Abstract generation method and system

Abstract

The disclosed embodiments disclose a summary generation system and method. The summary generation system includes: a data preprocessing module for obtaining an event dataset, wherein the event dataset includes multiple data items associated with the event; an event vector generation module for generating an event vector based on one or more of the multiple data items; a category vector generation module for generating a type vector based on one or more of the multiple data items; and a summary generation module for generating summary information of the event based on the category vector and event vector of the event. The disclosed embodiments incorporate the topic categories corresponding to the event into the summary generation model, thereby enabling the summary information generated for different topic categories to have different focuses and styles.

Description

Technical Field

[0001] The present disclosure relates to the field of Internet technology, and in particular to a summary generation method and system. Background Art

[0002] With the rapid development of the internet, a wide variety of online applications have emerged one after another. Social media applications are a very important and popular type of online application. Common social media applications include various social networking sites, Weibo, blogs, forums, and encyclopedias. Social media applications have a large number of users worldwide. Many events are sent to social media applications by users after they occur, and then spread across social networks. Therefore, some online applications use social media data to generate event summaries and display them to their users.

[0003] When generating event summaries from social media data, it's first necessary to aggregate related social media data into an event dataset (cluster), then generate an event summary based on this event dataset. Currently, there are various methods for generating event summaries from event datasets. For example, selecting the earliest published social media data from an event dataset to generate an event summary has the disadvantage that the earliest published social media data may be incomplete or not representative of the entire event. Another example is selecting the most forwarded or commented-on social media data to generate an event summary. This approach has the disadvantage that the most forwarded or commented-on social media data may be generated or forwarded by influential users, and its textual content may not be complete or representative of the entire event. A third approach involves calculating the importance of each word or string in the textual content of social media data and using this to determine the most core social media data for generating the event summary. However, this approach has the disadvantage that the textual content in social media data contains a lot of noise, making the resulting core social media data inaccurate. Summary of the Invention

[0004] In view of this, the purpose of the present disclosure is to provide a summary generation method and system to generate summary information that can reflect the different topic categories to which a social media dataset belongs.

[0005] To achieve this objective, according to a first aspect of the present disclosure, an embodiment of the present disclosure provides a method for generating a summary, comprising:

[0006] Obtaining an event data set, wherein the event data set includes a plurality of data associated with the event;

[0007] Obtaining a category vector and an event vector of the event according to one or more of the plurality of data;

[0008] Generate summary information of the event according to the category vector and the event vector of the event.

[0009] Optionally, after obtaining the event data set, the method further includes:

[0010] Filtering multiple pieces of data in the event data set to obtain filtered data;

[0011] Accordingly, obtaining the category vector and event vector of the event according to one or more of the multiple pieces of data includes:

[0012] According to one or more pieces of filtered data, a category vector and an event vector of the event are obtained.

[0013] Optionally, filtering the plurality of data in the event data set to obtain filtered data includes:

[0014] Calculate the importance of multiple pieces of data;

[0015] sorting the plurality of pieces of data according to the importance;

[0016] A predetermined number of data are obtained from the sorted plurality of data as filtered data.

[0017] Optionally, calculating the importance of each of the plurality of data pieces includes:

[0018] Select any one of the plurality of data;

[0019] Obtaining entity words and time expression words in any of the data;

[0020] The importance of any data is calculated based on the entity words and the time expression words.

[0021] Optionally, filtering the plurality of data in the event data set to obtain filtered data includes:

[0022] Clustering the plurality of data according to a time range;

[0023] Selecting one or more pieces of data within a predetermined time range as data representatives within the predetermined time range;

[0024] The data representatives within each predetermined time range are arranged in chronological order as the filtered data.

[0025] Optionally, obtaining the category vector and event vector of the event includes:

[0026] Using a pre-trained text classification model, training one or more of the multiple data to obtain a category vector of the event, where the category vector is used to represent the topic category of the event;

[0027] One or more of the multiple pieces of data are encoded using a pre-trained text encoding model to obtain an event vector of the event, where the event vector is used to represent the event itself.

[0028] Optionally, generating summary information of the event according to the category vector and the event vector of the event includes:

[0029] Inputting the event category vector and event vector into each layer of a multi-layer decoder of a text generation model, wherein the Nth layer of the multi-layer decoder receives the output of the N-1th layer and serves as the input of the N+1th layer;

[0030] Inputting the category vector of the event into the maximization layer of the text generation model; and inputting the output of the last layer of the multi-layer decoder into the maximization layer after passing through the linear connection layer of the text generation model;

[0031] The maximization layer outputs summary information of the event.

