CN116151235A

CN116151235A - Article generating method, article generating model training method and related equipment

Info

Publication number: CN116151235A
Application number: CN202211255662.4A
Authority: CN
Inventors: 吕乐宾; 蒋宁; 肖冰; 李宽; 丁隆耀
Original assignee: Mashang Xiaofei Finance Co Ltd
Current assignee: Mashang Xiaofei Finance Co Ltd
Priority date: 2022-10-13
Filing date: 2022-10-13
Publication date: 2023-05-23

Abstract

The embodiment of the application discloses an article generating method, an article generating model training method and related equipment. The method comprises the following steps: acquiring event key information of a target event; matching the event key information with a pre-established article database, and screening target historical sentence data matched with the event key information from the article database; the article database comprises a plurality of historical articles, and each historical article comprises at least one historical sentence data; and inputting the event key information and the target historical sentence data into a pre-trained article generation model to obtain a target article corresponding to the event key information. According to the technical scheme, the efficiency and the intelligence of article generation can be improved, and the semantic consistency and the semantic diversity of target articles are ensured.

Description

Article generating method, article generating model training method and related equipment

Technical Field

The present disclosure relates to the field of artificial intelligence language technologies, and in particular, to an article generating method, an article generating model training method, and related devices.

Background

Even the person who reads ten thousand rolls is hard to learn without knowing tiredness in the today of iterative information development, and the ability to collect and produce content anytime and anywhere is a common goal of each creator. The robot is expressed and authored like a person, is one of important prospects of artificial intelligence, and one of the core technical fields for realizing the prospect is intelligent authoring. Intelligent authoring has been developed rapidly in recent years and has also shown increasing value in applications, such as intelligent authoring of news events using robots.

In the related art, the following manner is generally adopted when the robot is used for intelligent manuscript writing: one is to train a deep neural network model using article related information and then generate articles using the trained deep neural network model. The method is completely dependent on the deep neural network, so that the content of the generated article is uncontrollable, and the phenomenon of unsmooth semantics is easy to occur. And the other is to set a specific template in a specific field manually, so that the article information is filled in a template mode, and the article content is perfected. Although the method has higher readability, the method has poorer diversity, and the generated article line mode is basically fixed every time, if the method relates to the new article field, a new template needs to be designed independently by relying on manpower, and the method is time-consuming and labor-consuming.

Disclosure of Invention

The embodiment of the application aims to provide an article generating method, an article generating model training method and related equipment, which are used for solving the problems of poor semantic consistency and poor diversity of articles generated intelligently in the prior art.

In order to solve the technical problems, the embodiment of the application is realized as follows:

in one aspect, an embodiment of the present application provides an article generating method, including:

Acquiring event key information of a target event;

matching the event key information with a pre-created article database, and screening target historical sentence data matched with the event key information from the article database; the article database comprises a plurality of history articles, and each history article comprises at least one history statement data;

and inputting the event key information and the target historical sentence data into a pre-trained article generation model to obtain a target article corresponding to the event key information. In another aspect, an embodiment of the present application provides an article generating model training method, including:

acquiring a sample article, carrying out semantic analysis on the sample article, and determining sample key information corresponding to the sample article;

carrying out matching processing on the sample key information and a pre-created article database, and screening sample historical sentence data matched with the sample key information from the article database; the article database comprises a plurality of history articles, and each history article comprises at least one history statement data;

taking the sample key information and the sample history statement data as input data of an article generation model to be trained to obtain predicted article data corresponding to the sample key information;

Training the article generating model to be trained based on the predicted article data and the sample articles, wherein the trained article generating model is used for generating target articles corresponding to the event key information according to the event key information and the target historical sentence information of the target event; and the target historical sentence data is the historical sentence data which is screened from the article database and is matched with the event key information by carrying out matching processing on the event key information and the article database.

In still another aspect, an embodiment of the present application provides an article generating apparatus, including:

the acquisition module is used for acquiring event key information of the target event;

the screening module is used for carrying out matching processing on the event key information and a pre-established article database, and screening target historical sentence data matched with the event key information from the article database; the article database comprises a plurality of history articles, and each history article comprises at least one history statement data;

and the generating module is used for inputting the event key information and the target historical sentence data into a pre-trained article generating model to obtain a target article corresponding to the event key information.

In still another aspect, an embodiment of the present application provides an article generating model training apparatus, including:

the acquisition module is used for acquiring a sample article, carrying out semantic analysis on the sample article and determining sample key information corresponding to the sample article;

the screening module is used for carrying out matching processing on the sample key information and a pre-established article database, and screening sample historical sentence data matched with the sample key information from the article database; the article database comprises a plurality of history articles, and each history article comprises at least one history statement data;

the prediction module is used for taking the sample key information and the sample historical sentence data as input data of an article generation model to be trained to obtain predicted article data corresponding to the sample key information;

the training module is used for training the article generating model to be trained based on the predicted article data and the sample articles, and the trained article generating model is used for generating target articles corresponding to the event key information according to the event key information of the target event and the target historical sentence information; and the target historical sentence data is the historical sentence data which is screened from the article database and is matched with the event key information by carrying out matching processing on the event key information and the article database.

In yet another aspect, an embodiment of the present application provides an electronic device, including a processor and a memory electrically connected to the processor, where the memory stores a computer program, and the processor is configured to call and execute the computer program from the memory to implement the above-mentioned article generation model training method, or the processor is configured to call and execute the computer program from the memory to implement the above-mentioned article generation model training method.

In yet another aspect, embodiments of the present application provide a computer readable storage medium storing a computer program executable by a processor to implement the above-described article generation model training method, or executable by a processor to implement the above-described article generation model training method.

By adopting the technical scheme of the embodiment of the application, the event key information of the target event is obtained, the event key information and the article database which is created in advance are subjected to matching processing, and the target historical sentence data matched with the event key information are screened out, wherein the article database comprises a plurality of historical articles, and each historical article comprises at least one historical sentence data. Because the sentence data in the article not only comprises text contents but also semantic information of the text, the event key information and the historical sentence data are matched, so that the screened target historical sentence data and the event key information not only have the text contents with high matching but also have the semantic information with high matching. Furthermore, the event key information and the screened target historical sentence data are input into a pre-trained article generation model, so that the article generation model can generate a corresponding target article based on the target historical sentence data with the text content and the semantic information which are highly matched with the event key information, the target article is matched with the historical sentence data in the historical article (such as sentence structure matched with the historical sentence data), and the article generation efficiency and intelligence are greatly improved. In addition, the article generating process depends on the highly matched historical sentence data and the article generating model, and semantic consistency and diversity of the target articles are ensured based on diversity of the historical sentence data and intellectualization of the article generating model.

Drawings

In order to more clearly illustrate one or more embodiments of the present specification or the prior art, the drawings that are required for the description of the embodiments or the prior art will be briefly described, and it is apparent that the drawings in the following description are only some embodiments described in one or more embodiments of the present specification, and other drawings may be obtained according to these drawings without inventive effort for a person of ordinary skill in the art.

FIG. 1 is a schematic flow chart diagram of an article generation method according to an embodiment of the present application;

FIG. 2 is a schematic diagram of an article generation method according to an embodiment of the present application;

FIG. 3 is a schematic flow chart diagram of an article generation model training method according to an embodiment of the present application;

FIG. 4 is a schematic block diagram of an article generation model according to an embodiment of the present application;

FIG. 5 is a schematic diagram of an article generation model training method according to an embodiment of the present application;

FIG. 6 is a schematic block diagram of an article generating apparatus according to an embodiment of the present application;

FIG. 7 is a schematic block diagram of an article generation model training apparatus according to an embodiment of the present application;

Fig. 8 is a schematic block diagram of an electronic device according to an embodiment of the present application.

