Semantic extraction method, report generation method, electronic device and storage medium

Abstract

The present invention provides a semantic extraction method, a report generation method, an electronic device and a storage medium. The semantic extraction method includes: determining the text to be extracted; performing semantic extraction on the text to be extracted based on the text semantics and semantic set of the sample text to obtain the text semantics of the text to be extracted, and the semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text. The method, device, electronic device and storage medium provided by the present invention can perform semantic extraction on the text to be extracted through the mapping relationship between the text and the text semantics obtained by comparative learning of the semantic set of the sample text momentum semantics and negative sample momentum semantics. The application of the negative sample momentum semantics enables the knowledge carried by the negative sample momentum semantics to be fully referred to during the semantic extraction process to strengthen the text semantic extraction ability, thereby improving the accuracy of semantic extraction and helping to ensure the accuracy and reliability of the subsequent report generation task execution.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to a semantic extraction method, a report generation method, an electronic device, and a storage medium. Background Art

[0002] As an online service, an Internet product needs to analyze consultation dialogues provided by customer service and comments on product orders to provide a professional user feedback analysis report to help optimize the functions of the product and build new services. The task of generating a user feedback analysis report based on consultation dialogues provided by customer service and comments on product orders is similar to the task of abstract generation. Currently, when using a neural network model to perform an abstract generation task, the generation task of the text abstract can be completed by extracting the semantic information at the high level of the text.

[0003] However, in the existing machine abstract generation tasks, most are based on token-level semantic extraction, that is, essentially more concerned with the character-level features in the language generation modeling task, and the effect of extracting global semantic features is not ideal, resulting in poor abstract generation effect. Summary of the Invention

[0004] The present invention provides a semantic extraction method, a report generation method, an electronic device, and a storage medium to solve the defect that the existing token-level semantic extraction method has an unsatisfactory effect in extracting global semantic features.

[0005] The present invention provides a semantic extraction method, including:

[0006] Determine the text to be extracted;

[0007] Based on the text semantics and semantic set of the sample text, perform semantic extraction on the text to be extracted to obtain the text semantics of the text to be extracted, where the semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text.

[0008] According to a semantic extraction method provided by the present invention, the text semantics and semantic set of the sample text are determined based on the following steps:

[0009] Perform semantic extraction on the current sample text to obtain the text semantics and positive sample momentum semantics of the sample text;

[0010] Based on the positive sample momentum semantics, update the negative sample cache queue, and determine the updated negative sample cache queue as the semantic set of the current sample text, and determine the next sample text of the current sample text as the current sample text.

[0011] A semantic extraction method provided by the present invention, updating the negative sample cache queue based on the positive sample momentum semantics, includes:

[0012] Based on the length of the negative sample cache queue, randomly select an insertion position;

[0013] Based on the insertion position, insert the positive sample momentum semantics into the negative sample cache queue, and remove the sample at the end of the negative sample cache queue to obtain an updated negative sample cache queue.

[0014] A semantic extraction method provided by the present invention, extracting the text semantics and positive sample momentum semantics of the sample text from the current sample text, includes:

[0015] Based on text semantic extraction parameters, perform text semantic extraction on the current sample text to obtain the text semantics of the sample text and the text backpropagation gradient of the text semantic extraction;

[0016] Based on momentum semantic extraction parameters, perform momentum semantic extraction on the current sample text to obtain the positive sample momentum semantics of the sample text and the momentum backpropagation gradient of the momentum semantic extraction;

[0017] Based on the text backpropagation gradient, update the text semantic extraction parameters, and based on the text backpropagation gradient and the momentum backpropagation gradient, update the momentum semantic extraction parameters.

[0018] A semantic extraction method provided by the present invention, extracting the text semantics of the text to be extracted based on the text semantics of the sample text and the semantic set, includes:

[0019] Based on the semantic mapping relationship, perform semantic extraction on the text to be extracted to obtain the text semantics of the text to be extracted;

[0020] The semantic mapping relationship is determined based on the similarity distribution of the text semantics of the sample text and each momentum semantics in the semantic set, and the positive and negative sample probability distributions of the semantic set.

[0021] A semantic extraction method provided by the present invention, the positive and negative sample probability distributions are determined based on the following steps:

[0022] Based on the positions of the positive sample momentum semantics and negative sample momentum semantics in the semantic set, determine the positive and negative sample label representations;

[0023] Based on the combinations of the text semantics and each momentum semantics in the semantic set respectively, determine the fusion features of each combination;

[0024] Determine the positive and negative sample probability distributions based on the similarities between the positive and negative sample label representations and the fusion features of each combination.

[0025] The present invention also provides a report generation method, including: based on the semantic extraction method described in any one of the above, perform semantic extraction on the original text to obtain the semantics of the original text;

[0026] Generate a report text corresponding to the original text based on the semantics of the original text.

[0027] The present invention also provides a semantic extraction device, including:

[0028] A determination module for determining the text to be extracted;

[0029] An extraction module for performing semantic extraction on the text to be extracted based on the text semantics and semantic set of the sample text to obtain the text semantics of the text to be extracted, where the semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text.

[0030] The present invention also provides a report generation device, including:

[0031] A semantic extraction module for performing semantic extraction on the original text based on the semantic extraction method described in any one of the above to obtain the semantics of the original text;

[0032] A report generation module for generating a report text corresponding to the original text based on the semantics of the original text.

[0033] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the program, it implements the steps of the semantic extraction method described in any one of the above or implements the steps of the report generation method provided above.

[0034] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the semantic extraction method described in any one of the above or implements the steps of the report generation method provided above.

[0035] The present invention also provides a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the steps of the semantic extraction method described in any one of the above or implements the steps of the report generation method provided above.

