Data extraction method and device, electronic equipment and computer readable medium

Abstract

The invention discloses a data extraction method and device, electronic equipment and a computer readable medium, and relates to the technical field of computers.The method comprises the steps that word segmentation is conducted on obtained single text data to obtain a word list, and the word list is converted into a semantic vector list; extracting subject terms in the target data set, and generating a subject term semantic vector list based on the subject terms; performing similarity calculation on each subject term semantic vector in the subject term semantic vector list and each word vector in the semantic vector list so as to determine the important proportion of each word in the word list, and further obtaining a word important proportion list; performing convolution calculation based on the word importance proportion list and the semantic vector list to obtain convolution calculation result data; and according to the convolution calculation result data, determining an abstract vector, and calling a theme information attention layer to determine target extraction data based on the abstract vector. And the accuracy of the determined target extraction data is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data extraction method, device, electronic device and computer-readable medium. Background Art

[0002] Automatic text summarization technology is used for data extraction, and many aspects such as data processing, model structure, and data source need to be considered as algorithm factors. Therefore, the current automatic text summarization technology may extract data that is not important data, the extracted data cannot meet the needs, and the data extraction accuracy is low. Summary of the invention

[0003] In view of this, the embodiments of the present application provide a data extraction method, device, electronic device and computer-readable medium, which can solve the problem of low accuracy of data extraction through existing automatic text summarization technology.

[0004] To achieve the above object, according to one aspect of an embodiment of the present application, a data extraction method is provided, comprising:

[0005] Segment the acquired single text data to obtain a word list, and convert the word list into a semantic vector list;

[0006] Extract the subject words in the target data set, and generate a list of subject word semantic vectors based on the subject words;

[0007] Calculate the similarity between each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtain a word importance ratio list;

[0008] Perform convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data;

[0009] According to the convolution calculation result data, the summary vector is determined, and the topic information attention layer is called to determine the target extraction data based on the summary vector.

[0010] Optionally, a convolution calculation is performed based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data, including:

[0011] Input the word importance ratio list into the one-way encoding layer to obtain the encoded word importance ratio list;

[0012] Input the semantic vector list into the bidirectional encoding layer to obtain the encoded semantic vector list;

[0013] The encoded semantic vectors in the encoded semantic vector list are convolved with the encoded importance ratios of the corresponding words in the encoded word importance ratio list to obtain convolution calculation result data.

[0014] Optionally, a topic information attention layer is called to determine target extraction data based on the summary vector, including:

[0015] Convolve each summary vector with each subject word semantic vector in the subject word semantic vector list to obtain each similarity;

[0016] According to each similarity and preset threshold, the target data is determined to be extracted.

[0017] Optionally, according to each similarity and a preset threshold, determining the target data to be extracted includes:

[0018] The words in the word list corresponding to the summary vector corresponding to the similarity greater than a preset threshold are determined as target extraction data.

[0019] Optionally, the acquired single text data is segmented to obtain a word list, including:

[0020] Stop words are deleted from a single text data, and the remaining single text data after the stop words are deleted are split based on the placeholders of the stop words to obtain a word list.

[0021] Optionally, extract the subject words in the target dataset, including:

[0022] The target data set is input into the topic word extraction model to identify and extract the topic words in the target data set through the topic word extraction model.

[0023] Optionally, based on the subject words, a subject word semantic vector list is generated, including:

[0024] Each topic word is input into the deep learning model to convert each topic word into a topic word semantic vector through the text conversion encoding layer in the deep learning model, thereby obtaining a topic word semantic vector list.

[0025] Optionally, determining the importance ratio of each word in the word list includes:

[0026] For each word in the word list, calculate the vector difference between the corresponding subject word semantic vector in the subject word semantic vector list and the corresponding word vector in the semantic vector list;

[0027] The importance ratio of the word is determined based on the vector difference and the vector modulus of the corresponding topic word semantic vector.