Detailed Description

The embodiment of the application provides an article generating method, an article generating model training method and related equipment, which are used for solving the problems of poor semantic consistency and poor diversity of an article generated intelligently in the prior art.

In order to better understand the technical solutions in the present application, the following description will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only some embodiments of the present application, not all embodiments. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without making any inventive effort, shall fall within the scope of the present application.

Intelligent authoring has been developed rapidly in recent years and has also shown increasing value in applications, such as intelligent authoring of news events using robots. The following briefly lists several related art intelligent contribution modes. Firstly, training a deep neural network model by using article related information, enabling the deep neural network model to learn the statement mode and logic of text data, and generating articles by using the trained deep neural network model. The method is completely dependent on the deep neural network, so that the content of the generated article is uncontrollable, and the phenomenon of unsmooth semantics is easy to occur. Secondly, setting a specific template in a specific field manually, so that the article information is filled in a template mode, and the article content is perfected. Although the readability of the method is higher, the variety is poor, the line text mode of each generated article is basically fixed, if the method relates to the new article field, a new template needs to be designed independently by relying on manpower, the time and the labor are consumed, and the content is single. Thirdly, searching information related to the vocabulary entries in the whole network through the keyword entries, sorting and arranging the searched information, and finally outputting the searched information in a manner of articles. The method can only be arranged according to the articles which are already released on the internet, and can not be used for writing original articles, so that the flexibility and the diversity are poor. According to the article generating method, the event key information of the target event is obtained, the event key information and the article database which is created in advance are subjected to matching processing, and target historical sentence data matched with the event key information are screened out, wherein the article database comprises a plurality of historical articles, and each historical article comprises at least one historical sentence data. Because the sentence data in the article not only comprises text contents but also semantic information of the text, the event key information and the historical sentence data are matched, so that the screened target historical sentence data and the event key information not only have the text contents with high matching but also have the semantic information with high matching. Furthermore, the event key information and the screened target historical sentence data are input into a pre-trained article generation model, so that the article generation model can generate a corresponding target article based on the target historical sentence data with the text content and the semantic information which are highly matched with the event key information, the target article is matched with the historical sentence data in the historical article (such as sentence structure matched with the historical sentence data), and the article generation efficiency and intelligence are greatly improved. In addition, the article generating process depends on the highly matched historical sentence data and the article generating model, and semantic consistency and diversity of the target articles are ensured based on diversity of the historical sentence data and intellectualization of the article generating model.

The article generating method and the article generating model training method provided by the embodiment of the application can be executed by the electronic device or software installed in the electronic device, and in particular, the electronic device can be a terminal device or a server. The terminal device may include a smart phone, a notebook computer, an intelligent wearable device, an on-board terminal, and the like, and the server may include an independent physical server, a server cluster composed of a plurality of servers, or a cloud server capable of performing cloud computing.

FIG. 1 is a schematic flow chart diagram of an article generating method according to an embodiment of the present application, as shown in FIG. 1, the method includes:

s102, acquiring event key information of a target event.

The event key information is core elements and knowledge which can represent key information of a target event.

Optionally, the event key information accords with a preset information structure, and the preset information structure is used for standardizing an information display mode of the event key information. For example, the preset information structure is "key information field: key information content ", the event key information listed below accords with the preset information structure.

Character: thirdly, stretching;

events: speaking;

Results: providing fundamental compliance for us.

In the event key information, the character, the event and the result are key information fields; "Zhang Sang", "speak-to-speak" and "provide us with key information content that follows at all" for each key information field individually.

Alternatively, when the event key information of the target event is acquired, the event content of the target event may be provided by the user, for example, the event content is stated in a simple statement manner, then the semantic information of the event content is analyzed by the electronic device, the event key information is extracted therefrom, and the event key information is organized according to a preset information structure. Alternatively, the user may directly provide the event key information of the target event, for example, the user arranges the event key information according to a preset information structure.

S104, matching the event key information with a pre-created article database, and screening target historical sentence data matched with the event key information from the article database; the article database comprises a plurality of history articles, and each history article comprises at least one history statement data.

The history articles in the article database can be obtained by the electronic device through one or more network channels (such as various large websites, forums and the like). Alternatively, the electronic device may periodically obtain published history articles from various network channels and store the obtained history articles in the article database. The history sentence data is sentence data included in a history article.

Optionally, the matching of the target historical sentence data with the event key information may include that a similarity between the target historical sentence data and the event key information reaches a preset similarity threshold. When the similarity between the target historical sentence data and the event key information is calculated, the similarity between the target historical sentence data and the event key information in at least one aspect of words, word ordering, semantic information, content tendency and the like can be calculated.

S106, inputting the event key information and the target historical sentence data into a pre-trained article generation model to obtain a target article corresponding to the event key information.

The article generating model is trained based on sample key information of a plurality of sample articles and sample historical sentence data matched with the sample key information. The sample key information is the core element and knowledge which can represent the key information in the sample article. Optionally, the sample key information accords with a preset information structure, and the preset information structure is used for standardizing an information display mode of the sample key information.

The training process of the article generating model will be described in detail in the following embodiments, which will not be repeated here.

In one embodiment, the event key information includes at least one key information element, and the target history statement data matched with the event key information includes sub-target history statement data corresponding to each of the at least one key information element. In the case where the event key information includes a plurality of key information elements, the element structure corresponding to each key information element may be the same or different. The following accords with the preset information structure 'key information field': the event key information of the key information content "is exemplified by:

character: thirdly, stretching;

events: speaking;

results: providing fundamental compliance for us.

In the listed event key information, each set of "key information fields: the key information content is a key information element, and it can be seen that the event key information includes three key information elements, and each key information element has the same preset information structure key information field: key information content).

In this embodiment, the matching process is performed on the event key information and the article database created in advance, and when the target historical sentence data matched with the event key information is screened from the article database, the following actions may be performed:

First, a first similarity between each key information element and each historical sentence data in the article database is calculated.

Wherein a first similarity between the key information element and the historical sentence data may be calculated in at least one of text, word ordering, semantic information, content trends, and the like. Alternatively, the key information elements and the historical sentence data may be semantically analyzed by using an existing semantic analysis algorithm, so as to analyze the first similarity of the key information elements and the historical sentence data in terms of text, word ordering, semantic information, content tendency, and the like.

Secondly, screening sub-target historical sentence data corresponding to each key information element from the article database according to the first similarity between each key information element and each historical sentence data in the article database; the first similarity between the sub-target historical sentence data corresponding to any one key information element and the any one key information element is larger than or equal to a first preset threshold value.

Optionally, the first similarities between the sub-target historical sentence data corresponding to any one key information element and the any one key information element are located in the first N (arranged in the order of the similarities from high to low) of all the first similarities, and N is an integer greater than or equal to 1.

In this embodiment, the event key information element is split according to the key information element, and the history article is split according to the history statement data, so that the split key information element and the history statement data are matched to obtain the history statement data with higher similarity to the key information element. By calculating the first similarity between each key information element and each history statement data in the article database, and screening sub-target history statement data with higher similarity to the key information elements according to the first similarity, the screened sub-target history statement data and the key information elements have higher similarity in at least one aspect of words, word ordering, semantic information, content tendency and the like, so that in the subsequent steps, the published sub-history statement data with higher similarity to the key information elements can be utilized to generate target articles, and the readability and semantic consistency of the generated target articles are stronger.