[0036] The semantic extraction method, report generation method, electronic device and storage medium provided by the present invention perform semantic extraction on the text to be extracted through the mapping relationship between the text and the text semantics. The application of the negative sample momentum semantics enables the knowledge carried by the negative sample momentum semantics to be fully referred to during the semantic extraction process, so as to strengthen the semantic extraction ability, thereby improving the accuracy of semantic extraction and helping to ensure the accuracy and reliability of the subsequent report generation task execution. Description of the Drawings

[0037] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0038] Figure 1 It is a schematic flowchart of the semantic extraction method provided by the present invention;

[0039] Figure 2 It is a schematic flowchart of the method for obtaining the text semantics and semantic set of the current sample text provided by the present invention;

[0040] Figure 3 It is a schematic flowchart of the negative sample cache queue update method provided by the present invention;

[0041] Figure 4 It is a schematic flowchart of the method for updating the text semantics and momentum semantics extraction parameters provided by the present invention;

[0042] Figure 5 It is a schematic flowchart of the method for obtaining the probability distribution of positive and negative samples provided by the present invention;

[0043] Figure 6 It is a schematic flowchart of the report generation method provided by the present invention;

[0044] Figure 7 It is a schematic structural diagram of the semantic extraction model provided by the present invention;

[0045] Figure 8 It is a schematic structural diagram of the report generation model provided by the present invention;

[0046] Figure 9 It is a schematic structural diagram of the semantic extraction device provided by the present invention;

[0047] Figure 10 It is a schematic structural diagram of the report generation device provided by the present invention;

[0048] Figure 11 It is a schematic structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The character-level semantic extraction currently used for text summarization cannot obtain global semantic features well, resulting in unsatisfactory summarization results. In addition, when generating experience reports based on text content such as customer service and user consultation conversations or product order comments, it is necessary to extract the key points of the text content from a global perspective.

[0051] Therefore, how to improve the ability to extract global semantic features at the sentence level is a technical problem that needs to be solved urgently in this field.

[0052] In view of the above situation, an embodiment of the present invention provides a semantic extraction method. Figure 1 FIG. 1 is a flow chart of the semantic extraction method provided by the present invention. Figure 1 As shown, the method includes:

[0053] Step 110, determining the text to be extracted;

[0054] Specifically, the text to be extracted is the text that needs to be extracted elements. The text to be recognized can be directly input by the user, or obtained by transcribing the collected audio, or obtained by collecting images through image acquisition devices such as scanners, mobile phones, cameras, etc., and performing OCR (Optical Character Recognition) on the images. The embodiments of the present invention do not limit this.

[0055] Step 120 , based on the text semantics and semantic set of the sample text, semantic extraction is performed on the text to be extracted to obtain the text semantics of the text to be extracted. The semantic set includes the positive sample momentum semantics and the negative sample momentum semantics of the sample text.

[0056] Specifically, the text semantics of the text to be extracted is obtained through semantic extraction. The semantic extraction of the text to be extracted here can be obtained by semantic mapping the text to be extracted through a pre-acquired mapping relationship. The mapping relationship here can be specifically reflected as a semantic extraction model obtained through model training, or as a semantic extraction rule obtained through association mining. The embodiment of the present invention does not make any specific limitation on this.

[0057] Here, for obtaining the mapping relationship, it can be obtained based on the text semantics of the sample text and the comparative learning of the semantic set of the sample text. In particular, considering that comparative learning is trained by automatically constructing similar positive samples and dissimilar negative samples, but in the process of comparative learning, the positive samples can see fewer negative samples. Therefore, in the process of comparative learning for extracting semantics, the current sample can obtain more negative sample momentum semantics, which helps to improve the accuracy of extracting sentence-level semantics. And the extracted semantics can dynamically update the negative sample momentum semantics during the comparative learning process, so that the samples in subsequent comparative learning can obtain new negative sample momentum semantics, which helps to further improve the accuracy of extracting sentence-level semantics.

[0058] Therefore, in the embodiments of the present invention, comparative learning is performed on the text semantics of the sample text and the semantic set corresponding to the sample text, which contains positive sample momentum semantics and negative sample momentum semantics, to obtain the process of the mapping relationship between the text and its text semantics, that is, the process of training the semantic extraction model or the process of rule mining. Here, the semantic set records positive sample momentum semantics and negative sample momentum semantics, and the semantic set is dynamically updated with the change of the sample text. That is, each sample text will have a corresponding semantic set. By adjusting the parameters of the semantic extraction model or the representation method of the corresponding relationship, the text semantics of the sample text can be made as similar as possible to the positive sample momentum semantics recorded in the semantic set and as different as possible from each negative sample momentum semantics recorded in the semantic set, so that the text semantics of the text to be extracted obtained thereby can be more reliable. Among them, the semantic set can contain one or more positive sample momentum semantics and multiple negative sample momentum semantics. Especially in the case of containing multiple negative sample momentum semantics, the semantic extraction model can learn more negative sample momentum semantics, which helps to improve the accuracy of extracting sentence-level semantics.

[0059] The mapping relationship obtained by training in this way has a more reliable semantic extraction ability. Applying the mapping relationship to perform sentence-level semantic extraction on the text to be extracted helps to improve the accuracy of extracting the semantics of the text to be extracted, thereby improving the accuracy of the subsequent report generation task.

