A method, device, system and server for generating original articles

Abstract

The present invention relates to the field of Internet technology, and discloses a method, device, system and server for generating original articles. The method for generating original articles includes: obtaining hot events; screening the obtained hot events through a pre-trained topic analysis model to determine hot events related to a preset topic; obtaining a news collection related to the hot events according to the hot events related to the preset topic; generating an event summary of the hot events related to the preset topic according to the news collection; obtaining a topic short article based on a pre-generated topic knowledge base, combining the event summary to generate the original article. By generating an event summary based on a news collection, the present invention can accurately summarize the event summary of the hot event, and, based on a pre-generated topic knowledge base, obtaining a topic short article based on the event summary to generate an original article, which can improve the contextual coherence of the article.

Description

Technical Field

[0001] The present invention relates to the field of Internet technology, and in particular to a method, device, system and server for generating original articles. Background Art

[0002] With the development of Internet technology, text generation technology has also emerged. Compared with the breakthroughs made in image generation technology, text generation technology still faces many challenges. In recent years, many article generators have emerged on the market. According to the technology they use, they are mainly divided into two categories: template-based and deep learning model-based.

[0003] However, the articles generated by template-based text generation technology have a single structure, while the text generation technology based on deep learning models finds it difficult to ensure the logical coherence of the context.

[0004] Based on this, the existing technology needs to be improved urgently. Summary of the invention

[0005] One purpose of the embodiments of the present invention is to provide a method, device, system and server for generating an original article, which can improve the contextual coherence of the article.

[0006] In a first aspect, an embodiment of the present invention provides a method for generating an original article, comprising:

[0007] Get hot events;

[0008] The acquired hot events are screened by a pre-trained topic analysis model to determine hot events related to a preset topic;

[0009] According to the hot events related to the preset topic, a news collection related to the hot events related to the preset topic is obtained;

[0010] Generating an event summary of the hot event according to the news collection;

[0011] Based on a pre-generated subject knowledge base, a subject short article is obtained, and combined with the event summary to generate the original article.

[0012] In some embodiments, determining a hot event related to a preset topic includes:

[0013] A topic category label is generated, and the hot event is identified by the topic category label.

[0014] In some embodiments, generating an event summary of the hot events related to the preset topic according to the news collection includes:

[0015] Based on a pre-trained multi-document summary extraction model, extracting original summaries related to the news event from the news collection to generate a candidate summary set;

[0016] An event summary of the hot event is generated according to the candidate summary set.

[0017] In some embodiments, generating an event summary of the hot event related to the preset topic according to the candidate summary set includes:

[0018] Determine a summary to be rewritten from the candidate summary set;

[0019] Based on a pre-trained synonymous sentence rewriting model, the summary to be rewritten is rewritten to generate an event summary of the hot event related to the preset topic.

[0020] In some embodiments, the acquiring of a subject short article based on a pre-generated subject knowledge base and combining the event summary to generate the original article includes:

[0021] According to the subject category label, obtaining subject short articles related to the subject category from the subject knowledge base;

[0022] The event summary and the subject essay are automatically spliced ​​together to generate the original article.

[0023] In some embodiments, after obtaining the subject short articles related to the subject category from the subject knowledge base according to the subject category label, the method further includes:

[0024] Select the best short article from at least two short articles on the subject;

[0025] Based on a pre-trained synonymous sentence rewriting model, the optimal short article is rewritten to generate an event short article of the hot event.

[0026] In some embodiments, the automatically splicing the event summary and the subject short article to generate the original article includes:

[0027] The event summary and the event essay are automatically spliced ​​together to generate the original article.

[0028] In some embodiments, the preset topics include insurance topics, and the topic knowledge base includes an insurance knowledge base.

[0029] In a second aspect, an embodiment of the present invention provides a device for generating an original article, including:

[0030] A hot event acquisition unit, used to acquire hot events;

[0031] A hot event determination unit, used to screen the acquired hot events through a pre-trained topic analysis model to determine hot events related to a preset topic;

[0032] A news collection acquisition unit, used to acquire a news collection related to the hot event according to the hot event;

[0033] An event summary generating unit, used for generating an event summary of the hot events related to the preset topic according to the news collection;

[0034] The original article generation unit obtains the subject short article based on the pre-generated subject knowledge base, combines the event summary, and generates the original article.

[0035] In some embodiments, the hot event determination unit is specifically used to:

[0036] A topic category label is generated, and the hot event is identified by the topic category label.

[0037] In some embodiments, the event summary generating unit includes:

[0038] A candidate summary set module is used to extract original summaries related to the news event from the news set based on a pre-trained multi-document summary extraction model to generate a candidate summary set;

[0039] The event summary generation module is used to generate an event summary of the hot event according to the candidate summary set.

[0040] In some embodiments, the event summary generation module is specifically used to:

[0041] Determine a summary to be rewritten from the candidate summary set;

[0042] Based on a pre-trained synonymous sentence rewriting model, the summary to be rewritten is rewritten to generate an event summary of the hot event.

[0043] In some embodiments, the original article generating unit includes:

[0044] A subject short article acquisition module, used for acquiring subject short articles related to the subject category from the subject knowledge base according to the subject category label;

[0045] The original article generation module is used to automatically splice the event summary and the subject short article to generate the original article.

[0046] In some embodiments, the original article generating unit further includes:

[0047] An optimal short article generation module is used to select a unique optimal short article from at least two acquired topic short articles;

[0048] The event short article generation module is used to rewrite the optimal short article based on a pre-trained synonymous sentence rewriting model to generate an event short article of the hot event.

[0049] In some embodiments, the original article generation module is specifically used to:

[0050] The event summary and the event essay are automatically spliced ​​together to generate the original article.

[0051] In some embodiments, the preset topics include insurance topics, and the topic knowledge base includes an insurance knowledge base.

[0052] In a third aspect, an embodiment of the present invention provides a server, including:

[0053] at least one processor; and

[0054] a memory communicatively connected to the at least one processor; wherein,

[0055] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned method for generating original articles.