In one embodiment, the event key information includes at least one key information element, and the target history statement data matched with the event key information includes sub-target history statement data corresponding to each of the at least one key information element. The definition of the key information element is the same as in the above embodiment, and will not be described here again. The event key information and a pre-created article database are subjected to matching processing, and when target historical sentence data matched with the event key information is screened from the article database, the following actions A1-A5 can be executed:

Action A1, determining a plurality of first history articles from the article database.

And (2) calculating second similarity between the event key information and a plurality of first historical articles in the article database.

The event key information and the historical articles are respectively taken as a whole, and similarity calculation is carried out on the event key information and a plurality of first historical articles in an article database. The plurality of first history articles may be some or all of the history articles in the article database. Optionally, the event key information and the first history article may be subjected to semantic analysis by using an existing semantic analysis algorithm, so as to analyze the first similarity of the event key information and the first history article in terms of text, word ordering, semantic information, content tendency, and the like.

Act A3, according to the event key information and the second similarity among the first historical articles, screen the second historical articles from the first historical articles; the second similarity between the second history article and the event key information is greater than or equal to a second preset threshold.

Optionally, the second similarity between the second history article and the event key information is located in the first M (in order of high-to-low similarity) of all the second similarities, and M is an integer greater than or equal to 1.

And (4) calculating a third similarity between each key information element and each history statement data in the second history article.

Wherein a third similarity between the key information element and each of the history sentence data in the second history article may be calculated in at least one of text, word ordering, semantic information, content trends, and the like. Optionally, the key information elements and the historical sentence data can be subjected to semantic analysis by using an existing semantic analysis algorithm, so that the third similarity of the key information elements and the historical sentence data in terms of characters, word ordering, semantic information, content tendency and the like is analyzed.

Action A5, screening sub-target historical sentence data corresponding to each key information element from the second historical article according to the third similarity between each key information element and each historical sentence data in the second historical article; and the third similarity between the sub-target historical sentence data corresponding to any one key information element and the any one key information element is larger than or equal to a third preset threshold value.

Optionally, the third similarity between the sub-target historical sentence data corresponding to any one key information element and the any one key information element is located in the first L (arranged in the order of the similarity from high to low) of all the third similarities, and L is an integer greater than or equal to 1.

In this embodiment, when the actions A2-A3 are executed, the event key information and the history articles are respectively used as a whole to perform similarity calculation, so that a second history article with higher similarity with the event key information is screened out, that is, the screening of the second history article is more focused on the similarity of the whole articles, and the similarity of the whole articles can reflect the similarity of the article types to a certain extent, so that the similarity calculation is performed by using the event key information and the history articles as a whole, and the article types with higher (or same) similarity corresponding to the screened second history article and the event key information can be ensured. When the actions A4-A5 are executed, similarity calculation is performed on each key information element in the event key information and the history statement data in the second history article respectively, so that sub-target history statement data with higher similarity with the key information elements are screened out, that is, screening of the sub-target history statement data (or target history statement data) is more focused on similarity on statements, and similarity of statement structures can be reflected to a certain extent due to similarity on the statements, such as a word ordering mode, a language description mode and the like, and therefore statement structures with higher similarity (or the same) between the screened target history statement data and the key information elements can be ensured by performing similarity calculation on the key information elements with smaller dimensions and the history statement data. Therefore, the article screening with higher similarity is performed on the overall dimension of the article, and then the sentence screening with higher similarity is performed on the sentence dimension, so that finally screened target historical sentence data are matched with event key information on the article category and the sentence structure, and the readability and semantic consistency of the generated target article are higher.

In one embodiment, each history article in the article database corresponds to a respective article category. When executing the action A1, the event category corresponding to the target event may be determined first, and then the event category of the target event and the article category of each history article in the article database are matched to determine the target article category matched with the event category of the target event. Further, it is determined that a history article corresponding to the target article category in the article database is a first history article.

In this embodiment, before calculating the second similarity between the event key information and the plurality of first history articles in the article database, a plurality of first history articles corresponding to the article types matched with the event types may be screened from the article database according to the event types of the target event, and then the event key information and the screened plurality of first history articles are further matched, so as to ensure that the matched history articles and the event key information correspond to the article types with higher similarity (or the same) and further make the semantic information of the finally generated target articles more accurate.

In one embodiment, published history articles may be obtained through one or more web channels (e.g., various large websites, forums, etc.), and the obtained history articles stored in an article database. Optionally, the obtained history articles may be initially screened, for example, duplicate articles, articles with low content, articles with low writing quality, articles with obsolete structure, etc., so that the history articles stored in the article database are all high quality articles.

After creating the article database, the article-related content to be updated, including the published articles and/or the specified types of characters, may continue to be collected. And when the article database meets the preset updating condition, updating the article database based on the article related content.

Wherein the preset updating condition comprises at least one of the following: the number of the released articles reaches a preset threshold value, the last update time of the article database reaches a preset duration from the current time, and at least one character of a designated type exists in the historical articles to be updated. The specified type of character may include at least one of: words for network, languages or words having a specific meaning in a specific period, hot words. Since the characters of the specified type may be updated over time, it is necessary to update the article database, for example, to synchronize the updated characters in the article database, as at least one of the characters of the specified type is updated. By updating the language, vocabulary and/or the article in the article database in time, the problems of over-time of the article structure, inaccurate meaning of the language or vocabulary and the like of the history article stored in the article database can be avoided, so that the history article in the article database has accurate semantic meaning. Especially in the news field, by updating the language, vocabulary and/or the articles in the article database in time, not only is the accurate semantic meaning of the news articles in the article database ensured, but also the frontier of the news articles is ensured.

In one embodiment, each key information element includes a key information field and key information content conforming to a preset information structure. When acquiring the event key information of the target event, the key information field matched with the event category of the target event can be determined first, and then the key information content corresponding to the target event and matched with the key information field can be acquired. And further, according to a preset information structure, each key information field and the corresponding key information content are subjected to structuring processing to obtain event key information.

Wherein, a plurality of key information fields such as time, place, person, event content, event result, event cause, person relationship, etc. of event occurrence can be preset. Since not all events have the same key information field, when acquiring the event key information of the target event, the key information field matched with the event category of the target event may be determined from a plurality of preset key information fields. Alternatively, the respective corresponding key information fields may be preset for the events of different event types, respectively. For example, for news-like events, corresponding key information fields may be set including people, places, and event results.

The preset information structure may be used to specify an information display manner of the event key information, and if the event key information includes a plurality of key information fields, the preset information structure may also be used to specify a sorting manner between the plurality of key information fields (or a sorting manner between a plurality of key information elements). The structuring of the key information fields and their corresponding key information content according to the preset information structure may include specifying and/or ordering the information structure of the key information fields and their corresponding key information content. For example, the preset information structure is "key information field: key information content ", and, for news-like events, the ordering of the corresponding key information fields is: the character, event and result are the key information of the event listed below accords with the preset information structure.

Character: thirdly, stretching;

events: speaking;

results: providing fundamental compliance for us.

In this embodiment, by determining the key information fields matched with the event category of the target event and performing structural processing on each key information field and the corresponding key information content according to the preset information structure, the event key information conforming to the preset information structure is obtained, so that the obtained event key information conforms to the preset information structure, and thus, when the trained article generation model is utilized to generate the article, the event key information is normalized, and the semantic information of the generated article is more accurate.

Fig. 2 is a schematic diagram of an article generating method according to an embodiment of the present application, and a specific implementation manner of the article generating method according to the embodiment will be described in detail below with reference to fig. 2.