[0060] In addition, before step 120 is executed, it is necessary to obtain the text semantics of the sample text and the corresponding semantic set. The positive sample momentum semantics in the semantic set are extracted based on this sample text. The text semantics of the sample text and the positive sample momentum semantics can be obtained by two different semantic extraction methods or by applying two different sets of semantic extraction parameters, but both reflect the semantics of the same text. The negative sample momentum semantics of the sample text are extracted based on other sample texts outside this sample text. The positive sample momentum semantics and the negative sample momentum semantics of the sample text are extracted based on the same semantic extraction method or by applying the same semantic extraction parameters for different sample texts. Therefore, the text semantics of the sample text and the negative sample momentum semantics reflect the semantics of different texts. Further, the extracted momentum semantics can be used as the positive sample momentum semantics in the semantic set corresponding to the sample text. The negative sample momentum semantics in the semantic set can be arbitrarily selected from multiple sample momentum semantics extracted from the sample momentum semantics of the recently trained historical sample texts as the negative sample momentum semantics of the current sample text and updated to the semantic set. It is also possible to use all the sample momentum semantics in the semantic set corresponding to the previous sample text as the negative sample momentum semantics of the current sample text. The embodiments of the present invention do not limit this.

[0061] The method provided by the embodiments of the present invention performs semantic extraction on the text to be extracted through the sample text semantics and the semantic set including positive sample momentum semantics and negative sample momentum semantics. The application of the negative sample momentum semantics enables the process of semantic extraction to fully refer to the knowledge carried by the negative sample momentum semantics to strengthen the text semantic extraction ability, thereby improving the accuracy of semantic extraction and contributing to ensuring the accuracy and reliability of the subsequent report generation task execution.

[0062] Based on the above embodiments, since in contrastive learning, in order to achieve better learning effects, it is necessary to enable the positive sample momentum features to see more negative sample momentum features, the present invention also provides a preferred embodiment. In this embodiment, the semantic set includes one positive sample momentum semantics and multiple negative sample momentum semantics.

[0063] Based on the above embodiments, Figure 2 is a schematic flowchart of the method for obtaining the text semantics and semantic set of the current sample text provided by the present invention. As Figure 2 shown, the text semantics of the sample text and the semantic set in step 120 are determined based on the following steps:

[0064] Step 210, perform semantic extraction on the current sample text to obtain the text semantics of the sample text and the positive sample momentum semantics;

[0065] Step 220: Update the negative sample cache queue based on the positive sample momentum semantics, and determine the updated negative sample cache queue as the semantic set of the current sample text. Then, determine the next sample text of the current sample text as the current sample text.

[0066] Considering that frequent allocation and release of memory during the update of the semantic set will increase the consumption of system resources and thus affect the efficiency of contrastive learning. Therefore, in the embodiments of the present invention, a memory space of a specified size is pre-allocated for the semantic set. At this time, dynamically updating the sample momentum semantics in the negative sample cache queue only requires reading and writing operations on this memory area, which helps reduce the consumption of system resources and improve the efficiency of contrastive learning.

[0067] Specifically, in the process of obtaining the mapping relationship between the text and its text semantics by performing contrastive learning on the text semantics of the sample text and the semantic set corresponding to the sample text that contains positive sample momentum semantics and negative sample momentum semantics, that is, during the training of the semantic extraction model, in step 210, the text semantics of the current sample text are extracted through the text semantic extraction parameters to obtain the text semantics of the sample text, and the positive sample momentum semantics are extracted from the current sample text through the momentum semantic extraction parameters. Here, the text semantic extraction parameters and the momentum semantic extraction parameters are two different semantic extraction parameters. In step 220, the negative sample cache queue is updated according to the positive sample momentum semantics obtained in step 210, and the updated negative sample cache queue is used as the semantic set of the current sample text. Here, the negative sample cache queue before the update is the semantic set corresponding to the previous sample text stored after the previous update, that is, it stores the positive sample momentum semantics of other sample texts. Since other sample texts and the current sample text are different texts, it can also be understood as the negative sample momentum semantics of the current sample text. That is, the updated negative sample cache queue stores the negative sample momentum semantics of the current sample text.

[0068] Then, adjust the parameters or the representation method of the corresponding relationship of the semantic extraction model according to the text semantics and the semantic set of the current sample text. After the adjustment is completed, determine the next sample text of the current sample text as the current sample text and perform the above operations until the entire contrastive learning is completed to obtain the final semantic extraction model, that is, the final mapping relationship between the text and its text semantics. Among them, the size of the memory space pre-set in the embodiments of the present invention can be fixed, that is, the size length of the negative sample cache queue is a fixed length, that is, it can store a fixed number of momentum semantics.

[0069] It should be noted that by using the positive sample momentum semantics to update the negative sample cache queue, based on the first-in-first-out (FIFO) characteristic of the queue, the positive sample can be inserted into the head of the negative sample cache queue. If the total number of momentum semantics in the negative sample cache queue is equal to the size length of the negative sample cache queue, then one momentum semantics is removed from the tail to complete the update of the negative sample cache queue. It is also possible to randomly find an insertion position in a specified area of the negative sample cache queue and insert the positive sample momentum semantics based on this position. If the total number of momentum semantics in the negative sample cache queue is equal to the size length of the negative sample cache queue, then one momentum semantics is removed from the tail. The embodiments of the present invention do not limit this.

[0070] Based on the above embodiments, Figure 3 is a schematic flowchart of the method for updating the negative sample cache queue provided by the present invention. As Figure 3 shown, in step 220, based on the positive sample momentum semantics, updating the negative sample cache queue includes:

[0071] Step 310, randomly select an insertion position based on the length of the negative sample cache queue;

[0072] Considering that in the process of obtaining the mapping relationship between the text and its text semantics by comparing and learning the text semantics of the sample text and the semantic set corresponding to the sample text that contains positive sample momentum semantics and negative sample momentum semantics, if the positive sample momentum semantics are inserted into the head of the negative sample cache queue, that is, inserting the momentum semantics at a fixed position, it will cause too much position information to be learned in the process of obtaining this mapping relationship, resulting in a decrease in the accuracy of the mapping relationship. Therefore, the embodiments of the present invention insert the positive sample momentum semantics in a way of randomly selecting positions, which can reduce the learning of position information in the process of comparative learning, so as to focus more on the learning of the mapping relationship between semantics.

