Multilingual automatic summary generation method, device, equipment and storage medium

Abstract

The present disclosure provides a multilingual automatic summary generation method, including: serializing and encoding each type of language document in a multilingual document to generate a first hidden layer vector sequence and a second hidden layer vector sequence; converting the first hidden layer vector sequence and the second hidden layer vector sequence into a first context semantic vector and a second context semantic vector; obtaining a multilingual context semantic vector based on the first context semantic vector and the second context semantic vector; obtaining a first type of language-second type of language translation probability distribution based on an attention distribution between a first encoder and a decoder and a bilingual dictionary, and obtaining a second type of language replication probability distribution based on the attention distribution between a second encoder and a decoder; obtaining a neural probability distribution and a decision probability distribution based on the multilingual context semantic vector; performing weighted summation on the translation probability distribution, the replication probability distribution, and the neural probability distribution to obtain a vocabulary distribution probability, and generate a summary of the multilingual document.

Description

Technical Field

[0001] The present invention relates to a multi-language automatic summary generation method, device, equipment and storage medium. Background Art

[0002] With the acceleration of global information exchange, the Internet is flooded with a large amount of content in multiple languages. Many users can absorb and understand information in their native language at a relatively fast speed, but it is a little difficult to browse news content in a non-native language. In addition, content in different languages ​​under the same news topic can provide different perspectives, thus helping users to better understand the overall picture of an event.

[0003] Multilingual automatic summarization can help users obtain core content from multilingual text information more quickly, which is particularly necessary and urgent in today's globalized environment. Summary of the invention

[0004] In order to solve at least one of the above technical problems, the present disclosure provides a multi-language automatic summary generation method, apparatus, device and storage medium.

[0005] The multi-language automatic summary generation method, device, equipment and storage medium disclosed in the present invention are implemented by the following technical solutions:

[0006] According to one aspect of the present disclosure, a multilingual automatic summary generation method is provided, comprising: S1, serializing and encoding each type of language document in a multilingual document, generating at least a first type of language hidden vector sequence and a second type of language hidden vector sequence, wherein the multilingual document at least includes a first type of language document and a second type of language document; S2, converting at least the first type of language hidden vector sequence into a first type of language context semantic vector and converting the second type of language hidden vector sequence into a second type of language context semantic vector; S3, fusing at least the first type of language context semantic vector and the second type of language context semantic vector to obtain a multilingual context semantic vector; S4, based at least on The attention distribution between the first language encoder and the decoder and the first language-second language dictionary obtain the translation probability distribution of the first language-second language, and the copy probability distribution of the second language is obtained based on the attention distribution between the second language encoder and the decoder; S5, based on the multilingual context semantic vector, a neural probability distribution and a decision probability distribution are obtained, and the decision probability distribution is the weight between the translation probability distribution, the copy probability distribution and the neural probability distribution; and S6, at least a weighted summation is performed on the translation probability distribution, the copy probability distribution and the neural probability distribution to obtain a vocabulary distribution probability, and at least based on the vocabulary distribution probability, a summary of the multilingual document is generated.

[0007] According to at least one embodiment of the multilingual automatic summary generation method of the present disclosure, step S1 includes: S11, based on a multilingual corpus, at least a first language high-frequency vocabulary and a second language high-frequency vocabulary are obtained; S12, based on the first language high-frequency vocabulary and the index of each word in the first language document, all words in the first language document are converted into word vectors, and based on the second language high-frequency vocabulary and the index of each word in the second language document, all words in the second language document are converted into word vectors; and S13, using a first language encoder to encode all word vectors of the first language document into a first language hidden layer vector sequence, and using a second language encoder to encode all word vectors of the second language document into a second language hidden layer vector sequence.

[0008] According to the multi-language automatic summary generation method of at least one embodiment of the present disclosure, step S2 includes: calculating the "first type of language encoder-decoder" attention distribution between the first type of language encoder and decoder, using the attention distribution as the weight of the first type of language hidden layer vector sequence to perform weighted summation on the first type of language hidden layer vector sequence, and obtaining the first type of language context semantic vector; calculating the "second type of language encoder-decoder" attention distribution between the second type of language encoder and decoder, using the attention distribution as the weight of the second type of language hidden layer vector sequence to perform weighted summation on the second type of language hidden layer vector sequence, and obtaining the second type of language context semantic vector.

[0009] According to the multilingual automatic summary generation method of at least one embodiment of the present disclosure, step S3 includes: using the matrix parameters of the first type of language context semantic vector and the matrix parameters of the second type of language context semantic vector as the linear weights of the first type of language context semantic vector and the second type of language context semantic vector respectively, to calculate the multilingual context semantic vector.

[0010] According to the multilingual automatic summary generation method of at least one embodiment of the present disclosure, in step S4, the first-category language-second-category language translation probability distribution is obtained based on at least the attention distribution between the first-category language encoder and the decoder and the first-category language-second-category language dictionary, including: multiplying the attention distribution between the first-category language encoder and the decoder by the vocabulary translation probability in the first-category language-second-category language dictionary to obtain the translation probability distribution.

[0011] According to at least one embodiment of the multilingual automatic summary generation method of the present disclosure, the attention distribution between the encoder and the decoder of the first language is multiplied by the vocabulary translation probability in the dictionary of the first language and the second language to obtain the translation probability distribution, including: segmenting the bilingual parallel corpus of machine translation from the first language to the second language, extracting a probabilistic bilingual dictionary from the bilingual parallel corpus after the segmentation using a word alignment tool, and multiplying the vocabulary translation probability in the probabilistic bilingual dictionary by the attention distribution between the encoder and the decoder of the first language to obtain the translation probability distribution.

[0012] According to the multilingual automatic summary generation method of at least one embodiment of the present disclosure, in step S4, the replication probability distribution of the second language is obtained based on the attention distribution between the encoder and decoder of the second language, including: converting the attention distribution of the second language source end into the replication probability distribution of the vocabulary.

