Key information extraction and live content processing method, device and equipment

Abstract

Embodiments of the present specification disclose a key information extraction and live content processing method, device and equipment. The scheme comprises: obtaining original text to be extracted key information; determining an attention weight vector for the original text, and correspondingly generating a first target word according to the attention weight vector; judging whether the first target word is included in the original text; if yes, obtaining the attention weight of the first target word when generating the first target word according to the attention weight vector, generating penalty data for the first target word according to the attention weight of the first target word, and continuing to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word; and generating key information according to the first target word and the second target word.

Description

TECHNICAL FIELD

[0001] The present specification relates to the technical field of machine learning, and particularly relates to a key information extraction and live content processing method, device and equipment. BACKGROUND

[0002] With the development of Internet technology, the amount of information faced by people in life is growing explosively.

[0003] At present, in order to facilitate quick browsing of information, information providers usually extract key information from provided information content, and summarize information content with large data volume in the form of a title, a brief introduction and other key information, so that users can quickly understand various information and save browsing time. In order to save manpower, key information can be automatically extracted through semantic analysis for some information content, but in actual application, redundant information often appears in the extracted key information, affecting user experience.

[0004] Therefore, a more accurate key information extraction scheme is needed. SUMMARY

[0005] One or more embodiments of the present specification provide a key information extraction and live content processing method, device, equipment and storage medium, to solve the technical problem that redundant information often appears in the extracted key information, affecting user experience, and a more accurate key information extraction scheme is needed.

[0006] To solve the above technical problems, one or more embodiments of the present specification are implemented as follows:

[0007] One or more embodiments of the present specification provide a key information extraction method, comprising:

[0008] obtaining original text of key information to be extracted;

[0009] determining an attention weight vector for the original text, and correspondingly generating a first target word according to the attention weight vector;

[0010] judging whether the first target word is included in the original text;

[0011] If yes, obtaining an attention weight of the first target word when the first target word is generated according to the attention weight vector, generating penalty data for the first target word according to the attention weight of the first target word, and continuing to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word is the same as the first target word;

[0012] generate the key information according to the first target word and the second target word.

[0013] One or more embodiments of the present specification provide a live content processing method based on a coverage mechanism, comprising:

[0014] obtaining a live video of an anchor on a live platform;

[0015] performing speech recognition and slicing processing on the live video to obtain a plurality of live segment texts;

[0016] determining an attention weight vector for the live segment texts, and correspondingly generating a first target word according to the attention weight vector;

[0017] judging whether the first target word is included in the live segment texts;

[0018] If yes, generating penalty data for the first target word to continue generating a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word;

[0019] generating key information of the live segment texts according to the first target word and the second target word as a segment title of a live video segment corresponding to the live segment texts;

[0020] publishing the live video segments and the segment titles generated therefor in combination so as to allow a user to select and play back.

[0021] One or more embodiments of the present specification provide a key information extraction device, comprising:

[0022] A first obtaining module obtains an original text for which key information is to be extracted;

[0023] A target word generation module determines an attention weight vector for the original text, and correspondingly generates a first target word according to the attention weight vector;

[0024] A first judging module judges whether the first target word is included in the original text;

[0025] A first penalty module, if yes, according to the attention weight vector, obtains an attention weight of the first target word when the first target word is generated, generates penalty data for the first target word according to the attention weight of the first target word, and continues to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word;

[0026] The first key information generation module generates the key information according to the first target word and the second target word.

[0027] One or more embodiments of the present specification provide a live content processing device based on a coverage mechanism, comprising:

[0028] The second acquisition module acquires a live video of an anchor on a live platform.

[0029] The processing module performs speech recognition and slicing processing on the live video to obtain a plurality of live segment texts.

[0030] The live target word generation module determines an attention weight vector for the live segment text, and correspondingly generates a first target word according to the attention weight vector.

[0031] The second judgment module judges whether the first target word is included in the live segment text.

[0032] The second penalty module, if so, generates penalty data for the first target word, so as to continue to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word continues to generate is the same as the first target word.

[0033] The second key information generation module generates key information of the live segment text according to the first target word and the second target word, as a segment title of a live video segment corresponding to the live segment text.

[0034] The publishing module publishes each live video segment and the segment title generated therefor in combination, so as to allow a user to select and play back.

[0035] One or more embodiments of the present specification provide a key information extraction device based on a coverage mechanism, comprising:

[0036] At least one processor; and

[0037] The memory is in communication connection with the at least one processor; wherein

[0038] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to:

[0039] Acquire an original text to be extracted key information;

[0040] Determine an attention weight vector for the original text, and correspondingly generate a first target word according to the attention weight vector.

[0041] determine whether the first target word is included in the original text;

[0042] If yes, according to the attention weight vector, an attention weight of the first target word in generating the first target word is obtained, a penalty data for the first target word is generated according to the attention weight of the first target word, and a second target word is generated according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated is the same as the first target word.

[0043] According to the first target word and the second target word, the key information is generated.

[0044] One or more embodiments of the present specification provide a live content processing device based on a coverage mechanism, comprising:

[0045] at least one processor; and,

[0046] a memory in communication connection with the at least one processor; wherein,

[0047] the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to:

[0048] obtain a live video of a host on a live platform;

[0049] perform speech recognition and slicing processing on the live video to obtain a plurality of live segment texts;

[0050] determine an attention weight vector for the live segment text, and correspondingly generate a first target word according to the attention weight vector;

[0051] determine whether the first target word is included in the live segment text;

[0052] If yes, a penalty data for the first target word is generated, and a second target word is generated according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated is the same as the first target word.

[0053] According to the first target word and the second target word, the key information of the live segment text is generated as a segment title of a live video segment corresponding to the live segment text;

[0054] The live video segments and the segment titles generated therefor are combined and published so as to be played back by a user.