[0056] In a fourth aspect, an embodiment of the present invention provides a system for generating original articles, including:

[0057] the above mentioned servers;

[0058] A third-party news platform, which is communicatively connected to the server, and includes a third-party news library for storing hot news information so that the server can obtain the hot news information;

[0059] An article publishing platform is communicatively connected to the server and is used to publish original articles generated by the server.

[0060] In a fourth aspect, an embodiment of the present invention provides a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a server to execute the above-mentioned method for generating original articles.

[0061] In the method for generating original articles provided in various embodiments of the present invention, first, hot events are obtained; the hot events obtained are screened through a pre-trained topic analysis model to determine hot events related to a preset topic; based on the hot events related to the preset topic, a news collection related to the hot events related to the preset topic is obtained; based on the news collection, an event summary of the hot events related to the preset topic is generated; based on a pre-generated topic knowledge base, a topic short article is obtained, and the original article is generated in combination with the event summary. On the one hand, since hot news information is screened through a pre-trained topic analysis model to determine hot events related to a preset topic, and on the other hand, by obtaining a news collection related to the hot events and generating an event summary based on the news collection, it can accurately summarize the event summary of the hot event, and also generate an original article based on a pre-generated topic knowledge base in combination with the event summary, which can improve the contextual coherence of the article. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] One or more embodiments are exemplarily described by pictures in the corresponding drawings, and these exemplified descriptions do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, and unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0063] Figure 1 It is a structural schematic diagram of a system for generating original articles provided by an embodiment of the present invention;

[0064] Figure 2 is a flow chart of a method for generating an original article provided by an embodiment of the present invention;

[0065] Figure 3 is a schematic diagram of a topic analysis model based on a BERT pre-training model provided in an embodiment of the present invention;

[0066] Figure 4 yes Figure 2 A detailed flowchart of step S40 in FIG.

[0067] Figure 5 yes Figure 4 A detailed flowchart of step S42 in FIG.

[0068] Figure 6 yes Figure 2 A detailed flowchart of step S50 in FIG.

[0069] Figure 7 is a schematic diagram of a Pointer-Generator Networks network structure provided by an embodiment of the present invention;

[0070] Figure 8 is another flow chart of a method for generating an original article provided by an embodiment of the present invention;

[0071] Fig. 9 It is an interactive schematic diagram of a system for generating original articles provided by an embodiment of the present invention;

[0072] Fig.10 It is a structural schematic diagram of a device for generating original articles provided by an embodiment of the present invention;

[0073] Fig.11 It is a structural diagram of a server provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0075] It should be noted that, if there is no conflict, the various features in the embodiments of the present invention can be combined with each other, all within the scope of protection of the present invention. In addition, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flow chart, in some cases, the steps shown or described can be performed in a different order from the module division in the device or the flow chart. Furthermore, the words "first", "second", "third", etc. used in the present invention do not limit the data and execution order, but only distinguish the same items or similar items with basically the same functions and effects.

[0076] Compared with the breakthroughs made in image generation technology, text generation technology still faces many challenges. In recent years, many article generators have emerged on the market, which can be divided into two categories according to the technology they use: template-based and deep learning model-based.

[0077] The template-based text generation technology solution is mainly suitable for fields with simple article structures and rich structured data, such as weather forecasts, financial news, sports news, etc. The generated articles are highly readable. The specific principle is that for articles in a certain field, a series of article templates are sorted out in advance through machine mining or manual methods, and the article generation process is to fill the structured data into the template.

[0078] Text generation technology based on deep learning models uses a well-designed neural network structure. The network parameters are usually as high as millions or even tens of billions. Large-scale corpus is required to train the model. It is widely used in scenarios such as picture title generation, speech-to-text, chatbots, synonymous sentence rewriting, article summaries, and translation. In addition to requiring large-scale training samples and powerful computing resources, deep learning-based methods have uncontrollable model effects, especially in long text scenarios such as paragraphs and chapters, where it is difficult to ensure the logical coherence of the context.

[0079] The template-based method generates articles with a single structure, which is not suitable for insurance popular science articles. The deep learning-based method adopts a completely end-to-end method, which makes it difficult to ensure the coherence between sentences.

[0080] In view of the above problems, the present invention proposes a new article generation solution.

[0081] Before describing the present invention in detail, the nouns and terms involved in the embodiments of the present invention are described. The nouns and terms involved in the embodiments of the present invention are subject to the following explanations.

[0082] 1) Abstract, which refers to a shorter description that compresses an article into a shorter description containing the main information of the original text.

[0083] 2) Out of Vocabulary (OOV) refers to words that have never appeared in the vocabulary.

[0084] 3) Attention mechanism, also known as Attention mechanism, refers to the addition of Attention model in the encoding segment to perform data weighted transformation on the source data sequence, or the introduction of Attention model at the decoding end to perform weighted changes on the target data within the encoder-decoder framework, which can effectively improve the system performance in a natural sequence-to-sequence manner, breaking the limitation that the traditional encoder-decoder structure relies on a fixed-length vector internally during encoding and decoding. By retaining the intermediate output results of the LSTM encoder for the input sequence, a model is trained to selectively learn these inputs and associate the output sequence with them when the model outputs.

[0085] See also Figure 1 , Figure 1 Schematic diagram of the structure of a system for generating original articles provided by an embodiment of the present invention. Figure 1 As shown, the original article generation system 100 includes: a server 11, a third-party news platform 12 and an article publishing platform 13.

[0086] The server 11 is communicated with the third-party news platform 12 and the article publishing platform 13 to complete the test of the auto insurance business. For example, the third-party news platform 12 is connected to the server 11, and the server 11 sends a news request to the third-party news platform 12 to obtain third-party news. The third-party news platform 12 processes the news request and sends the news corresponding to the news request to the server 11.

[0087] In some embodiments, the number of third-party news platforms 12 may be one or more, and the third-party news platform 12 may be a computer terminal, a server, a mobile terminal, and other devices. The third-party news platform includes a third-party news library for storing hot news information so that the server can obtain the hot news information. Preferably, the third-party news platform is a news website, and the news website includes a news library for storing hot news information.