[0013] According to at least one embodiment of the multilingual automatic summary generation method of the present disclosure, the attention distribution of the second language source end is converted into the replication probability distribution of the vocabulary, including: indexing the second language vocabulary contained in the source end according to a preset second language vocabulary, treating the words not in the vocabulary as out-of-vocabulary words, and marking the out-of-vocabulary words with an additional index sequence; establishing a mapping dictionary according to the index of the out-of-vocabulary words, in which the index sequence number is the key value and the vocabulary is the value; and using the attention distribution between the second language encoder and decoder as the replication probability distribution of the second language source end.

[0014] According to at least one embodiment of the multilingual automatic summary generation method of the present disclosure, in step S5, obtaining a neural probability distribution based on the multilingual context semantic vector includes:

[0015] On the basis of the multilingual context semantic vector, a linear layer is added, and the neural probability distribution is obtained by the Softmax function. The calculation formula is as follows:

[0016] P N =softmax(W V Z MC +b); where W V a matrix parameter representing the importance of the multilingual context semantic vector when predicting the neural probability distribution, b is a bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, and Z MC represents the multilingual context semantic vector, P N represents a neural probability distribution.

[0017] According to at least one embodiment of the present disclosure, the multilingual automatic summary generation method obtains the decision probability distribution based on the multilingual context semantic vector, including: adding two linear layers on the basis of the multilingual context semantic vector, and obtaining the decision probability distribution by the Softmax function, and the calculation formula is as follows:

[0018] [p n , p c , p t ]=softmax(W2(W1Z MC +b1)+b2); wherein W1 represents the first matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b1 is the first bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, W2 represents the second matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b2 is the second bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, [p n ,p c ,p t ] is the decision probability distribution, p n represents the probability that the final vocabulary generation comes from the neural probability distribution, p c represents the probability that the final vocabulary generation comes from the above-mentioned replication probability distribution, p t Represents the probability that the final vocabulary generation comes from the translation probability distribution.

[0019] According to the multilingual automatic summary generation method of at least one embodiment of the present disclosure, in step S6, at least the translation probability distribution, the copy probability distribution and the neural probability distribution are weightedly summed to obtain the vocabulary distribution probability, including: fusing the out-of-collection words with the preset second category language vocabulary to construct a temporary extended vocabulary; performing weighted summation of the translation probability distribution, the copy probability distribution and the neural probability distribution according to the decision probability distribution, and corresponding the probability distribution to the extended vocabulary to obtain the vocabulary distribution probability.

[0020] According to the multilingual automatic summary generation method of at least one embodiment of the present disclosure, in step S6, a summary of the multilingual document is generated based at least on the vocabulary distribution probability, including: using a beam search algorithm to predict summary vocabulary based on the obtained vocabulary distribution probability, thereby decoding and generating a summary.

[0021] According to the multilingual automatic summary generation method of at least one embodiment of the present disclosure, the first type of language document is a Chinese language document, and the second type of language document is an English language document.

[0022] According to the multilingual automatic summary generation method of at least one embodiment of the present disclosure, the language type of the summary of the multilingual document is the second type of language.

[0023] According to the multilingual automatic summary generation method of at least one embodiment of the present disclosure, a word segmentation tool (such as Jieba tool) is used to segment the language documents that need word segmentation.

[0024] According to the multi-language automatic summary generation method of at least one embodiment of the present disclosure, the first type of language encoder is a self-attention mechanism encoder, and the second type of language encoder is a self-attention mechanism encoder.

[0025] According to another aspect of the present disclosure, a multilingual automatic summary generation device is provided, including: a hidden vector sequence generation module, the hidden vector sequence generation module serializes and encodes each type of language document in a multilingual document, and generates at least a first type of language hidden vector sequence and a second type of language hidden vector sequence, wherein the multilingual document at least includes a first type of language document and a second type of language document; a hidden vector sequence-context semantic vector conversion module, the hidden vector sequence-context semantic vector conversion module at least converts the first type of language hidden vector sequence into a first type of language context semantic vector and converts the second type of language hidden vector sequence into a second type of language context semantic vector; a multilingual context semantic vector generation module, the multilingual context semantic vector generation module at least fuses the first type of language context semantic vector and the second type of language context semantic vector to obtain a multilingual context semantic vector; a translation probability acquisition module, the translation probability acquisition module The extraction module obtains the translation probability distribution of the first type of language-the second type of language based on at least the attention distribution between the first type of language encoder and the decoder and the first type of language-the second type of language dictionary; the copy probability acquisition module, the copy probability acquisition module obtains the copy probability distribution of the second type of language based on the attention distribution between the second type of language encoder and the decoder; the neural probability distribution and decision probability distribution acquisition module, the neural probability distribution and decision probability distribution acquisition module obtains the neural probability distribution and the decision probability distribution based on the multilingual context semantic vector, the decision probability distribution is the weight between the translation probability distribution, the copy probability distribution and the neural probability distribution; the vocabulary distribution probability acquisition module, the vocabulary distribution probability acquisition module at least performs weighted summation on the translation probability distribution, the copy probability distribution and the neural probability distribution to obtain the vocabulary distribution probability; and the summary generation module, the summary generation module generates the summary of the multilingual document based on at least the vocabulary distribution probability.

[0026] According to another aspect of the present disclosure, an electronic device is provided, comprising: a memory storing execution instructions; and a processor executing the execution instructions stored in the memory, so that the processor executes any one of the methods described above.

[0027] According to another aspect of the present disclosure, a readable storage medium is provided, wherein the readable storage medium stores execution instructions, and the execution instructions are used to implement any of the above-mentioned methods when executed by a processor. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings illustrate exemplary embodiments of the present disclosure and together with the description serve to explain the principles of the present disclosure. These drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification.

[0029] Figure 1 The present invention is a flowchart of a multi-language automatic summary generation method according to an embodiment of the present invention.

[0030] Figure 2 The present invention is a flowchart of a method for generating a multilingual summary according to another embodiment of the present invention.

[0031] Figure 3 The present invention is a block diagram showing the structure of a multi-language automatic summary generation device according to an embodiment of the present invention.