[0055] The one or more embodiments of the specification provide a non-volatile computer storage medium storing computer executable instructions configured to:

[0056] Obtain original text to be extracted key information;

[0057] Determine an attention weight vector for the original text, and correspondingly generate a first target word according to the attention weight vector;

[0058] Determine whether the first target word is included in the original text;

[0059] If yes, obtain the attention weight of the first target word when generating the first target word according to the attention weight vector, generate penalty data for the first target word according to the attention weight of the first target word, and continue to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word;

[0060] Generate the key information according to the first target word and the second target word.

[0061] The one or more embodiments of the specification provide a non-volatile computer storage medium storing computer executable instructions configured to:

[0062] Obtain live video of a host on a live broadcast platform;

[0063] Perform speech recognition and slicing processing on the live video to obtain a plurality of live segment texts;

[0064] Determine an attention weight vector for the live segment text, and correspondingly generate a first target word according to the attention weight vector;

[0065] Determine whether the first target word is included in the live segment text;

[0066] If yes, generate penalty data for the first target word, and continue to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word;

[0067] Generate key information of the live segment text according to the first target word and the second target word, as a segment title of a live video segment corresponding to the live segment text;

[0068] Combine each live video segment and the segment title generated therefor, and publish the combination so as to allow a user to play back a selected segment.

[0069] The above at least one technical solution adopted by one or more embodiments of the present specification can achieve the following beneficial effects: focusing on the target word generated by the real generation, by judging the relationship between the generated target word and the original text, generating penalty data for the target word already included in the original text, realizing accurate punishment, reducing the attention of the prediction model to the target word, reducing the possibility of repeated appearance of the target word, and solving the problem of repeated content of key information. At the same time, the target word not included in the original text is not punished, which avoids the problem of mistakenly reducing the appearance probability of the target word that should be generated caused by possible excessive punishment of all words, causing the omission of key information, so as to improve the accuracy of key information extraction and increase the effective information in the key information while removing redundancy. BRIEF DESCRIPTION OF DRAWINGS

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present specification or the prior art, the drawings needed to be used in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present specification, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0071] Figure 1 A flowchart of a key information extraction method provided by one or more embodiments of the present specification;

[0072] Figure 2 A detailed flowchart of a method under an application scenario provided by one or more embodiments of the present specification; Figure 1

[0073] Figure 3 A structure diagram of a key information extraction device provided by one or more embodiments of the present specification;

[0074] Figure 4 A structure diagram of a key information extraction device provided by one or more embodiments of the present specification;

[0075] Figure 5 A flowchart of a live content processing method provided by one or more embodiments of the present specification;

[0076] Figure 6 A structure diagram of a live content processing device provided by one or more embodiments of the present specification;

[0077] Figure 7 A structure diagram of a live content processing device provided by one or more embodiments of the present specification. DETAILED DESCRIPTION ​

[0078] The embodiment of the present specification provides a key information extraction and live content processing method, device and equipment, and a storage medium.

[0079] In order to enable personnel in the technical field to better understand the technical solutions in the present specification, the technical solutions in the present specification will be described clearly and completely in the following with reference to the drawings in the embodiment of the present specification. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments of the present specification, all other embodiments obtained by those of ordinary skill in the art without creative labor should belong to the scope of protection of the present application.

[0080] In one or more embodiments of the present specification, a coverage mechanism is proposed to solve the problems in the background art. This mechanism generates a copy probability for each target word, and punishes the words that have been paid attention to before, so as to avoid these words from appearing frequently in the generated results, thereby alleviating the problem of redundancy caused by the repetition of the generated results. However, this coverage mechanism still has problems. It does not distinguish whether the word that is currently punished has really appeared in the generated results (in actual applications, the words in the generated results are often included in the original text) or not. It only punishes all the words based on the historical cumulative copy probability. This will punish the copy probability of the word that has a relatively high copy probability before the current time, but has not actually appeared in the generated results, to a very low value at the current time. However, this word is very likely to be copied at the current time, which will lead to the omission of this word, resulting in the lack of content in the generated results. On the short title generation task, the problem caused by this defect is particularly serious, because the number of words in the short title is small. The opportunity of the word that should appear but is punished to appear is even smaller. In order to make up for this defect, under the premise of reducing redundancy, alleviate the mis-punishment of other words, and further avoid the lack of key information in the generated results, the above scheme is further improved to realize accurate punishment.

[0081] In one or more embodiments of the present specification, in the process of generating each target word according to the attention weight vector to further constitute the key information, it is determined whether the target word generated each time is included in the original text. If yes, the target word is punished according to the attention weight of the target word to reduce the possibility of generating the target word again. In this way, the repetition in the process of extracting key information can be avoided, and only the actually generated target word is punished, which is beneficial to prevent missing the target word that should be generated and improve the accuracy of key information extraction. Moreover, after the punishment, the punishment effect is dynamically focused on, and the subsequent punishment strength is more accurately adjusted according to the punishment effect, so that more accurate punishment is realized, which will be described in detail later. The scheme can be applied to various application scenarios, such as live broadcast publishing field (such as generating a title for live broadcast content), service integration field (such as regenerating an overall description for integrated multiple different services, etc.), and the like. Taking the live broadcast publishing field as an example, the present specification also processes the live broadcast video based on the above scheme to determine the title corresponding to the live broadcast video segment, so that the user can more accurately and efficiently review the segment of interest without browsing the entire live broadcast video. The following will be described in detail based on this idea.

[0082] Figure 1 A flowchart of a key information extraction method provided by one or more embodiments of the present specification is shown. The method can be applied to different business fields, such as Internet financial business field, e-commerce business field, instant messaging business field, game business field, public service business field, etc. The flowchart can be executed by a computing device in the corresponding field, and some input parameters or intermediate results in the flowchart allow manual intervention to adjust to help improve accuracy.

[0083] Figure 1 The flowchart in the present specification can include the following steps:

[0084] S102: Obtain an original text whose key information needs to be extracted.

[0085] In one or more embodiments of the present specification, the original text represents the content that needs to be summarized and key information extracted, which can be complete content such as an article, or partial content such as a section of an article. The original text is in the form of text, but the present scheme is also applicable to key information extraction of non-text content such as video and audio. For non-text content, it needs to be converted into text form through voice recognition or the like to obtain the corresponding original text, and then key information extraction is performed.