[0073] Specifically, within the range of the length of the negative sample cache queue, randomly select an insertion position as the insertion position for the subsequent positive sample momentum semantics.

[0074] It should be noted that the insertion position is the index position of the sample in the negative sample cache queue. For example: if the length of the negative sample cache queue is 10, then its index values are [0, 9], and the randomly selected insertion position is any index position represented by a value in [0, 9].

[0075] Step 320, based on the insertion position, insert the positive sample momentum semantics into the negative sample cache queue, and remove the sample at the tail of the negative sample cache queue to obtain an updated negative sample cache queue.

[0076] It should be noted that considering that removing the momentum semantics from the end of the queue may remove the positive sample momentum semantics of the previous round or the negative sample momentum semantics of the previous round, and the positive sample momentum semantics of the sample text of the previous round is different from the positive sample momentum semantics of the current sample text, that is, the positive sample momentum semantics of the sample text of the previous round can be used as the negative sample momentum semantics of the current sample text. Therefore, in the embodiment of the present invention, before inserting the positive sample momentum semantics of the current sample text into the negative sample cache queue, the tags of all the momentum semantics in the negative sample cache queue can be set to negative sample momentum semantics tags, and the momentum semantics tags of other positions except the inserted position can also be set to negative sample momentum semantics tags after inserting the positive sample momentum semantics and removing the momentum semantics at the end of the queue. The embodiment of the present invention does not limit this.

[0077] Particularly, when the negative sample cache queue is empty, before the number of stored samples reaches the memory size, the position of the inserted momentum samples can be identified by a mask. When the mask identifies the position where the momentum samples have been inserted, this position will not be queried during the next insertion until the momentum samples in the negative sample cache queue are full, and then the logic for updating the negative sample cache queue as described above is started. The embodiment of the present invention does not limit this.

[0078] Based on the above embodiments, Figure 4 is a schematic flow diagram of the method for updating the text semantics and momentum semantics extraction parameters provided by the present invention. As Figure 4 shown, step 210 includes:

[0079] Step 211, based on the text semantics extraction parameters, perform text semantics extraction on the current sample text to obtain the text semantics of the sample text and the text backpropagation gradient of the text semantics extraction;

[0080] Step 212, based on the momentum semantics extraction parameters, perform momentum semantics extraction on the current sample text to obtain the positive sample momentum semantics of the sample text and the momentum backpropagation gradient of the momentum semantics extraction;

[0081] Step 213, update the text semantics extraction parameters based on the text backpropagation gradient, and update the momentum semantics extraction parameters based on the text backpropagation gradient and the momentum backpropagation gradient.

[0082] Considering that the stability of the momentum feature set in the negative sample cache queue in contrastive learning helps to improve the accuracy of contrastive learning, therefore, in the embodiment of the present invention, the momentum semantics extraction parameters are updated by means of momentum update.

[0083] Specifically, the text backpropagation gradient returned after performing text semantic extraction on the current sample text by extracting parameters through text semantics is used to update the text semantic extraction parameters. The momentum backpropagation gradient returned after performing momentum semantic extraction on the current sample text by extracting parameters through momentum semantics is combined with the text backpropagation gradient to update the momentum semantic extraction parameters.

[0084] It should be noted that step 211 and step 212 can be executed synchronously or sequentially in order, and the embodiments of the present invention do not limit this.

[0085] Based on the above embodiments, step 120 includes:

[0086] Based on the semantic mapping relationship, perform semantic extraction on the text to be extracted to obtain the text semantics of the text to be extracted;

[0087] The semantic mapping relationship is determined based on the similarity distribution of the text semantics of the sample text and each momentum semantics in the semantic set, as well as the positive and negative sample probability distributions of the semantic set.

[0088] Considering that when comparing and learning the text semantics of the sample text and the semantic set corresponding to the sample text that contains positive sample momentum semantics and negative sample momentum semantics to extract the mapping relationship between the text and its text semantics, combining the probability distributions of the labels of the positive and negative samples in this comparative learning helps to converge more stably when obtaining this mapping relationship, and thus helps to improve the accuracy of the semantics extracted by this mapping relationship.

[0089] Specifically, the text semantics of the text to be extracted are obtained by performing semantic extraction on the text to be extracted according to the semantic mapping relationship. Here, the semantic mapping relationship is determined by respectively performing similarity matching between the text semantics of the sample text and each momentum semantics in the corresponding semantic set, obtaining the similarity distribution of the text semantics of the sample text and each momentum semantics in the semantic set, and combining the probability distributions of the positive and negative sample labels in the semantic set. Specifically, it can be embodied as a semantic extraction model obtained through model training or as a semantic extraction rule obtained through association mining. The embodiments of the present invention do not make specific limitations on this.

[0090] It should be noted that in the embodiments of the present invention, for the text semantics of the sample text, a contrastive learning is performed on the semantic set corresponding to the sample text, which contains positive sample momentum semantics and negative sample momentum semantics, to extract the acquisition process of the mapping relationship between the text and its text semantics, that is, in the process of training the semantic extraction model or rule mining, the similarity distribution of the text semantics of the sample text and each momentum semantics in the semantic set, as well as the positive and negative sample probability distributions of the semantic set are applied. Here, the similarity distribution and the probability distributions of positive and negative samples are calculated for the degree of difference through the method of Kullback-Leibler divergence (KL-Divergency), which can gradually reduce the degree of difference between the similarity distribution and the probability distributions of positive and negative samples, so that the text semantics of the text to be extracted obtained thereby can be more reliable.