[0032] Optionally, the method further includes:

[0033] The first layer of the multi-layer decoder receives the event summary of the training data;

[0034] The multi-layer decoder then decodes the category vector and event type of the event according to the event summary of the training data.

[0035] Optionally, obtaining the category vector of the event according to one or more of the multiple pieces of data includes:

[0036] generating a topic category in text form based on one or more of the plurality of data;

[0037] And generating a category vector in numerical form according to the topic category in text form.

[0038] Optionally, after obtaining the event data set, the method further includes:

[0039] Noise reduction processing is performed on each piece of data in the event data set.

[0040] In a second aspect, an embodiment of the present disclosure provides a summary generation system, including:

[0041] A data preprocessing module, configured to obtain an event data set, wherein the event data set includes a plurality of data items associated with the event;

[0042] an event vector generating module, configured to generate an event vector according to one or more pieces of data;

[0043] a category vector generating module, configured to generate a type vector according to one or more pieces of data;

[0044] The summary generation module is used to generate summary information of the event according to the category vector and event vector of the event.

[0045] Optionally, the data preprocessing module includes:

[0046] Filtering multiple pieces of data in the event data set to obtain filtered data;

[0047] Accordingly, the event vector generation module includes: obtaining the event vector of the event according to one or more pieces of filtered data,

[0048] The category vector generation module includes: obtaining the category vector of the event according to one or more pieces of filtered data.

[0049] Optionally, filtering the plurality of data in the event data set to obtain filtered data includes:

[0050] Calculate the importance of multiple pieces of data;

[0051] sorting the plurality of pieces of data according to the importance;

[0052] A predetermined number of data are obtained from the sorted plurality of data as filtered data.

[0053] Optionally, calculating the importance of each of the plurality of data pieces includes:

[0054] Select any one of the plurality of data;

[0055] Obtaining entity words and time expression words in any of the data;

[0056] The importance of any data is calculated based on the entity words and the time expression words.

[0057] Optionally, filtering the plurality of data in the event data set to obtain filtered data includes:

[0058] Clustering the plurality of data according to a time range;

[0059] Selecting one or more pieces of data within a predetermined time range as data representatives within the predetermined time range;

[0060] The data representatives within each predetermined time range are arranged in chronological order as the filtered data.

[0061] Optionally, the event vector generation module includes:

[0062] Encode one or more of the multiple pieces of data using a pre-trained text encoding model to obtain an event vector for the event, where the event vector is used to represent the event itself;

[0063] The category vector generation module includes: using a pre-trained text classification model to train one or more of the multiple data to obtain a category vector of the event, and the category vector is used to represent the topic category of the event.

[0064] Optionally, the data preprocessing module further includes performing noise reduction processing on the data in the event data set.

[0065] Optionally, the summary generation module generates summary information using a text generation model, including:

[0066] Inputting the event category vector and event vector into each layer of a multi-layer decoder of a text generation model, wherein the Nth layer of the multi-layer decoder receives the output of the N-1th layer and serves as the input of the N+1th layer;

[0067] Inputting the category vector of the event into the maximization layer of the text generation model; and inputting the output of the last layer of the multi-layer decoder into the maximization layer after passing through the linear connection layer of the text generation model;

[0068] The maximization layer outputs summary information of the event.

[0069] Optionally, the first layer of the multi-layer decoder receives an event summary of training data, and the multi-layer decoder decodes the category vector and event type of the event according to the event summary of the training data.

[0070] In a third aspect, an embodiment of the present disclosure provides an electronic device, including a memory and a processor, wherein the memory further stores computer instructions executable by the processor, and when the computer instructions are executed, any of the above-mentioned summary generation methods is implemented.

[0071] In a fourth aspect, an embodiment of the present disclosure provides a computer-readable medium storing computer instructions executable by an electronic device, wherein when the computer instructions are executed, the summary generation method described above is implemented.

[0072] The method provided by the embodiments of the present disclosure uses category vectors and event vectors as input data for a trained summary generation model, and generates summaries based on the different topic categories to which each social media dataset belongs. This ensures that the summary content generated for different topic categories has a corresponding appropriate style and captures the key information of the topic.