[0086] Specifically, when processing the live video of the anchor on the live broadcast platform, the live video can be subjected to speech recognition and slicing processing to obtain a plurality of live video segment texts as original texts for which key information is to be extracted, and then the extracted key information is taken as a segment title corresponding to the live video segment text.

[0087] The key information can be a relatively short phrase or group of words, such as a title, or a relatively long sentence, such as a brief introduction. The key information is composed of a plurality of target words, which can exist in the original text, i.e., directly copied from the original text, or can not exist in the original text, i.e., summarized from the content of the original text.

[0088] In the process of extracting the key information, the content existing in the original text can be directly copied into the key information through a copy mechanism. For each word in the original text, a copy probability can be used to represent the possibility of being directly copied into the key information, and the copy probability and the possibility are in a positive correlation.

[0089] S104: determining an attention weight vector for the original text, and correspondingly generating a first target word according to the attention weight vector.

[0090] According to the attention mechanism, the content of the original text is analyzed to determine an attention weight vector for the original text. The attention weight vector is used to represent the importance of each word in the original text through an attention weight, and the attention weight and the importance are in a positive correlation. The higher the attention weight, the stronger the importance, and the lower the attention weight, the weaker the importance.

[0091] Through the attention weight vector, the more important information in the original text can be mainly focused on to correspondingly generate the first target word. The key information to be extracted usually contains a plurality of words, which can be extracted one by one in the extraction process. Each extracted word is referred to as a target word of this time. In order to facilitate the description, in some embodiments below, one target word is extracted each time. Under the attention mechanism, the extraction of the target word each time is affected by the previously extracted target words, and therefore the attention weight vector each time also changes.

[0092] The first target word represents a target word extracted according to the original text at a certain time (which can be any time). It is generally related to the content represented by a higher attention weight in the attention weight vector. The target word extracted at a time subsequent to the certain time is referred to as a second target word. It should be noted that the second target word can be the same as the first target word, which can cause redundancy in the key information. The present solution is committed to eliminating such redundancy as much as possible or keeping it within a reasonable range while reducing the processing burden as much as possible.

[0093] S106: Determine whether the first target word is included in the original text.

[0094] The first target word can be directly copied from the original text, or can be summarized according to the original text. In actual application, the copied word is more likely to be repeatedly generated. If the first target word is included in the original text, the word is punished in the subsequent generation process to avoid repeated attention to the word.

[0095] S108: If yes, according to the attention weight vector, the attention weight of the first target word in generating the first target word is obtained, the penalty data for the first target word is generated according to the attention weight of the first target word, and the second target word is generated according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated is the same as the first target word.

[0096] In one or more embodiments of the present specification, even if the penalty mechanism is not introduced, in order to extract key information, it is still necessary to determine the attention weight vector of the original text, which reflects the attention weight of each word in the original text. Therefore, under the idea of "punishing the word according to its attention weight", the attention weight which needs to be determined can be used at the same time, so as to not only expand the use of such attention weight, but also facilitate the implementation and reduce the implementation cost of the scheme.

[0097] Since such attention weight belongs to the word in the original text, the punishment can be mainly performed on the word included in the original text. Further, among the words in the original text, the word that has already contributed (has appeared in the stage generation result and has contributed corresponding semantic information) is punished to reduce the possibility of repeated contribution (generated again next time) of the word, so as to give more opportunities to other words that have not yet contributed, which is more reasonable and helps to make the generation result balance accuracy, completeness and conciseness. This paragraph and the previous paragraph also explain why step S106 is performed, and why steps S106 and S108 are cooperated in this way.

[0098] In the process of generating the target word in stages, since part of the key information has been extracted, in order to select other key information, the attention weight vector corresponding to the original text is updated, and the next target word is generated according to the updated attention weight vector.

[0099] In one or more embodiments of the present specification, if it is determined that the first target word does not exist in the original text, no punishment is needed, and the updated attention weight vector corresponding to the second target word to be generated can be obtained without generating the penalty data for the first target word, so as to continue generating the second target word.

[0100] S110: generating the key information according to the first target word and the second target word.

[0101] By generating each target word in sequence, the key information extracted from the original text can be determined. In a simple way, each target word is connected in sequence to form the key information, and of course the key information can be further purified based on each target word.

[0102] According to different specific punishment methods, the above scheme can be applied to the training stage of the prediction model, or can be applied to the application stage of the prediction model. In some embodiments below, a loss-based punishment method is proposed, which is relatively more suitable for the training stage, and the punishment effect is integrated into the calculation of the loss and the correction of the model parameters in the training process.

[0103] In one or more embodiments of the present specification, by judging the relationship between the generated target word and the original text, penalty data is generated for the target word that already exists in the original text to reduce the attention of the prediction model to the target word and reduce the possibility of repeated appearance of the target word, thereby solving the problem of repeated content in the key information. At the same time, the target word that does not exist in the original text is not punished, so as to avoid reducing the probability of occurrence of the target word that should be generated, causing omission of the key information, so that more accurate key information can be obtained.

[0104] Based on the method of Figure 1 The present specification also provides some specific embodiments and extension schemes of the method, which will be described below.

[0105] In one or more embodiments of the present specification, when generating the penalty data for the first target word, at least one dimension related to the prediction model (for generating the target word) can be determined as a punishment dimension, such as at least one of the attention weight, the loss function, and the duplication probability.

[0106] According to the punishment dimension and the attention weight of the first target word, the penalty data for the first target word is generated to reduce the possibility of duplicating the first target word as the second target word. The punishment dimension includes, for example, the attention weight dimension, the loss function dimension of the prediction model (for generating the target word), the duplication probability dimension, etc.