[0028] In addition, the present application also provides a data extraction device, comprising:

[0029] A word segmentation unit is configured to segment the acquired single text data to obtain a word list, and convert the word list into a semantic vector list;

[0030] An extraction unit is configured to extract the subject words in the target data set, and generate a subject word semantic vector list based on the subject words;

[0031] a similarity calculation unit configured to perform similarity calculation on each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtain a word importance ratio list;

[0032] A convolution calculation unit is configured to perform convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data;

[0033] The target extraction data determination unit is configured to determine a summary vector according to the convolution calculation result data, and call the topic information attention layer to determine the target extraction data based on the summary vector.

[0034] Optionally, the convolution computing unit is further configured to:

[0035] Input the word importance ratio list into the one-way encoding layer to obtain the encoded word importance ratio list;

[0036] Input the semantic vector list into the bidirectional encoding layer to obtain the encoded semantic vector list;

[0037] The encoded semantic vectors in the encoded semantic vector list are convolved with the encoded importance ratios of the corresponding words in the encoded word importance ratio list to obtain convolution calculation result data.

[0038] Optionally, the target extraction data determination unit is further configured to:

[0039] Convolve each summary vector with each subject word semantic vector in the subject word semantic vector list to obtain each similarity;

[0040] According to each similarity and preset threshold, the target data is determined to be extracted.

[0041] Optionally, the target extraction data determination unit is further configured to:

[0042] The words in the word list corresponding to the summary vector corresponding to the similarity greater than a preset threshold are determined as target extraction data.

[0043] Optionally, the word segmentation unit is further configured as:

[0044] Stop words are deleted from a single text data, and the remaining single text data after the stop words are deleted are split based on the placeholders of the stop words to obtain a word list.

[0045] Optionally, the extraction unit is further configured to:

[0046] The target data set is input into the topic word extraction model to identify and extract the topic words in the target data set through the topic word extraction model.

[0047] Optionally, the extraction unit is further configured to:

[0048] Each topic word is input into the deep learning model to convert each topic word into a topic word semantic vector through the text conversion encoding layer in the deep learning model, thereby obtaining a topic word semantic vector list.

[0049] Optionally, the similarity calculation unit is further configured to:

[0050] For each word in the word list, calculate the vector difference between the corresponding subject word semantic vector in the subject word semantic vector list and the corresponding word vector in the semantic vector list;

[0051] The importance ratio of the word is determined based on the vector difference and the vector modulus of the corresponding topic word semantic vector.

[0052] In addition, the present application also provides a data extraction electronic device, including: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by one or more processors, the one or more processors implement the data extraction method as described above.

[0053] In addition, the present application also provides a computer-readable medium on which a computer program is stored, and when the program is executed by a processor, the data extraction method as described above is implemented.

[0054] To achieve the above objective, according to another aspect of the embodiments of the present application, a computer program product is provided.

[0055] A computer program product according to an embodiment of the present application includes a computer program, and when the program is executed by a processor, the data extraction method provided by the embodiment of the present application is implemented.

[0056] One embodiment of the above invention has the following advantages or beneficial effects: the present application performs word segmentation on the acquired single text data to obtain a word list, and converts the word list into a semantic vector list; extracts the subject words in the target data set, and generates a subject word semantic vector list based on the subject words; performs similarity calculation on each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtains a word importance ratio list; performs convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data; determines the summary vector based on the convolution calculation result data, and calls the subject information attention layer to determine the target extraction data based on the summary vector. Thereby improving the accuracy of the determined target extraction data.

[0057] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The accompanying drawings are used to better understand the present application and do not constitute an improper limitation on the present application.

[0059] Figure 1 is a schematic diagram of the main process of a data extraction method according to an embodiment of the present application;

[0060] Figure 2 is a schematic diagram of the main process of a data extraction method according to an embodiment of the present application;

[0061] Figure 3 It is a main flow chart of a data extraction method according to an embodiment of the present application;

[0062] Figure 4 is a schematic diagram of main units of a data extraction device according to an embodiment of the present application;

[0063] Figure 5 is an exemplary system architecture diagram to which the embodiments of the present application can be applied;

[0064] Figure 6 It is a structural diagram of a computer system of a terminal device or server suitable for implementing an embodiment of the present application. DETAILED DESCRIPTION