[0032] Figure 4 An example diagram of a multi-language automatic summary generation apparatus using a hardware implementation of a processing system is shown.

[0033] Description of Reference Numerals

[0034] 1000 Multi-language automatic abstract generation device

[0035] 1001 Hidden layer vector sequence generation module

[0036] 1002 Hidden layer vector sequence-context semantic vector conversion module

[0037] 1003 Multilingual contextual semantic vector generation module

[0038] 1004 Translation Probability Acquisition Module

[0039] 1005 Copy Probability Acquisition Module

[0040] 1006 Neural Probability Distribution and Decision Probability Distribution Acquisition Module

[0041] 1007 Vocabulary Distribution Probability Acquisition Module

[0042] 1008 Summary Generation Module

[0043] 1100 Bus

[0044] 1200 processor

[0045] 1300 Memory

[0046] 1400 Other circuits. DETAILED DESCRIPTION

[0047] The present disclosure is further described in detail below in conjunction with the accompanying drawings and implementations. It is understood that the specific implementations described herein are only used to explain the relevant content, rather than to limit the present disclosure. It should also be noted that, for ease of description, only the parts related to the present disclosure are shown in the accompanying drawings.

[0048] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in the present disclosure can be combined with each other. The technical solution of the present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0049] Unless otherwise specified, the exemplary embodiments / embodiments shown will be understood as providing exemplary features of various details of some ways in which the technical concept of the present disclosure can be implemented in practice. Therefore, unless otherwise specified, the features of the various embodiments / embodiments can be combined, separated, interchanged and / or rearranged without departing from the technical concept of the present disclosure.

[0050] The use of cross-hatching and / or shading in the accompanying drawings is generally used to make the boundaries between adjacent components clear. As such, unless otherwise specified, the presence or absence of cross-hatching or shading does not convey or indicate any preference or requirement for the specific materials, material properties, dimensions, proportions, commonalities between the components shown, and / or any other characteristics, attributes, properties, etc. of the components. In addition, in the accompanying drawings, the sizes and relative sizes of the components may be exaggerated for clarity and / or descriptive purposes. When the exemplary embodiments can be implemented differently, the specific process sequence can be performed in a different order than described. For example, two successively described processes can be performed substantially simultaneously or in an order opposite to the described order. In addition, the same figure numbers represent the same components.

[0051] When a component is referred to as being "on" or "over," "connected to," or "coupled to" another component, the component may be directly on, directly connected to, or directly coupled to the other component, or intervening components may be present. However, when a component is referred to as being "directly on," "directly connected to," or "directly coupled to" another component, there are no intervening components. For this purpose, the term "connected" may refer to a physical connection, an electrical connection, etc., with or without intervening components.

[0052] For descriptive purposes, the present disclosure may use spatially relative terms such as "under," "beneath," "under," "down," "over," "upper," "above," "higher," and "side (e.g., as in "sidewall")," to describe the relationship of one component to another (other) component as shown in the accompanying drawings. The spatially relative terms are intended to encompass different orientations of the device in use, operation, and / or manufacture in addition to the orientation depicted in the accompanying drawings. For example, if the device in the accompanying drawings is turned over, components described as "under" or "beneath" other components or features would subsequently be positioned "over" the other components or features. Thus, the exemplary term "under" can encompass both the "above" and "below" orientations. Furthermore, the device may be otherwise oriented (e.g., rotated 90 degrees or at other orientations), and the spatially relative descriptors used herein should be interpreted accordingly.

[0053] The terms used here are for the purpose of describing specific embodiments, and are not intended to be restrictive. As used here, unless the context clearly indicates otherwise, the singular forms "one (kind, person)" and "said (the)" are also intended to include plural forms. In addition, when the terms "comprise" and / or "include" and their variations are used in this specification, it is explained that there are stated features, integral bodies, steps, operations, parts, assemblies and / or their groups, but it is not excluded that there are or add one or more other features, integral bodies, steps, operations, parts, assemblies and / or their groups. It should also be noted that, as used here, the terms "substantially", "approximately" and other similar terms are used as approximate terms and not as degree terms, so that they are used to explain the inherent deviations of the measured values, calculated values ​​and / or the values ​​provided that will be recognized by those of ordinary skill in the art.

[0054] Figure 1 The present invention is a flowchart of a multi-language automatic summary generation method according to an embodiment of the present invention.

[0055] like Figure 1 As shown, the multi-language automatic summary generation method includes:

[0056] S1. Serialize and encode each language document in a multilingual document to generate at least a first language hidden layer vector sequence and a second language hidden layer vector sequence, wherein the multilingual document includes at least the first language document and the second language document;

[0057] S2, at least converting the first type of language hidden vector sequence into the first type of language context semantic vector and converting the second type of language hidden vector sequence into the second type of language context semantic vector;

[0058] S3, fusing at least the first type of language context semantic vector and the second type of language context semantic vector to obtain a multilingual context semantic vector;

[0059] S4, obtaining a translation probability distribution of the first language-the second language based at least on the attention distribution between the first language encoder and the decoder and the first language-the second language dictionary, and obtaining a copy probability distribution of the second language based on the attention distribution between the second language encoder and the decoder;

[0060] S5. Obtaining a neural probability distribution and a decision probability distribution based on the multilingual context semantic vector, where the decision probability distribution is a weight among the translation probability distribution, the copy probability distribution, and the neural probability distribution; and

[0061] S6. Perform a weighted summation on at least the translation probability distribution, the copy probability distribution, and the neural probability distribution to obtain a vocabulary distribution probability, and generate a summary of the multilingual document based at least on the vocabulary distribution probability.

[0062] The multilingual document may include two-language documents, three-language documents, or more-language documents, for example, a multilingual document may include both Chinese documents and English documents, or a multilingual document may include Chinese documents, Japanese documents, and English documents, and so on.

[0063] Through the multilingual automatic summary generation method disclosed in the present invention, the finally generated monolingual summary can integrate the semantic information of multiple languages ​​at the source end.