[0088] There are multiple servers 12, and multiple servers 12 can constitute a server cluster. For example, the server cluster includes a first server, a second server, ..., an Nth server. Alternatively, the server cluster can be a cloud computing service center, which includes several servers. The server serves as an insurance company server and is used to connect to the insurance company's business personnel or developers.

[0089] In some embodiments, the server 11 may be pre-configured with a multi-layer software architecture. For example, the software architecture of the server 11 includes: a business layer, a test layer, and an interface layer, wherein the business layer is used to determine the business cooperation model, the test layer is used to confirm the test points and generate corresponding data, and the interface layer is used to connect to the server of a third-party news platform or article publishing platform.

[0090] See also Figure 2 , Figure 2 is a flow chart of a method for generating an original article provided by an embodiment of the present invention;

[0091] like Figure 2 As shown, the method for generating the original article includes:

[0092] Step S10: Obtain hot events;

[0093] Specifically, the search engine uses a web crawler to periodically capture popular search lists such as Baidu, Weibo, and Toutiao, and automatically obtains popular events, such as popular news information, from web pages of news websites, entertainment websites, and the like, and generates a list of popular events based on the popular news information, or obtains the list of popular events based on the list of popular events included in the web pages of news websites, entertainment websites, and the like, and based on the list of popular events, obtains the popular news information in the list of popular events, such as headlines, entertainment events, sports news, national events, international news, and other popular news information.

[0094] In an embodiment of the present invention, the step of obtaining hot news information includes the following steps:

[0095] (1) Crawl web pages from news websites, entertainment websites, etc.

[0096] Specifically, the search engine uses a web crawler to regularly crawl popular search lists such as Baidu, Weibo, and Toutiao, automatically obtains popular events from the web pages of news websites or entertainment websites such as Baidu, Weibo, and Toutiao, and collects the popular events to generate a popular event list. First, a part of the seed URLs are selected and these URLs are put into the queue of URLs to be crawled; the URLs to be crawled are taken out, the DNS is parsed to obtain the IP of the host, and the web pages corresponding to the URLs are downloaded and stored in the downloaded web page library, and these URLs are put into the queue of crawled URLs. The URLs in the queue of crawled URLs are analyzed, and other URLs therein are analyzed, and the URLs are put into the queue of URLs to be crawled, thereby entering the next cycle.

[0097] (2) Data storage;

[0098] Specifically, the search engine crawls the web pages and stores the data in the original page database. The page data is exactly the same as the HTML obtained by the user's browser. When crawling pages, the search engine spider also performs certain duplicate content detection. Once it encounters a website with a low access weight that contains a large amount of plagiarized, collected or copied content, it is likely to stop crawling.

[0099] (3) Pretreatment;

[0100] Specifically, the search engine pre-processes the pages captured by the crawler in various steps, such as: extracting text, Chinese word segmentation, eliminating noise (such as copyright statement text, navigation bars, advertisements, etc.), indexing, link relationship calculation, special file processing, etc.

[0101] (4) Provide search services and website rankings;

[0102] Specifically, after organizing and processing the information, the search engine provides keyword search services for users and displays the relevant information to users. At the same time, it ranks websites according to the PageRank value of the page (the ranking of the number of visits to the link).

[0103] In the embodiment of the present invention, hot news information is automatically obtained by means of a web crawler, which can improve the speed of obtaining hot news information and can obtain a large amount of hot news information, thereby ensuring that data support can be provided for the subsequent acquisition of news collections related to the hot events.

[0104] Step S20: Screening the acquired hot events through a pre-trained topic analysis model to determine hot events related to a preset topic;

[0105] Specifically, the preset topic includes an insurance topic, and the topic analysis model includes an insurance topic analysis model. The insurance topic analysis model generated by offline pre-training filters the obtained hot news information to determine hot events related to the preset topic, for example, the obtained hot news information is filtered to determine hot events related to the insurance topic, for example, the obtained hot events are: the death of a famous poet, a worker who does not enjoy hospitalization subsidies because he did not buy social insurance, etc. Among them, the pre-trained topic analysis model filters the obtained hot news information by judging the relevance of the hot events with insurance, thereby determining hot events related to insurance, for example, the death of a famous poet, a worker who does not enjoy hospitalization subsidies because he did not buy social insurance, etc.

[0106] In the embodiment of the present invention, determining the hot events related to the preset topic includes:

[0107] A topic category label is generated, and the hot event is identified by the topic category label.

[0108] Specifically, after screening the acquired popular news information and determining the popular events related to the preset theme, for the popular events, determine the theme type corresponding to the popular events, and generate a theme type label, and identify the popular events by the theme type label. For example, if the popular event is related to insurance, then the popular event is determined to be an insurance-related event. At this time, according to the theme type corresponding to the insurance-related event, the insurance-related event is identified with the insurance type label corresponding to the theme type. For example, the insurance-related event is marked with an insurance type label, for example, "the famous poet XX is labeled with a life insurance label."

[0109] The following takes the preset topic as the insurance topic and the topic analysis model as the insurance topic analysis model as an example to illustrate the pre-trained topic analysis model:

[0110] Among them, this topic analysis model is a topic analysis model based on the BERT pre-trained model. The BERT model is a pre-trained model of word vectors, which is a speech model built based on the bidirectional Transformer. The BERT model provides two sizes of models on the English dataset, Base and Large. Uncased means that the input words will be converted to lowercase, and cased means that the input words will be preserved in their uppercase (required in projects such as named entity recognition).

[0111] See also Figure 3 , Figure 3 is a schematic diagram of a topic analysis model based on a BERT pre-training model provided in an embodiment of the present invention;

[0112] like Figure 3 As shown in the figure, by inputting a sentence (Single Sentence), the short text classifier (Classifier) ​​is trained, and the BERT model is allowed to change as little as possible during the training process. This training stage is called fine-tuning. The first input token is filled with a special [CLS] token. CLS stands for Classification. Like the simple and pure encoder of Transformer, BERT takes a sequence of words as input. These words will continue to flow upward in its stack. Each layer will apply a self-attention mechanism and pass its results through a feedforward network, which will then be handed over to the next encoder for training to generate a topic (Class Label). Among them, BERT is a self-supervised method for pre-training deep transformer encoders. After pre-training, it can be fine-tuned for different downstream tasks. BERT optimizes two training objectives - masked language modeling (MLM) and next sentence prediction (NSP). Its training only requires the use of large datasets without labels.

