scenario may be determined. For example, in a business scenario of enterprise operation, the target parameter analysis function of the new index type is collected, and the target parameter analysis function of the new index duty ratio, the target parameter analysis function of the new index reference value, and the target parameter analysis function of the maximum value of the new index are collected. The business scenario of the business report corresponds to a target parameter analysis function for counting various index types and also corresponds to a target parameter analysis function for counting various index trends.
Further, the target parameter profiling function characterizes the operational relationship of the target parameter to known parameters. In the target parameter parsing function, the target parameter is unknown, and the known parameter can be obtained according to the element data of the corresponding chart object and the chart type of the corresponding chart object. For example, sample parameter information may be determined from element data of the chart object, and information such as a sample linear trend, a sample duty ratio, etc. may be determined from a chart type of the chart object.
S150: inputting element data corresponding to the chart object, the chart type corresponding to the chart object and the target parameter into the pre-training interpretation model to obtain a chart interpretation text.
It will be appreciated that the pre-trained interpretation model is represented in a specific application scenario as a generative deep language model, preferably a BART deep language model or a T5 deep language model. The pre-trained interpretation model also stores a number of interpretation templates. And the interpretation templates have an association relation with the business scene. The interpretation template includes at least an interpretation text presentation structure. The interpretation text presentation structure has an exchangeable entity. The replaceable entity can fill in the target parameter or the word entity according to the element data of the corresponding chart object, the chart type of the corresponding chart object and the target parameter.
The pre-training interpretation model can determine an interpretation template corresponding to the service scene according to the element data of the corresponding chart object, the chart type of the corresponding chart object and the target parameter; and filling the replaceable entity with the target parameter or entity based on the sentence-level semantics in the interpretation template to form the graph interpretation text.
The training process of the pre-training interpretation model is described below, including:
inputting training chart objects formed by training service scenes, training element data and training chart types into an original interpretation model to obtain a first chart interpretation text;
updating data corresponding to the training service scene in the first chart interpretation text, and generating a chart interpretation text set with elements at least including a second chart interpretation text and a third chart interpretation text;
performing cross-dimensional interference evaluation on the elements in the chart interpretation text set respectively to generate element scoring values with corresponding relations with the elements in the chart interpretation text set;
determining the element with the highest element grading value as an optimized graph interpretation text;
and taking the training chart and the optimizing chart interpretation text as rewarding results, and updating the original interpretation model into a pre-training interpretation model by adopting a near-end strategy optimization algorithm.
It can be understood that in the specific application scenario provided in the present application, the present application uses a fixed rule template in the prior art to generate different chart interpretation texts for different training chart objects, and uses a data set size of more than 2 ten thousand.
For example, when the training chart object is a cross table, the fixed rule template is expressed as [ index 1 header ] total [ index 1 total row number ], and [ index 2 header ] total [ index 2 total row number ]. The ranking index header is the first row of values of the ranking index.
When the training chart object is a line graph, the fixed rule template is represented as the highest [ index 1 title ] of the largest dimension item value of the [ index 1], the largest value of the [ index 1], and the lowest [ index 1 title ] of the smallest dimension item value of the [ index 1], the smallest value of the [ index 1 ]. The reference line is of the type of the reference line, the specific numerical value of the reference line is higher than the specific numerical value of the reference line, the dimensional value is …, and the dimensional value is lower than the specific numerical value of the reference line. Overall, the trend is linear.
Next, the application uses a hierarchical sampling method to further perform service scene marking and service field marking on the graph interpretation dataset, and randomly adopts about 10% of graphs and graph interpretation texts (i.e. more than 2 thousand), and then uses relevant service personnel to rewrite and optimize the original interpretation template, and replaces the original graph interpretation texts. In this way, the interpretation text of the dataset will have better language diversity. Finally, a chart interpretation data set D1 is obtained, wherein the scale size is more than 2 ten thousand, and 2 thousand chart interpretation texts after manual rewriting are obtained.
Then, the present application adjusts the setting parameters of the generated depth language model based on the data set D1, so as to obtain the original interpretation model M1.