[0091] The semantic extraction method provided by the present invention realizes the extraction of the text semantics of the text to be extracted according to the semantic mapping relationship by training to reduce the degree of difference between the similarity distribution of the text semantics of the sample text and each momentum semantics in the semantic set and the positive and negative sample probability distributions of the semantic set, further improves the accuracy of semantic extraction, and further improves the accuracy of the subsequent report generation task.

[0092] Based on the above embodiments, Figure 5 is a schematic flowchart of the method for obtaining the positive and negative sample probability distributions provided by the present invention. As Figure 5 shown, the positive and negative sample probabilities in step 121 are determined based on the following steps:

[0093] Step 510, determine the positive and negative sample label representations based on the positions of the positive sample momentum semantics and negative sample momentum semantics in the semantic set;

[0094] Specifically, perform label semantic extraction on the positive and negative sample labels corresponding to the positive sample momentum semantics and negative sample momentum semantics in the semantic set to obtain the positive and negative sample label representations. Among them, the positive and negative sample labels can be one-hot labels.

[0095] Step 520, determine the fusion features of each combination based on the combinations of the text semantics with each momentum semantics in the semantic set;

[0096] Specifically, each momentum semantics in the semantic set corresponding to the current sample text is combined with the text semantics of the current sample text respectively, and the features after combination are subjected to secondary semantic extraction to obtain the fusion features of each combination with the same dimension as the positive and negative sample label representations.

[0097] Step 530, determine the positive and negative sample probability distributions based on the similarities between the positive and negative sample label representations and the fusion features of each combination respectively.

[0098] Specifically, according to the positions of the positive and negative sample momentum semantics, the similarity between each positive and negative sample label representation and the combined fusion feature corresponding to its position is calculated in sequence, obtaining the similarity between the positive and negative sample label representations and the fusion features of each combination, and determining the positive and negative sample probability distributions based on this similarity.

[0099] It should be noted that considering that the positive and negative sample probability distributions are calculated after multiple semantic extractions, in order to prevent the degradation problem caused by too deep a network, after obtaining the similarity between the positive and negative sample label representations and the fusion features of each combination, weighting is performed using a preset weight value and the original positive and negative sample labels to obtain the positive and negative sample probability distributions.

[0100] Based on the above embodiments, Figure 6 is a schematic flow diagram of the report generation method provided by the present invention. As Figure 6 shown, an embodiment of the present invention further provides a report generation method, which includes:

[0101] Step 610, based on the semantic extraction method provided in any of the above embodiments, perform semantic extraction on the original text to obtain the semantics of the original text;

[0102] Step 620, based on the semantics of the original text, generate a report text corresponding to the original text.

[0103] Specifically, the mapping relationship obtained by the semantic extraction method provided in any of the above embodiments is embedded into the report generation method for semantic extraction of the original text to obtain the semantics of the original text, and based on the semantics of the original text, semantic extraction and summary are performed on the original text to generate a report text of the original text. Among them, when performing semantic extraction and summary on the original text, it can be directly extracted and summarized based on the semantics of the original text obtained in step 610, or preliminary extraction can be performed based on the semantics of the original text in step 610 to obtain a draft report text, and then the draft report text is again based on the mapping relationship obtained by the semantic extraction method provided in any of the above embodiments to perform semantic extraction on the draft report text, and according to the semantics of the draft report text, extraction and summary are performed on the draft report text to generate the final report text of the original text.

[0104] It should be noted that the original text is the text for which a report needs to be generated. The original text can be directly input by the user, or obtained by voice transcription of the collected audio, or obtained by collecting an image through an image acquisition device such as a scanner, mobile phone, or camera and performing OCR (Optical Character Recognition) on the image. The embodiments of the present invention do not limit this.

[0105] The reporting method provided by the present invention extracts the semantics of the original text by using the semantic extraction method provided in any of the above embodiments, realizes the accurate extraction of the sentence-level semantics of the original report, and further improves the accuracy of the report generation task.

[0106] Based on the above embodiments, Figure 7 is a schematic structural diagram of the semantic extraction model provided by the present invention. As Figure 7 shown, the model training process includes:

[0107] Step 710, extract text semantics and momentum semantics, specifically including:

[0108] Determine the sample text, and use the entire sample text as a query to input into the semantic extraction encoder of BERT (Bidirectional Encoder Representation from Transformers) to extract semantics, obtain the text semantics (query features) of the sample text, and input the entire sample text into the semantic extraction encoder of MoUp-BERT (Momentum Update BERT) to extract semantics to obtain the momentum semantics (key features) of the sample text, and use the key features of this sample as the positive sample momentum semantics (k + semantic features) and insert them into the negative sample cache queue (dictionary) through step 720.

[0109] Among them, the semantic extraction encoder of BERT, without special processing, is used as the encoder of the query features in the contrast learning; and in the task of generating downstream user experience reports, sentence-level semantic features are provided;

[0110] The semantic extraction encoder of MoUp-BERT uses the gradient transmitted back by the semantic extraction encoder of BERT that extracts the query features to update the momentum parameters; it is used to encode and generate the generation task of the k + semantic features entering the dictionary. The semantic extraction encoder of MoUp-BERT and the semantic extraction encoder of BERT are independent encoders, which respectively complete the generation of query features and the search for key features in the dictionary, so that the feature set in the dictionary in the contrast learning is stable and will not change greatly in the feature space due to the change of the query feature encoding. The form of the momentum update parameter, that is, the formula:

[0111] θ′ k = mθ k +(1 - m)θ q

[0112] In the formula, θ qThe gradient returned by the semantic extraction encoder of BERT in this round, θ k The gradient returned by the semantic extraction encoder of MoUp-BERT in this round, where m is a hyperparameter of the momentum coefficient, usually set to 0.999, θ′ k The gradient of MoUp-BERT weighted by the momentum coefficient, and θ′k is used to update the momentum parameter of MoUp-BERT.