[0073] In a further embodiment, the event vector is obtained by using a text encoding model constructed using the BERT model. Since the BERT model is a relatively mature pre-trained model, it can be used as the text encoding model in this embodiment through simple adjustments, which not only saves model training time but also fully utilizes the powerful natural language processing capabilities of the BERT model. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The above and other objects, features and advantages of the present disclosure will become more apparent through description of the embodiments of the present disclosure with reference to the following drawings, in which:

[0075] Figure 1 is an exemplary scene diagram of an embodiment of the present disclosure;

[0076] Figure 2 is a structural block diagram of a summary generation system provided by an embodiment of the present disclosure;

[0077] Figure 3a This is a schematic diagram of the input and output of the BERT model;

[0078] Figure 3b This is the architectural diagram of the BERT model;

[0079] Figure 4 is a flowchart of a summary generation method provided by an embodiment of the present disclosure;

[0080] Figure 5 It is a structural diagram of an electronic device for implementing an embodiment of the present disclosure. DETAILED DESCRIPTION

[0081] The present disclosure is described below based on examples, but the present disclosure is not limited to these examples. Certain specific details are described in detail in the detailed description of the present disclosure below. Those skilled in the art will appreciate that the present disclosure is fully understood without these details. To avoid obscuring the essence of the present disclosure, well-known methods, processes, and procedures have not been described in detail. The accompanying drawings are not necessarily drawn to scale.

[0082] Before introducing various aspects of the embodiments of the present disclosure, background information is first introduced.

[0083] Social media applications are applications that allow users to access and share various information. Through social media applications, a user can send messages, either created or obtained elsewhere, to one or more users, and simultaneously receive messages from one or more users. These messages can be in various formats, such as text, images, videos, audio, and animations. In this article, social media data refers to messages carried in text. A collection of multiple pieces of social media data organized according to different events is called an event dataset. In other words, an event dataset contains multiple pieces of social media data associated with a specific event.

[0084] Figure 1 As shown in the figure, the system includes a plurality of terminal devices 10 and a server cluster 30 coupled via a network 20 .

[0085] Each terminal device 10 can be built based on various processor models currently available on the market and include various components such as memory and input / output devices. Various operating systems and application software are installed on the terminal device 10. Operating systems include, for example, Windows™, UNIX, Linux, and RTOS. Application software includes, for example, browsers, video playback software, communication software, and various R&D tools.

[0086] Each terminal device 10 can communicate with the network 20 in various ways. The network 20 can be implemented using one or a combination of various communication technologies implemented by exchanging signals, including but not limited to wired technologies using electrical and / or optical cables, and wireless technologies using infrared, radio frequency, and / or other forms. In different application scenarios, the network 20 can be the Internet, a wide area network, or a local area network, such as a company's proprietary network. The network 20 can also be a wired network or a wireless network.

[0087] Each terminal device 10 is deployed with a client program 11 of a social media application. A server cluster 30 is deployed with a server program 31 of the social media application, a storage system 32, a processing system 33, and a portal website 34. When multiple terminal devices 10 interact through the server program 31, the relevant social media data is stored in the storage system 32. During this process, the storage system 32 is used to store the social media data.

[0088] Summary generation system 33 can read social media data 321 from storage system 32 and generate summary information 322 for an event based on one or more pieces of social media data. Summary information 322 can be stored in storage system 32 and pushed to portal website 34 for display via storage system 32. Portal website 34 can display summary information 322 based on user-entered filtering criteria. Summary information 322 can also be pushed to users based on user subscriptions.

[0089] Of course, this application scenario is only an exemplary application scenario. There are many such application scenarios, which are not limited to this example. For example, a search engine service can crawl social media data from the storage system 32 or crawl summary information to its own storage system through a crawler program for user retrieval.

[0090] Figure 2 yes Figure 1 The structural block diagram of the summary generation system 33 is shown in FIG. Figure 2 As shown, the input of system 33 is Figure 2 The social media data 321 shown is output as Figure 2 The summary information 322 is shown. The social media data 321 includes multiple pieces of data related to various events. The system 33 is used to integrate the social media data according to different events and obtain summary information of various events.

[0091] As described above, the processing system 33 includes a data pre-processing module 203 , an event vector generating module 201 , a category vector generating module 204 and a summary generating module 202 , which are divided according to processing logic.

[0092] Data preprocessing module 203 is used to obtain social media data 321 and organize it according to different events to generate event datasets 323 for each event. Each event dataset 323 includes multiple pieces of social media data related to the specific event. Event datasets 323 are provided to category vector generation module 204 and event vector generation module 201.