Further, in order to facilitate the accurate filling of the original interpretation model M1 or the pre-trained interpretation model into the target parameters, the configuration input format of the present application is expressed as:
the [ hint ] s [ chart data profile ] s [ chart raw data ]
Wherein [ hint ] is: reading the following steps: "
[ Chart raw data ] is business data extracted from chart billboard objects, which are spliced together according to the following format:
"field 1", "field 2", "field N", "row 1", "row 2", and "the like.
[ Row x ] is: "[ number 1], [ number 2 ], [ number N ]," is shown in the specification;
[ Chart data parsing ] is the data parsing result of the business data, and the format is: "[ maximum ]: [ number 1 ]; [ reference value ]: [ number 2 ]; ...".
Further, after inputting training chart objects composed of training service scenes, training element data and training chart types into the original interpretation model M1, the original interpretation model M1 performs semantic recognition and then fills in an original interpretation template to generate a first chart interpretation text. The first chart interpretation text is typically generated by populating the original interpretation template with target parameters or entities.
At this time, the first chart interpretation text is dead in the text language, and when facing the changeable business scene, the interpretation pertinence of the first chart interpretation text generated by the original interpretation template is weaker.
For this reason, the present application updates the data corresponding to the training service scenario in the first chart interpretation text, and generates a chart interpretation text set with elements of at least the second chart interpretation text and the third chart interpretation text.
Specifically, by changing the value of the service scenario in the prompt, the original interpretation model M1 may interpret the text with respect to the first chart, and then generate T interpretation texts by deforming. Wherein, according to anticipated input manual work volume select T value, T value is 2 at least. Preferably, t=4.
And then pairing a plurality of interpretation texts in pairs, and randomly distributing each pair to P business personnel for cross-dimension interference evaluation. Business personnel score 2 for the interpretation text with more comfortable sense of language in the two interpretation texts according to own sense of language preference, and score 0 for the other interpretation text with poorer sense of language. If the two interpreted texts have a more or less pronounced feel, they are scored as 1. Wherein, the P value is selected according to the expected input manual quantity, and the application preferably carries out the experiment by p=3.
And collecting scoring results of all business personnel, and respectively summarizing the scoring values of each interpretation. And sequencing the plurality of interpretations according to the total score to finally obtain a data set D2, wherein the data scale is more than 2 kilos.
Next, the data set D2 is input to the bonus model M2, and a bonus score is calculated. The bonus model M2 inputs the following description, and outputs the bonus points (interval 0, 1) that are selected for interpretation text.
Input: the "hint" s "is the" chart data profile "is the" s "is the" chart raw data "is the" s "is the" interpretation ".
Wherein [ hint ] is: "pair [ chart type ]: ".
From the interpretation ranking list, the training process needs to ensure that the reward points remain substantially the same ranking, i.e., a good interpretation score should be greater than a bad interpretation.
In general, the higher the interpretation text reward score the highest the element score value. The interpretation text with the highest element score value is taken as the interpretation text of the optimization chart.
Finally, the application adopts a near-end strategy optimization algorithm (Proximal Policy Optimization, PPO) as a reinforcement learning strategy, and updates the original interpretation model M1 into a pre-training interpretation model M3.
Further, in a preferred embodiment provided in the present application, the chart interpretation method further includes the steps of:
and playing the chart interpretation text through a voice synthesis tool.
Thus, the voice interpretation work of the chart billboard can be automatically completed, so that more business scenes are supported.
In summary, the graph interpretation method determines the target parameter analysis function of the corresponding service scene according to the service scene of the corresponding graph object, thereby improving the reading efficiency. And the pre-training interpretation model selects an interpretation template with better sense of language to generate a chart interpretation text according to the business scene of the corresponding chart object, so that the interpretation pertinence of the chart interpretation text is improved, and the matching property of the chart interpretation and the business scene is further improved.