[0107] If the attention weight dimension is selected, after the attention weight corresponding to the first target word is determined, the attention weight can be directly reduced (the more it is reduced, the greater the punishment degree), and the punishment data for the first target word is generated accordingly. In the subsequent process of generating target words, the prediction model will directly reduce the attention to the first target word. It should be noted that the relationship between the punishment degree and the attention weight is not necessarily a simple positive correlation, but a more complex relationship, especially when multiple dimensions are combined for punishment. In the later examples of loss functions, it can be seen that this more complex relationship helps to more accurately and flexibly punish.

[0108] If the loss function dimension of the prediction model is selected, the generation and use of the punishment data can be performed in the training stage of the prediction model used to extract the key information. Since the loss function of the prediction model represents the error (i.e., loss) that the prediction model may produce when selecting a certain word as a target word, and the prediction model is trained to make the loss as small as possible so that it can be more accurate in prediction, when the first target word is generated, the loss corresponding to the current first target word is expected to be as small as possible. Therefore, in order to make the prediction model reduce the attention to the first target word in the next generation of target words, an additional loss (the greater the additional loss, the greater the punishment degree) can be introduced for the first target word as punishment to make the prediction model correct towards the direction of not repeating the generation of the first target word.

[0109] If the copy probability dimension is selected, when the target word is generated again, intervention can be performed to reduce the copy probability of the first target word, so as to avoid copying the first target word from the original text as the generated target word.

[0110] The following embodiments mainly utilize the attention weight and loss function dimensions, and the attention dimension is embodied in the loss function.

[0111] Based on such an idea, in addition to the conventional loss term in the loss function, a penalty loss term related to the attention weight of the generated target word can be defined in the loss function to represent the punishment of the target word. When it is determined that the target word needs to be punished, the penalty loss term defined in the loss function of the prediction model is obtained. According to the attention weight of the first target word and the penalty loss term, the penalty loss is generated, that is, the penalty loss term is valued, the loss corresponding to the first target word is increased, as the punishment data for the first target word, to train the prediction model to reduce the attention to the first target word later.

[0112] In this way, the first target word is punished, so that the prediction model can naturally reduce the attention to the generated target word in the application process, which is conducive to more scientifically and reasonably determining the key information corresponding to the original text and improving the accuracy of the extracted key information.

[0113] Further, in the process of extracting key information from the original text, the first target word may have been generated several times before the current generation of the first target word (assuming that the first target word is a word in the original text). This may indicate that the previous punishment is not enough, and the punishment can be increased accordingly by considering the accumulation of the punishment. Based on this, in this case, the historical attention weights of several times of generation of the first target word can be obtained. According to the attention weight of the first target word in this time and the historical attention weights, the accumulated attention weight is generated. According to the accumulated attention weight and the punishment loss term, the punishment loss is generated.

[0114] In this way, the case of generating the first target word several times is comprehensively considered, and the first target word generated this time is punished, which can make the punishment of the first target word more reasonable and help to enhance the accuracy of key information extraction.

[0115] Further, after the current generation of the first target word, the generation phase of the next target word is entered, and the punishment of the first target word is performed in this phase (that is, if the current occurrence decides to punish, the next time the punishment actually takes effect, similarly, if the last occurrence decides to punish, the current time the punishment actually takes effect, and so on, each new resulting punishment actually takes effect in the next time and may continue to affect more times afterwards), and the second target word is generated. Figure 1 For example, assuming that it is judged that the first target word already exists in the original text, a new punishment will be caused, which actually takes effect in the next time, which is specifically embodied in the process of generating the second target word, introducing a new loss, so that the probability of the prediction model generating the first target word again is reduced.

[0116] Further, specific implementations of accurately controlling the punishment strength based on the loss are also provided, including: the accumulated attention weight corresponding to the first target word and the current attention weight can be compared, and the smaller one of the two is used to determine the punishment loss again, which is used to generate the second target word.

[0117] The principle is as follows: in the second target word generation stage, the attention weight of the first target word is determined as the current attention weight, and the first target word has been punished in the case of multiple generations of the first target word. If the attention weight of the current first target word is small, it indicates that the punishment of the first target word has taken effect, and the attention of the first target word has been reduced in the prediction model, in which case the punishment intensity can be reduced, for example, when determining the punishment loss of the first target word, it is more reasonable to determine according to the smaller current attention weight (at this time, the cumulative attention weight may be large), which helps to prevent over-punishment and is conducive to the training and rapid convergence of the prediction model. If the attention weight of the current first target word is still large after multiple punishments, it is assumed to be greater than the cumulative attention weight, which indicates that the punishment effect may not be good and the punishment intensity needs to be continuously accumulated and increased, in which case the attention weight of the first target word is likely to be unreliable, and the cumulative attention weight can be considered to determine the punishment loss instead of the attention weight of the first target word, in order to prevent the risk of introducing a large deviation of the punishment loss; of course, in some cases, it is also possible to indicate that the first target word is indeed important enough, in which case the cumulative attention weight may continue to increase, but the attention weight is limited (when the cumulative attention weight is greater than the attention weight, the attention weight is selected), so that the punishment loss does not increase too much.

[0118] In this way, according to the size relationship change between the current attention weight and the cumulative attention weight, the dynamic expansion and contraction of the punishment intensity is realized more reasonably, making the punishment more accurate and the training process more efficient.

[0119] Specifically, according to the cumulative attention weight and the current attention weight of the first target word, the principle of determining the punishment loss can be represented by formula one to formula three, which are described below with the first target word as an example.

[0120]

[0121] In formula one, word represents the first target word generated in the t-th prediction (for ease of description, the t-th prediction is also referred to as time t, and t is assumed to start from 0), and represents the attention weight of the first target word in the t-th prediction. represents the attention weight of the first target word in the t-th prediction, and p represents the generation probability (such as based on the copying probability or further combined with other probability dimensions to determine), i==argmax(p t ) represents the generation probability when the i-th prediction is the first target word, i.e., the first target word is generated. This formula indicates that when the first target word is generated, I takes 1, otherwise, I takes 0.