Before generating an article for a target event, firstly, published historical articles are acquired, and an article database is created according to the acquired historical articles.

Wherein published history articles may be obtained periodically through one or more network channels (e.g., various large websites, forums, etc.). In order to facilitate management of the historical articles stored in the article database, the historical articles can be classified and stored according to article types of the historical articles. The article category can be flexibly divided according to different dimensions, for example, according to different article structures, the articles can be divided into news articles, story articles, debate articles and the like; articles can be classified into music-like articles, sports-like articles, entertainment-like articles, etc. according to the field of events. After creating the article database, article-related content, including published articles, specified types of characters, etc., may continue to be periodically collected from one or more web channels (e.g., various large websites, forums, etc.), and the article database may be updated in time based on the article-related content.

When a corresponding target article needs to be generated for a target event, acquiring event key information of the target event. The event key information comprises at least one key information element, and each key information element accords with a preset information structure. The preset information structure is used for standardizing an information display mode of the event key information and a sequencing mode among key information elements. Optionally, the key information field matched with the event category of the target event can be determined first, then the key information content corresponding to the target event and matched with the key information field is obtained, and then each key information field and the corresponding key information content are structured according to the preset information structure to obtain the event key information.

Table 1 below schematically shows event key information conforming to a preset information structure. The preset information structure standardizes the following: the first column in the table is the key information field, and the second column is the key information content corresponding to the key information field. As shown in table 1, the first column of subject behaviors, time, event 1, event 2, event 3, event 4, and event 5 belongs to key information fields matching the event category of the target event, and the second column of X country publishes interest rate, 2022 years, adding 25 base points to 0.75%, one quarter 0.94%, two quarter 1.51%, three quarter 1.86%, and four quarter 2.14% belongs to key information contents corresponding to the key information fields.

TABLE 1

Subject behavior	Published interest rate in the X country
		Time	2022 years
Event 1	Adding 25 base points to 0.75%
		Event 2	0.94% in one quarter,
event 3	Second quarter 1.51%
		Event 4	Three quarters 1.86%
Event 5	Four seasons 2.14%

After the event key information of the target event is acquired, matching the event key information with a plurality of first historical articles in an article database, and calculating the similarity between the event key information and each first historical article. Wherein the plurality of first history articles may be part or all of the history articles in the article database. If the first plurality of history articles are all history articles in the article database, the similarity between the event key information and each history article in the article database can be directly calculated. If the plurality of first history articles are part of the history articles in the article database, the plurality of first history articles are determined first, and then the similarity between the event key information and each first history article is calculated.

In this step, optionally, an event category corresponding to the target event may be determined first, and then the event category of the target event and the article category of each history article in the article database are matched to determine a target article category that matches the event category of the target event. Further, it is determined that a history article corresponding to the target article category in the article database is a first history article. For example, music articles, sports articles and entertainment articles are stored in the article database according to the article types, and if the event type corresponding to the target event is an entertainment event, the entertainment article in the article database is determined to be a first historical article.

And screening second historical articles from the plurality of first historical articles according to the similarity between the event key information and each first historical article, wherein the similarity between the second historical articles and the event key information is greater than or equal to a second preset threshold value, and/or the similarity between the second historical articles and the event key information is positioned in the first M (arranged according to the sequence of the similarity from high to low) of all the similarities, and M is an integer greater than or equal to 1.

In this embodiment, assuming that m=10, by calculating the similarity between the event key information and each first history article, the first history articles with the similarity being in the first 10 are selected as the second history articles according to the order of the similarity from high to low. Taking the event key information shown in table 1 as an example, by executing the above steps, 10 history articles (i.e., the second history article) corresponding to the event key information shown in table 1 can be screened out. The second history articles (the actual retrieved history articles include 10, and only 3 of them are listed here), it should be noted that only the article content of the retrieved history articles is listed here, and the article release (such as pictures, font sizes, etc.) of the history articles is not limited here.

Article one: the national statistical office published today at 18 years that the GDP quarter increase will take on a "decreasing season by season" situation. This has been documented in the two quarters prior to this year: according to the data published by the national statistical office, the GDP is increased by 18.3% in a quarter by the same ratio, and the GDP is increased by 5.0% in two years on average; the GDP is increased by 7.9% in the same proportion in the second quarter and is increased by 5.5% in average in two years.

Article two: the national economy population in the first three seasons keeps a recovery situation (2021, 10, 18 days in 10 months) and the national statistical bureau in the first three seasons, the first quarter is increased by 18.3% in comparison with the first quarter, and the average two years are increased by 5.0%; the quarters are increased by 7.9% in a same ratio, and the average growth is 5.5% in two years; the three quarters are increased by 4.9% in a same ratio, and the average increase in two years is 4.9%.

Article three: the correction data published by the business department of China, 24 days, shows that the total actual domestic production (GDP) of China, third quarter, is increased by 2.1% according to the annual rate calculation, and is slightly higher than the last estimated 2%, but lower than the increase of 6.7% of the second quarter.

It can be seen that the second history article retrieved has a high similarity with the event key information shown in table 1.

Then, a similarity between each key information element and each history sentence data in the second history article is calculated. And screening sub-target historical sentence data corresponding to each key information element from the second historical articles according to the similarity between each key information element and each historical sentence data in the second historical articles, wherein the similarity between the sub-target historical sentence data corresponding to any one key information element and any one key information element is greater than or equal to a third preset threshold value, and/or the similarity between the sub-corresponding to any one key information element and any one key information element is positioned in the first L (arranged according to the sequence of the similarity from high to low) of all the similarities, and L is an integer greater than or equal to 1.

In this embodiment, assuming that l=3, by calculating the similarity between each key information element and each history statement data in the second history article, the history statement data with the similarity being located in the first 3 are screened out and used as 3 sub-target history statement data corresponding to each key information element respectively. Taking the event key information shown in table 1 as an example, by executing the above steps, 3 sub-target history sentence data corresponding to each key information element (or key information field) shown in table 1 can be screened. Sub-target historical sentence data corresponding to the retrieved partial key information elements is listed below.

Key information element "body behavior: the sub-objective historical sentence data corresponding to the published interest rate of the X country comprises: the national statistics office publishes that the GDP quarter increase in the present year at 18 days will take a 'decreasing season by season' situation; the national economy in the first three seasons is kept in a recovery situation as a whole; data previously published by the national statistical office.

Key information element "event 2: the sub-target historical sentence data corresponding to quarter 0.94% "includes: GDP increased by 18.3% in the same proportion for one quarter; a quarter of 18.3% of the same; average growth was 16.3% over a quarter.

It can be seen that the retrieved sub-target historical sentence data has a higher similarity with the corresponding key information element.

After screening the sub-target historical sentence data corresponding to each key information element, inputting each key information element and the sub-target historical sentence data corresponding to each key information element into a pre-trained article generating model, so that the pre-trained article generating model generates sentence data corresponding to each key information element according to each key information element and the sub-target historical sentence data corresponding to each key information element, and then generates a target article corresponding to event key information based on the sentence data corresponding to each key information element.

In this embodiment, the target article generated by the article generation model is exemplified as follows: the X nations store today's published interest rate resolution declaring 25 base points of interest to 0.75%. The aggregate forecast that the average level of official cash interest will be 0.94% for one quarter, 1.51% for two quarters, 1.86% for three quarters and 2.14% for four quarters for 2022 years.