Referring to fig. 2, to support the chart interpretation method, the present application further provides a chart interpretation system 100, including:
an acquisition module 11 for acquiring a chart object;
the interpretation module 12 is configured to determine, according to the chart object, a service scenario, element data, and a chart type of the corresponding chart object in the database; the target parameter analysis function is used for determining a target parameter analysis function corresponding to the business scene according to the business scene corresponding to the chart object; the method is also used for calculating and obtaining target parameters by adopting a target parameter analysis function of the corresponding business scene according to the element data of the corresponding chart object and the chart type of the corresponding chart object; and the method is also used for inputting element data corresponding to the chart object, the chart type corresponding to the chart object and the target parameter into the pre-training interpretation model to obtain a chart interpretation text.
It will be appreciated that the chart object at least includes a statistical chart or table, and in a specific application scenario, the chart object may be represented as a chart billboard.
In a scene of displaying data characteristics facing to clients, a client selects a chart object, a server determines element data and chart types of the corresponding chart object in a database according to the chart object selected by the client, and builds and renders a chart billboard, and then visually displays the chart billboard at the client.
Different service scenes need to be displayed and have different interpretation texts, for example, the service scenes of enterprise operation are often provided with a detail table and an index card, and the newly added index type, the newly added index duty ratio, the newly added index reference value and the newly added index maximum value are displayed on a side without paying attention to various index trends. The business report is often matched with a grouping table and a current situation diagram, and the types and the trends of various indexes are displayed on the emphasis instead of paying attention to the reference value and the maximum value of the newly added index.
Therefore, the chart objects displayed by different service scenes are different, and the required acquisition, statistics and calculation target parameters are also different. In the prior art, a generation mode of the interpretation text is only applied by adopting a fixed rule template, so that the generated text language is dead. When facing to a changeable business scene, the interpretation pertinence of the interpretation text generated by the fixed rule template is weaker. Therefore, in the specific application scenario provided in the present application, the chart object has data information such as a service scenario, element data, a chart type, and the like, and these data information are stored in a database. The method and the device are specially recorded in the business scenes of the chart object and stored in the database so as to generate different interpretation texts according to different business scenes.
Specifically, the business scenario corresponding to the chart object includes at least one of supervision report, enterprise operation analysis, marketing analysis, business report and lead cockpit.
Further, in a preferred embodiment provided in the present application, the database further stores a business field corresponding to the chart object. The business field at least comprises at least one of finance, electric power, medical treatment, manufacturing and environmental protection. Therefore, the interpretation text content of the corresponding chart object can be ensured to be matched with the service scene by comprehensively considering the various aspects such as different service fields, service scenes, chart types and the like.
The interpretation module 12 determines a target parameter parsing function of the corresponding business scenario according to the business scenario of the corresponding chart object. The interpretation module 12 calculates the target parameters by using the target parameter analysis function of the corresponding business scene according to the element data of the corresponding chart object and the chart type of the corresponding chart object.
It can be appreciated that the target parameter profiling function has an association with a business scenario. In a specific application scenario, different traffic scenarios correspond to different objective parameter profiling functions. The target parameter profiling function defines target parameters that need to be collected, counted, calculated. For example, the target parameter profiling function may be represented as a SUM summation function for calculating an index SUM. The target parameter profiling function may be represented as an AVERAGE function for calculating an index AVERAGE. The target parameter profiling function may be represented as a MAX maximum function for collecting index maxima. The target parameter profiling function may be represented as a COUNT statistics function for counting each index type.
Of course, the interpretation module 12 may determine a plurality of target parameter profiling functions corresponding to the service scenario according to the service scenario. For example, in a business scenario of enterprise operation, the target parameter analysis function of the new index type is collected, and the target parameter analysis function of the new index duty ratio, the target parameter analysis function of the new index reference value, and the target parameter analysis function of the maximum value of the new index are collected. The business scenario of the business report corresponds to a target parameter analysis function for counting various index types and also corresponds to a target parameter analysis function for counting various index trends.
Further, the target parameter profiling function characterizes the operational relationship of the target parameter to known parameters. In the target parameter parsing function, the target parameter is unknown, and the known parameter can be obtained according to the element data of the corresponding chart object and the chart type of the corresponding chart object. For example, the interpretation module 12 may determine sample parameter information from the element data of the chart object, and the interpretation module 12 may determine information such as a sample linear trend, a sample duty cycle, and the like from the chart type of the chart object.