[0122]

[0123] The parameter explanations of Formula Two refer to those of Formula One. In Formula Two, if t is the current prediction, then represents the current attention weight of the first target word, respectively represent the historical attention weights of the first target word, and the accumulated cov t represents the accumulated attention weight of the first target word.

[0124]

[0125] The parameter explanations of Formula Three refer to those of Formula One and Formula Two. In Formula Three, loss represents the loss function, w represents the first target word, -log(p t (w)) represents the conventional loss term of the loss function, represents the penalty loss term, that is, the smaller value between the accumulated attention weight and the current attention weight of the first target word is selected as the penalty loss, and λ represents a constant.

[0126] The above takes time 0-t-1 as the history, which is before the current time, and takes time t as the current time. Similarly, at time t+1 (next time), the history becomes time 0-t, and the above-mentioned formulas can be used after adding 1 to the corresponding time. In some embodiments above, it is understood that the first target word is generated at time t, and the second target word is generated at time t+1.

[0127] In one or more embodiments of the present specification, the extracted key information can be used as a short title of the original text, and the specific limit is that the number of short titles can be set according to actual conditions, for example, set to within 10 words or within 5 words, etc. Since the number of words in the short title is usually short, it is necessary to accurately condense the key information of the original text, and therefore the accuracy of key information extraction is more important. The present scheme can effectively reduce the repetition, omission, etc. in the extraction of key information, and increase the accuracy of key information extraction.

[0128] In combination with the foregoing description, one or more embodiments of the present specification provide a kind of application scene, Figure 1 A detailed flowchart of the method in the embodiment is shown in the following figure. Figure 2 As shown in the figure.

[0129] In the flowchart of Figure 2 , for the target word w generated at time t, it is judged whether the target word w exists in the original text. If not, the target word w is not punished. If it exists, the attention weight corresponding to the target word w is determined according to the attention weight of each word in the original text calculated in advance. The target word w is punished according to the attention weight corresponding to the target word w, and the punishment is effective at time t+1.

[0130] Based on the same idea, one or more embodiments of the present specification also provide a device and equipment corresponding to the above method, as shown in Figure 3 、 Figure 4

[0131] Figure 3 A structural schematic diagram of a key information extraction device based on a coverage mechanism provided by one or more embodiments of the present specification, the device comprising:

[0132] The first acquisition module 302 acquires the original text to be extracted as key information.

[0133] The target word generation module 304 determines the attention weight vector for the original text, and correspondingly generates the first target word according to the attention weight vector.

[0134] The first judgment module 306 judges whether the first target word is included in the original text.

[0135] The first penalty module 308, if yes, according to the attention weight vector, acquires the attention weight of the first target word when generating the first target word, generates penalty data for the first target word according to the attention weight of the first target word, and continues to generate the second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word.

[0136] The first key information generation module 310 generates the key information according to the first target word and the second target word.

[0137] Optionally, the first penalty module 308 determines at least one dimension related to the prediction model used to extract the key information as a penalty dimension, the penalty dimension including at least one of attention weight, loss function, and copying probability; and generates the penalty data for the first target word according to the penalty dimension and the attention weight of the first target word.

[0138] Optionally, the generation of the penalty data for the first target word according to the attention weight of the first target word is performed in the training stage of the prediction model used to extract the key information, and the loss function of the prediction model includes a penalty loss term related to the attention weight of the generated target word; and the first penalty module 308 generates a penalty loss as the penalty data for the first target word according to the attention weight of the first target word and the penalty loss term, for training the prediction model.

[0139] ​Optionally, the first penalty module 308, before generating the first target word this time, if the first target word has been generated for the original text for several times, obtains historical attention weight when the first target word is generated for the several times; generates accumulated attention weight according to the attention weight for the first target word and the historical attention weight; generates a penalty loss according to the accumulated attention weight and the penalty loss term.

[0140] Optionally, the first penalty module 308, enters the next target word generation stage after this time; in the next target word generation stage, determines the attention weight of the first target word as the current attention weight; compares the accumulated attention weight and the current attention weight, and re-determines the penalty loss according to the smaller one of the two, for generating a second target word.

[0141] Optionally, the apparatus further comprises a non-penalty module 312, if the result of the judgment is no, obtains an attention weight vector corresponding to a second target word to be generated without generating penalty data for the first target word, to continue generating the second target word.

[0142] Optionally, the first obtaining module 302 obtains a live video of an anchor on a live platform; performs speech recognition and slicing processing on the live video to obtain a plurality of live segment texts as original texts for extracting key information, and the key information extracted is used as a segment title corresponding to the live segment text.

[0143] Optionally, the key information is a short title of the original text.

[0144] Figure 4 A structural schematic diagram of a key information extraction device based on a coverage mechanism is provided for one or more embodiments of the present specification, and the device comprises:

[0145] at least one processor; and,

[0146] a memory in communication connection with the at least one processor; wherein,

[0147] the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to:

[0148] obtain an original text for extracting key information;

[0149] determine an attention weight vector for the original text, and correspondingly generate a first target word according to the attention weight vector;

[0150] determine whether the first target word is included in the original text;

[0151] If yes, according to the attention weight vector, obtain the attention weight of the first target word when the first target word is generated, generate penalty data for the first target word according to the attention weight of the first target word, and continue to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated is the same as the first target word;

[0152] generate the key information according to the first target word and the second target word.

[0153] Based on the same idea, one or more embodiments of the present specification also provide a non-volatile computer storage medium corresponding to the above method, which stores computer executable instructions, and the computer executable instructions are arranged to:

[0154] obtain the original text to be extracted key information;

[0155] determine the attention weight vector for the original text, and correspondingly generate a first target word according to the attention weight vector;

[0156] determine whether the first target word is included in the original text;

[0157] If yes, according to the attention weight vector, obtain the attention weight of the first target word when the first target word is generated, generate penalty data for the first target word according to the attention weight of the first target word, and continue to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated is the same as the first target word;

[0158] generate the key information according to the first target word and the second target word.