[0064] According to a preferred embodiment of the present disclosure, Figure 2 As shown, the multi-language automatic summary generation method includes:

[0065] S11, obtaining at least a first language high frequency vocabulary list and a second language high frequency vocabulary list based on the multilingual corpus;

[0066] S12, based on the high-frequency vocabulary of the first language and the index of each word in the document of the first language, convert all words in the document of the first language into word vectors, and based on the high-frequency vocabulary of the second language and the index of each word in the document of the second language, convert all words in the document of the second language into word vectors;

[0067] S13, using a first language encoder to encode all word vectors of a first language document into a first language hidden vector sequence, and using a second language encoder to encode all word vectors of a second language document into a second language hidden vector sequence;

[0068] S2, at least converting the first type of language hidden vector sequence into the first type of language context semantic vector and converting the second type of language hidden vector sequence into the second type of language context semantic vector;

[0069] S3, fusing at least the first type of language context semantic vector and the second type of language context semantic vector to obtain a multilingual context semantic vector;

[0070] S4, obtaining a translation probability distribution of the first language-the second language based at least on the attention distribution between the first language encoder and the decoder and the first language-the second language dictionary, and obtaining a copy probability distribution of the second language based on the attention distribution between the second language encoder and the decoder;

[0071] S5. Obtaining a neural probability distribution and a decision probability distribution based on the multilingual context semantic vector, where the decision probability distribution is a weight among the translation probability distribution, the copy probability distribution, and the neural probability distribution; and

[0072] S6. Perform a weighted summation on at least the translation probability distribution, the copy probability distribution, and the neural probability distribution to obtain a vocabulary distribution probability, and generate a summary of the multilingual document based at least on the vocabulary distribution probability.

[0073] In each of the above-mentioned embodiments, preferably, step S2 of the multilingual automatic summary generation method includes: calculating the "first type of language encoder-decoder" attention distribution between the first type of language encoder and decoder, using the attention distribution as the weight of the first type of language hidden layer vector sequence to perform weighted summation on the first type of language hidden layer vector sequence, and obtaining the first type of language context semantic vector; calculating the "second type of language encoder-decoder" attention distribution between the second type of language encoder and decoder, using the attention distribution as the weight of the second type of language hidden layer vector sequence to perform weighted summation on the second type of language hidden layer vector sequence, and obtaining the second type of language context semantic vector.

[0074] In each of the above-mentioned embodiments, preferably, step S3 of the multilingual automatic summary generation method includes: using the matrix parameters of the first type of language context semantic vector and the matrix parameters of the second type of language context semantic vector as the linear weights of the first type of language context semantic vector and the second type of language context semantic vector, respectively, to calculate a multilingual context semantic vector.

[0075] In each of the above-mentioned embodiments, preferably, in step S4, the first-type language-second-type language translation probability distribution is obtained based at least on the attention distribution between the first-type language encoder and the decoder and the first-type language-second-type language dictionary, including: multiplying the attention distribution between the first-type language encoder and the decoder with the vocabulary translation probability in the first-type language-second-type language dictionary to obtain the translation probability distribution.

[0076] In the above implementation, preferably, the attention distribution between the encoder and the decoder of the first language is multiplied by the vocabulary translation probability in the first language-second language dictionary to obtain the translation probability distribution, including:

[0077] The bilingual parallel corpus of machine translation from the first language to the second language is segmented, and a probabilistic bilingual dictionary is extracted from the bilingual parallel corpus after segmentation using a word alignment tool. The vocabulary translation probability in the probabilistic bilingual dictionary is multiplied by the attention distribution between the encoder and decoder of the first language to obtain the translation probability distribution.

[0078] In each of the above-mentioned embodiments, preferably, in step S4 of the multilingual automatic summary generation method, the replication probability distribution of the second language is obtained based on the attention distribution between the second language encoder and the decoder, including: converting the second language source-end attention distribution into the replication probability distribution of vocabulary.

[0079] In the above implementation, the attention distribution of the second language source is converted into the replication probability distribution of the vocabulary, including: indexing the second language vocabulary contained in the source according to a preset second language vocabulary, treating the words not in the vocabulary as out-of-vocabulary words, and marking the out-of-vocabulary words with an additional index sequence; establishing a mapping dictionary according to the index of the out-of-vocabulary words, in which the index number is the key value and the vocabulary is the value; and using the attention distribution between the second language encoder and decoder as the replication probability distribution of the second language source.

[0080] In each of the above implementations, preferably, in step S5 of the multilingual automatic summary generation method, obtaining a neural probability distribution based on the multilingual context semantic vector includes:

[0081] On the basis of the multilingual context semantic vector, a linear layer is added, and the neural probability distribution is obtained by the Softmax function. The calculation formula is as follows:

[0082] P N =softmax(W V Z MC +b); where W V ∈ R is a matrix parameter representing the importance of the multilingual contextual semantic vector when predicting the neural probability distribution, b is a bias parameter representing the importance of the multilingual contextual semantic vector when predicting the neural probability distribution, and Z MC represents the multilingual context semantic vector, P N represents a neural probability distribution.

[0083] In each of the above implementations, preferably, in step S5 of the multilingual automatic summary generation method, obtaining a decision probability distribution based on the multilingual context semantic vector includes:

[0084] On the basis of the multilingual context semantic vector, two linear layers are added, and the decision probability distribution is obtained by the Softmax function. The calculation formula is as follows:

[0085] [p n , p c , p t ]=softmax(W2(W1Z MC +b1)+b2); wherein W1 represents the first matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b1 is the first bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, W2 represents the second matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b2 is the second bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, [p n ,p c ,p t ] is the decision probability distribution, p n represents the probability that the final vocabulary generation comes from the neural probability distribution, p c represents the probability that the final vocabulary generation comes from the copy probability distribution, p t Represents the probability that the final vocabulary generation comes from the translation probability distribution.