The interpretation module 12 inputs the element data corresponding to the chart object, the chart type corresponding to the chart object, and the target parameter to the pre-training interpretation model, and obtains the chart interpretation text.
It will be appreciated that the pre-trained interpretation model is represented in a specific application scenario as a generative deep language model, preferably a BART deep language model or a T5 deep language model. The pre-trained interpretation model also stores a number of interpretation templates. And the interpretation templates have an association relation with the business scene. The interpretation template includes at least an interpretation text presentation structure. The interpretation text presentation structure has an exchangeable entity. The replaceable entity can fill in the target parameter or the word entity according to the element data of the corresponding chart object, the chart type of the corresponding chart object and the target parameter.
The pre-training interpretation model can determine an interpretation template corresponding to the service scene according to the element data of the corresponding chart object, the chart type of the corresponding chart object and the target parameter; and filling the replaceable entity with the target parameter or entity based on the sentence-level semantics in the interpretation template to form the graph interpretation text.
The training process of the pre-training interpretation model is described below, including:
establishing an association relation between a service scene and an interpretation template; the interpretation template at least comprises an interpretation text presentation structure and a target parameter analysis function;
inputting training chart objects formed by training service scenes, training element data and training chart types into an original interpretation model to obtain a first chart interpretation text;
updating data corresponding to the training service scene in the first chart interpretation text, and generating a chart interpretation text set with elements at least including a second chart interpretation text and a third chart interpretation text;
performing cross-dimensional interference evaluation on the elements in the chart interpretation text set respectively to generate element scoring values with corresponding relations with the elements in the chart interpretation text set;
determining the element with the highest element grading value as an optimized graph interpretation text;
and taking the training chart and the optimizing chart interpretation text as rewarding results, and updating the original interpretation model into a pre-training interpretation model by adopting a near-end strategy optimization algorithm.
It can be understood that in a specific application scenario, the application adopts a fixed rule template in the prior art, generates different chart interpretation texts for different training chart objects, and adopts a data set size of more than 2 ten thousand.
For example, when the training chart object is a cross table, the fixed rule template is expressed as [ index 1 header ] total [ index 1 total row number ], and [ index 2 header ] total [ index 2 total row number ]. The ranking index header is the first row of values of the ranking index.
When the training chart object is a line graph, the fixed rule template is represented as the highest [ index 1 title ] of the largest dimension item value of the [ index 1], the largest value of the [ index 1], and the lowest [ index 1 title ] of the smallest dimension item value of the [ index 1], the smallest value of the [ index 1 ]. The reference line is of the type of the reference line, the specific numerical value of the reference line is higher than the specific numerical value of the reference line, the dimensional value is …, and the dimensional value is lower than the specific numerical value of the reference line. Overall, the trend is linear.
Next, the application uses a hierarchical sampling method to further perform service scene marking and service field marking on the graph interpretation dataset, and randomly adopts about 10% of graphs and graph interpretation texts (i.e. more than 2 thousand), and then uses relevant service personnel to rewrite and optimize the original interpretation template, and replaces the original graph interpretation texts. In this way, the interpretation text of the dataset will have better language diversity. Finally, a chart interpretation data set D1 is obtained, wherein the scale size is more than 2 ten thousand, and 2 thousand chart interpretation texts after manual rewriting are obtained.
Then, the present application adjusts the setting parameters of the generated depth language model based on the data set D1, so as to obtain the original interpretation model M1.
Further, in order to facilitate the accurate filling of the original interpretation model M1 or the pre-trained interpretation model into the target parameters, the configuration input format of the present application is expressed as:
the [ hint ] s [ chart data profile ] s [ chart raw data ]
Wherein [ hint ] is: reading the following steps: "
[ Chart raw data ] is business data extracted from chart billboard objects, which are spliced together according to the following format:
"field 1", "field 2", "field N", "row 1", "row 2", and "the like.
[ Row x ] is: "[ number 1], [ number 2 ], [ number N ]," is shown in the specification;
[ Chart data parsing ] is the data parsing result of the business data, and the format is: "[ maximum ]: [ number 1 ]; [ reference value ]: [ number 2 ]; ...".