[0159] Figure 5 A flowchart of a live content processing method based on a coverage mechanism is provided for one or more embodiments of the present specification. The method can be applied to different business fields, such as the Internet financial business field, the e-commerce business field, the instant messaging business field, the game business field, the public service business field, etc. The flowchart can be executed by a computing device in the corresponding field, and some input parameters or intermediate results in the flowchart allow manual intervention to adjust to help improve accuracy.

[0160] Figure 5 The flowchart in the above embodiment can include the following steps:

[0161] S502: Obtain a live video of an anchor on a live platform.

[0162] S504: Perform speech recognition and slicing processing on the live video to obtain a plurality of live segment texts.

[0163] For a live video on a live platform, in order to facilitate user playback and viewing, the live video can be sliced and processed into a plurality of small video segments, and speech recognition is performed to obtain live segment texts (i.e., original texts in Figure 1 ).

[0164] S506: Determine an attention weight vector for the live segment text, and correspondingly generate a first target word according to the attention weight vector.

[0165] Through a prediction model, an attention weight vector of a live segment text is determined, and a first target word is generated according to the attention weight vector as part of the key information.

[0166] S508: Determine whether the first target word is included in the live segment text.

[0167] S510: If yes, generate penalty data for the first target word to continue generating a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word.

[0168] Determine whether the first target word is directly copied from the live segment text, and if so, penalize the first target word to reduce the possibility that the prediction model generates the first target text again, and continue to generate a second target word.

[0169] S512: Generate key information of the live segment text according to the first target word and the second target word as a segment title of a live video segment corresponding to the live segment text.

[0170] S514: Combine each live video segment and the segment title generated therefor to publish, so as to facilitate user selection and playback.

[0171] According to the obtained target word, the key information is generated as a segment title of a live video segment, which facilitates the user to understand the content of the corresponding live video segment directly through the segment title, so as to conveniently and quickly playback and view the live video segment of interest.

[0172] In one or more embodiments of the present specification, according to whether the generated target word exists in the live segment text, penalty data is generated for the target word already existing in the live segment text to reduce the attention of the prediction model to the target word, reduce the possibility of repeated appearance of the target word, solve the problem of repeated content of key information, improve the accuracy of the obtained title, facilitate service to users, and improve user experience.

[0173] It should be noted that, Figure 5 The scheme shown and Figure 1 The scheme shown is essentially similar, Figure 5 The parts not described in detail in the steps can refer to the related description of Figure 1 This specification will not be repeated here.

[0174] Figure 1 In addition to being applied to the live content processing described above, the scheme shown can also be applied to information processing in various scenarios, such as integrating information of key information of different sources and different formats, and unifying the key information.

[0175] Based on the same idea, one or more embodiments of the present specification also provide a device and equipment corresponding to the above method, as shown in Figure 6 、 Figure 7

[0176] Figure 6 A structure diagram of a live content processing device based on a coverage mechanism provided by one or more embodiments of the present specification is shown. The device includes:

[0177] The second acquisition module 602 acquires the live video of the host on the live platform.

[0178] The processing module 604 performs speech recognition and slicing processing on the live video to obtain a plurality of live segment texts.

[0179] The live target word generation module 606 determines an attention weight vector for the live segment text, and correspondingly generates a first target word according to the attention weight vector.

[0180] The second judgment module 608 judges whether the first target word is included in the live segment text.

[0181] The second penalty module 610 generates penalty data for the first target word if the first target word is included in the live segment text, so as to continue to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word.

[0182] ​The second key information generation module 612 generates key information of the live broadcast clip text according to the first target word and the second target word, as a clip title of a live broadcast video clip corresponding to the live broadcast clip text.

[0183] The publishing module 614 publishes the live broadcast video clips and the clip titles generated for the live broadcast video clips in combination, so as to enable a user to play back selected clips.

[0184] Figure 7 A structural schematic diagram of a live broadcast content processing device based on a coverage mechanism is provided for one or more embodiments of the present specification, and the device comprises:

[0185] at least one processor; and

[0186] a memory in communication connection with the at least one processor; wherein

[0187] the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to:

[0188] obtain a live broadcast video of a host on a live broadcast platform;

[0189] perform speech recognition and slicing processing on the live broadcast video to obtain a plurality of live broadcast clip texts;

[0190] determine an attention weight vector for the live broadcast clip texts, and correspondingly generate a first target word according to the attention weight vector;

[0191] determine whether the first target word is included in the live broadcast clip text;

[0192] if yes, generate penalty data for the first target word, to continue to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word;

[0193] generate key information of the live broadcast clip text according to the first target word and the second target word, as a clip title of a live broadcast video clip corresponding to the live broadcast clip text;

[0194] publish the live broadcast video clips and the clip titles generated for the live broadcast video clips in combination, so as to enable a user to play back selected clips.

[0195] Based on the same idea, one or more embodiments of the present specification also provide a non-volatile computer storage medium corresponding to the above method, which stores computer executable instructions, and the computer executable instructions are set to:

[0196] acquire a live video of an anchor on a live broadcast platform;

[0197] perform voice recognition and slicing processing on the live video to obtain a plurality of live segment texts;

[0198] determine an attention weight vector for the live segment texts, and correspondingly generate a first target word according to the attention weight vector;

[0199] determine whether the first target word is included in the live segment texts;

[0200] if yes, generate penalty data for the first target word, so as to continue to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word continues to be generated is the same as the first target word;

[0201] generate key information of the live segment texts as a segment title of a live video segment corresponding to the live segment texts according to the first target word and the second target word;

[0202] publish the live video segments and the segment titles generated therefor in combination, so as to allow a user to play back a selected segment.