It can be seen that, by adopting the article generating method provided by the embodiment of the application, by acquiring the event key information of the target event and performing matching processing on the event key information and the article database created in advance, target historical sentence data matched with the event key information is screened out, wherein the article database comprises a plurality of historical articles, and each historical article comprises at least one historical sentence data. Because the sentence data in the article not only comprises text contents but also semantic information of the text, the event key information and the historical sentence data are matched, so that the screened target historical sentence data and the event key information not only have the text contents with high matching but also have the semantic information with high matching. Furthermore, the event key information and the screened target historical sentence data are input into a pre-trained article generation model, so that the article generation model can generate a corresponding target article based on the target historical sentence data with the text content and the semantic information which are highly matched with the event key information, the target article is matched with the historical sentence data in the historical article (such as sentence structure matched with the historical sentence data), and the article generation efficiency and intelligence are greatly improved. In addition, the article generating process depends on the highly matched historical sentence data and the article generating model, and semantic consistency and diversity of the target articles are ensured based on diversity of the historical sentence data and intellectualization of the article generating model. In addition, when the target historical sentence data is screened, the article screening with high similarity (namely the second historical article) is firstly carried out on the overall dimension of the article, and then the sentence screening with high similarity is carried out on the sentence dimension, so that the finally screened target historical sentence data is matched with the event key information on the article category and the sentence structure, and the readability and the semantic consistency of the generated target article are further enhanced.

FIG. 3 is a schematic flow chart diagram of an article generation model training method according to an embodiment of the present application, as shown in FIG. 3, the method comprising:

s302, acquiring a sample article, carrying out semantic analysis on the sample article, and determining sample key information corresponding to the sample article.

The sample key information is core elements and knowledge capable of representing key information in the sample article. Optionally, the sample key information accords with a preset information structure, and the preset information structure is used for standardizing an information display mode of the sample key information. For example, the preset information structure is "key information field: key information content ", the sample key information listed below accords with the preset information structure.

Character: thirdly, stretching;

events: speaking;

results: providing fundamental compliance for us.

In the sample key information, the character, the event and the result are key information fields; "Zhang Sang", "speak-to-speak" and "provide us with key information content that follows at all" for each key information field individually.

Optionally, any existing semantic analysis algorithm may be used to perform semantic analysis on the sample article, so as to analyze sample key information in the sample article.

S304, matching the sample key information with a pre-created article database, and screening sample historical sentence data matched with the sample key information from the article database; the article database comprises a plurality of history articles, and each history article comprises at least one history statement data.

Optionally, the sample history statement data is matched with the sample key information, which may include that the similarity between the sample history statement data and the sample key information reaches a preset similarity threshold. When the similarity between the sample historical sentence data and the sample key information is calculated, the similarity between the sample historical sentence data and the sample key information in at least one aspect of words, word ordering, semantic information, content tendency and the like can be calculated.

S306, taking the sample key information and the sample history statement data as input data of an article generation model to be trained, and obtaining predicted article data corresponding to the sample key information.

S308, training an article generating model to be trained based on the predicted article data and the sample articles.

The trained article generation model is used for generating target articles corresponding to the event key information according to the event key information of the target event and the target historical sentence information, and the target historical sentence data is historical sentence data which is screened from the article database and is matched with the event key information by carrying out matching processing on the event key information and the article database.

By adopting the technical scheme of the embodiment of the application, the sample key information corresponding to the sample article is determined by carrying out semantic analysis on the sample article, the sample key information is matched with the article database which is created in advance, and the sample historical sentence data matched with the sample key information is screened out from the article database, wherein the article database comprises a plurality of historical articles, and each historical article comprises at least one historical sentence data. Because the sentence data in the article not only comprises text content but also semantic information of the text, the sample historical sentence data and the sample historical sentence data are subjected to matching processing, so that the screened sample historical sentence data and the sample key information not only have the text content with high matching, but also have the semantic information with high matching. Furthermore, the sample key information and the screened sample historical sentence data are used as input data of an article generating model to be trained, predicted article data corresponding to the sample key information is obtained, the article generating model to be trained is trained based on the predicted article data and the sample article, the article generating model to be trained can be trained based on the sample historical sentence data with text content and semantic information which are highly matched with the sample key information, and the trained article generating model at least has the following capabilities based on the diversity of the sample article: the method comprises the steps of accurately analyzing semantic information and text content of sample key information, accurately analyzing matching between sample historical sentence data and the sample key information, and accurately generating articles based on the sample key information and the sample historical sentence data. Further, when an article is generated using the trained article generation model, semantic consistency and diversity of the generated article can be ensured.

In one embodiment, the sample key information comprises at least one sample key information element, each sample key information element comprising a sample key information field and a sample key information content conforming to a preset information structure. When the sample key information corresponding to the sample article is determined by carrying out semantic analysis on the sample article, the sample key information field and the corresponding sample key information content can be extracted from the sample article according to the semantic analysis result of the sample article. And then carrying out structuring treatment on each sample key information field and the corresponding sample key information content according to a preset information structure to obtain sample key information.

Wherein, a plurality of key information fields such as time, place, person, event content, event result, event cause, person relationship, etc. of the sample event can be preset. Because not all events have the same key information field, when determining the sample key information corresponding to the sample article, the sample key information field matching (such as the same or similar) with the preset key information field and the sample key information content corresponding to each sample key information field can be extracted from the sample article.

The preset information structure may be used to specify an information display manner of the sample key information, and if the sample key information includes a plurality of sample key information fields, the preset information structure may also be used to specify a sorting manner between the plurality of sample key information fields (or a sorting manner between a plurality of sample key information elements). The structuring of the sample key information fields and their corresponding sample key information content according to the preset information structure may include specifying and/or ordering the information structure of the sample key information fields and their corresponding sample key information content. For example, the preset information structure is "sample key information field: sample key information content ", and, for news-like events, the ordering of the corresponding sample key information fields is: the key information of the sample listed below accords with the preset information structure.

Character: thirdly, stretching;

events: speaking;

results: providing fundamental compliance for us.

In one embodiment, as shown in FIG. 4, the article generation model to be trained includes: statement structure analysis layer, statement generation layer, article generation layer and full connection layer. When sample key information and sample historical sentence data are used as input data of an article generation model to be trained to obtain predicted article data corresponding to the sample key information, each level respectively executes the following actions:

And determining statement structure information corresponding to each sample key information element through a statement structure analysis layer.

Wherein the sentence structure information includes at least one of the following information: word ordering, connection relationships between adjacent words, semantic information, and the like. The sentence structure analysis layer can be executed based on any existing semantic analysis algorithm, namely, semantic analysis is carried out on the input sample historical sentence data by utilizing the semantic analysis algorithm, and sentence structure information corresponding to the sample key information elements can be determined according to semantic analysis results.

And generating sample sentence data corresponding to each sample key information element according to sentence structure information corresponding to the sample key information element through a sentence generation layer.

According to statement structure information of the sample historical statement data, statement modes of the sample historical statement data can be obtained, and sample key information elements are stated by the statement modes, so that sample statement data corresponding to the sample key information elements can be generated. Taking the key information element and the corresponding sub-target history statement data in the above embodiment as an example, specifically, the key information element "main body behavior: the X country publishes interest rate as a sample key information element, and the sub-target historical statement data is published by the country statistics bureau at 18 days that the GDP quarter amplification will show a situation of decreasing from quarter to quarter; the national economy in the first three seasons is kept in a recovery situation as a whole; the data previously published by the national statistical office is used as sample history statement data. According to the statement mode of the sample historical statement data, sample statement data corresponding to the sample key information element can be generated as follows: "X nationally store today's published interest rate resolution".