[0113] Step 720, insert the k + semantic features generated in step 710 into the negative sample cache queue (dictionary), specifically including:

[0114] In the contrastive learning task, a set (dictionary) for supplying query features to find key features needs to be constructed. The negative sample cache (Negtive-Memory-Buffer) in the figure is to apply for a fixed amount of memory in advance in the model content. In the form of a queue data structure, it is used to store the set of key features for similarity matching with query features, that is, the dictionary mentioned above. During the training process, the generation process of the negative sample cache queue dictionary is as follows:

[0115] <1> In the dictionary, randomly select the position vector where the new key feature needs to be inserted;

[0116] <2> Remove the old key feature at the end of the queue from the dictionary;

[0117] <3> Insert the new key feature into the position vector randomly selected in step <1> to complete the generation of the dictionary in this round of search task;

[0118] <4> According to the position vector of the inserted new key feature, correspondingly generate the hard label (one-hot encoded hard label) of the pre-training task of whether the key feature sequence corresponding to the dictionary in the contrastive learning is similar to the query feature; that is, the position vector of the new key feature is marked as similar to the query feature, and the other position vectors in the dictionary are marked as not similar to the query feature.

[0119] The design of the negative sample cache queue only needs to store key feature vectors, and can accumulate and retain the key features of a large number of negative samples in the memory of the model. Isolating the size of the negative samples in the contrastive learning from the batch_size of the language model input can enable the positive samples in the training process to see a larger number of negative samples.

[0120] It can be seen from the loss function of contrastive learning that during the training process, if a single sample can see more negative samples, that is, increasing the denominator, it will greatly increase the difficulty of solving the loss function, thereby further exploring the ability of the BERT language model to capture sentence-level semantic features. This patent proposes a negative memory caching mechanism to give full play to the potential of the language model and better complete the task of text semantic feature extraction.

[0121] By using the randomness of the position where the new key feature is inserted into the dictionary, the recall label of the query feature constructed can avoid the situation where the model only learns the information of the fixed position where the new query feature semantic is inserted when predicting the recall label distribution. At the same time, a constructed label classification task makes the model essentially focus on the label classification task and reduces the capture of fine-grained semantics.

[0122] Step 730: Generate enhanced labels according to the updated dictionary obtained in step 720, specifically including:

[0123] Convert the hard label in one-hot encoding generated in step 720 into the probability distribution of the enhanced label. The loss function of the contrastive learning task can select the KL divergence, that is, by calculating the difference between the probability distribution of the similarity between the query feature and each key feature in the dictionary and the enhanced label, so as to more accurately measure the performance of the model. The following introduces the generation process of the enhanced label:

[0124] <1>The hard label in one-hot encoding is encoded by a label semantic encoder (Label-Representaion-Encoder) built by a DNN network structure to obtain the label representation in the high-dimensional semantic space;

[0125] <2>Append the semantics of the query feature to the back of each key feature stored in the dictionary to form a {key, query} feature matrix;

[0126] <3>In the label similarity layer (Label-Similarity-Layer) in the figure: the {key, query} feature matrix passes through a stacked DNN structure to obtain a key-query fusion feature with the same dimension as the label representation; calculate the semantic similarity between the label representation and the corresponding key-query fusion feature. Multiply this similarity coefficient by a certain weight and add it to the original hard label in one-hot encoding to obtain the enhanced label.

[0127] Finally, the model will use the enhanced label after semantic enhancement and the predicted similarity distribution of query-key to calculate the KL contrast divergence as the loss of the model to guide the update of the model parameters.

[0128] Step 740: Calculate the similarity between the query feature of the sample and each key feature in the dictionary to obtain the similarity distribution of query-key, specifically including:

[0129] The query feature obtained by the BERT semantic extraction encoder and the semantic set of key features in the dictionary updated by the MoUp-BERT encoder are all fed into the Representation-Similarity-Layer. The similarity calculation layer designed in this step calculates the similarity between the query feature and each key feature in the dictionary through the neural network structure of DNN, and then obtains the probability distribution of the predicted labels recalled by the sample in the dictionary through the softmax layer, that is, the similarity distribution of query-key.

[0130] Step 750: The query-key similarity predicted in Step 740, after being processed by the softmax layer, is a probability distribution, denoted by y s And the enhanced label generated in Step 730 is also a probability distribution, denoted by y p The deviation between the model and the label can be measured by the difference between the probability distribution of the predicted similarity and the probability distribution of the soft label. To measure the difference between the two distributions, the Kullback-Leibler divergence (KL-Divergency) is selected for calculation, and the formula is as follows:

[0131]

[0132] In the formula, C represents the number of probabilities in the probability distribution, represents the probability at the c-th position in the similarity distribution of query-key, represents the probability at the c-th position in the enhanced label.

[0133] Using the above-designed pre-trained language model framework, in the corpus of customer service conversation records accumulated in the industry, the pre-training task can be completed, and better sentence-level semantic features of the conversation can be captured for downstream tasks.

[0134] Based on the above embodiments, Figure 8 is a schematic structural diagram of the report generation model provided by the present invention. As Figure 8 shown, the process of the model generating a report includes:

[0135] The original text extracts document-level semantic embeddings (Document Embedding) through the pre-trained semantic extraction model (BERT language model in the figure) above. The report draft (Report-Draft-Output) in the figure is a draft of the user experience report generated by an existing mature abstract generation model. The draft generation model is designed based on the mechanism of writing text from left to right, and can only capture one-way semantic features. The one-way generation model can ensure the smoothness of the generated sentences. Taking the document-level semantic embeddings obtained by BERT and the semantic features of the report draft as inputs, they are fed into the report draft encoder (Report-Draft-Decoder) to obtain the smooth semantic features and document semantic features that can be captured in the draft.