Further, after inputting training chart objects composed of training service scenes, training element data and training chart types into the original interpretation model M1, the original interpretation model M1 performs semantic recognition and then fills in an original interpretation template to generate a first chart interpretation text. The first chart interpretation text is typically generated by populating the original interpretation template with target parameters or entities.
At this time, the first chart interpretation text is dead in the text language, and when facing the changeable business scene, the interpretation pertinence of the first chart interpretation text generated by the original interpretation template is weaker.
For this reason, the present application updates the data corresponding to the training service scenario in the first chart interpretation text, and generates a chart interpretation text set with elements of at least the second chart interpretation text and the third chart interpretation text.
Specifically, by changing the value of the service scenario in the prompt, the original interpretation model M1 may interpret the text with respect to the first chart, and then generate T interpretation texts by deforming. Wherein, according to anticipated input manual work volume select T value, T value is 2 at least. Preferably, t=4.
And then pairing a plurality of interpretation texts in pairs, and randomly distributing each pair to P business personnel for cross-dimension interference evaluation. Business personnel score 2 for the interpretation text with more comfortable sense of language in the two interpretation texts according to own sense of language preference, and score 0 for the other interpretation text with poorer sense of language. If the two interpreted texts have a more or less pronounced feel, they are scored as 1. Wherein, the P value is selected according to the expected input manual quantity, and the application preferably carries out the experiment by p=3.
And collecting scoring results of all business personnel, and respectively summarizing the scoring values of each interpretation. And sequencing the plurality of interpretations according to the total score to finally obtain a data set D2, wherein the data scale is more than 2 kilos.
Next, the data set D2 is input to the bonus model M2, and a bonus score is calculated. The bonus model M2 inputs the following description, and outputs the bonus points (interval 0, 1) that are selected for interpretation text.
Input: the "hint" s "is the" chart data profile "is the" s "is the" chart raw data "is the" s "is the" interpretation ".
Wherein [ hint ] is: "pair [ chart type ]: ".
From the interpretation ranking list, the training process needs to ensure that the reward points remain substantially the same ranking, i.e., a good interpretation score should be greater than a bad interpretation.
In general, the higher the interpretation text reward score the highest the element score value. The interpretation text with the highest element score value is taken as the interpretation text of the optimization chart.
Finally, the application adopts a near-end strategy optimization algorithm (Proximal Policy Optimization, PPO) as a reinforcement learning strategy, and updates the original interpretation model M1 into a pre-training interpretation model M3.
Further, in a preferred embodiment provided herein, the chart interpretation system 100 further comprises:
and the playing module 13 is used for playing the chart interpretation text through a voice synthesis tool.
In this way, the chart interpretation system 100 is able to automatically complete the speech interpretation of the chart board, thereby supporting more business scenarios.
In summary, the chart interpretation system 100 determines the target parameter parsing function of the corresponding business scenario according to the business scenario of the corresponding chart object, so as to improve the reading efficiency. And the pre-training interpretation model selects an interpretation template with better sense of language to generate a chart interpretation text according to the business scene of the corresponding chart object, so that the interpretation pertinence of the chart interpretation text is improved, and the matching property of the chart interpretation and the business scene is further improved.
It should be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, the statement "comprises" or "comprising" an element defined by … … does not exclude the presence of other identical elements in a process, method, article or apparatus that comprises the element.
It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The foregoing is merely exemplary of the present application and is not intended to limit the present application. Various modifications and changes may be made to the present application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. which are within the spirit and principles of the present application are intended to be included within the scope of the claims of the present application.