And combining sample sentence data corresponding to each sample key information element according to the element position of each sample key information element in the sample key information by the article generation layer to obtain prediction article data corresponding to the sample key information.

The article data are article data output based on current model parameters when the model is trained in each iteration; and the sample article is the standard output data of the article generation model. There is a difference between the predicted article data and the sample article as the standard output data that determines whether the article generation model stops iterating.

For example, sample key information element "body behavior: the corresponding sample sentence data of the published interest rate of the X country is "the published interest rate resolution of the X country by the present day". Sample key information element "event 1: the sample sentence data corresponding to the addition of 25 base points to 0.75% "is" the addition of 25 base points to 0.75% "sample key information element" event 2: the corresponding sample sentence data for a quarter of 0.94% "is" average official cash interest level 2022 will be 0.94% ", a quarter. The sample sentence data are combined, and the predicted article data corresponding to the sample key information are obtained as follows: the X nationally stores today's published interest rate resolution declaring 25 base points for addition to 0.75%, the average level of official cash interest rate will be 0.94% for one quarter 2022.

When training an article generation model to be trained based on predicted article data and sample articles, the following actions may be performed: and carrying out iterative training on the article generation model to be trained according to the predicted article data, the sample articles and the preset convergence condition through the full connection layer.

Optionally, the preset convergence condition may include at least one of: the loss value of the article generating model is smaller than or equal to a preset loss threshold value, and the iteration number reaches the preset number threshold value.

Taking the example that the preset convergence condition includes that the loss value of the article generating model is smaller than or equal to a preset loss threshold value, optionally, the sample article includes a plurality of sample characters, and the predicted article data includes a probability value of each sample character at each character position in the sample article. If the probability value is represented by P, u _i Representing the character at the i-th position, P (u) _i ) A probability value representing that the character is located at the i-th position.

When the article generation model is iteratively trained according to the predicted article data, the sample articles and the preset convergence condition, the following actions can be executed:

first, a loss value of an article generation model is calculated from probability values of each sample character at character positions in a sample article.

And secondly, stopping iteration if the loss value is smaller than or equal to a preset loss threshold value. And if the loss value is greater than the preset loss threshold, adjusting model parameters of the article generation model, and performing iterative training based on the adjusted model parameters until the loss value is less than or equal to the preset loss threshold.

In this embodiment, the loss function of the article generating model is not limited, and the following equation (1) lists a loss function used by a common language model. In the following loss function, for the sample article u= [ U1, U2, …, un]U1, u2, un represent each sample character in the sample article, P (u) _i |u _i-k ,…，u _i-1 ) The representation is: with the sample characters at the previous positions i-1, i-2, i-3 and … … known, each sample character in the sample article is located at the i-th positionProbability values over locations. θ is a fixed parameter of the likelihood function, L is a loss value,

L ₁ (U)＝∑ _i logP(u _i |u _i-k ,…，u _i-1 ；θ) (1)

it can be seen that, under the condition of calculating the probability value of each sample character at each character position in the sample article, the loss value of the article generating model to be trained can be calculated according to the formula (1), and then whether to continue iterative training is determined according to the magnitude relation between the loss value and the preset loss threshold value, or the trained article generating model is obtained after stopping iteration.

Fig. 5 is a schematic diagram of an article generation model training method according to an embodiment of the present application, and with reference to fig. 5, a specific implementation manner of the article generation model training method in the embodiment will be described in detail below.

Before training an article generation model, firstly, published historical articles are acquired, and an article database is created according to the acquired historical articles.

And then, acquiring a plurality of sample articles, carrying out semantic analysis on each sample article, and determining sample key information corresponding to each sample article. The definition and determination manners of the sample key information are similar to those of the event key information in the above embodiment, so that the description is omitted.

And then, carrying out matching processing on the sample key information and the article database, and screening sample historical sentence data matched with the sample key information from the article database. Specifically, for each sample article, sample key information corresponding to the sample article is first matched with a first sample historical article in an article database, and the similarity between the sample key information and each first sample historical article is calculated. Wherein the plurality of first sample history articles may be some or all of the history articles in the article database. If the first plurality of sample history articles are all history articles in the article database, the similarity between the sample key information and each history article in the article database can be directly calculated. If the plurality of first sample history articles are part of the history articles in the article database, the plurality of first sample history articles are determined first, and then the similarity between the sample key information and each first sample history article is calculated.

In this step, optionally, the sample article category corresponding to the sample article may be determined first, then the sample article category and the sample article category of each sample article in the article database are subjected to matching processing, and a history article matched with the sample article category is determined as the first sample history article. For example, music articles, sports articles and entertainment articles are stored in the article database according to the article types, and if the sample article types are entertainment articles, the entertainment articles in the article database are determined to be the first sample historical articles.

And screening second sample historical articles from the plurality of first sample historical articles according to the similarity between the sample key information and each first sample historical article, wherein the similarity between the second sample historical articles and the sample key information is greater than or equal to a second preset threshold value, and/or the similarity between the second sample historical articles and the sample key information is positioned in the first M (arranged in the order of the similarity from high to low) of all the similarities, and M is an integer greater than or equal to 1.

Then, a similarity between each sample key information element and each history sentence data in the second sample history article is calculated. And screening sample historical sentence data corresponding to each sample key information element from the second sample historical articles according to the similarity between each sample key information element and each historical sentence data in the second sample historical articles, wherein the similarity between the sample historical sentence data and the sample key information element is greater than or equal to a third preset threshold value, and/or the similarity between the sample historical sentence data and the sample key information element is located in the first L (arranged according to the sequence of the similarity from high to low) of all the similarities, and L is an integer greater than or equal to 1.

After screening out sample historical sentence data corresponding to each sample key information element, inputting each sample key information element and the corresponding sample historical sentence data thereof into an article generating model to be trained, so that the article generating model to be trained generates sample sentence data corresponding to each sample key information element according to each sample key information element and the corresponding sample historical sentence data thereof, and then combines the sample sentence data according to the element position of each sample key information element in the sample key information to obtain predicted article data corresponding to the sample key information.

And finally, carrying out iterative training on the article generating model according to the predicted article data, the sample article and the preset convergence condition to obtain a trained article generating model. The sample article can be regarded as label data or standard output data of the article generating model and is used for comparing with predicted article data output by the article generating model, so that a loss value of the article generating model is calculated. Alternatively, the loss value of the article generation model may be calculated using the above formula (1). And stopping iteration if the loss value is smaller than or equal to the preset loss threshold value. And if the loss value is greater than the preset loss threshold, adjusting model parameters of the article generation model, and performing iterative training based on the adjusted model parameters until the loss value is less than or equal to the preset loss threshold.

In addition, whether iteration is stopped or not can be determined according to the iteration number of the article generation model, and if the iteration number reaches a preset number threshold, iteration is stopped.

According to the embodiment, the sample key information elements and the screened sample historical sentence data are used as input data of the article generating model to be trained, so that predicted article data corresponding to the sample key information is obtained, the article generating model to be trained is trained based on the predicted article data and the sample article, the article generating model to be trained can be trained based on the sample historical sentence data with the text content and the semantic information which are highly matched with the sample key information elements, and the trained article generating model has at least the following capability based on the diversity of the sample articles: the method comprises the steps of accurately analyzing semantic information and text content of sample key information, accurately analyzing matching between sample historical sentence data and the sample key information, and accurately generating articles based on the sample key information and the sample historical sentence data. Further, when an article is generated using the trained article generation model, semantic consistency and diversity of the generated article can be ensured.