[0093] Optionally, data preprocessing module 203 filters event dataset 323 and accordingly provides filtered event dataset 323 to category vector generation module 204 and event vector generation module 201. Filtered event dataset 323 only contains one or more pieces of social media data that are closely associated with a specific event, while filtering out duplicate and unimportant social media data. We can employ various methods to filter event dataset 323 for a specific event.

[0094] As a first method, the importance of each piece of social media data in the event dataset can be calculated, and the multiple pieces of social media data can be sorted according to their importance. A predetermined number of pieces of data from the sorted pieces of social media data can be obtained as the data in the filtered event dataset. The importance of each piece of data can be calculated in the following manner. Since the description of an event requires elements such as the subject word, time expression words, and location expression words of the event, the subject words (personal names, place names, unit names), time expressions, and location expressions in the social media data are first identified. The importance of a piece of social media data is then calculated based on these elements. For example, different scores are assigned to each element, and the scores of the elements contained in each piece of social media data are then accumulated to obtain a total score for the social media data. This total score represents the importance of the social media data. Sorting by importance is equivalent to sorting by the total score. To distinguish the importance of different elements, a weight coefficient can also be set for each element. The total score of each social media data piece is obtained by multiplying the weight coefficient by the element score and then summing them.

[0095] As a second method, multiple data items within the event dataset are clustered according to time range. One or more data items within the predetermined time range are selected as representative data items within the predetermined time range. The data representatives within each predetermined time range are then arranged in chronological order to form a filtered event dataset. Data representatives can be selected from the data set based on semantic analysis or the elements they contain.

[0096] Optionally, the data preprocessing module 203 performs noise reduction processing on the various data in the event dataset 323. While social media applications allow users to freely create, edit, and share their own original content online, this can encourage users to publish their own original content. However, it can also lead to various problems in the disseminated content, such as grammatical errors and incorrect speech. Furthermore, users can use various icons and pronouns to circumvent social media application checks on social media data. This makes social media data significantly noisier than formal news and information text. This noise can affect the analysis of natural language, making it necessary to perform noise reduction processing first. Noise reduction processing, for example, includes removing stop words. Stop words, such as various pronouns, adverbs, and prepositions in Chinese and English, do not provide useful information for semantic analysis and can interfere with semantic analysis. Therefore, these words need to be removed before information extraction. Noise reduction processing also includes stemming and lemmatization. Stemming and lemmatization are commonly used techniques in text processing. Stemming unifies words with different parts of speech but the same meaning. For example, stemming "cats" becomes "cat" and "effective" becomes "effect." Lemmatization unifies words from different eras, such as "drove" becomes "drive" and "driving" becomes "drive." Noise reduction also includes correcting spelling errors and removing duplicate data. We won't go into detail here.

[0097] By means of data filtering and data noise reduction, the provided category vector 213 and event vector 212 will be made more accurate, thereby making the generated summary information more accurate and more logical.

[0098] The category vector generation module 204 receives the event dataset 323 and obtains a category vector 213 for the event based on one or more data items in the event dataset 323. The category vector generation module 204 is used to generate a category vector for the text content. An optional processing method for the category vector generation module 204 is to first generate textual topic categories based on the event dataset, and then obtain a category vector corresponding to the topic category based on the topic category. The topic category is the category to which the event corresponding to the event dataset 323 belongs. For example, an event dataset can belong to different categories, such as medical accidents, explosions, floods, and commercial mergers and acquisitions. The final classification requires further analysis. For example, a car explosion can be classified as a traffic accident or a suicide attack. In implementation, topic terms can be first extracted from the event dataset, and then the topic category of the event dataset can be determined based on the topic terms. To extract topic terms, each social media data item is first segmented. Then, the word frequency, document frequency, or inverse document frequency (TFIDF) of each word in the vocabulary is determined. If the word frequency, document frequency, or inverse document frequency falls within a predetermined range, the word is extracted as a topic term. After determining the topic category, the topic category is converted into a category vector. A category vector is a set of numerical arrays that represent an event. Converting the textual topic category into a fixed-dimensional numerical category vector facilitates calculation by the summary generation module 202. Another optional processing method is to use a pre-trained text classification model to train one or more of the multiple data items to obtain the category vector of the event.