[0086] In each of the above-mentioned embodiments, preferably, in step S6 of the multilingual automatic summary generation method, at least a weighted summation is performed on the translation probability distribution, the copy probability distribution and the neural probability distribution to obtain a vocabulary distribution probability, including: fusing the out-of-collection words with a preset second category language vocabulary to construct a temporary extended vocabulary; performing a weighted summation of the translation probability distribution, the copy probability distribution and the neural probability distribution according to the decision probability distribution, and corresponding the probability distribution to the extended vocabulary to obtain a vocabulary distribution probability.

[0087] In each of the above-mentioned embodiments, preferably, in step S6 of the multilingual automatic summary generation method, a summary of the multilingual document is generated based at least on the vocabulary distribution probability, including: using a beam search algorithm to predict summary vocabulary based on the obtained vocabulary distribution probability, thereby decoding and generating a summary.

[0088] In the above-mentioned various implementation modes, the first type of language document is, for example, a Chinese language document, and the second type of language document is, for example, an English language document.

[0089] In each of the above-mentioned embodiments, the language type of the summary of the multilingual document is the second language type.

[0090] In each of the above implementations, a word segmentation tool (such as Jieba tool) is used to segment the language document that needs word segmentation.

[0091] In each of the above-mentioned embodiments, the first type of language encoder is a self-attention mechanism encoder, and the second type of language encoder is a self-attention mechanism encoder.

[0092] The following is a specific example to illustrate the multilingual automatic summary generation method disclosed in the present invention:

[0093] For example, given an input English document and an input Chinese document:

[0094] "It's just an example for illustration. There are 56 nationalities in China."

[0095] "This example is for illustration only. China consists of 56 ethnic groups."

[0096] First, the Chinese documents and English documents are serialized and encoded by two language-specific encoders (Chinese encoder and English encoder) to generate Chinese hidden vector sequences and English hidden vector sequences. Specifically, the following steps can be used:

[0097] Use Jieba tool to segment Chinese text and convert all the contents in English text into lowercase. You can use open source Chinese segmentation tool to segment Chinese text and use Python language to directly convert English text into lowercase. Chinese segmentation tool belongs to the prior art.

[0098] After word segmentation and lowercase conversion, the input English document and Chinese document can be represented as follows:

[0099] "it's just an example for illustration.there are 56nationalities inchina."

[0100] "This example is for illustration only. China consists of 56 ethnic groups."

[0101] The N (eg, N=100,000) words with the highest frequency in each language are calculated from all Chinese and English texts, and these N words are used as the vocabulary lists of Chinese and English, respectively.

[0102] Based on the Chinese and English vocabulary tables, these words are converted into word vectors according to the index of each word in the Chinese and English input documents.

[0103] Two self-attention mechanism encoders are used to encode Chinese and English word vectors into Chinese and English hidden vector sequences respectively.

[0104] Next, the Chinese and English hidden vector sequences are converted into Chinese and English contextual semantic vectors through the attention mechanism. Specifically, the following steps can be taken:

[0105] The attention distribution of the "Chinese encoder-decoder" is calculated between the Chinese encoder and decoder, and the attention distribution is used as the weight of the Chinese hidden layer vector sequence to perform weighted summation on the vector sequence to obtain the Chinese context semantic vector. The calculation formula is as follows:

[0106]

[0107]

[0108]

[0109] Where Q, K, and V represent the matrices of "query", "key", and "value" respectively, which constitute the calculation method of the attention mechanism. k Represents the dimension of the hidden layer vector sequence; represents the current state vector of the decoder (English), Represents the matrix composed of the hidden vector sequence of the Chinese encoder; represents the attention distribution obtained for the Chinese source, H Zh It represents the Chinese document context semantic vector.

[0110] The attention distribution of the "English encoder-decoder" is calculated between the English encoder and decoder, and the attention distribution is used as the weight of the English hidden layer vector sequence to perform weighted summation on the vector sequence to obtain the English context semantic vector. The calculation formula is as follows:

[0111]

[0112]

[0113] in represents the current state vector of the decoder (English), Represents the matrix composed of the hidden vector sequence of the English encoder; represents the obtained attention distribution for the English source, H En It represents the English document context semantic vector.

[0114] The Chinese and English context semantic vectors are fused using the multilingual context layer to obtain a multilingual context semantic vector. Specifically, the following steps can be used:

[0115] Two matrix parameters are used as the linear weights of the Chinese context semantic vector and the English context semantic vector to calculate the multilingual context semantic vector. The calculation formula is as follows:

[0116] Z MC =W En H En +W Zh H Zh

[0117] Where W En Represents the matrix parameter of the contextual semantic vector of the English document, which is used to measure the importance of the semantic vector. Zh The matrix parameter representing the contextual semantic vector of the Chinese document is used to measure the importance of the semantic vector; Z MC Represents the calculated multilingual context semantic vector.

[0118] The decoder calculates a translation probability distribution based on the attention distribution between the Chinese encoder and decoder and a bilingual dictionary. Specifically, the following steps can be used:

[0119] The Chinese-to-English machine translation bilingual parallel corpus is segmented according to the Chinese and English segmentation granularity in the above steps; a probabilistic bilingual dictionary is extracted from the segmented Chinese-English parallel corpus using a word alignment tool; the translation probability of the vocabulary in the probabilistic bilingual dictionary is multiplied by the attention distribution between the Chinese encoder and the decoder to obtain the translation probability distribution. The specific calculation formula is as follows:

[0120]

[0121] Where P T (w) represents the value of w in the translation probability distribution in the vocabulary, Indicates w i The value of attention distribution on the English source side, represents the word w in the probabilistic bilingual dictionary in the above step src The probability of translating to w.

[0122] The decoder calculates a replication probability distribution based on the attention distribution between the English encoder and the decoder. Specifically, the following steps can be used:

[0123] The English source attention distribution is converted into the word copy probability distribution. The calculation formula is as follows:

[0124]

[0125] Where P C (w i ) represents the English source word w i The probability of replication, Indicates w i The value of attention distribution on the English source side.