In summary, particular embodiments of the present subject matter have been described. Other embodiments are within the scope of the following claims. In some cases, the actions recited in the claims can be performed in a different order and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may be advantageous.

The above method and device for generating an article and training an article generating model are based on the same ideas.

Fig. 6 is a schematic block diagram of an article generating apparatus according to an embodiment of the present application, as shown in fig. 6, the apparatus includes:

an acquisition module 61, configured to acquire event key information of a target event;

the screening module 62 is configured to perform matching processing on the event key information and a pre-created article database, and screen target historical sentence data matched with the event key information from the article database; the article database comprises a plurality of history articles, and each history article comprises at least one history statement data;

And the generating module 63 is configured to input the event key information and the target historical sentence data into a pre-trained article generating model, so as to obtain a target article corresponding to the event key information.

In one embodiment, the event key information includes at least one key information element; the target historical sentence data matched with the event key information comprises sub-target historical sentence data corresponding to each key information element in the at least one key information element;

the filtering module 62 performs the following operations when performing matching processing on the event key information and a pre-created article database, and filtering target historical sentence data matched with the event key information from the article database:

calculating a first similarity between each key information element and each history statement data;

screening sub-target historical sentence data corresponding to each key information element from the article database according to the first similarity between each key information element and each historical sentence data; the first similarity between the sub-target historical sentence data corresponding to any one key information element and the any one key information element is larger than or equal to a first preset threshold value.

the filtering module 62 performs the following steps when performing matching processing on the event key information and a pre-created article database, and filtering target historical sentence data matched with the event key information from the article database:

determining a plurality of first history articles from the article database;

calculating a second similarity between the event key information and the plurality of first history articles;

screening second historical articles from the plurality of first historical articles according to the event key information and the second similarity between the plurality of first historical articles; the second similarity between the second historical article and the event key information is greater than or equal to a second preset threshold;

calculating a third similarity between each key information element and each history statement data in the second history article;

screening sub-target historical sentence data corresponding to each key information element from the second historical article according to a third similarity between each key information element and each historical sentence data in the second historical article; and the third similarity between the sub-target historical sentence data corresponding to any one key information element and any one key element information is larger than or equal to a third preset threshold value.

In one embodiment, each history article corresponds to a respective article category; the filtering module 62, when determining a plurality of first history articles from the articles database, performs the steps of:

determining event categories corresponding to the target events;

matching the event category with the article category of each historical article in the article database, and determining a target article category matched with the event category;

and determining the historical articles corresponding to the target article category in the article database as a plurality of first historical articles.

In one embodiment, the apparatus further comprises:

the collection module is used for collecting the related content of the article to be updated; the article related content comprises published articles and/or characters of a specified type;

the updating module is used for updating the article database based on the article related content when the article database meets the preset updating condition; the preset updating condition comprises at least one of the following: the number of the released articles reaches a preset threshold value, the last update time of the article database reaches a preset duration from the current time, and at least one character of the appointed type exists in the historical articles to be updated.

In one embodiment, each key information element includes a key information field and key information content conforming to a preset information structure;

the acquiring module 61 performs the following steps when acquiring event key information of a target event:

determining a key information field matched with the event category of the target event;

acquiring key information content which corresponds to the target event and is matched with the key information field;

and carrying out structuring processing on each key information field and the corresponding key information content according to the preset information structure to obtain the event key information.

By adopting the device of the embodiment of the application, the event key information of the target event is obtained, the event key information and the article database which is created in advance are subjected to matching processing, and the target historical sentence data matched with the event key information are screened out, wherein the article database comprises a plurality of historical articles, and each historical article comprises at least one historical sentence data. Because the sentence data in the article not only comprises text contents but also semantic information of the text, the event key information and the historical sentence data are matched, so that the screened target historical sentence data and the event key information not only have the text contents with high matching but also have the semantic information with high matching. Furthermore, the event key information and the screened target historical sentence data are input into a pre-trained article generation model, so that the article generation model can generate a corresponding target article based on the target historical sentence data with the text content and the semantic information which are highly matched with the event key information, the target article is matched with the historical sentence data in the historical article (such as sentence structure matched with the historical sentence data), and the article generation efficiency and intelligence are greatly improved. In addition, the article generating process depends on the highly matched historical sentence data and the article generating model, and semantic consistency and diversity of the target articles are ensured based on diversity of the historical sentence data and intellectualization of the article generating model.

It should be understood by those skilled in the art that the article generating apparatus in fig. 6 can be used to implement the article generating method described above, and the detailed description thereof should be similar to the description of the method section above, so as to avoid complexity and avoid redundancy.

FIG. 7 is a schematic block diagram of an article-generating model training apparatus, as shown in FIG. 7, according to an embodiment of the present application, the apparatus comprising:

the acquiring module 71 is configured to acquire a sample article, perform semantic analysis on the sample article, and determine sample key information corresponding to the sample article;

a screening module 72, configured to perform matching processing on the sample key information and a pre-created article database, and screen sample historical sentence data matched with the sample key information from the article database; the article database comprises a plurality of history articles, and each history article comprises at least one history statement data;

a prediction module 73, configured to take the sample key information and the sample history statement data as input data of an article generating model to be trained, to obtain predicted article data corresponding to the sample key information;

the training module 74 is configured to train the article generating model to be trained based on the predicted article data and the sample article, where the trained article generating model is configured to generate a target article corresponding to the event key information according to the event key information and the target historical sentence information of the target event; and the target historical sentence data is the historical sentence data which is screened from the article database and is matched with the event key information by carrying out matching processing on the event key information and the article database.

In one embodiment, the sample key information comprises at least one sample key information element; each sample key information element comprises a sample key information field and sample key information content which accord with a preset information structure;

the obtaining module 71 performs the following operations when performing semantic analysis on the sample article and determining sample key information corresponding to the sample article:

carrying out semantic analysis on the sample article, and extracting sample key information fields and corresponding sample key information contents from the sample article;

and carrying out structuring treatment on each sample key information field and the corresponding sample key information content according to the preset information structure to obtain the sample key information.

In one embodiment, the article generation model to be trained includes: a sentence structure analyzing layer, a sentence generating layer, an article generating layer and a full connection layer;

the prediction module 73 is configured to:

determining statement structure information corresponding to each sample key information element through the statement structure analysis layer;

generating sample sentence data corresponding to each sample key information element according to sentence structure information corresponding to each sample key element through the sentence generation layer;

Combining sample sentence data corresponding to each sample key information element according to the element position of each sample key information element in the sample key information by the article generation layer to obtain predicted article data corresponding to the sample key information;

the training module 74 performs the following operations when the sample key information and the sample history statement data are used as input data of an article generating model to be trained to obtain predicted article data corresponding to the sample key information:

and carrying out iterative training on the article generation model through the full connection layer according to the predicted article data, the sample articles and preset convergence conditions.

In one embodiment, the sample article includes a plurality of sample characters; the predicted article data includes: a probability value for each sample character at a respective character position in the sample article; the preset convergence condition includes: the loss value of the article generation model is smaller than or equal to a preset loss threshold value;

the training module 74 performs the following operations when performing iterative training on the article generation model according to the predicted article data, the sample article, and a preset convergence condition:

Calculating a loss value of the article generating model according to the probability value of each sample character on each character position in the sample article;

if the loss value is smaller than or equal to the preset loss threshold value, stopping iteration; and if the loss value is larger than the preset loss threshold, adjusting model parameters of the article generation model, and performing iterative training based on the adjusted model parameters until the loss value is smaller than or equal to the preset loss threshold.