[0099] The event vector generation module 201 receives the event dataset 323 and obtains an event vector for the event based on one or more data items in the event dataset 323. The event vector 212 is a semantic representation of the event itself corresponding to the event dataset. The event vector generation module 201 must be able to distinguish event descriptions. For events that are similar in description but substantively different, it can output different event vectors. For events that are substantially identical but have significantly different descriptions, it can output the same or only slightly different event vectors.

[0100] Optionally, the event vector generation module 201 utilizes a pre-trained text encoding model to encode one or more of the multiple data items to obtain an event vector for the event. The training process of the text encoding model gradually enables the text encoding model to accurately describe events. When constructing training data for the text encoding model, several event datasets with similar descriptions but essentially different, as well as several event datasets with significantly different descriptions but essentially the same, can be prepared to train the text encoding model. The text encoding model can be a neural network model, such as a recurrent neural network, a convolutional neural network, or a recursive neural network. Each event dataset is converted into an event vector by this module, thereby facilitating subsequent input into the summary generation module 202 for calculation.

[0101] The text encoding model can be built based on the BERT model. BERT (Bidirectional Encoder Representations from Transformers) is a neural network language model based on the bidirectional encoding Transformer released in 2019. The BERT model is a pre-trained model that converts the input sentence or paragraph into a feature vector. As we all know, model training usually requires a lot of resources, including but not limited to a large amount of sample data that needs to be labeled and computer resources to perform training, so model training is not easy to perform. The pre-trained model provides a better set of weight parameter values. R&D personnel can directly apply it to actual scenarios, or before applying it to actual scenarios, they only need to simply modify some output layers, and then perform incremental training with their own data to fine-tune the weight parameter values. The following is through Figure 3a and 3b Further explanation of the BERT model. Figure 3a The initial input of the BERT model is the original text. The BERT model first converts each word in the text into a word vector by querying the word vector table, and generates a text vector and a position vector. The BERT model then calculates based on the word vector, text vector, and position vector, and outputs the text semantics. Figure 3bThe BERT model mainly includes 12 stacked processing layers, each of which includes a multi-head self-attention mechanism layer 302, a normalization layer 303, a feedforward neural network layer 305 and a normalization layer 304. The input data is a vector 301 (character vector, text vector and position vector representing the current text), and the output is a prediction result for the current text: predicting a covered word and predicting the next sentence in the text (the final result is composed of the results of each processing layer). BERT can be applied to different natural language processing tasks. When it comes to a specific language processing task, the input and output can be fine-tuned. For example, when processing single text (this embodiment is single text processing), the event data set is first composed of an input text of no more than 512 characters (BERT's input requirement is no more than 512 characters), and then a [CLS] symbol is inserted before the text. The output vector corresponding to the symbol of the BERT model is the semantic expression of the input text (that is, the event vector in the present invention).

[0102] The summary generation module 202 receives the category vector 213 and the event vector 212 and generates summary information of the event. The summary generation module 202 can use a trained text generation model to generate summary information. The text generation model includes a multi-layer decoder, a linear connection layer and a maximization layer. The Nth layer of multiple decoding layers receives the output of the N-1th layer and serves as the input of the N+1th layer. N is a positive integer. The category vector and the event vector of the event are input to each layer of the multi-layer decoder of the text generation model, the category vector of the event is input to the multi-layer decoder and the maximization value layer, and the output of the last layer of the multi-layer decoder is input to the maximization value layer after passing through the linear connection layer. Thus, the category vector 213 will play a role in the probability corresponding to the various summary information output by the maximization layer. Ultimately, the summary information with the highest probability is selected as the summary information of the event.

[0103] The figure shows an exemplary text generation model, which includes a six-layer Transformer decoding structure, including six decoding layers decode1-decode6, a linear connection layer Lconx, and a maximization layer Stmx. The six decoding layers are stacked one on top of the other, and the output of the previous decoding layer is the input of the next decoding layer. When the summary information is generated using the summary generation module 202, the event vector 212 obtained by the event vector generation module 201 will be sent to each of the six decoding layers of the decoder. The category vector 213 is sent to each of the six decoding layers and the maximization layer Stmx, so that the category vector 213 can affect the decoding results of each decoding layer and play a role in maximizing the probability of each summary information output by the layer Stmx.

[0104] The summary generation system provided in this embodiment adds the category targeted by the event into the summary generation model, so that the summary content generated for different topic categories has corresponding emphasis and appropriate style, and captures the key information of the topic.