[0113] For each word or subunit sequence X = (x1, ..., xn), BERT generates its context-based vector representation through the encoder: x1, ..., xn = enc(x1, ..., xn). Since BERT uses this encoder through a deep transformer structure, the model uses its position embedding p1, ..., pn to identify the absolute position of each word in the sequence.

[0114] Among them, the masked language modeling (MLM), also known as the fill-in-the-blank test, is to predict the missing word at a certain position in a sequence. This step samples a subset Y from the word set X and replaces it with another word set. In BERT, Y accounts for 15% of X. In Y, 80% of the words are replaced with [MASK], 10% of the words are replaced with random words based on the unigram distribution, and 10% remain unchanged. The task is to use these replaced words to predict the original words in Y.

[0115] Specifically, the recognition of popular insurance events is modeled as a short text classification problem. The insurance industry defines 10 common insurance categories, including health insurance, accident insurance, critical illness insurance, life insurance, property insurance, unemployment insurance, pension insurance, auto insurance, children's insurance, and travel insurance. A large number of popular event libraries are annotated to obtain training samples. Considering the small number of training samples, in order to improve the generalization ability of the model, the short text classifier is fine-tuned based on Google's BERT Chinese pre-training model. Finally, the short text classifier has an accuracy rate of 95% in the verification set and an insurance event recall rate of 93%.

[0116] In an embodiment of the present invention, by pre-training a topic analysis model and screening popular news information through the topic analysis model, it is possible to filter out irrelevant popular events and only retain popular events related to a preset topic, which is beneficial to reducing popular events irrelevant to the preset topic and improving the processing speed of the system.

[0117] Step S30: according to the hot events related to the preset topic, obtaining a news collection related to the hot events related to the preset topic;

[0118] Specifically, according to the hot event, a news collection related to the hot event is obtained from the original page database stored in the hot news information obtained by the web crawler, or a collection of news articles related to the hot event is captured from a third-party news library using a web crawler. For example, using query = "famous poet XXX passed away", multiple related news collections are obtained from a news website.

[0119] In the embodiment of the present invention, the obtaining of the news set related to the hot events related to the preset topic includes:

[0120] A news collection that is strongly related to the hot event is obtained. Specifically, according to the title of the hot event, a web crawler is used to obtain news, and news information that is strongly related to the hot event is obtained, and then multiple news information is collected to generate a news collection that is strongly related to the hot event. The obtaining of news information that is strongly related to the hot event includes: calculating a correlation coefficient between the hot event and the news information, and the correlation coefficient is used to reflect the degree of linear correlation between the hot event and the news information. The strong correlation means that the correlation coefficient between the hot event and the news information is greater than a preset threshold. If the correlation coefficient between the hot event and the news information is greater than the preset threshold, it is determined that the news information is strongly related to the hot event. For example, if the preset threshold is 0.9, if the correlation coefficient between the hot event and the news information is greater than 0.9, it is determined that the news information is strongly related to the hot event. At this time, the news information is obtained, and multiple news information are collected to generate a news collection.

[0121] In the embodiment of the present invention, by acquiring a news collection related to the hot event, it is convenient to filter the news information in the news collection, which is conducive to better generation of articles.

[0122] Step S40: generating an event summary of the hot events related to the preset topic according to the news collection;

[0123] Specifically, the news collection includes multiple news articles and / or news information. Through the news articles and / or news information, summary information in the news articles and / or news information is obtained to further generate an event summary of the hot event.

[0124] For details, please refer to Figure 4 , Figure 4 yes Figure 2 A detailed flowchart of step S40 in FIG.

[0125] Since the generative summaries generated by the summary generation technology based on the deep model generally have the problem of context incoherence, in order to improve the coherence of the context, the present invention obtains the article summary by adopting an extractive method.

[0126] like Figure 4 As shown, step S40: generating an event summary of the hot event according to the news collection, including:

[0127] Step S41: extracting original summaries related to the news event from the news collection based on a pre-trained multi-document summary extraction model to generate a candidate summary set;

[0128] The news collection includes multiple news articles and / or news information, and article summaries are extracted through a multi-document summary extraction model. Specifically, the pre-trained multi-document summary extraction model extracts original summaries related to the news event from the news collection to generate a candidate summary set, including the following steps:

[0129] (1) obtaining keywords of the news event from the news collection based on a TF-IDF keyword extraction algorithm;

[0130] (2) adding weights to the keywords of the news event to generate a weighted keyword set;

[0131] (3) For each news article or news information, extract the key sentences of the article based on the TextRank algorithm;

[0132] (4) taking the continuous text segments containing all the key sentences as the article summary of the news article or news information, and generating a candidate summary set of the news event;

[0133] Step S42: Generate an event summary of the hot event according to the candidate summary set.

[0134] Specifically, the candidate summary set includes multiple candidate summaries, so it is necessary to determine a unique event summary based on the multiple candidate summaries. For details, please refer to Figure 5 , Figure 5 yes Figure 4 A detailed flowchart of step S42 in FIG.

[0135] like Figure 5 As shown, step S42: generating an event summary of the hot event according to the candidate summary set includes:

[0136] Step S421: determining a summary to be rewritten from the candidate summary set;

[0137] Specifically, determining the summary to be rewritten from the candidate summary set includes: counting the weighted cumulative sum of keywords contained in the article summary of each news article or news information, sorting the weighted cumulative sums in descending order, and using the article summary with the highest score as the summary to be rewritten for the hot event.

[0138] Step S422: rewriting the summary to be rewritten based on the pre-trained synonymous sentence rewriting model to generate an event summary of the hot event.