[0136] Summarize the output of the draft encoder. After masking out the dirty data through the masking operation of the upper triangular matrix, it is fed into another semantic extraction model (BERT language model in the figure) again to encode and obtain the embedded representation (Draft-Embedding) of the document-level semantics of the summary draft.

[0137] The user experience report generation layer (Generate-Report-Layer) takes the original document features and draft features as inputs, and is designed with a seq-to-seq model structure to generate the user experience report (the report text of the original report).

[0138] Next, the semantic extraction device and report generation device provided by the present invention will be described. The semantic extraction device described below can be mutually corresponding and referenced to the semantic extraction method described above, and the report generation device can be mutually corresponding and referenced to the report generation method described above.

[0139] Figure 9 It is a schematic structural diagram of the semantic extraction device provided by the present invention. As Figure 9 shown, the device includes: a determination module 910 and an extraction module 920;

[0140] Among them,

[0141] The determination module 910 is used to determine the text to be extracted;

[0142] The extraction module 920 is used to perform semantic extraction on the text to be extracted based on the text semantics and semantic set of the sample text, and obtain the text semantics of the text to be extracted. The semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text.

[0143] In an embodiment of the present invention, a determination module 910 is configured to determine text to be extracted; an extraction module 920 is configured to perform semantic extraction on the text to be extracted based on the text semantics and semantic set of the sample text, so as to obtain the text semantics of the text to be extracted. The semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text. The application of the negative sample momentum semantics enables the process of semantic extraction to fully refer to the knowledge carried by the negative sample momentum semantics, so as to strengthen the extraction ability of text semantics, thereby improving the accuracy of semantic extraction, and contributing to ensuring the accuracy and reliability of the subsequent report generation task execution.

[0144] Based on any of the above embodiments, the extraction module 920 further includes:

[0145] A semantic extraction sub-module is configured to perform semantic extraction on the current sample text to obtain the text semantics and positive sample momentum semantics of the sample text;

[0146] A queue update sub-module is configured to update the negative sample cache queue based on the positive sample momentum semantics, and determine the updated negative sample cache queue as the semantic set of the current sample text, and determine the next sample text of the current sample text as the current sample text.

[0147] Based on any of the above embodiments, the queue update sub-module includes:

[0148] An insertion position determination sub-module is configured to randomly select an insertion position based on the length of the negative sample cache queue;

[0149] An insertion queue sub-module is configured to insert the positive sample momentum semantics into the negative sample cache queue based on the insertion position, and remove the sample at the tail of the negative sample cache queue to obtain an updated negative sample cache queue.

[0150] Based on any of the above embodiments, the semantic extraction sub-module includes:

[0151] A text gradient acquisition sub-module is configured to perform text semantic extraction on the current sample text based on text semantic extraction parameters to obtain the text semantics of the sample text and the text backpropagation gradient of the text semantic extraction;

[0152] A momentum gradient acquisition sub-module is configured to perform momentum semantic extraction on the current sample text based on momentum semantic extraction parameters to obtain the positive sample momentum semantics of the sample text and the momentum backpropagation gradient of the momentum semantic extraction;

[0153] An extraction parameter update sub-module is configured to update the text semantic extraction parameters based on the text backpropagation gradient, and update the momentum semantic extraction parameters based on the text backpropagation gradient and the momentum backpropagation gradient.

[0154] Based on any of the above embodiments, the extraction module 920 further includes:

[0155] A semantic extraction sub-module, configured to perform semantic extraction on the text to be extracted based on the semantic mapping relationship to obtain the text semantics of the text to be extracted;

[0156] A semantic mapping relationship determination sub-module: configured to determine the semantic mapping relationship based on the similarity distribution of the text semantics of the sample text and each momentum semantics in the semantic set, and the positive and negative sample probability distributions of the semantic set.

[0157] Based on any of the above embodiments, the semantic mapping relationship determination sub-module includes:

[0158] A label representation acquisition sub-module, configured to determine the positive and negative sample label representations based on the positions of the positive sample momentum semantics and negative sample momentum semantics in the semantic set;

[0159] A fusion feature acquisition sub-module, configured to determine the fusion features of each combination based on the combination of the text semantics with each momentum semantics in the semantic set;

[0160] A probability distribution acquisition sub-module, configured to determine the positive and negative sample probability distributions based on the similarity between the positive and negative sample label representations and the fusion features of each combination.

[0161] Figure 10 It is a schematic structural diagram of the report generation device provided by the present invention. As Figure 10 shown, the device includes: a semantic extraction module 1010 and a report generation module 1020;

[0162] Among them,

[0163] The semantic extraction module 1010 is configured to perform semantic extraction on the original text based on the semantic extraction method provided in any of the above embodiments to obtain the semantics of the original text;

[0164] The report generation module 1020 is configured to generate a report text corresponding to the original text based on the semantics of the original text.

[0165] In the embodiments of the present invention, through the semantic extraction module 1010, which is configured to perform semantic extraction on the original text based on the semantic extraction method provided in any of the above embodiments to obtain the semantics of the original text; and the report generation module 1020, which is configured to generate a report text corresponding to the original text based on the semantics of the original text, accurate extraction of the sentence-level semantics of the original report is achieved, thereby improving the accuracy of the report generation task.