[0203] In the 1990s, it was quite obvious to distinguish whether an improvement in a technology was in hardware (e.g., improvement in circuit structures of diodes, transistors, switches, etc.) or in software (improvement in method flow). However, as technology has evolved, many improvements in method flow today can be considered as direct improvements in hardware circuit structures. Designers almost always obtain the corresponding hardware circuit structures by programming the improved method flow into hardware circuits. Therefore, it cannot be said that an improvement in a method flow cannot be implemented by hardware entity modules. For example, a programmable logic device (PLD) (e.g., a field programmable gate array (FPGA)) is an integrated circuit whose logic function is determined by user programming of the device. A digital system is "integrated" on a PLD by the designer programming it, rather than by asking a chip manufacturer to design and fabricate a custom integrated circuit chip. Moreover, instead of manually fabricating integrated circuit chips, this programming is now mostly implemented by "logic compiler" software, which is similar to software compilers used in program development, and the original code to be compiled is written in a specific programming language, which is called a hardware description language (HDL), and there are many such languages, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc., and the most commonly used are VHDL (Very-High-Speed Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should be aware that, as long as the method flow is logically programmed in the above-mentioned hardware description languages and programmed into an integrated circuit, a hardware circuit implementing the logical method flow can be easily obtained.

[0204] The controller can be implemented in any suitable way, for example, the controller can take the form of a microprocessor or processor and a computer readable medium storing computer readable program code, such as software or firmware, executable by the (micro)processor, logic gates, switches, an application specific integrated circuit (ASIC), a programmable logic controller and an embedded microcontroller, examples of which include but are not limited to the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20 and Silicone Labs C8051F320, the memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also know that, in addition to being implemented in pure computer readable program code, the controller can equally well be implemented to perform the same functions using logic gates, switches, an application specific integrated circuit, a programmable logic controller and an embedded microcontroller, etc. by means of a logical programming of the method steps. The controller can thus be considered as a hardware component, and the means comprised therein for performing the various functions can be considered as structures within the hardware component. Alternatively, the means for performing the various functions can even be considered as both a software module implementing the method and a structure within the hardware component.

[0205] The systems, apparatuses, modules or units illustrated by the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0206] For the sake of description, the above apparatuses are described in various units by functions respectively. Of course, the functions of the units can be implemented in one or more software and / or hardware in the implementation of the present specification.

[0207] Those skilled in the art will understand that the embodiments of the present specification can be provided as a method, a system or a computer program product. Therefore, the embodiments of the present specification can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present specification can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0208] The specification is presented with reference to flow diagrams and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the specification. It will be understood that each block of the flow diagrams and / or block diagrams, and combinations of blocks in the flow diagrams and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing element or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flow diagrams and / or block diagrams block or blocks. Figure 1 one or more flow or multiple flows and / or blocks Figure 1 one or more flow or multiple flows and / or blocks

[0209] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flow diagrams and / or block diagrams block or blocks. Figure 1 one or more flow or multiple flows and / or blocks Figure 1 one or more flow or multiple flows and / or blocks

[0210] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow diagrams and / or block diagrams block or blocks. Figure 1 one or more flow or multiple flows and / or blocks Figure 1 one or more flow or multiple flows and / or blocks

[0211] In one typical configuration, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0212] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory, non-volatile memory, such as read-only memory (ROM), EPROM, and / or flash memory, etc. The memory is an example of computer readable media.

[0213] Computer-readable media includes permanent and non-permanent, movable and non-movable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer-readable media does not include transitory media such as modulated data signals and carriers.

[0214] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to cover non-exclusive inclusion, such that processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed, or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or device that includes the element.

[0215] The specification can be described in the general context of computer-executable instructions, such as program modules, executed by computers. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The specification can also be practiced in a distributed computing environment, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0216] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other, and each embodiment focuses on the difference from other embodiments. In particular, for device, equipment, non-volatile computer storage medium embodiments, since they are basically similar to method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

[0217] The above-described embodiments of the application have special structure and can achieve the desired results. Other embodiments can have different structures and achieve the same results. The purpose of the above-described embodiments is to illustrate the principles of the application and not to limit the scope of the application. The scope of the application is defined by the claims and their equivalents. Other embodiments are within the scope of the claims.

[0218] The above description is merely illustrative of the embodiments of the present application and is not intended to limit the scope of the present application. Various modifications can be made by those skilled in the art based upon the teachings disclosed herein. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present application shall fall within the scope of the claims of the present application.

Claims

1. A method for extracting key information, comprising: Obtain the original text from which the key information to be extracted; Determine the attention weight vector for the original text, and generate the first target word accordingly based on the attention weight vector; Determine whether the first target word is included in the original text; If the first target word is included in the original text, then according to the attention weight vector, the attention weight of the first target word when it is generated is obtained, and according to the attention weight of the first target word, penalty data for the first target word is generated, so as to continue to generate the second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated is the same as the first target word; If the first target word is not included in the original text, then without generating penalty data for the first target word, the attention weight vector corresponding to the second target word to be generated is obtained, so as to continue generating the second target word; The key information is generated based on the first target word and the second target word. The generation of penalty data for the first target word includes: Determine whether the first target word has been generated before this generation; and If the first target word has already been generated before this generation, the penalty is increased compared to if the first target word has not been generated before this generation, in order to further reduce the possibility that the second target word generated in the future will be the same as the first target word.

2. The method as described in claim 1, wherein generating penalty data for the first target word based on the attention weight of the first target word specifically includes: Determine at least one dimension associated with the predictive model used to extract the key information as a penalty dimension, the penalty dimension including at least one of attention weight, loss function, and replication probability; Based on the penalty dimension and the attention weight of the first target word, penalty data for the first target word is generated.

3. The method as described in claim 1, wherein generating penalty data for the first target word based on the attention weight of the first target word is performed during the training phase of a prediction model for extracting the key information, and the loss function of the prediction model includes a penalty loss term related to the attention weight of the generated target word; The step of generating penalty data for the first target word based on its attention weight specifically includes: Based on the attention weights for the first target word and the penalty loss term, a penalty loss is generated as penalty data for the first target word, which is then used to train the prediction model.

4. The method as described in claim 3, wherein generating the penalty loss based on the attention weight for the first target word and the penalty loss term specifically includes: If the first target word has been generated for the original text several times before this generation, then the historical attention weights of the several times the first target word was generated are obtained. Based on the attention weight for the first target word and the historical attention weight, a cumulative attention weight is generated; A penalty loss is generated based on the cumulative attention weight and the penalty loss term.