[0126] The decoder calculates a neural probability distribution and a decision probability distribution based on the multilingual context semantic vector. The decision probability distribution represents the weights between the translation probability distribution, the copy probability distribution, and the neural probability distribution. Specifically, the following steps can be used:

[0127] On the basis of the multilingual context semantic vector, a linear layer is added, and a neural probability distribution is obtained by the Softmax function. The calculation formula is as follows:

[0128] P N =softmax(W V Z MC +b)

[0129] Where W V The matrix parameter representing the importance of the multilingual contextual semantic vector when predicting the neural probability distribution, b is the bias parameter of the importance of the multilingual contextual semantic vector when predicting the neural probability distribution, PN represents a neural probability distribution.

[0130] On the basis of the multilingual context semantic vector, two linear layers are added, and a decision probability distribution is obtained by the Softmax function. The calculation formula is as follows:

[0131] [p n , p c , p t =softmax(W2(W1Z MC +b1)+b2)

[0132] Wherein, W1 represents the first matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b1 is the first bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, W2 represents the second matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b2 is the second bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, [p n ,p c ,p t ] is the decision probability distribution, p n represents the probability that the final vocabulary generation comes from the neural probability distribution, p c represents the probability that the final vocabulary generation comes from the copy probability distribution, p t Represents the probability that the final vocabulary generation comes from the translation probability distribution.

[0133] The translation probability distribution, the copy probability distribution and the neural probability distribution are weighted and summed to predict the summary words at each moment and decode them. Specifically, the following steps can be adopted:

[0134] The out-of-vocabulary words in the above steps are merged with the preset vocabulary to construct a temporary extended vocabulary, whose size is the sum of the original vocabulary size and the additional index sequence table size; the translation probability distribution, copy probability distribution and neural probability distribution are weighted and summed according to the decision probability distribution, and the probability distribution is mapped to the extended vocabulary; the calculation formula is as follows:

[0135] P(w)=p n P N (w)+p c P C (w)+p t P T (w)

[0136] Where P(w) represents the probability of word w in the final vocabulary generation prediction distribution, P N (w) represents the probability of word w in the neural probability distribution, P C(w) represents the probability of word w in the replication distribution, P T (w) represents the probability of word w in the translation probability distribution.

[0137] The beam search algorithm is used to predict summary vocabulary based on the obtained vocabulary distribution probability, thereby decoding and generating a summary.

[0138] Figure 3 It is a schematic block diagram of the structure of a multi-language automatic summary generation device according to an embodiment of the present disclosure.

[0139] like Figure 3 As shown, the multi-language automatic summary generation device 1000 includes:

[0140] A hidden vector sequence generation module 1001, wherein the hidden vector sequence generation module 101 performs serialization encoding on each type of language document in the multilingual document, and generates at least a first type of language hidden vector sequence and a second type of language hidden vector sequence, wherein the multilingual document at least includes the first type of language document and the second type of language document;

[0141] A hidden vector sequence-context semantic vector conversion module 1002, which converts at least a first-category language hidden vector sequence into a first-category language context semantic vector and a second-category language hidden vector sequence into a second-category language context semantic vector;

[0142] A multilingual context semantic vector generation module 1003, which fuses at least the first type of language context semantic vector and the second type of language context semantic vector to obtain a multilingual context semantic vector;

[0143] A translation probability acquisition module 1004, which acquires a translation probability distribution between the first language and the second language based on at least the attention distribution between the first language encoder and the decoder and the first language-second language dictionary;

[0144] A copy probability acquisition module 1005, which acquires a copy probability distribution of the second language based on the attention distribution between the encoder and the decoder of the second language;

[0145] A neural probability distribution and decision probability distribution acquisition module 1006, wherein the neural probability distribution and decision probability distribution acquisition module 1006 obtains a neural probability distribution and a decision probability distribution based on a multilingual context semantic vector, wherein the decision probability distribution is a weight among a translation probability distribution, a copy probability distribution and a neural probability distribution;

[0146] A vocabulary distribution probability acquisition module 1007, which performs a weighted summation on at least the translation probability distribution, the copy probability distribution and the neural probability distribution to obtain the vocabulary distribution probability; and

[0147] The summary generation module 1008 generates a summary of the multilingual document based at least on the vocabulary distribution probability.

[0148] Figure 4 An exemplary diagram of a multi-language automatic summary generation apparatus 1000 using a hardware implementation of a processing system is shown.

[0149] The device 1000 may include corresponding modules for executing each or several steps in the above flowchart. Therefore, each step or several steps in the above flowchart may be executed by the corresponding module, and the device may include one or more modules in these modules. The module may be one or more hardware modules specially configured to execute the corresponding steps, or implemented by a processor configured to execute the corresponding steps, or stored in a computer-readable medium for implementation by a processor, or implemented by some combination.

[0150] The hardware structure can be implemented using a bus architecture. The bus architecture can include any number of interconnecting buses and bridges, depending on the specific application and overall design constraints of the hardware. The bus 1100 connects various circuits including one or more processors 1200, memory 1300, and / or hardware modules together. The bus 1100 can also connect various other circuits 1400 such as peripherals, voltage regulators, power management circuits, external antennas, etc.

[0151] The bus 1100 may be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the figure only uses one connecting line, but does not mean that there is only one bus or one type of bus.

[0152] Any process or method description in the flowchart or otherwise described herein can be understood as a module, fragment or portion of a code representing one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present disclosure includes other implementations, in which the functions may not be performed in the order shown or discussed, including performing the functions in a substantially simultaneous manner or in a reverse order according to the functions involved, which should be understood by a person skilled in the art of the present disclosure. The processor performs the various methods and processes described above. For example, the method implementation in the present disclosure can be implemented as a software program, which is tangibly contained in a machine-readable medium, such as a memory. In some embodiments, part or all of the software program can be loaded and / or installed via a memory and / or a communication interface. When the software program is loaded into the memory and executed by the processor, one or more steps in the method described above can be performed. Alternatively, in other embodiments, the processor can be configured to perform one of the above methods in any other appropriate manner (e.g., by means of firmware).

[0153] The logic and / or steps represented in the flowchart or otherwise described herein may be embodied in any readable storage medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other system that can fetch instructions from and execute instructions on an instruction execution system, apparatus, or device), or in combination with such instruction execution systems, apparatuses, or devices.