[0105] Furthermore, the summary information obtained can be a shorter title or a longer text, and the length can be different according to different topic categories. In addition, different topic categories can target different users and generate different summary information.

[0106] Corresponding to the above embodiment, Figure 4 The flowchart of the summary generation method provided by the embodiment of the present disclosure includes the following steps.

[0107] In step S401, an event dataset is obtained.

[0108] The event data set includes multiple pieces of data associated with the event. The event data set comes from social media data. In other words, social media data is sorted into multiple event data sets that represent specific events. Social media data is a collection of disorganized data. Most of this data is not related to the event itself. Therefore, it is usually necessary to filter the obtained social media data. The filtered event data set 323 only contains one or more social media data that are relatively closely related to the specific event, and duplicate and unimportant social media data will be filtered out. The filtering method is as described above and will not be repeated here. Similarly, the data contained in the event data set may contain a large amount of information (i.e., noise) that is irrelevant to the event itself, such as emoticons and various greeting words. These information are mixed with the event description, so the data contained in the data set is subjected to noise reduction processing.

[0109] In step S402 , a category vector and an event vector of the event are obtained according to one or more pieces of data.

[0110] Event vector 212 is the semantic representation of the event itself, corresponding to the corresponding event dataset. This step requires the ability to distinguish event descriptions. For events with similar descriptions but substantively different, different event vectors can be output. However, for events with significantly different descriptions but substantively the same, identical or identical event vectors can be output. Optionally, a pre-trained text encoding model can be used to encode one or more of the multiple data items to obtain the event vectors. The training process of the text encoding model gradually enables the text encoding model to accurately describe events. As previously mentioned, the text encoding model can be constructed based on the BERT model. The BERT model offers significant flexibility in processing different natural language tasks. For different natural language processing tasks, simple adjustments to the input, output, and weight parameters of the pre-trained BERT model are sufficient. For example, a modified BERT model can be used to determine the semantic expression of the input text, or the emotional expression of the input text, and so on.

[0111] The event category vector is a numerical representation of the event category. An optional processing method for this step is to first generate topic categories in text form based on the event dataset, and then obtain a category vector corresponding to the topic category based on the topic category. The topic category is the category to which the event corresponding to the event dataset belongs. For example, an event dataset can belong to different categories such as medical accidents, explosions, floods, and commercial mergers and acquisitions. The final classification into which category requires further analysis. For example, a car explosion can be classified as a traffic accident or a suicide attack. In implementation, the subject words of the event dataset can be first extracted, and then the topic category of the event dataset can be determined based on the subject words. When extracting the subject words, each social media data is first segmented, and then the word frequency, document frequency, or inverse document frequency (TFIDF) of each word in the vocabulary is determined. When the word frequency, document frequency, or inverse document frequency is within a predetermined range, the word is extracted as a subject word. After determining the topic category, the topic category is converted into a category vector. Another optional processing method is to use a pre-trained text classification model to train one or more of the multiple data points to obtain the event category vector.

[0112] In step S403, summary information of the event is generated based on the category vector and event vector of the event. Optionally, the category vector and event vector of the event are output to a trained text generation model to generate summary information. The text generation model includes a multi-layer decoder, a linear connection layer, and a maximization layer. The Nth layer of the multiple decoding layers receives the output of the N-1th layer and serves as the input of the N+1th layer. N is a positive integer. The category vector and event vector of the event are input to each layer of the multi-layer decoder of the text generation model. The category vector of the event is input to the multi-layer decoder and the maximization layer. The output of the last layer of the multi-layer decoder is input to the maximization layer after passing through the linear connection layer. The category vector thus plays a role in the probability corresponding to the various summary information output by the maximization layer. Ultimately, the summary information with the highest probability is selected as the summary information of the event. The text generation model can be determined through a training process. During the training process, the first layer of the multi-layer decoder receives the event summary of the training data. The multi-layer decoder then decodes the category vector and event type of the event based on the event summary of the training data. The event summary generated by the model is compared with the input event summary to determine whether the difference between the two has been reduced. The number of layers in a multi-layer decoder can also be determined during training. For example, Figure 2 A six-layer decoder was determined.

[0113] The summary generation method provided in this embodiment combines the event category with the event itself to generate summary information, thereby generating summary information for different topic categories and making the summary information have different focuses and styles.

[0114] For example, for a shooting incident, the resulting summary might focus on the time, location, circumstances of the shooting, casualties, and suspects. For a corporate acquisition, the resulting summary might focus on the acquiring parties, the acquisition amount, and the acquisition date.