[0139] Specifically, by pre-training a synonymous sentence rewriting model based on deep learning, the summary to be rewritten is rewritten to generate an event summary of the hot event. For example, pre-training a synonymous sentence rewriting model based on a pointer generation network, rewriting the summary to be rewritten, and generating an event summary of the hot event.

[0140] In an embodiment of the present invention, the original summaries in the news collection are extracted based on a pre-trained multi-document summary extraction model to generate a candidate summary set, and the summaries to be rewritten are determined from the candidate summary set, and the summaries to be rewritten are rewritten based on a pre-trained synonym rewriting model, thereby generating an event summary of the hot event, which can ensure the coherence of the context.

[0141] Step S50: Based on the pre-generated subject knowledge base, the subject short article is obtained and combined with the event summary to generate the original article.

[0142] Specifically, the subject knowledge base includes an insurance knowledge base, which includes multiple subject short articles, such as popular science short articles, which are popular science short articles on insurance, such as popular science short articles on insurance payment procedures, deadlines, precautions, etc., wherein the popular science short articles in the subject knowledge base carry subject category tags, such as insurance type tags. For details, please refer to Figure 6 , Figure 6 yes Figure 2 The detailed flow chart of step S20 in FIG. Figure 6 As shown, step S50: generating the original article based on the pre-generated subject knowledge base and in combination with the event summary, includes:

[0143] Step S51: acquiring topic essays related to the topic category from the topic knowledge base according to the topic category label;

[0144] Specifically, a subject knowledge base is pre-generated, and the subject knowledge base includes multiple subject essays, and the subject essays carry subject category tags. According to the subject category tags corresponding to the popular events, subject essays related to the subject categories corresponding to the subject category tags are obtained from the subject knowledge base, that is, subject essays carrying the subject category tags in the subject knowledge base are obtained. For example, if the subject knowledge base is an insurance knowledge base, and the subject category tags corresponding to the popular events are insurance and auto insurance tags, then according to the insurance and auto insurance tags, subject essays carrying the insurance and auto insurance tags in the insurance knowledge base are obtained.

[0145] Step S52: automatically splicing the event summary and the subject short article to generate the original article;

[0146] Specifically, the subject short articles carrying the subject category label are screened out from the subject knowledge base, and the event summary is spliced ​​with the subject short article to generate the original article.

[0147] It is understandable that the original article also includes a title, and the title of the original article is the title of the hot event, for example: the title of hot news information obtained from a third-party news platform.

[0148] In the embodiment of the present invention, after acquiring the subject short articles related to the subject category from the subject knowledge base according to the subject category label, the method further includes:

[0149] Select the best short article from at least two short articles on the subject;

[0150] It is understandable that the subject knowledge base may include multiple subject essays related to the subject category, that is, the subject knowledge base stores at least two subject essays with subject category tags corresponding to the subject category. At this time, it is necessary to determine a unique one for splicing with the event summary to generate the original article. Therefore, it is necessary to filter out a unique optimal essay from the at least two acquired subject essays. Specifically, the method of filtering out a unique optimal essay from the at least two acquired subject essays includes: obtaining all subject essays related to the subject category, determining keywords, performing keyword search on the contents of all subject essays, and sorting the contents of the subject essays from high to low according to the frequency of occurrence of the keywords in the contents of the subject essays, and using the subject essay with the highest ranking as the unique optimal essay. In an embodiment of the present invention, the keywords may include one or more keywords, and the keywords may further include weights. For example, the keywords include a first keyword A and a second keyword B, wherein the weight of A is 60% and the weight of B is 40%. According to the number of occurrences of the first keyword A and the second keyword B in the subject essay, combined with the weights of the first keyword A and the second keyword B, a weighted sum is performed, and the subject essay with the largest sum obtained by the weighted sum is determined as the unique optimal essay.

[0151] Based on a pre-trained synonymous sentence rewriting model, the optimal short article is rewritten to generate an event short article of the hot event.

[0152] Specifically, the synonymous sentence rewriting model is a synonymous sentence rewriting model based on deep learning. The synonymous sentence rewriting model based on pre-training rewrites the optimal short article to generate the event short article of the hot event, including the following steps:

[0153] (1) Construct a parallel corpus of synonymous sentences;

[0154] Specifically, with the help of a translation application, a large number of synonymous sentence pairs are constructed for each sentence in the public news library through a secondary translation mode of converting Chinese to English and then from English to Chinese, wherein the synonymous sentence pairs are noisy synonymous sentence pairs.

[0155] (2) Generate a synonymous sentence rewriting model based on a pointer generation network;

[0156] See also Figure 7 , Figure 7 is a schematic diagram of a Pointer-Generator Networks network structure provided by an embodiment of the present invention;

[0157] like Figure 7 As shown, the left side is the encoder side, and the right side is the decoder side. The encoder side still does emdebing at each step, and then inputs it into the LSTM after embedding. The output of each step in the LSTM is used, and a weight is applied to the output of each step. Each encoder has a coefficient α (α can be a scalar or a vector. If it is a vector, it is dot-multiplied with the vector output by the LSTM. If it is a scalar, it is directly multiplied with the vector. α needs to be normalized). The output of each step of the LSTM on the encoder side is multiplied by α, and all the multiplications are finally added up. This is the process of weighted averaging. After the weighted averaging process, it will be input into the second unit of the decoder. By using the information of the first step, the first step output of the decoder participates in the calculation process of α, and will be operated with each LSTM output in the encoder to calculate the value of α. The value of α is weighted averaged with the vectors of the 5 LSTMs to obtain a vector, which is input into the second step of the decoder. Similarly, the second step output of the decoder participates in the α calculation to obtain the α value, which is dot-multiplied with the vectors in the 5 LSTMs on the encoder, and the weighted average is obtained to obtain a vector, which is input into the third step of the decoder.