[0154] For the purposes of this specification, a "readable storage medium" may be any device that can contain, store, communicate, propagate or transmit a program for use with or in conjunction with an instruction execution system, device or apparatus. More specific examples of readable storage media (a non-exhaustive list) include the following: an electrical connection with one or more wirings (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and editable read-only memory (EPROM or flash memory), a fiber optic device, and a portable read-only memory (CDROM). In addition, the readable storage medium may even be paper or other suitable medium on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a memory.

[0155] It should be understood that the various parts of the present disclosure can be implemented in hardware, software, or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented in software stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0156] A person skilled in the art may understand that all or part of the steps of implementing the above-mentioned implementation method may be completed by instructing related hardware through a program, and the program may be stored in a readable storage medium, which, when executed, includes one or a combination of the steps of the implementation method.

[0157] In addition, each functional unit in each embodiment of the present disclosure may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a readable storage medium. The storage medium may be a read-only memory, a disk or an optical disk, etc.

[0158] The present disclosure also provides an electronic device, including: a memory, the memory storing execution instructions; and a processor or other hardware module, the processor or other hardware module executes the execution instructions stored in the memory, so that the processor or other hardware module executes the above method.

[0159] The present disclosure also provides a readable storage medium, in which execution instructions are stored. When the execution instructions are executed by a processor, they are used to implement the above method.

[0160] In the description of this specification, the description with reference to the terms "one embodiment / method", "some embodiments / methods", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment / method or example are included in at least one embodiment / method or example of the present application. In this specification, the schematic representations of the above terms are not necessarily the same embodiment / method or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments / methods or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments / methods or examples described in this specification and the features of the different embodiments / methods or examples, unless they are contradictory.

[0161] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0162] Those skilled in the art should understand that the above embodiments are only for the purpose of clearly illustrating the present disclosure, and are not intended to limit the scope of the present disclosure. For those skilled in the art, other changes or modifications may be made based on the above disclosure, and these changes or modifications are still within the scope of the present disclosure.

Claims

1. A multi-language automatic summary generation method, characterized in that: include: S1. Serializing and encoding each language document in a multilingual document to generate at least a first language hidden layer vector sequence and a second language hidden layer vector sequence, wherein the multilingual document at least includes the first language document and the second language document; S2, at least converting the first type of language hidden vector sequence into the first type of language context semantic vector and converting the second type of language hidden vector sequence into the second type of language context semantic vector; S3, fusing at least the first category language context semantic vector and the second category language context semantic vector to obtain a multilingual context semantic vector; S4, obtaining a translation probability distribution of the first language-the second language based at least on the attention distribution between the first language encoder and the decoder and the first language-the second language dictionary, and obtaining a copy probability distribution of the second language based on the attention distribution between the second language encoder and the decoder; The attention distribution between the encoder and the decoder of the first language is multiplied by the vocabulary translation probability in the first language-second language dictionary to obtain the translation probability distribution, including: Segment the bilingual parallel corpus from the first language to the second language, extract a probabilistic bilingual dictionary from the bilingual parallel corpus after segmentation using a word alignment tool, and multiply the vocabulary translation probability in the probabilistic bilingual dictionary by the attention distribution between the encoder and decoder of the first language to obtain a translation probability distribution; The method converts the attention distribution of the second language source into the probability distribution of word replication, including: indexing the second language words contained in the source according to a preset second language vocabulary, treating words not in the vocabulary as out-of-vocabulary words, and marking the out-of-vocabulary words with an additional index sequence; establishing a mapping dictionary according to the index of the out-of-vocabulary words, wherein the index sequence number is the key value and the vocabulary is the value; and using the attention distribution between the second language encoder and the decoder as the probability distribution of replication of the second language source; S5. Obtaining a neural probability distribution and a decision probability distribution based on the multilingual context semantic vector, wherein the decision probability distribution is a weight among the translation probability distribution, the copy probability distribution and the neural probability distribution; Obtain decision probability distribution based on multilingual context semantic vectors, including: On the basis of the multilingual context semantic vector, two linear layers are added, and the decision probability distribution is obtained by the Softmax function. The calculation formula is as follows: [p n , p c , p t ]=softmax(W2(W1Z MC +b1)+b2); wherein W1 represents the first matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b1 is the first bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, W2 represents the second matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b2 is the second bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, [p n ,p c ,p t ] is the decision probability distribution, p n represents the probability that the final vocabulary generation comes from the neural probability distribution, p c represents the probability that the final vocabulary generation comes from the copy probability distribution, p t represents the probability that the final vocabulary generation comes from the translation probability distribution; S6, performing a weighted summation on at least the translation probability distribution, the copy probability distribution and the neural probability distribution to obtain a vocabulary distribution probability, and generating a summary of the multilingual document based at least on the vocabulary distribution probability; Obtain neural probability distribution based on multilingual contextual semantic vectors, including: On the basis of the multilingual context semantic vector, a linear layer is added, and the neural probability distribution is obtained by the Softmax function. The calculation formula is as follows: P N =softmax(W V Z MC +b); where W V The matrix parameter representing the importance of the multilingual contextual semantic vector when predicting the neural probability distribution, b is the bias parameter of the importance of the multilingual contextual semantic vector when predicting the neural probability distribution, and Z MC Represents the multilingual context semantic vector, P N represents a neural probability distribution.

2. The multilingual automatic summary generation method according to claim 1, characterized in that: Step S1 includes: S11, obtaining at least a first language high frequency vocabulary list and a second language high frequency vocabulary list based on the multilingual corpus; S12: based on the first language high frequency vocabulary and the index of each word in the first language document, convert all words in the first language document into word vectors; based on the second language high frequency vocabulary and the index of each word in the second language document, convert all words in the second language document into word vectors; and S13. Use a first language encoder to encode all word vectors of the first language document into a first language hidden layer vector sequence, and use a second language encoder to encode all word vectors of the second language document into a second language hidden layer vector sequence.