Claims (6)
1. A chart interpretation method, characterized by comprising the steps of:
obtaining a chart object;
according to the chart object, determining a business scene, element data and a chart type of the corresponding chart object in a database;
determining a target parameter analysis function of the corresponding business scene according to the business scene of the corresponding chart object;
according to the element data of the corresponding chart object and the chart type of the corresponding chart object, a target parameter analysis function of the corresponding business scene is adopted, and a target parameter is obtained through calculation;
inputting element data corresponding to the chart object, the chart type corresponding to the chart object and the target parameter into a pre-training interpretation model to obtain a chart interpretation text;
playing the chart interpretation text through a voice synthesis tool;
the training process of the pre-training interpretation model comprises the following steps:
inputting training chart objects formed by training service scenes, training element data and training chart types into an original interpretation model to obtain a first chart interpretation text;
updating data corresponding to the training service scene in the first chart interpretation text, and generating a chart interpretation text set with elements at least including a second chart interpretation text and a third chart interpretation text;
performing cross-dimensional interference evaluation on the elements in the chart interpretation text set respectively to generate element scoring values with corresponding relations with the elements in the chart interpretation text set;
determining the element with the highest element grading value as an optimized graph interpretation text;
and taking the training chart and the optimizing chart interpretation text as rewarding results, and updating the original interpretation model into a pre-training interpretation model by adopting a near-end strategy optimization algorithm.
2. The chart interpretation method of claim 1, wherein the business scenario of the corresponding chart object includes at least one of a supervisory report, an enterprise operation analysis, a marketing analysis, a business report, and a lead cockpit.
3. The graph interpretation method of claim 1, wherein the graph type of the corresponding graph object includes at least one of a detail table, a cross table, a grouping table, a fixed table, a bar graph, a current situation graph, a pie chart, a multi-dimensional pie chart, a dashboard, a duty graph, a radar graph, a bubble graph, an index card, a word cloud graph, a pyramid graph, and a rendering map.
4. A chart interpretation system, comprising the steps of:
the acquisition module is used for acquiring the chart object;
the interpretation module is used for determining a business scene, element data and a chart type of the corresponding chart object in the database according to the chart object; the target parameter analysis function is used for determining a target parameter analysis function corresponding to the business scene according to the business scene corresponding to the chart object; the method is also used for calculating and obtaining target parameters by adopting a target parameter analysis function of the corresponding business scene according to the element data of the corresponding chart object and the chart type of the corresponding chart object; the method comprises the steps of obtaining a chart interpretation text, wherein the chart interpretation text is used for inputting element data corresponding to a chart object, a chart type corresponding to the chart object and target parameters into a pre-training interpretation model;
the playing module is used for playing the chart interpretation text through a voice synthesis tool;
the training process of the pre-training interpretation model comprises the following steps:
inputting training chart objects formed by training service scenes, training element data and training chart types into an original interpretation model to obtain a first chart interpretation text;
updating data corresponding to the training service scene in the first chart interpretation text, and generating a chart interpretation text set with elements at least including a second chart interpretation text and a third chart interpretation text;
performing cross-dimensional interference evaluation on the elements in the chart interpretation text set respectively to generate element scoring values with corresponding relations with the elements in the chart interpretation text set;
determining the element with the highest element grading value as an optimized graph interpretation text;
and taking the training chart and the optimizing chart interpretation text as rewarding results, and updating the original interpretation model into a pre-training interpretation model by adopting a near-end strategy optimization algorithm.
5. The chart interpretation system of claim 4, wherein the business scenario of the corresponding chart object includes at least one of a supervisory report, an enterprise operation analysis, a marketing analysis, a business report, and a lead cockpit.