[0158] The Pointer-Generator Networks is a hybrid of a baseline and a point network. Since it allows copying words by pointing, and can also generate words from a fixed vocabulary, the synonymous sentence rewriting model based on the Pointer-Generator Network has the following advantages over other synonymous sentence rewriting models:

[0159] First, the Attention mechanism is continued, which can encode long texts;

[0160] Second, in addition to the Attention mechanism, a Convergence mechanism is added. That is, Attention does not consider the situation of repeated sentences, while Convergence adds redundancy processing on the basis of Attention. For the existing words in the sentence, their attention in the future is reduced to avoid the word duplication problem that often occurs in other models.

[0161] Third, for the case of Out of Vocabulary (OOV), the pointer mechanism and the generator mechanism are combined. The pointer mechanism is to copy words from the original text, and the generator mechanism is to select the most suitable words from the vocabulary. The pointer mechanism effectively solves the problem of incorrect rewriting of proper nouns such as names of people, organizations, code numbers, time dates, and place names, while the generator mechanism ensures that high-frequency words are reasonably rewritten with a greater probability. The synonym rewriting model based on the pointer generation network has a good model effect. The sample evaluation of the rewritten sentences shows that the proportion of high-quality rewriting is as high as 95%, and the problem of poor rewriting effect is concentrated in the small rewriting granularity.

[0162] In an embodiment of the present invention, the optimal essay is rewritten by adopting a synonymous sentence rewriting model based on a pointer generation network to generate an event essay of the hot event, and the event summary and event essay of the hot event are spliced ​​to generate an original article, which can better rewrite and generate high-quality event essays.

[0163] Please refer to Figure 8 , Figure 8 is another flow chart of a method for generating an original article provided by an embodiment of the present invention;

[0164] like Figure 8 As shown, the method for generating the original article includes:

[0165] Step S801: monitoring hot events;

[0166] Specifically, with the help of web crawlers, popular search lists such as Baidu, Weibo, and Toutiao are captured regularly, and popular news information is automatically obtained from web pages of news websites, entertainment websites, and the like, and a hot event list is generated based on the hot news information, or, based on the hot event list included in the web pages of news websites, entertainment websites, and the like, the hot event list is obtained, and based on the hot event list, hot news information in the hot event list is obtained, for example: hot news information such as headlines, entertainment events, sports news, national events, international news, etc.

[0167] Step S802: identifying popular insurance events;

[0168] Specifically, the offline pre-trained insurance topic model is used to determine the relevance of popular events to insurance, and insurance-related events are labeled with insurance type labels.

[0169] Step S803: Capture popular insurance news;

[0170] Specifically, a web crawler is used to retrieve a collection of news articles related to hot insurance events from a third-party news library.

[0171] Step S804: extracting summary of popular insurance news;

[0172] Specifically, an offline pre-trained multi-document summary extraction model is used to obtain an event summary that can accurately summarize the news event from a news collection of hot events.

[0173] Step S805: rewriting the abstract for originality;

[0174] Specifically, a synonymous sentence rewriting model pre-trained offline is used to rewrite synonyms for each sentence in the event summary, and the fluency of each sentence and the semantics of each sentence are ensured to be the same as the original sentence, thereby obtaining an original summary of the insurance event.

[0175] Step S806: article splicing;

[0176] Specifically, the insurance type label of the event is used to randomly select a popular science essay that matches the label from the insurance knowledge base, and the original abstract is spliced ​​with the popular science essay to obtain an original insurance popular science article based on the hot news event.

[0177] In an embodiment of the present invention, a method for generating an original article is provided, including: obtaining hot news information; filtering the obtained hot news information through a pre-trained topic analysis model to determine hot events related to a preset topic; obtaining a news collection related to the hot event according to the hot event; generating an event summary of the hot event according to the news collection; generating the original article based on a pre-generated topic knowledge base and in combination with the event summary. On the one hand, by obtaining hot news information, filtering the obtained hot news information through a pre-trained topic analysis model to determine hot events related to the preset topic, the present invention can keep up with current hot topics and easily resonate with readers. After the article is published, compared with knowledge-based insurance popular science articles, the user reading volume and forwarding volume have obvious advantages. On the other hand, by generating an event summary of a hot event, the present invention can ensure that the generated article is original and can pass the originality detection of platforms such as Baidu and Toutiao. Finally, the present invention automatically generates original articles. Compared with manual editing, the tool is efficient and low-cost, and can seize news hotspots in the first place.

[0178] Please refer to Fig. 9 , Fig. 9 It is an interactive schematic diagram of a system for generating original articles provided by an embodiment of the present invention;

[0179] like Fig. 9 As shown, the original article generation system includes a server, a third-party news platform and an article publishing platform, wherein the interaction process includes:

[0180] Step S901: Send a news information request;

[0181] Specifically, the server sends a news information request to the third-party news platform, so that the third-party news platform returns corresponding news information to the server based on the news information request.

[0182] Step S902: Exporting popular news information;

[0183] Specifically, the third-party news platform exports hot news information according to the news information request, and sends the hot news information to the server.

[0184] Step S903: Generate original article;

[0185] Specifically, the server obtains hot news information;

[0186] The acquired popular news information is screened through a pre-trained topic analysis model to determine popular events related to a preset topic; based on the popular events, a news collection related to the popular events is acquired; based on the news collection, an event summary of the popular events is generated; based on a pre-generated topic knowledge base and in combination with the event summary, the original article is generated.

[0187] Step S904: Send the original article;

[0188] Specifically, the server sends the original article to the article publishing platform, so that the article publishing platform publishes the original article on its own platform based on the original article.

[0189] Step S905: Publish the original article;

[0190] Specifically, the article publishing platform publishes the original article according to the original article sent by the server. The article publishing platform includes: WeChat public account, Toutiao account, Baijia account, Penguin account and other self-media platforms or other insurance information platforms.

[0191] The original article is published to the article publishing platform through the server, and the generated insurance article can be directly published to self-media platforms such as WeChat public account, Toutiao account, Baijia account, Penguin account, or other insurance information platforms. The generated article focuses on the current news hotspots that are strongly related to insurance, which is real and tangible, and easy to resonate with readers. Compared with empty and boring preaching, it has important value in improving users' insurance awareness and spreading insurance knowledge.