3. The multilingual automatic summary generation method according to claim 2, characterized in that: Step S2 includes: The "first type language encoder-decoder" attention distribution is calculated between the first type language encoder and decoder, and the attention distribution is used as the weight of the first type language hidden layer vector sequence to perform weighted summation on the first type language hidden layer vector sequence to obtain the first type language context semantic vector; the "second type language encoder-decoder" attention distribution is calculated between the second type language encoder and decoder, and the attention distribution is used as the weight of the second type language hidden layer vector sequence to perform weighted summation on the second type language hidden layer vector sequence to obtain the second type language context semantic vector.

4. The multilingual automatic summary generation method according to any one of claim 3, characterized in that: Step S3 includes: The multilingual context semantic vector is calculated by using the matrix parameters of the first type of language context semantic vector and the matrix parameters of the second type of language context semantic vector as the linear weights of the first type of language context semantic vector and the second type of language context semantic vector respectively.

5. The multilingual automatic summary generation method according to any one of claim 4, characterized in that: In step S4, the translation probability distribution between the first language and the second language is obtained based at least on the attention distribution between the first language encoder and the decoder and the first language-second language dictionary, including: The attention distribution between the first language encoder and the decoder is multiplied by the vocabulary translation probability in the first language-second language dictionary to obtain the translation probability distribution.

6. The multilingual automatic summary generation method according to claim 5, characterized in that: The attention distribution between the first language encoder and the decoder is multiplied by the vocabulary translation probability in the first language-second language dictionary to obtain the translation probability distribution, including: The bilingual parallel corpus of machine translation from the first language to the second language is segmented, a probabilistic bilingual dictionary is extracted from the bilingual parallel corpus after the segmentation using a word alignment tool, and the vocabulary translation probability in the probabilistic bilingual dictionary is multiplied by the attention distribution between the encoder and the decoder of the first language to obtain the translation probability distribution.

7. The multilingual automatic summary generation method according to any one of claim 6, characterized in that: In step S4, the copy probability distribution of the second language is obtained based on the attention distribution between the encoder and the decoder of the second language, including: The second type of language source attention distribution is converted into the replication probability distribution of the vocabulary.

8. A multi-language automatic summary generation device, characterized in that: include: A hidden vector sequence generation module, wherein the hidden vector sequence generation module performs serialization encoding on each type of language document in the multilingual document to generate at least a first type of language hidden vector sequence and a second type of language hidden vector sequence, wherein the multilingual document at least includes the first type of language document and the second type of language document; A hidden layer vector sequence-context semantic vector conversion module, wherein the hidden layer vector sequence-context semantic vector conversion module at least converts a first type of language hidden layer vector sequence into a first type of language context semantic vector and converts a second type of language hidden layer vector sequence into a second type of language context semantic vector; a multilingual context semantic vector generation module, which fuses at least the first type of language context semantic vector and the second type of language context semantic vector to obtain a multilingual context semantic vector; A translation probability acquisition module, wherein the translation probability acquisition module acquires a translation probability distribution between the first language and the second language based on at least the attention distribution between the first language encoder and the decoder and the first language-second language dictionary, wherein obtaining the translation probability distribution includes: performing word segmentation on a bilingual parallel corpus for machine translation from the first language to the second language, extracting a probabilistic bilingual dictionary from the bilingual parallel corpus after word segmentation using a word alignment tool, and multiplying the vocabulary translation probability in the probabilistic bilingual dictionary by the attention distribution between the first language encoder and the decoder to obtain the translation probability distribution; A replication probability acquisition module, wherein the replication probability acquisition module acquires the replication probability distribution of the second language based on the attention distribution between the second language encoder and the decoder, wherein the attention distribution of the second language source end is converted into the replication probability distribution of the vocabulary, including: indexing the second language vocabulary contained in the source end according to a preset second language vocabulary table, and the words not in the vocabulary table are regarded as out-of-vocabulary words, and the out-of-vocabulary words are marked with an additional index sequence; establishing a mapping dictionary according to the index of the out-of-vocabulary words, wherein the index sequence number is the key value and the vocabulary is the value; and using the attention distribution between the second language encoder and the decoder as the replication probability distribution of the second language source end; A neural probability distribution and decision probability distribution acquisition module, wherein the neural probability distribution and decision probability distribution acquisition module obtains a neural probability distribution and a decision probability distribution based on the multilingual context semantic vector, wherein the decision probability distribution is a weight among the translation probability distribution, the copy probability distribution and the neural probability distribution; On the basis of the multilingual context semantic vector, a linear layer is added, and the neural probability distribution is obtained by the Softmax function. The calculation formula is as follows: PN = softmax (WVZMC + b); where WV represents the matrix parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, b is the bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, ZMC represents the multilingual context semantic vector, and PN represents the neural probability distribution; On the basis of the multilingual context semantic vector, two linear layers are added, and the decision probability distribution is obtained by the Softmax function. The calculation formula is as follows: [pn, pc, pt] = softmax (W2 (W1ZMC + b1) + b2); wherein W1 represents the first matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b1 is the first bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, W2 represents the second matrix parameter of the importance of the multilingual context semantic vector when predicting the decision probability distribution, b2 is the second bias parameter of the importance of the multilingual context semantic vector when predicting the neural probability distribution, [pn, pc, pt] is the decision probability distribution, pn represents the probability that the final vocabulary generation comes from the neural probability distribution, pc represents the probability that the final vocabulary generation comes from the copy probability distribution, and pt represents the probability that the final vocabulary generation comes from the translation probability distribution; a vocabulary distribution probability acquisition module, wherein the vocabulary distribution probability acquisition module performs a weighted summation on at least the translation probability distribution, the copy probability distribution and the neural probability distribution to obtain a vocabulary distribution probability; and A summary generation module is configured to generate a summary of the multilingual document based at least on the vocabulary distribution probability.

9. An electronic device, characterized in that: include: A memory storing execution instructions; as well as A processor, wherein the processor executes the execution instructions stored in the memory, so that the processor executes the method according to any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium stores execution instructions, which are used to implement the method according to any one of claims 1 to 7 when executed by a processor.

Images