[0115] Furthermore, as an optional embodiment, the raw social media data can be filtered to reduce irrelevant social media data and noise can be reduced through noise reduction, thereby generating more accurate event category vectors and event vectors, and thus more accurate event summary information. As another optional embodiment, the data in the event dataset can be sorted by time. When generating the event summary information, the summary information can be output based on time. For example, the summary content can be in the following order: explosion -> casualty information -> tracking suspect. This can make the summary content more logical and the description of the event more accurate.

[0116] Corresponding to the above embodiment, Figure 5As shown, the present disclosure also provides an electronic device 50, which includes a memory 502 and a processor 501 at the hardware level. In addition, in some cases, it also includes input and output devices 503 and other hardware 504. The memory 502 is, for example, a high-speed random access memory (RAM), or it may be a non-volatile memory (non-volatile memory), such as at least one disk storage, etc. The input and output devices 503 are, for example, devices such as a display, keyboard, mouse, network controller, etc. The processor 501 can be built based on various models of processors currently on the market. The processor 501, memory 502, input and output devices 503 and other hardware 504 are interconnected through a bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 The fact that only one line is used does not mean that there is only one bus or only one type of bus.

[0117] The memory 502 is used to store programs. Specifically, the program may include program code, which includes computer instructions. The memory may include internal memory and non-volatile memory, and provides computer instructions and data to the processor 501. The processor 501 reads the corresponding computer program from the memory 502 into the internal memory and then runs it. At the logical level, the summary generation method specifically includes the following steps: obtaining an event data set, wherein the event data set includes multiple data associated with the event; obtaining a category vector and an event vector for the event based on one or more of the multiple data; and generating summary information for the event based on the category vector and event vector of the event.

[0118] Those skilled in the art will appreciate that the present disclosure can be implemented as systems, methods, and computer program products. Therefore, the present disclosure can be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, and microcode), or in a combination of software and hardware. Furthermore, in some embodiments, the present disclosure can also be implemented in the form of a computer program product in one or more computer-readable media containing computer-readable program code.

[0119] Any combination of one or more computer-readable media can be used. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium is, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared or semiconductor system, device or component, or any combination thereof. More specific examples of computer-readable storage media include: an electrical connection of one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this article, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by a processing unit, device or device, or used in combination with it.

[0120] A computer-readable signal medium may include a data signal propagated in baseband or as part of a chopped wave, which carries computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any other suitable combination. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transfer a program for use by or in conjunction with an instruction system, apparatus, or device.

[0121] Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., and any suitable combination of the foregoing.

[0122] The computer program code for executing the disclosed embodiments can be written in one or more programming languages or combinations. The programming languages include object-oriented programming languages, such as JAVA, C++, and can also include conventional procedural programming languages, such as C. The program code can be executed entirely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (e.g., using an Internet service provider to connect via the Internet).

[0123] The foregoing description is merely a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present disclosure shall be included within the scope of protection of the present disclosure.

Claims

1. A method for generating a summary, comprising: Obtaining an event data set, wherein the event data set includes a plurality of data associated with the event; Obtaining a category vector and an event vector of the event according to one or more of the plurality of data; Input the category vector and event vector of the event into each layer of the multi-layer decoder of the text generation model, wherein the Nth layer of the multi-layer decoder receives the output of the N-1th layer and serves as the input of the N+1th layer; Inputting the category vector of the event into the maximization layer of the text generation model; And, the output of the last layer of the multi-layer decoder is input into the maximization layer after passing through the linear connection layer of the text generation model; The maximum value layer outputs summary information of the event.

2. The method for generating a summary according to claim 1, wherein: After obtaining the event data set, the method further includes: Filtering multiple pieces of data in the event data set to obtain filtered data; Accordingly, obtaining the category vector and event vector of the event according to one or more of the multiple pieces of data includes: According to one or more pieces of filtered data, a category vector and an event vector of the event are obtained.

3. The method for generating a summary according to claim 2, wherein: The filtering of the plurality of data in the event data set to obtain filtered data includes: Calculate the importance of multiple pieces of data; sorting the plurality of pieces of data according to the importance; A predetermined number of data are obtained from the sorted plurality of data as filtered data.