[0166] Figure 11 Illustrates a schematic structural diagram of the entity of an electronic device, as Figure 11As shown in the figure, the electronic device may include: a processor 1110, a communications interface 1120, a memory 1130, and a communication bus 1140. Among them, the processor 1110, the communications interface 1120, and the memory 1130 complete mutual communication through the communication bus 1140. The processor 1110 may call the logical instructions in the memory 1130 to execute a semantic extraction method, which includes: determining the text to be extracted; based on the text semantics and semantic set of the sample text, performing semantic extraction on the text to be extracted to obtain the text semantics of the text to be extracted. The semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text, or to execute a report generation method, which includes: based on the semantic extraction method provided in any of the above embodiments, performing semantic extraction on the original text to obtain the semantics of the original text; based on the semantics of the original text, generating a report text corresponding to the original text.

[0167] In addition, when the logical instructions in the above-mentioned memory 1130 are implemented in the form of software functional units and sold or used as an independent product, they may be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, may be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0168] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the semantic extraction method provided by the above-mentioned various methods. The method includes: determining the text to be extracted; based on the text semantics and semantic set of the sample text, performing semantic extraction on the text to be extracted to obtain the text semantics of the text to be extracted. The semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text, or executing the report generation method provided by the above-mentioned various methods. The method includes: based on the semantic extraction method provided in any of the above embodiments, performing semantic extraction on the original text to obtain the semantics of the original text; based on the semantics of the original text, generating a report text corresponding to the original text.

[0169] In another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the semantic extraction method provided by each of the above methods. The method includes: determining the text to be extracted; based on the text semantics and semantic set of the sample text, performing semantic extraction on the text to be extracted to obtain the text semantics of the text to be extracted. The semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text. Or it implements the report generation method provided by each of the above methods. The method includes: based on the semantic extraction method provided in any of the above embodiments, performing semantic extraction on the original text to obtain the semantics of the original text; based on the semantics of the original text, generating a report text corresponding to the original text.

[0170] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative effort.

[0171] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. A semantic extraction method, characterized in that, it includes: Determine the text to be extracted; Based on the text semantics and semantic set of the sample text, perform semantic extraction on the text to be extracted to obtain the text semantics of the text to be extracted, where the semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text; The text semantics and semantic set of the sample text are determined based on the following steps: Perform semantic extraction on the current sample text to obtain the text semantics and positive sample momentum semantics of the sample text; Based on the positive sample momentum semantics, update the negative sample cache queue, and determine the updated negative sample cache queue as the semantic set of the current sample text, and determine the next sample text of the current sample text as the current sample text.

2. The semantic extraction method according to claim 1, characterized in that, The updating of the negative sample cache queue based on the positive sample momentum semantics includes: Randomly select an insertion position based on the length of the negative sample cache queue; Based on the insertion position, insert the positive sample momentum semantics into the negative sample cache queue, and remove the tail sample of the negative sample cache queue to obtain an updated negative sample cache queue.

3. The semantic extraction method according to claim 1, characterized in that, The performing semantic extraction on the current sample text to obtain the text semantics and positive sample momentum semantics of the sample text includes: Based on text semantic extraction parameters, perform text semantic extraction on the current sample text to obtain the text semantics of the sample text and the text backpropagation gradient of the text semantic extraction; Based on momentum semantic extraction parameters, perform momentum semantic extraction on the current sample text to obtain the positive sample momentum semantics of the sample text and the momentum backpropagation gradient of the momentum semantic extraction; Based on the text backpropagation gradient, update the text semantic extraction parameters, and based on the text backpropagation gradient and the momentum backpropagation gradient, update the momentum semantic extraction parameters.

4. The semantic extraction method according to any one of claims 1 to 3, characterized in that, The performing semantic extraction on the text to be extracted based on the text semantics and semantic set of the sample text to obtain the text semantics of the text to be extracted includes: Based on the semantic mapping relationship, perform semantic extraction on the text to be extracted to obtain the text semantics of the text to be extracted; The semantic mapping relationship is determined based on the similarity distribution of the text semantics of the sample text and each momentum semantics in the semantic set, and the positive and negative sample probability distributions of the semantic set.

5. The semantic extraction method according to claim 4, characterized in that, The positive and negative sample probability distributions are determined based on the following steps: Based on the positions of the positive sample momentum semantics and negative sample momentum semantics in the semantic set, determine the positive and negative sample label representations; Based on the combinations of the text semantics with each momentum semantics in the semantic set, determine the fusion features of each combination; Based on the similarities between the positive and negative sample label representations and the fusion features of each combination, determine the positive and negative sample probability distributions.

6. A report generation method, characterized in that, comprising: performing semantic extraction on the original text based on the semantic extraction method described in any one of claims 1 to 5 to obtain the semantics of the original text; generating a report text corresponding to the original text based on the semantics of the original text.

7. A semantic extraction device, characterized in that, comprising: a determination module for determining the text to be extracted; an extraction module for performing semantic extraction on the text to be extracted based on the text semantics and semantic set of the sample text to obtain the text semantics of the text to be extracted, where the semantic set includes the positive sample momentum semantics and negative sample momentum semantics of the sample text; the text semantics and semantic set of the sample text are determined based on the following steps: performing semantic extraction on the current sample text to obtain the text semantics and positive sample momentum semantics of the sample text; updating the negative sample cache queue based on the positive sample momentum semantics, and determining the updated negative sample cache queue as the semantic set of the current sample text, and determining the next sample text of the current sample text as the current sample text.

8. A report generation device, characterized in that, comprising: a semantic extraction module for performing semantic extraction on the original text based on the semantic extraction method described in any one of claims 1 to 5 to obtain the semantics of the original text; a report generation module for generating a report text corresponding to the original text based on the semantics of the original text.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the steps of the semantic extraction method described in any one of claims 1 to 5 or implements the steps of the report generation method described in claim 6.

10. A non-transitory computer-readable storage medium, on which a computer program is stored, characterized in that, when the computer program is executed by a processor, it implements the steps of the semantic extraction method described in any one of claims 1 to 5 or implements the steps of the report generation method described in claim 6.

Images