[0192] In an embodiment of the present invention, a method for generating an original article is provided, including: obtaining hot news information; screening the obtained hot news information through a pre-trained topic analysis model to determine hot events related to a preset topic; obtaining a news collection related to the hot event based on the hot event; generating an event summary of the hot event based on the news collection; and generating the original article based on a pre-generated topic knowledge base and in combination with the event summary. By generating an event summary based on a news collection, the present invention can accurately summarize the event summary of the hot event, and by generating an original article based on a pre-generated topic knowledge base and in combination with the event summary, the contextual coherence of the article can be improved.

[0193] Please refer to Fig.10 , Fig.10 It is a structural schematic diagram of a device for generating original articles provided by an embodiment of the present invention;

[0194] like Fig.10 As shown, the original article generation device 80 includes:

[0195] A hot event acquisition unit 801 is used to acquire hot events;

[0196] A hot event determination unit 802 is used to screen the acquired hot events through a pre-trained topic analysis model to determine hot events related to a preset topic;

[0197] A news collection acquisition unit 803 is used to acquire a news collection related to the hot events related to the preset topic according to the hot events related to the preset topic;

[0198] An event summary generating unit 804 is used to generate an event summary of the hot events related to the preset topic according to the news collection;

[0199] The original article generating unit 805 obtains the subject short article based on the pre-generated subject knowledge base, combines the event summary, and generates the original article.

[0200] In the embodiment of the present invention, the hot event determination unit 802 is specifically used to:

[0201] A topic category label is generated, and the hot event is identified by the topic category label.

[0202] In some embodiments, the event summary generating unit 804 includes:

[0203] A candidate summary set module 8041 is used to extract original summaries related to the news event from the news set based on a pre-trained multi-document summary extraction model to generate a candidate summary set;

[0204] The event summary generating module 8042 is used to generate an event summary of the hot event according to the candidate summary set.

[0205] In some embodiments, the event summary generation module 8042 is specifically used to:

[0206] Determine a summary to be rewritten from the candidate summary set;

[0207] Based on a pre-trained synonymous sentence rewriting model, the summary to be rewritten is rewritten to generate an event summary of the hot event.

[0208] In some embodiments, the original article generating unit 805 includes:

[0209] The subject short article acquisition module 8051 is used to acquire subject short articles related to the subject category from the subject knowledge base according to the subject category label;

[0210] The original article generation module 8054 is used to automatically splice the event summary and the subject essay to generate the original article.

[0211] In some embodiments, the original article generating unit 805 further includes:

[0212] The optimal short article generation module 8052 is used to select a unique optimal short article from the at least two acquired topic short articles;

[0213] The event essay generating module 8053 is used to rewrite the optimal essay based on a pre-trained synonymous sentence rewriting model to generate an event essay of the hot event.

[0214] In some embodiments, the original article generation module 8054 is specifically used to:

[0215] The event summary and the event essay are automatically spliced ​​together to generate the original article.

[0216] In some embodiments, the preset topics include insurance topics, and the topic knowledge base includes an insurance knowledge base.

[0217] In an embodiment of the present invention, a device for generating original articles is provided, which is applied to a server. The device for generating original articles includes: a hot event acquisition unit, which is used to acquire hot events; a hot event determination unit, which is used to screen the acquired hot events through a pre-trained topic analysis model to determine hot events related to a preset topic; a news collection acquisition unit, which is used to acquire a news collection related to the hot events related to the preset topic according to the hot events related to the preset topic; an event summary generation unit, which is used to generate an event summary of the hot events related to the preset topic according to the news collection; and an original article generation unit, which acquires a topic short article based on a pre-generated topic knowledge base, combines the event summary, and generates the original article. On the one hand, since the hot news information is screened through a pre-trained topic analysis model to determine the hot events related to the preset topic, on the other hand, by acquiring a news collection related to the hot events and generating an event summary based on the news collection, it can accurately summarize the event summary of the hot event, and also generates an original article based on the pre-generated topic knowledge base and the event summary, which can improve the contextual coherence of the article.

[0218] Please refer to Fig.11 , Fig.11 It is a structural diagram of a server provided by an embodiment of the present invention;

[0219] like Fig.11 As shown, the server 110 includes one or more processors 111 and a memory 112. Fig.11 A processor 111 is taken as an example.

[0220] The processor 111 and the memory 112 may be connected via a bus or other means. Fig.11 The example of connecting through bus is taken in the following.

[0221] The memory 112 is a non-volatile computer-readable storage medium that can be used to store non-volatile software programs, non-volatile computer executable programs and modules, such as program instructions / modules corresponding to the vehicle insurance data testing method in the embodiment of the present invention. The processor 111 executes various functional applications and data processing by running the non-volatile software programs, instructions and modules stored in the memory 112, that is, the method for generating original articles provided in the above method embodiment and the functions of each module or unit in the above device embodiment.

[0222] The memory 112 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 112 may optionally include a memory remotely arranged relative to the processor 111, and these remote memories may be connected to the processor 111 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0223] The program instructions / modules are stored in the memory 112 , and when executed by the one or more processors 111 , the method for generating an original article in any of the above method embodiments is executed.

[0224] The server 110 of the embodiment of the present invention exists in various forms. Figure 2 The various steps shown in the figure can also realize the functions of the various units described above, and the server 110 includes but is not limited to:

[0225] (1) Tower Server

[0226] A general tower server case is similar to the PC case we usually use, but a large tower case is much larger. In general, there is no fixed standard for the external dimensions.

[0227] (2) Rack-mounted servers

[0228] Rack-mounted servers are servers with a standard width of 19-inch racks, and the height ranges from 1U to several U, in order to meet the needs of dense deployment in enterprises. Placing servers on racks is not only conducive to daily maintenance and management, but also may avoid unexpected failures. First, placing servers does not take up too much space. Rack servers are neatly arranged in the rack, without wasting space. Secondly, connecting wires can also be neatly stored in the rack. Power cords and LAN cables can all be laid out in the cabinet, which can reduce the number of connecting wires piled on the ground, thereby preventing accidents such as kicking off the wires. The specified dimensions are the width (48.26cm = 19 inches) and height (multiples of 4.445cm) of the server. Since the width is 19 inches, racks that meet this requirement are sometimes called "19-inch racks."

[0229] (3) Blade Server

[0230] Blade server is a low-cost server platform with HAHD (High Availability High Density), which is specially designed for special application industries and high-density computer environments. Each "blade" is actually a system motherboard, similar to an independent server. In this mode, each motherboard runs its own system and serves different designated user groups, and there is no connection between them. However, these motherboards can be combined into a server cluster using system software. In cluster mode, all motherboards can be connected to provide a high-speed network environment, share resources, and serve the same user group.

[0231] (4) Cloud Server

[0232] Cloud server (Elastic Compute Service, ECS) is a simple, efficient, secure, reliable, and elastically scalable computing service. Its management method is simpler and more efficient than that of physical servers. Users can quickly create or release any number of cloud servers without having to purchase hardware in advance. The distributed storage of cloud servers is used to integrate a large number of servers into a supercomputer, providing a large amount of data storage and processing services. Distributed file systems and distributed databases allow access to common storage resources and achieve IO sharing of application data files. Virtual machines can break through the limitations of a single physical machine, and dynamic resource adjustment and allocation eliminate single points of failure of servers and storage devices to achieve high availability.

[0233] An embodiment of the present invention further provides a non-volatile computer storage medium, wherein the computer storage medium stores computer executable instructions, and the computer executable instructions are executed by one or more processors, such as Fig.11 A processor 111 in the embodiment may enable the one or more processors to execute the method for generating an original article in any of the above method embodiments.

[0234] An embodiment of the present invention also provides a computer program product, which includes a computer program stored on a non-volatile computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a server, the server executes the above-mentioned method for generating original articles.

[0235] The above described device or equipment embodiments are merely illustrative, wherein the unit modules described as separate components may or may not be physically separated, and the components displayed as module units may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple network module units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0236] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0237] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may also be combined, the steps may be implemented in any order, and there are many other changes in different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for generating an original article, characterized in that: include: Get hot events; The acquired hot events are screened by a pre-trained topic analysis model to determine hot events related to a preset topic; According to the hot events related to the preset topic, a news collection related to the hot events related to the preset topic is obtained, wherein the news collection includes a plurality of news information; Generating an event summary of the hot events related to the preset topic according to the news collection; Based on a pre-generated subject knowledge base, a subject short article is obtained, and combined with the event summary to generate the original article; Wherein, obtaining a news set related to the hot events related to the preset topic according to the hot events related to the preset topic includes: Calculate the correlation coefficient between the hot events and news information related to the preset topic; If the correlation coefficient is greater than a preset threshold, it is determined that the news information is strongly correlated with the hot event; The news information is obtained, and a plurality of the news information is collected to generate a news collection.

2. The method according to claim 1, characterized in that The determining of hot events related to the preset topic includes: A topic category label is generated, and the hot event is identified by the topic category label.

3. The method according to claim 1, characterized in that Generating an event summary of the hot events related to the preset topic according to the news collection includes: Based on a pre-trained multi-document summary extraction model, extracting original summaries related to the news event from the news collection to generate a candidate summary set; An event summary of the hot event related to the preset topic is generated according to the candidate summary set.

4. The method according to claim 3, characterized in that Generating, according to the candidate summary set, an event summary of the hot event related to the preset topic, includes: Determine a summary to be rewritten from the candidate summary set; Based on a pre-trained synonymous sentence rewriting model, the summary to be rewritten is rewritten to generate an event summary of the hot event related to the preset topic.

5. The method according to claim 2, characterized in that: The method of obtaining a subject short article based on a pre-generated subject knowledge base and combining the event summary to generate the original article includes: According to the subject category label, obtaining subject short articles related to the subject category from the subject knowledge base; The event summary and the subject essay are automatically spliced ​​together to generate the original article.

6. The method according to claim 5, characterized in that After acquiring the subject short articles related to the subject category from the subject knowledge base according to the subject category label, the method further comprises: Select the best short article from at least two short articles on the subject; Based on a pre-trained synonymous sentence rewriting model, the optimal short article is rewritten to generate an event short article of the hot event.

7. The method according to claim 6, characterized in that The automatic splicing of the event summary and the subject short article to generate the original article includes: The event summary and the event essay are automatically spliced ​​together to generate the original article.

8. The method according to any one of claims 1 to 7, characterized in that: The preset topics include insurance topics, and the topic knowledge base includes an insurance knowledge base.

9. A device for generating original articles, characterized in that: include: Hot news information acquisition unit, used to obtain hot events; A hot event determination unit, used to screen the acquired hot events through a pre-trained topic analysis model to determine hot events related to a preset topic; A news collection acquisition unit, configured to acquire a news collection related to the hot events related to the preset topic according to the hot events related to the preset topic, wherein the news collection includes a plurality of news information; An event summary generating unit, configured to generate an event summary of the hot events related to the preset topic according to the news collection; An original article generating unit, based on a pre-generated subject knowledge base, obtains a subject short article, combines the event summary, and generates the original article; The news collection acquisition unit is specifically used to: Calculate the correlation coefficient between the hot events and news information; If the correlation coefficient is greater than a preset threshold, it is determined that the news information is strongly correlated with the hot event; The news information is obtained, and a plurality of the news information is collected to generate a news collection.

10. A server, characterized in that: include: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method for generating an original article as described in any one of claims 1-8.

11. A system for generating original articles, characterized in that: include: The server according to claim 10; A third-party news platform, which is communicatively connected to the server, and includes a third-party news library for storing hot news information so that the server can obtain the hot news information; An article publishing platform is communicatively connected to the server and is used to publish original articles generated by the server.

12. A computer program product, characterized in that The computer program product comprises a computer program stored on a non-volatile computer-readable storage medium, wherein the computer program comprises program instructions, and when the program instructions are executed by a server, the server executes the method for generating an original article as described in any one of claims 1 to 8.

Images