4. The method for generating a summary according to claim 3, wherein: The calculating of the importance of each of the plurality of data includes: Select any one of the plurality of data; Obtaining entity words and time expression words in any of the data; The importance of any data is calculated based on the entity words and the time expression words.

5. The method for generating a summary according to claim 2, wherein: The filtering of the plurality of data in the event data set to obtain filtered data includes: Clustering the plurality of data according to a time range; Selecting one or more pieces of data within a predetermined time range as data representatives within the predetermined time range; The data representatives within each predetermined time range are arranged in chronological order as the filtered data.

6. The method for generating a summary according to claim 1, wherein: The obtaining of the category vector and the event vector of the event includes: Using a pre-trained text classification model, training one or more of the multiple data to obtain a category vector of the event, where the category vector is used to represent the topic category of the event; One or more of the multiple pieces of data are encoded using a pre-trained text encoding model to obtain an event vector of the event, where the event vector is used to represent the event itself.

7. The method for generating a summary according to claim 1, wherein: The method further comprises: The first layer of the multi-layer decoder receives the event summary of the training data; The multi-layer decoder then decodes the category vector and event type of the event according to the event summary of the training data.

8. The method for generating a summary according to claim 1, wherein: Obtaining the category vector of the event according to one or more of the plurality of data includes: generating a topic category in text form based on one or more of the plurality of data; and A category vector in numerical form is generated according to the topic category in the text form.

9. The method for generating a summary according to claim 1, wherein: After obtaining the event data set, the method further includes: Noise reduction processing is performed on each piece of data in the event data set.

10. A summary generation system, comprising: A data preprocessing module, configured to obtain an event data set, wherein the event data set includes a plurality of data items associated with the event; an event vector generating module, configured to generate an event vector according to one or more pieces of data; A category vector generating module, configured to generate a type vector according to one or more pieces of data; A summary generation module is configured to input the category vector and event vector of the event into each layer of a multi-layer decoder of a text generation model, wherein the Nth layer of the multi-layer decoder receives the output of the N-1th layer and serves as the input of the N+1th layer; Inputting the category vector of the event into the maximization layer of the text generation model; And, the output of the last layer of the multi-layer decoder is input into the maximization layer after passing through the linear connection layer of the text generation model; The maximum value layer outputs summary information of the event.

11. The summary generation system according to claim 10, wherein: The data preprocessing module includes: Filtering multiple pieces of data in the event data set to obtain filtered data; Accordingly, the event vector generation module includes: obtaining the event vector of the event according to one or more pieces of filtered data, The category vector generation module includes: obtaining the category vector of the event according to one or more pieces of filtered data.

12. The summary generation system according to claim 11, wherein: The filtering of the plurality of data in the event data set to obtain filtered data includes: Calculate the importance of multiple pieces of data; sorting the plurality of pieces of data according to the importance; A predetermined number of data are obtained from the sorted plurality of data as filtered data.

13. The summary generation system according to claim 12, wherein: The calculating of the importance of each of the plurality of data includes: Select any one of the plurality of data; Obtaining entity words and time expression words in any of the data; The importance of any data is calculated based on the entity words and the time expression words.

14. The summary generation system according to claim 11, wherein: The filtering of the plurality of data in the event data set to obtain filtered data includes: Clustering the plurality of data according to a time range; Selecting one or more pieces of data within a predetermined time range as data representatives within the predetermined time range; The data representatives within each predetermined time range are arranged in chronological order as the filtered data.

15. The summary generation system according to claim 10, wherein: The event vector generation module includes: Encode one or more of the multiple pieces of data using a pre-trained text encoding model to obtain an event vector for the event, where the event vector is used to represent the event itself; The category vector generation module includes: using a pre-trained text classification model to train one or more of the multiple data to obtain a category vector of the event, and the category vector is used to represent the topic category of the event.

16. The summary generation system according to claim 10, wherein: The data preprocessing module also includes performing noise reduction processing on the data in the event data set.

17. The summary generation system according to claim 10, wherein: A first layer of the multi-layer decoder receives an event summary of training data, and the multi-layer decoder decodes a category vector and an event type of the event according to the event summary of the training data.

18. An electronic device comprising a memory and a processor, wherein the memory further stores computer instructions executable by the processor, and when the computer instructions are executed, the method for generating a summary according to any one of claims 1 to 9 is implemented.

19. A computer-readable medium storing computer instructions executable by an electronic device, wherein when the computer instructions are executed, the method for generating a summary according to any one of claims 1 to 9 is implemented.

Images