By adopting the device of the embodiment of the application, the sample key information corresponding to the sample article is determined by carrying out semantic analysis on the sample article, the sample key information is matched with the article database which is created in advance, and the sample historical sentence data matched with the sample key information is screened out from the article database, wherein the article database comprises a plurality of historical articles, and each historical article comprises at least one historical sentence data. Because the sentence data in the article not only comprises text content but also semantic information of the text, the sample historical sentence data and the sample historical sentence data are subjected to matching processing, so that the screened sample historical sentence data and the sample key information not only have the text content with high matching, but also have the semantic information with high matching. Furthermore, the sample key information and the screened sample historical sentence data are used as input data of an article generating model to be trained, predicted article data corresponding to the sample key information is obtained, the article generating model to be trained is trained based on the predicted article data and the sample article, the article generating model to be trained can be trained based on the sample historical sentence data with text content and semantic information which are highly matched with the sample key information, and the trained article generating model at least has the following capabilities based on the diversity of the sample article: the method comprises the steps of accurately analyzing semantic information and text content of sample key information, accurately analyzing matching between sample historical sentence data and the sample key information, and accurately generating articles based on the sample key information and the sample historical sentence data. Further, when an article is generated using the trained article generation model, semantic consistency and diversity of the generated article can be ensured.

It should be understood by those skilled in the art that the article generating model training apparatus in fig. 7 can be used to implement the foregoing article generating model training method, and the detailed description thereof should be similar to the foregoing method part description, so as to avoid complexity and avoid redundancy.

Based on the same thought, the embodiment of the application also provides electronic equipment, as shown in fig. 8. The electronic device may vary considerably in configuration or performance and may include one or more processors 801 and memory 802, where the memory 802 may store one or more stored applications or data. Wherein the memory 802 may be transient storage or persistent storage. The application programs stored in the memory 802 may include one or more modules (not shown), each of which may include a series of computer-executable instructions for use in an electronic device. Still further, the processor 801 may be configured to communicate with a memory 802 and execute a series of computer executable instructions in the memory 802 on an electronic device. The electronic device may also include one or more power supplies 803, one or more wired or wireless network interfaces 804, one or more input/output interfaces 805, one or more keyboards 806.

In particular, in this embodiment, an electronic device includes a memory, and one or more programs, where the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for the electronic device, and the one or more programs configured to be executed by one or more processors include instructions for:

acquiring event key information of a target event;

and inputting the event key information and the target historical sentence data into a pre-trained article generation model to obtain a target article corresponding to the event key information.

In particular, in another embodiment, an electronic device includes a memory, and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for the electronic device, and the one or more programs configured to be executed by one or more processors comprise instructions for:

The embodiments of the present application also provide a storage medium storing one or more computer programs, where the one or more computer programs include instructions, which when executed by an electronic device including a plurality of application programs, enable the electronic device to perform the respective processes of the above-described article generating method embodiments, and specifically are configured to perform:

acquiring event key information of a target event;

The embodiments of the present application also provide a storage medium storing one or more computer programs, where the one or more computer programs include instructions, which when executed by an electronic device including a plurality of application programs, enable the electronic device to perform the respective processes of the above-described article generation model training method embodiments, and specifically are configured to perform:

The system, apparatus, module or unit set forth in the above embodiments may be implemented in particular by a computer chip or entity, or by a product having a certain function. One typical implementation is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

For convenience of description, the above devices are described as being functionally divided into various units, respectively. Of course, the functions of each element may be implemented in one or more software and/or hardware elements when implemented in the present application.

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 present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In one typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include volatile memory in a computer-readable medium, random Access Memory (RAM) and/or nonvolatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of computer-readable media.

Computer readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of storage media for a computer include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium, which can be used to store information that can be accessed by a computing device. Computer-readable media, as defined herein, does not include transitory computer-readable media (transmission media), such as modulated data signals and carrier waves.

It should also 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, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article or apparatus that comprises the element.

The application may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The application may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

In this specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, as relevant to see a section of the description of method embodiments.

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

1. An article generation method, comprising:

acquiring event key information of a target event;

2. The method of claim 1, wherein the event critical information comprises at least one critical information element; the target historical sentence data matched with the event key information comprises sub-target historical sentence data corresponding to each key information element in the at least one key information element;

the matching processing is performed on the event key information and a pre-created article database, and target historical sentence data matched with the event key information is screened from the article database, including:

3. The method of claim 1, wherein the event critical information comprises at least one critical information element; the target historical sentence data matched with the event key information comprises sub-target historical sentence data corresponding to each key information element in the at least one key information element;

determining a plurality of first history articles from the article database;

4. The method of claim 3, wherein each history article corresponds to a respective article category;

the determining a plurality of first history articles from the article database includes:

determining event categories corresponding to the target events;

5. The method according to claim 1, wherein the method further comprises:

collecting the related content of the article to be updated; the article related content comprises published articles and/or characters of a specified type;

when the article database meets preset updating conditions, updating the article database based on the article related content; the preset updating condition comprises at least one of the following: the number of the released articles reaches a preset threshold value, the last update time of the article database reaches a preset duration from the current time, and at least one character of the appointed type exists in the historical articles to be updated.

6. A method according to claim 2 or 3, wherein each key information element comprises a key information field and a key information content conforming to a preset information structure;

the acquiring the event key information of the target event comprises the following steps:

7. An article generation model training method, comprising:

8. The method of claim 7, wherein the sample key information comprises at least one sample key information element; each sample key information element comprises a sample key information field and sample key information content which accord with a preset information structure;

the semantic analysis is performed on the sample article, and determining sample key information corresponding to the sample article includes:

9. The method of claim 8, wherein the article generation model to be trained comprises: a sentence structure analyzing layer, a sentence generating layer, an article generating layer and a full connection layer;

the step of obtaining predicted article data corresponding to the sample key information by using the sample key information and the sample history statement data as input data of an article generation model to be trained, includes:

The training the article generating model to be trained based on the predicted article data and the sample article comprises the following steps:

10. The method of claim 9, wherein the sample article comprises a plurality of sample characters; the predicted article data includes: a probability value for each sample character at a respective character position in the sample article; the preset convergence condition includes: the loss value of the article generation model is smaller than or equal to a preset loss threshold value;

the performing iterative training on the article generating model according to the predicted article data, the sample article and the preset convergence condition includes:

if the loss value is smaller than or equal to the preset loss threshold value, stopping iteration;

and if the loss value is larger than the preset loss threshold, adjusting model parameters of the article generation model, and performing iterative training based on the adjusted model parameters until the loss value is smaller than or equal to the preset loss threshold.

11. An article generating apparatus, comprising:

12. An article generation model training device, comprising:

the training module is used for training the article generating model to be trained based on the predicted article data and the sample articles, and the article generating model after training is used for generating target articles corresponding to the event key information according to the event key information of the target event and the target historical sentence information; and the target historical sentence data is the historical sentence data which is screened from the article database and is matched with the event key information by carrying out matching processing on the event key information and the article database.

13. An electronic device comprising a processor and a memory electrically connected to the processor, the memory storing a computer program, the processor being configured to invoke and execute the computer program from the memory to implement the article generation method of any of claims 1-6, or the processor being configured to invoke and execute the computer program from the memory to implement the article generation model training method of any of claims 7-10.

14. A computer readable storage medium storing a computer program executable by a processor to implement the article generation method of any one of claims 1-6 or executable by a processor to implement the article generation model training method of any one of claims 7-10.