5. The method as described in claim 4, further generating a second target word based on the penalty data, specifically including: Proceed to the next target word generation stage after this one; During the next target word generation stage, the attention weight of the first target word is determined and used as the current attention weight; The cumulative attention weight and the current attention weight are compared, and the penalty loss is re-determined based on the smaller of the two to generate the second target word.

6. A method for processing live streaming content, specifically including: Obtain live stream videos from hosts on live streaming platforms; The live video is subjected to speech recognition and segmentation processing to obtain multiple live video segment texts; Determine the attention weight vector for the live stream text segment, and generate the first target word accordingly based on the attention weight vector; Determine whether the first target word is included in the text of the live broadcast segment; If the first target word is included in the text of the live broadcast segment, penalty data for the first target word is generated, and a second target word is generated based on the penalty data, wherein the penalty data is used to reduce the probability that the second target word generated is the same as the first target word; If the first target word is not included in the text of the live broadcast segment, then without generating penalty data for the first target word, the attention weight vector corresponding to the second target word to be generated is obtained, so as to continue generating the second target word; Based on the first target word and the second target word, key information of the live broadcast segment text is generated, which serves as the segment title of the live broadcast video segment corresponding to the live broadcast segment text; The live video segments and their generated titles are combined and published so that users can select segments for playback. The generation of penalty data for the first target word includes: Determine whether the first target word has been generated before this generation; and If the first target word has already been generated before this generation, the penalty is increased compared to if the first target word has not been generated before this generation, in order to further reduce the possibility that the second target word generated in the future will be the same as the first target word.

7. A key information extraction device, comprising: The first acquisition module acquires the original text of the key information to be extracted. The target word generation module determines the attention weight vector for the original text and generates the first target word accordingly based on the attention weight vector. The first judgment module determines whether the first target word is included in the original text; The first penalty module, if the first target word is included in the original text, obtains the attention weight of the first target word when it is generated according to the attention weight vector, generates penalty data for the first target word according to the attention weight of the first target word, and continues to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated in the future is the same as the first target word; The non-penalty module, if the first target word is not included in the original text, obtains the attention weight vector corresponding to the second target word to be generated without generating penalty data for the first target word, so as to continue generating the second target word; The first key information generation module generates the key information based on the first target word and the second target word. The generation of penalty data for the first target word includes: Determine whether the first target word has been generated before this generation; and If the first target word has already been generated before this generation, the penalty is increased compared to if the first target word has not been generated before this generation, in order to further reduce the possibility that the second target word generated in the future will be the same as the first target word.

8. A live streaming content processing device, comprising: The second acquisition module acquires live videos from the streamers on the live streaming platform; The processing module performs speech recognition and segmentation on the live video to obtain multiple live video segment texts; The live stream target word generation module determines the attention weight vector for the live stream segment text and generates the first target word accordingly based on the attention weight vector. The second judgment module determines whether the first target word is included in the live broadcast segment text; The second penalty module generates penalty data for the first target word if the first target word is included in the live broadcast segment text, and continues to generate a second target word based on the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated in the future is the same as the first target word; The non-penalty module, if the first target word is not included in the live broadcast segment text, obtains the attention weight vector corresponding to the second target word to be generated without generating penalty data for the first target word, so as to continue generating the second target word; The second key information generation module generates key information of the live broadcast segment text based on the first target word and the second target word, which serves as the segment title of the live broadcast video segment corresponding to the live broadcast segment text. The publishing module combines each of the live video segments and the generated segment titles to publish them, allowing users to select segments for playback. The generation of penalty data for the first target word includes: Determine whether the first target word has been generated before this generation; and If the first target word has already been generated before this generation, the penalty is increased compared to if the first target word has not been generated before this generation, in order to further reduce the possibility that the second target word generated in the future will be the same as the first target word.

9. A key information extraction device, comprising: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: Obtain the original text from which the key information to be extracted; Determine the attention weight vector for the original text, and generate the first target word accordingly based on the attention weight vector; Determine whether the first target word is included in the original text; If the first target word is included in the original text, then according to the attention weight vector, the attention weight of the first target word when it is generated is obtained, and according to the attention weight of the first target word, penalty data for the first target word is generated, so as to continue to generate the second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated is the same as the first target word; If the first target word is not included in the original text, then without generating penalty data for the first target word, the attention weight vector corresponding to the second target word to be generated is obtained, so as to continue generating the second target word; The key information is generated based on the first target word and the second target word. The generation of penalty data for the first target word includes: Determine whether the first target word has been generated before this generation; and If the first target word has already been generated before this generation, the penalty is increased compared to if the first target word has not been generated before this generation, in order to further reduce the possibility that the second target word generated in the future will be the same as the first target word.

10. A live streaming content processing device, comprising: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: Obtain live stream videos from hosts on live streaming platforms; The live video is subjected to speech recognition and segmentation processing to obtain multiple live video segment texts; Determine the attention weight vector for the live stream text segment, and generate the first target word accordingly based on the attention weight vector; Determine whether the first target word is included in the text of the live broadcast segment; If the first target word is included in the text of the live broadcast segment, penalty data for the first target word is generated, and a second target word is generated based on the penalty data, wherein the penalty data is used to reduce the probability that the second target word generated is the same as the first target word; If the first target word is not included in the text of the live broadcast segment, then without generating penalty data for the first target word, the attention weight vector corresponding to the second target word to be generated is obtained, so as to continue generating the second target word; Based on the first target word and the second target word, key information of the live broadcast segment text is generated, which serves as the segment title of the live broadcast video segment corresponding to the live broadcast segment text; The live video segments and their generated titles are combined and published so that users can select segments for playback. The generation of penalty data for the first target word includes: Determine whether the first target word has been generated before this generation; and If the first target word has already been generated before this generation, the penalty is increased compared to if the first target word has not been generated before this generation, in order to further reduce the possibility that the second target word generated in the future will be the same as the first target word.

Images