6. The chart interpretation system of claim 4, wherein the chart type of the corresponding chart object comprises at least one of a detail table, a cross table, a grouping table, a fixed table, a bar chart, a current situation chart, a pie chart, a multi-dimensional pie chart, a dashboard, a duty cycle chart, a radar chart, a bubble chart, an index card, a word cloud chart, a pyramid chart, a rendered map.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310464984.8A CN116186331B (en) | 2023-04-27 | 2023-04-27 | Graph interpretation method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310464984.8A CN116186331B (en) | 2023-04-27 | 2023-04-27 | Graph interpretation method and system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116186331A CN116186331A (en) | 2023-05-30 |
| CN116186331B true CN116186331B (en) | 2023-08-04 |
Family
ID=86452604
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310464984.8A Active CN116186331B (en) | 2023-04-27 | 2023-04-27 | Graph interpretation method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116186331B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117745875A (en) * | 2024-02-19 | 2024-03-22 | 粤港澳大湾区数字经济研究院(福田) | Graph analysis method, device, equipment and storage medium |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104866567A (en) * | 2015-05-22 | 2015-08-26 | 国家计算机网络与信息安全管理中心 | Method and apparatus for presenting business data |
| US11068477B1 (en) * | 2018-06-06 | 2021-07-20 | Gbt Travel Servces Uk Limited | Natural language processing with pre-specified SQL queries |
| CN114118026A (en) * | 2020-08-28 | 2022-03-01 | 北京仝睿科技有限公司 | Automatic document generation method and device, computer storage medium and electronic equipment |
| CN114595272A (en) * | 2022-02-28 | 2022-06-07 | 珠海金山办公软件有限公司 | Method and device for obtaining recommended chart type, electronic equipment and storage medium |
| CN115408499A (en) * | 2022-11-02 | 2022-11-29 | 思创数码科技股份有限公司 | Automatic analysis and interpretation method and system for government affair data analysis report chart |
| CN115587581A (en) * | 2022-10-26 | 2023-01-10 | 卫宁健康科技集团股份有限公司 | Medical quality report generation method and device, computer equipment and storage medium |
-
2023
- 2023-04-27 CN CN202310464984.8A patent/CN116186331B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104866567A (en) * | 2015-05-22 | 2015-08-26 | 国家计算机网络与信息安全管理中心 | Method and apparatus for presenting business data |
| US11068477B1 (en) * | 2018-06-06 | 2021-07-20 | Gbt Travel Servces Uk Limited | Natural language processing with pre-specified SQL queries |
| CN114118026A (en) * | 2020-08-28 | 2022-03-01 | 北京仝睿科技有限公司 | Automatic document generation method and device, computer storage medium and electronic equipment |
| CN114595272A (en) * | 2022-02-28 | 2022-06-07 | 珠海金山办公软件有限公司 | Method and device for obtaining recommended chart type, electronic equipment and storage medium |
| CN115587581A (en) * | 2022-10-26 | 2023-01-10 | 卫宁健康科技集团股份有限公司 | Medical quality report generation method and device, computer equipment and storage medium |
| CN115408499A (en) * | 2022-11-02 | 2022-11-29 | 思创数码科技股份有限公司 | Automatic analysis and interpretation method and system for government affair data analysis report chart |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116186331A (en) | 2023-05-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108363821A (en) | A kind of information-pushing method, device, terminal device and storage medium | |
| CN110292775B (en) | Method and device for acquiring difference data | |
| US7730023B2 (en) | Apparatus and method for strategy map validation and visualization | |
| CN110909986A (en) | Suspected actual controller risk identification method and system based on knowledge graph | |
| JP2010079657A (en) | Information processor, information processing method, and program | |
| CN116186331B (en) | Graph interpretation method and system | |
| CN109508441B (en) | Method and device for realizing data statistical analysis through natural language and electronic equipment | |
| CN111428503B (en) | Identification processing method and processing device for homonymous characters | |
| EP4131031A1 (en) | Recommendation method and device based on automatic feature grouping | |
| CN116737906A (en) | Information display method, device, electronic equipment and storage medium | |
| CN116563028A (en) | AI-based report data verification method, system and storage medium | |
| CN115167965A (en) | Transaction progress bar processing method and device | |
| CN114840530A (en) | Multi-element visual data model construction method and system | |
| CN114971962A (en) | Student homework evaluation method and device, electronic device and storage medium | |
| CN110543636B (en) | Training data selection method for dialogue system | |
| CN110413899B (en) | Storage resource optimization method and system for server storage news | |
| CN108763565B (en) | Deep learning-based data automatic association matching construction method | |
| CN113407827A (en) | Information recommendation method, device, equipment and medium based on user value classification | |
| CN117808124B (en) | Llama 2-based text simplification method | |
| CN111046300A (en) | Method and device for determining crowd attributes of users | |
| CN117235249B (en) | Intelligent creation method and system based on knowledge and data dual drive | |
| CN117891826A (en) | Method and device for constructing large vertical field model based on 12345 data | |
| CN117725191B (en) | Guide information generation method and device of large language model and electronic equipment | |
| CN114722163B (en) | Data query method and device, electronic equipment and storage medium | |
| Azuar et al. | Interactive Dashboard For Tracking System Dashboard Using Power Bi |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |