Abstract generation method, device and storage medium
By combining the fact-aware abstract generation method of element words, factor words such as time, place, object and numerals are extracted and integrated, the problem of factual consistency of the generative abstract system is solved, and the factuality and credibility of the abstract is improved.
Patent Information
- Application Number
- CN202210360857.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-07
- Publication Date
- 2025-03-28
- Estimated Expiration
- 2042-04-07
AI Technical Summary
The existing generative digest systems have problems with factual consistency, resulting in a decrease in reference and credibility of generated digests.
The fact-aware abstract generation method combined with element words is adopted, and the element words with four attributes, including time, place, object and numerals, are trained to extract element words with four types of attributes, namely time, place, object and numerals, and incorporate them into the abstract generation model to improve the factual quality of the abstract.
It improves the accuracy and practicality of factor word extraction, improves the factual quality in the abstract generation results, and is suitable for the Chinese abstract generation system.
Smart Images
Figure CN115081438B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method, device and storage medium for generating a summary, and in particular to a method, device and storage medium for generating a fact-aware summary combining element words. Background Art
[0002] Generative summarization is a tool that helps people in modern society understand the core content of long text information quickly and efficiently. It can compress and repeat the input long text and transform it into a shorter summary text with higher information density to summarize the important information in the input. As a key subfield in the field of natural language processing, generative summarization has been able to generate text with better readability after more than ten years of rapid development, but it still faces the problem of factual consistency. The reference and credibility of summaries with factual consistency problems will be greatly reduced, which is a huge blow to the usability of generative summary systems.
[0003] At present, there are two main methods to improve the factual consistency problem:
[0004] 1) Summary fact checking;
[0005] 2) Fact-aware summary generation.
[0006] Summary fact detection can detect whether the summary is consistent with the facts described in the original text, but it cannot solve the factual errors in the summary generation model. Fact-aware summary generation integrates additional factual information into the summary generation model to improve the factual errors in the model generation results during the generation stage.
[0007] One of the current mainstream fact-aware summary generation methods is to combine factor word information. For traditional fact-aware summary generation solutions that combine keywords, the words that overlap between the manual summary and the original text are generally used as target extracted keywords. However, this will cause the extracted words to be too redundant, and it is easy to extract some words that are not so critical to the fact description, and some redundant words that are not important to the core facts. At the same time, the sequence composed of the extracted keywords is very similar to the manual summary, making the keyword extraction task more difficult, thereby reducing the extraction effect.
[0008] For example, for the material "The city's auto show is in full swing, and this auto show has invited many brands and manufacturers to participate.", it is easy to extract words such as "this time", "please come", and "in full swing". Moreover, too many keywords will make the keyword extraction task a vassal of manual summarization, increase the difficulty of the task, and make the difficulty of the keyword extraction task close to the summary generation task, thus reducing the extraction effect.
[0009] In order to solve a series of shortcomings in the prior art, the present invention proposes a fact-aware summary generation method combining element words, which is more suitable for Chinese summary generation systems than traditional methods. In addition, the method can more specifically extract element words related to the basic elements when describing facts, improve the accuracy and practicality of the extracted element words, and then integrate these element words into the input of the summary generation model, thereby effectively improving the factual quality of the summary generation results. Summary of the invention
[0010] In order to solve or alleviate some or all of the above technical problems, the present invention is implemented through the following technical solutions:
[0011] A summary generation method, the summary generation method comprising a training step and a prediction step,
[0012] The training phase comprises the following steps:
[0013] a) Using word segmentation and part-of-speech tagging, we extract the key words of four attributes, namely time, place, object and numeral, from the manual summary, and use the prompt as the training target of the key word extraction model to train the key word extraction model;
[0014] b) extracting key words from the training data set of the summary generation model using the key word extraction model and part-of-speech tagging;
[0015] c) After extracting the key words, the key words are concatenated to the head of the original text in combination with the prompt to obtain a training set of the key words;
[0016] d) training the summary generation model using the training set combined with the key words;
[0017] The prediction stage step includes the following steps:
[0018] e) Using the trained feature word extraction model to extract feature words of four attributes: time, place, object, and numeral from the original text;
[0019] f) The extracted key words are combined with the prompt and concatenated with the original text, and then input into the trained summary generation model to obtain the final summary result.
[0020] In a certain type of embodiment, the element word extraction module is directly used to extract the element words of the four attributes, and then the element words of the four attributes and the original text are concatenated to form a training input data combined with the element words.
[0021] In a certain embodiment, the word frequency of each element word extracted by the element word extraction model in the manual summary is calculated, and finally the word frequency of all element words is summed and averaged to obtain the overall relevance score of the element words of the four attributes and the manual summary; the calculation formula of the relevance score is as follows:
[0022]
[0023] Among them, S(k) represents the relevance score between the element word set k and the manual summary in all four attributes; w i represents the i-th element word in the four element word set k; n is the total number of element words contained in the element word set k of the four attributes; Count represents the frequency of the element word appearing in the manual summary; ε is a very small smooth number to prevent division by zero.
[0024] In certain embodiments, if the relevance score between the element word set and the manual summary is lower than a threshold value of 0.25, the sample will be discarded.
[0025] In a certain embodiment, after extracting the essential words in step b), if one of the essential words of the four attributes of time, place, object and numeral is missing, an essential word extraction model is used in the original text to extract the missing type of essential words, and finally the essential words from these two different sources are combined to form a set of four complete types of essential words, which are then spliced with the original text as training samples.
[0026] In certain embodiments, the element word extraction model is a UniLM model.
[0027] A summary generation device comprises an element word extraction model acquisition module and a generative summary module, and the two modules are configured to execute any of the above-mentioned summary generation methods.
[0028] A storage medium stores computer instructions, and the computer instructions are loaded to execute any of the above-mentioned methods for generating a summary.
[0029] Some or all of the embodiments of the present invention have the following beneficial technical effects:
[0030] 1) The present invention draws on the basic elements of Chinese factual narration to design element words. By targeting four types of element words, such as "time", "place", "object", and "numeral", an element word extraction model is trained based on prompt to extract element words from the original text, so as to provide additional factual information for the generative summary model. This not only improves the accuracy and practicality of the extracted element words, but also improves the factual quality of the summary generation results.
[0031] 2) The present invention proposes a generative summary model based on prompt training, and proposes two different training modes for combining element words. These two modes can respectively improve the robustness / denoising ability of the generative summary model when combining element word information for generation, and improve the correlation between the summary model results and element words during generation.
[0032] 3) Combining the basic elements of Chinese factual narration, only four key words that are more critical to factual description are extracted. While being able to better describe the key information in the event, it also reduces the difficulty of the key word extraction task and improves the accuracy of key word extraction.
[0033] 4) The present invention is particularly suitable for generating Chinese abstracts.
[0034] More beneficial effects will be further introduced in the preferred embodiments. The technical solutions / features disclosed above are intended to summarize the technical solutions and technical features described in the specific implementation method section, so the scope of the records may not be exactly the same. However, these new technical solutions disclosed in this section are also part of the many technical solutions disclosed in the invention document. The technical features disclosed in this section and the technical features disclosed in the subsequent specific implementation method section and some contents in the drawings that are not explicitly described in the specification disclose more technical solutions in a reasonable combination with each other. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein:
[0036] Figure 1 A schematic diagram showing the steps of obtaining the element word extraction model of the present invention is shown;
[0037] Figure 2 A specific example of tagging a sentence according to the part-of-speech tagging specification of the present invention is shown;
[0038] Figure 3 An example of the prompt-based feature word extraction model of the present invention is shown;
[0039] Figure 4 A schematic diagram showing the training of a generative summary module in a first embodiment of the present invention is shown;
[0040] Figure 5 A schematic diagram showing the training of a generative summary module in a second embodiment of the present invention is shown;
[0041] Figure 6 A schematic diagram of training a generative summary module in a third embodiment of the present invention is shown. DETAILED DESCRIPTION
[0042] Since it is impossible to describe all alternatives, the following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the key points of the technical solutions in the embodiments of the present invention. For other technical solutions and details not disclosed in detail below, they generally belong to technical goals or technical features that can be achieved by conventional means in the art. Due to space limitations, the present invention will not introduce them in detail.
[0043] The present invention will describe various key points for combining into various specific embodiments, and these key points will be combined into various methods and products. In the present invention, even if the key points are only described when introducing the method / product solution, it means that the corresponding product / method solution also explicitly includes the technical features.
[0044] When describing the existence or inclusion of a certain step, module, or feature at any position in the present invention, it does not imply that such existence is exclusive and unique. Those skilled in the art can obtain other embodiments based on the technical solution disclosed in the present invention and supplemented by other technical means; based on the key points described in the specific embodiments of the present invention, those skilled in the art can replace, delete, add, combine, and change the order of certain technical features to obtain a technical solution that still follows the concept of the present invention. These solutions that do not deviate from the technical concept of the present invention are also within the protection scope of the present invention.
[0045] The present invention discloses a summary generation method, device and storage medium. The summary generation method mainly includes a step of acquiring a factor word extraction model executed by a factor word extraction model acquisition module, a step of training a summary generation model executed by a generative summary module, and a step of generating a summary. According to time division, it can be divided into a training phase step and a prediction phase step.
[0046] refer to Figure 1 , which shows the steps of obtaining the element word extraction model. The steps of obtaining the element word extraction model include word segmentation and part-of-speech tagging steps, element word extraction steps, splicing steps, training steps and generation steps.
[0047] The word segmentation and part-of-speech tagging steps include: performing word segmentation and part-of-speech tagging on the manual summary. The inventors have studied and believe that extracting four types of element words, namely, time words, place words, object words, and numeral words, is the most suitable four element words for summary generation. Specifically, the part-of-speech tag of the time word is t, the part-of-speech tag of the place word is ns or s, the part-of-speech tag of the object word is nr, nt, nz or n, and the part-of-speech tag of the numeral is m.
[0048] refer to Figure 2, which shows a specific example of tagging a sentence according to the part-of-speech tagging specification. For time element words, in the part-of-speech tagging specification, only the time word t and the time morpheme tg are related to time, and the time morpheme is not a complete word in the strict sense. Therefore, only the words with the part-of-speech of time word t can be selected as the element words of this type.
[0049] For place words, the first thing that comes into consideration is the place name entity in part-of-speech recognition, such as "China" or "Sichuan", but this type of place name entity cannot cover all location information, because not all places are an entity. Therefore, another type of word with part-of-speech is added to the place element words, namely "place words". The part-of-speech label is s, and it represents words that describe the space where the current object is located, rather than a specific place. For example, the common Chinese words "roadside" and "outside the school gate" are very typical place words. Place words can be combined with place name entities to reflect more complete location information.
[0050] For object words, considering that the "person" mentioned in the element is often just an abstract concept, the main participants in many events may not be just people, but objects, organizations or things pointed to by other proper nouns. For example, in the short sentence "The city's auto show is in full swing, and this auto show has invited many brands and manufacturers to participate.", the main participant of the entire fact is obviously the auto show rather than a specific person. Based on the above considerations, we decided to redefine the "person" in the element as "object". Because the main participants of an event can be any object, the part of speech corresponding to the object word is also broader. All words that can express important objects should be included. Therefore, we selected the entity words of the person name, with the part of speech tag nr; the entity words of the organization name, with the part of speech tag nt, representing the organization group, such as "Animal Protection Association" or "Guangdong Development Bank" and other proper nouns, with the part of speech tag nz. Other proper nouns represent names that cannot be used, such as names of people, places and organizations. For example, the names of celestial bodies such as "star" and "neutron star" are generally not used in other contexts. They are words that are specific to celestial bodies and are classified as other proper nouns. In addition to the words of these three parts of speech, common nouns can also become important participants in the entire fact and event. For example, the car show in the above example is a common noun.
[0051] However, the selection of common nouns must be cautious and their number must be limited, because common nouns are not the main component of the abstract in most cases, and words with the common noun part of speech account for a large part of the words in the manual abstract. Therefore, if too many common nouns are selected, it is easy for the extracted element word sequence to become a vassal of the artificial summary. Therefore, for the definition of common nouns as element words, we consider the following: First, if there are two or more names of people, names of organizations and other proper nouns in the artificial summary, it is considered that the common nouns with lower priority no longer need to take on the task of being the participants of the event, because two or more names of people, names of organizations and other proper nouns can already support the main participants in the facts; and if there are only one or less names of people, names of organizations and other proper nouns in the artificial summary, then common nouns are needed to fill in the main participants in the facts, because generally speaking, most summaries will have the most basic subject and object and their relationship as in the triple. Therefore, if there is only one or less names of people, names of organizations and other proper nouns, then common nouns are likely to play the role of the subject or object of the event. Therefore, in this case, we select all the words of the common noun part of speech in the summary, calculate the word frequency of each word in the original text, and then use Text The Rank algorithm calculates the importance score of each word, and then jointly calculates the word frequency and the importance score based on the TextRank algorithm to obtain the total score of each common noun. The calculation method is shown in formula (1).
[0052] Score i =TR(w i )+λ·Freq(w i ) (1)
[0053] In formula (1), Score i represents the total score of the ith common noun; w i represents the ith common noun; TR represents the importance score calculation of the word using the Text Rank algorithm; λ is a fixed parameter set based on experience, and its value is 0.001; Freq represents the frequency of the word, that is, the number of times it appears in the original text.
[0054] By using the method in formula (1), the total importance score of each common noun can be obtained, and then all common nouns in the manual summary can be sorted based on importance based on the score, so as to select the most important common noun. For words of the common noun part of speech, in the preferred embodiment, only the most important one is selected, because there may be a certain gap between the results obtained by the mechanical method and the real important common nouns. The more you select, the more errors may be caused, thereby increasing the complexity of the entire factor word extraction task, which is not conducive to the extraction accuracy of the final factor word extraction model.
[0055] The step of extracting the element words includes: after determining the labels of the element words to be extracted in the part-of-speech tagging specification, selecting the words of the corresponding part-of-speech in the manual summary as the element words to be extracted, and checking whether these words to be extracted have appeared in the original text. If they have appeared in the original text, they will be formally extracted, otherwise they will not be used as element words.
[0056] Figure 3 An example of the prompt-based factor word extraction model of the present invention is shown. By extracting time, place, object and number of times, the corresponding extraction results are obtained.
[0057] In order to train the feature word extraction model, different prompt contexts are designed for each type of feature word. These prompt contexts will be used together with these feature words as the generation targets of the sequence-to-sequence feature word extraction model. Each different prompt context is as follows: Figure 3 For example, if the location key words extracted from the manual summary are "Beijing" and "Shanghai", then by adding a special character "[PART]" to the model vocabulary as a symbol to separate each key word, and using this symbol to splice the key words of the same type into a sequence, then this location key word sequence is "Beijing [PART] Shanghai", and the entire target generation sequence is "The key location of this article is: Beijing [PART] Shanghai".
[0058] The splicing step includes: splicing at least the above-mentioned prompt context and the corresponding four-element words into a target generation sequence; and connecting multiple different element words of the same type through special symbols.
[0059] The training step includes: training the element word extraction model based on the aforementioned artificial summary and the target generation sequence extracted therefrom. For example, the element word extraction model is a UniLM model, and the training step enables the UniLM model to gradually learn how to generate the expected target generation sequence based on the artificial summary. The specific technology for training the UniLM model belongs to the prior art in this field and will not be described here.
[0060] Through training, four different types of feature words can be trained on the same model to obtain a trained feature word extraction model, which is used for feature word extraction.
[0061] The generation step includes: generating four different types of element words through the same element word extraction model. For example, in the generation stage, only the original text and the corresponding prompt preceding text need to be composed into a sentence pair and input into the UniLM model, which is equivalent to using the original text as the input of the encoder and the prompt preceding text as the starting input of the decoder, so that the UniLM model continues to write based on the prompt preceding text, thereby obtaining the element words of the corresponding type in the original text.
[0062] The present invention also proposes three different training schemes for the summary generation model training step performed by the generative summary module.
[0063] The first option: Reference Figure 4 For each original text, the element word extraction module is directly used to extract all four types of element words, and then the four types of element words and the original text are concatenated to form an input data combining element words. The input data combining element words is used to obtain the training input sample of the generative summary module.
[0064] The concatenation of the element words and the original text is also combined with the prompt. After extracting the element words, combine the prompt to concatenate the element words to the head of the original text. For the four different types of element words, add the corresponding prompt in front of them: for the time type element words, add "original text key time:" in front of them, for the location type element words, add "original text key location:" in front of them, for the object type element words, add "original text key object:" in front of them, and for the quantity type element words, add "original text key number:" in front of them. Each type of element word and its prompt are separated by a semicolon and end with a period.
[0065] After the key word sequence, the original text is concatenated, and "the original text is:" is added before the original text to separate the original text from the key word sequence, and also as a hint of the beginning of the original text. Finally, before the artificial summary as the prediction target, "the summary is:" is also added as a hint for the summary generation task.
[0066] Finally, the summary generation model is trained using the training set combined with the key words.
[0067] The second option: Reference Figure 5 ,As a further improvement of the first solution, it can improve the relevance between key words and manual summaries, thereby reducing the noise of training data and improving the final training effect.
[0068] In this scheme, the training samples based entirely on the feature word extraction model are filtered for relevance. The frequency of each feature word extracted by the feature word extraction model in the manual summary is calculated, and finally the frequency of all feature words is summed up and averaged to obtain the approximate relevance scores of the four types of feature words and the manual summary as a whole. The calculation method is shown in formula (2):
[0069]
[0070] In formula (2), S(k) represents the relevance score of all four element word sets k and the manual summary; w irepresents the ith element word in the four element word set k; n is the total number of element words contained in the four element word set k; Count represents the frequency of the element word in the manual summary; ε is a minimum smooth number to prevent division by zero. Adding a logarithm based on the number of element words to the denominator of formula (2) is to further reduce the advantage of the increase in the number of element words in the evaluation score, and also impose a penalty on the number of element words. If more element words are extracted, more element words that are strongly correlated with the manual summary are needed to obtain a higher score.
[0071] Preferably, the threshold of the relevance score is set to 0.25. If the relevance score between the element word set and the manual summary is lower than this threshold, then the sample will be discarded. By observing different relevance thresholds, setting this value to 0.25 can effectively retain samples with high relevance between element words and manual summaries, while not filtering out too much data, resulting in too few samples in the training stage. The training mode of the generative summary model combined with relevance filtering is as follows: Figure 5 As shown in the figure.
[0072] This training mode is completely based on the factor words extracted by the factor word extraction model. Although it has undergone relevance filtering, there are inevitably some differences between the factor words and the manual summary. Although this small difference reduces the correlation between the summary generated by the summary model in the prediction stage and the given factor words, it also improves the robustness of the generative summary model and its denoising and resolution capabilities on factor words. It can select truly important factor words from several factor words as an aid to generate summaries, and is less affected by useless factor words, thereby generating summaries containing more key information.
[0073] The other contents of this embodiment are the same as those of the first embodiment, which are described here by reference and will not be repeated here.
[0074] refer to Figure 6 , which shows the third solution. In order to further improve the correlation between the abstract and the key words generated in the prediction stage, it is proposed to combine the words in the manual abstract and the words extracted by the key word extraction model as the training method of the key words.
[0075] In this mode, we first use the part-of-speech tagging method mentioned above to define and extract element words from the manual summary. At this time, the element words provided by the manual summary must have the highest correlation with the manual summary. This part of the element words may cover four types of element words, or one or more types of element words may be missing. At this time, we can use the element word extraction model to extract the missing type of element words. Finally, we combine these two types of element words from different sources to form a set of four complete types of element words, which are then spliced with the original text as training samples.
[0076] Because the types of element words contained in manual summaries are mostly relatively complete, this training mode that combines mechanical extraction methods and element word extraction models can produce training samples that are more relevant to manual summaries. Therefore, in the prediction stage, we can generate machine summaries that are more relevant to the element words based on the element words given by the element word extraction model, making the entire summary generation task "more like" relying on the assistance of element words.
[0077] The other contents of this embodiment are the same as those of the first embodiment, which are described here by reference and will not be repeated here.
[0078] The above steps complete the training of the feature word extraction model and the summary generation model, which is the training phase step. After that, the prediction phase step is executed.
[0079] The steps in the forecasting phase include:
[0080] 1) Use the key word extraction model to extract four types of key words from the original text.
[0081] 2) The key words are combined with the prompt and the original text, and then input into the summary generation model to obtain the final summary result.
[0082] The steps in the prediction phase are similar to those in the training phase, and are incorporated into the description of this section by reference and will not be repeated here.
[0083] In addition, the present invention also discloses a summary generation device, which includes an element word extraction model acquisition module and a generative summary module, and the module is configured to perform the above steps.
[0084] A storage medium stores computer instructions, and any one of the above-mentioned summary generation methods is executed by loading the computer instructions.
[0085] The entire solution process is completed by training the above-mentioned feature word extraction model, extracting feature words from the training data set of the summary generation model, training the summary generation model with the training set combined with the feature words, and finally generating the final artificial summary based on the trained summary generation model.
[0086] Although the present invention has been described with reference to specific features and embodiments of the present invention, various modifications, combinations, and substitutions may be made thereto without departing from the present invention. The scope of protection of the present invention is not intended to be limited to the specific embodiments of the processes, machines, manufactures, compositions of matter, devices, methods, and steps described in the specification, and these methods and modules may also be implemented in one or more products and methods that are associated, interdependent, mutually coordinated, and preceding / following stages.
[0087] Some of the technical features mentioned in the attached claims may have alternative technical features, or the order of certain technical processes and the order of material organization may be reorganized. After knowing the present invention, ordinary technicians in this field can easily think of these replacement means, or change the order of technical processes and the order of material organization, and then use basically the same means to solve basically the same technical problems and achieve basically the same technical effects. Therefore, even if the above-mentioned means and / or order are clearly defined in the claims, these modifications, changes, and substitutions should fall within the scope of protection of the claims based on the principle of equivalents.
[0088] In conjunction with the various method steps or modules described in the embodiments disclosed herein, they can be implemented in hardware, software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the steps and components of each embodiment have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application or design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered outside the scope of protection claimed by the present invention.
Claims
1. A summary generation method, comprising a training step and a prediction step, characterized in that: The training phase comprises the following steps: a) Using word segmentation and part-of-speech tagging, we extract the key words of four attributes, namely time, place, object and numeral, from the manual summary, and use the prompt as the training target of the key word extraction model to train the key word extraction model; b) extracting key words from the training data set of the summary generation model using the key word extraction model and part-of-speech tagging; c) After extracting the key words, the key words are concatenated to the head of the original text in combination with the prompt to obtain a training set of the key words; d) training the summary generation model using the training set combined with the key words; The prediction stage step includes the following steps: e) Using the trained feature word extraction model to extract feature words of four attributes: time, place, object, and numeral from the original text; f) The extracted key words are combined with the prompt and concatenated with the original text, and then input into the trained summary generation model to obtain the final summary result.
2. The method for generating a summary according to claim 1, characterized in that: The element word extraction module is directly used to extract the element words of the four attributes, and then the element words of the four attributes and the original text are concatenated to form a training input data combined with the element words.
3. The method for generating a summary according to claim 1, characterized in that: The frequency of each element word extracted by the element word extraction model in the manual summary is calculated, and finally the frequency of all element words is summed up and averaged to obtain the overall relevance score of the element words of the four attributes and the manual summary; the calculation formula of the relevance score is as follows: Among them, S(k) represents the relevance score between the element word set k and the manual summary in all four attributes; w i represents the i-th element word in the four element word set k; n is the total number of element words contained in the element word set k of the four attributes; Count represents the frequency of the element word appearing in the manual summary; ε is a very small smooth number to prevent division by zero.
4. The method for generating a summary according to claim 3, characterized in that: If the relevance score of the feature word set to the manual summary is lower than the threshold of 0.25, the sample will be discarded.
5. The method for generating a summary according to claim 1, characterized in that: After extracting the essential words in step b), if one of the essential words of the four attributes of time, place, object and numeral is missing, the essential word extraction model is used in the original text to extract the missing type of essential words. Finally, the essential words from these two different sources are combined to form a set of four complete types of essential words, which are then spliced with the original text as training samples.
6. The method for generating a summary according to claim 1, characterized in that: The element word extraction model is the UniLM model.
7. A summary generation device, comprising an element word extraction model acquisition module and a generative summary module, and the two modules are configured to execute the summary generation method according to any one of claims 1 to 6.
8. A storage medium storing computer instructions, wherein the computer instructions are loaded to execute the method for generating a summary according to any one of claims 1 to 6.