[0065] The following is an explanation of the exemplary embodiments of the present application in conjunction with the accompanying drawings, including various details of the embodiments of the present application to facilitate understanding, which should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Similarly, for the sake of clarity and conciseness, the description of known functions and structures is omitted in the following description. It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of the present application are in compliance with the relevant provisions of national laws and regulations. It should be noted that in the embodiments of the present application, certain software, components, models, and other existing solutions in the industry may be mentioned, which should be considered as exemplary, and their purpose is only to illustrate the feasibility of the implementation of the technical solution of the present application, but it does not mean that the applicant has or must have used the solution. In the technical solution of the present application, the collection, collection, update, analysis, processing, use, transmission, storage, etc. of the user personal information involved are in compliance with the provisions of relevant laws and regulations, are used for legal and reasonable purposes, and do not violate public order and good customs, are not shared, leaked or sold outside these legal uses, and are subject to supervision and management by regulatory authorities. Take necessary measures to prevent illegal access to such user personal information data, maintain the security of user personal information, network security and national security, ensure that persons who have access to personal information data comply with relevant laws and regulations, and ensure the security of user personal information. Once such user personal information data is no longer needed, the risk should be minimized by limiting or even prohibiting data collection and / or deleting data.

[0066] When used, including in certain related applications, user privacy is protected by de-identifying data, such as by removing specific identifiers when used, controlling the amount or specificity of data stored, controlling how data is stored, and / or other methods of de-identification.

[0067] Figure 1 is a schematic diagram of the main process of the data extraction method according to an embodiment of the present application, such as Figure 1 As shown, the data extraction method mainly includes the following steps S101 to S105.

[0068] Step S101, segmenting the acquired single text data to obtain a word list, and converting the word list into a semantic vector list.

[0069] Specifically, the acquired single text data is segmented to obtain a word list, including: deleting stop words from the single text data, and splitting the remaining single text data after deleting the stop words based on the placeholders of the stop words to obtain a word list.

[0070] The single text data may be the detailed data of the commodity to be recommended. The execution subject (for example, a server) may segment the single text data according to the separators in the single text data, such as "#", "*", etc., to accurately capture the important words in the single text data and obtain a word list. The execution subject may embed each word in the word list to obtain each word vector, and compose the obtained each word vector into a semantic vector list.

[0071] Step S102, extracting the subject words in the target data set, and generating a subject word semantic vector list based on the subject words.

[0072] The target data set may be a data set consisting of transaction data, operation data, recommendation data of various commodities, and other data related to the field corresponding to the acquired single text data. For example, the field corresponding to the single text data may be the e-commerce field, the financial field, and the like.

[0073] Specifically, extracting the subject words in the target data set includes: inputting the target data set into a subject word extraction model to identify and extract the subject words in the target data set through the subject word extraction model.

[0074] For example, a topic word extraction model is an LDA model. Specifically, the LDA (Latent Dirichlet Allocation) model is a document topic generation model, also known as a three-layer Bayesian probability model, which contains a three-layer structure of words, topics and documents. The LDA model believes that each word in an article is obtained through a process of "selecting a topic with a certain probability, and selecting a word from this topic with a certain probability". Documents to topics follow a multinomial distribution, and topics to words also follow a multinomial distribution. The LDA model is an unsupervised machine learning technology that can be used to identify hidden topic information in large-scale document sets or corpora. The bag of words method is used to treat each document as a word frequency vector, thereby converting text information into digital information that is easy to model. Each document represents a probability distribution composed of some topics, and each topic represents a probability distribution composed of many words. Through a topic word extraction model, such as the LDA model, the words in the word list can be represented as word frequency vectors, which can be easily classified and clustered to accurately identify and extract topic words in the word list.

[0075] In addition, the keyword extraction model can also be a QDA model, a PLSA model or a LSA model.

[0076] Among them, the QDA (Quadratic Discriminant Analysis) model is a discriminant analysis method, similar to LDA, but used when the covariance matrices of samples of different categories are different. QDA and LDA are both a type of discriminant analysis, which aims to predict which category an item belongs to based on known categories.

[0077] The PLSA (Probabilistic Latent Semantic Analysis) model and the LSA (Latent Semantic Analysis) model are also topic models that discover the topic structure in documents by analyzing document collections. PLSA is a probabilistic model, while LSA focuses more on finding low-order approximations by building a "word-document" matrix and performing singular value decomposition (SVD) to achieve dimensionality reduction and denoising.

[0078] Specifically, based on the subject words, a subject word semantic vector list is generated, including: inputting each subject word into a deep learning model, converting each subject word into each subject word semantic vector through a text conversion encoding layer in the deep learning model, and then obtaining a subject word semantic vector list.

[0079] A deep learning model, for example, can be a BERT model, which is a language representation model. BERT stands for Bidirectional Encoder Representations from Transformers. BERT is designed to pre-train deep bidirectional representations by jointly adjusting the left and right contexts in all layers. The text transformation encoding layer in the deep learning model can quickly and accurately convert the topic words into a list of topic word semantic vectors.

[0080] In addition, deep learning models can also be GPT models, Word2Vec models, etc. The GPT model, based on the Transformer architecture, is pre-trained through a large-scale corpus and can generate coherent text. It can generate semantic vectors of subject words corresponding to subject words and is suitable for a variety of natural language processing tasks. The Word2Vec model is a word embedding model that maps words to vector space through two training methods (CBOW and Skip-gram). Words with similar semantics are closer in the vector space.

[0081] Step S103, calculating the similarity between each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtaining a word importance ratio list.

[0082] Specifically, determining the importance ratio of each word in the word list includes: for each word in the word list, calculating the vector difference between the corresponding subject word semantic vector in the subject word semantic vector list and the corresponding word vector in the semantic vector list; determining the importance ratio of the word based on the vector difference and the vector modulus of the corresponding subject word semantic vector.

[0083] For example, the execution subject may determine the number of topic words extracted from the target data set, for example, n, and may determine the number of words obtained by segmenting the acquired single text data, for example, i. For each word in the word list (for example, word i), the corresponding topic word semantic vector in the topic word semantic vector list is calculated (for example, the topic word vector j corresponding to the jth topic word extracted from the target data set, for example, topic_word j ) and the corresponding word vector in the semantic vector list (for example, the word vector corresponding to the i-th word obtained by segmenting the acquired single text data, such as word i ) vector difference (e.g. topic_word j -word i ); according to the vector difference (for example, topic_word j -word i ) and the vector modulus of the corresponding topic word semantic vector (e.g. |topic_word j |), determine the importance ratio of the word.

[0084] For example, the execution entity can perform vector difference (such as topic_word j -word i ) is squared to obtain a square value, and then the sum of the square values ​​within a preset numerical range (e.g., a range from j=1 to j=n, when j=1, it represents the first subject word extracted from the target data set, and when j=n, it represents the nth subject word extracted from the target data set) is calculated, and the sum of the vector moduli of the corresponding subject word semantic vector within the preset numerical range (e.g., a range from j=1 to j=n, when j=1, it represents the first subject word extracted from the target data set, and when j=n, it represents the nth subject word extracted from the target data set) is calculated, and then the ratio of the sum of the vector moduli to the sum of the square values ​​is calculated, and the ratio is determined as the importance ratio of the word, and then a word importance ratio list is obtained. Thereby improving the efficiency and accuracy of determining the importance ratio of each word in the word list, and then improving the efficiency and accuracy of obtaining the word importance ratio list.

[0085] For example, the formula for the importance ratio of the word can be as follows:

[0086]

[0087] Among them, j is the jth subject word extracted from the target data set, i can mean the i-th word obtained by segmenting the acquired single text data, n is the number of subject words extracted from the target data set, and when j=n, it represents the nth subject word extracted from the target data set.

[0088] Step S104, performing convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data.

[0089] Specifically, a convolution calculation is performed based on the word importance ratio list and the semantic vector list to obtain the convolution calculation result data, including: inputting the word importance ratio list into a unidirectional encoding layer to obtain an encoded word importance ratio list; inputting the semantic vector list into a bidirectional encoding layer to obtain an encoded semantic vector list; and convolving the encoded semantic vector in the encoded semantic vector list with the encoded importance ratio of the corresponding word in the encoded word importance ratio list to obtain the convolution calculation result data.

[0090] The word importance ratio list is input into the one-way encoding layer to obtain an encoded word importance ratio list, which may include Figure 3 encoder_WIRL1, encoder_WIRL2, encoder_WIRL3, ..., encoder_WIRLl in .

[0091] The semantic vector list is input into the bidirectional encoding layer to obtain an encoded semantic vector list, which may include Figure 3 ESL1, ESL2, ESL3, …, ESL1.

[0092] The encoded semantic vectors in the encoded semantic vector list, such as Figure 3 The encoded importance ratios of ESL1, ESL2, ESL3, ..., ESL1 and the corresponding words in the encoded word importance ratio list, for example Figure 3 The encoder_WIRL1, encoder_WIRL2, encoder_WIRL3, ..., encoder_WIRL1 are convolved to obtain the convolution calculation result data.

[0093] For example, the convolution result data can be content_vector t :

[0094]

[0095] Among them, the meaning of i can be the i-th word obtained by segmenting the acquired single text data, the range of i can be 1 to m, m is the maximum value that i can reach, m is an integer, t can be any value from 1 to m, and can represent the t-th word obtained by segmenting the acquired single text data, m can represent the m-th word obtained by segmenting the acquired single text data, 1≤t≤m, and t is an integer.

[0096] Step S105, determining a summary vector according to the convolution calculation result data, and calling the topic information attention layer to determine the target extraction data based on the summary vector.

[0097] The decoding network is used to decode the data obtained through convolution calculation. The decoding network is built using a single-layer LSTM network. The convolution calculation result data is decoded through the decoding network to obtain a summary vector, such as y. The personalized attention mechanism (IN-ATT) that incorporates the idea of ​​attention is used to call the subject information attention layer to perform convolution calculation on the summary vector and the subject word semantic vector corresponding to the subject word of the extracted target data set, thereby determining the target extracted data. In order to make important information get attention, the summary vector output by the decoding network is processed with the subject word extracted from the target data set as a convolution, and then the proportion of key information in a single text data can be obtained. For a target data set, the overall target data set background is considered, and the single text data is processed to focus on the more important subject words of the single text data. Therefore, using this approach, the designed model not only considers the domain of the overall target data set, but also considers the differences of the single text data. With the overall parameters, personalized information extraction is performed for the single.

[0098] This embodiment performs word segmentation on the acquired single text data to obtain a word list, and converts the word list into a semantic vector list; extracts the subject words in the target data set, and generates a subject word semantic vector list based on the subject words; performs similarity calculation on each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtains a word importance ratio list; performs convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data; determines the summary vector based on the convolution calculation result data, and calls the subject information attention layer to determine the target extraction data based on the summary vector. Thereby, the accuracy of data extraction can be improved.

[0099] Figure 2 is a schematic diagram of the main flow of a data extraction method according to an embodiment of the present application, such as Figure 2 As shown, the data extraction method mainly includes the following steps S201 - S205.

[0100] Step S201: Segment the obtained single text data to obtain a word list, and convert the word list into a semantic vector list.

[0101] Delete stop words from the single text data, and based on the placeholders of the stop words, split the remaining single text data after deleting the stop words to obtain a word list.

[0102] In the embodiments of the present application, stop words refer to words without actual meanings, such as words or phrases like "de", "hao", "de", "yiban", etc. The embodiments of the present application do not make specific limitations on stop words. After deleting stop words from the single text data, the positions where there were originally stop words in the single text data become blanks, which are the placeholders of the stop words. Based on these placeholders, splitting the remaining single text data after deleting the stop words can conveniently and quickly obtain a word list. And the word list can be converted into a semantic vector list by means of word embedding.

[0103] In some embodiments, the single text data can be the detailed data of the product to be recommended. The execution entity can segment the single text data according to preset symbols in the single text data, such as "@", "&", etc., to accurately capture important words in the single text data and obtain a word list. The execution entity can perform word embedding on each word in the word list to obtain each word vector, and form a semantic vector list with the obtained word vectors.

[0104] Step S202: Extract the topic words in the target dataset, and based on the topic words, generate a topic word semantic vector list.

[0105] In some embodiments, the execution entity can extract the topic words in the target dataset through a pre-trained topic word extraction model or based on the topic word portrait. Specifically, when extracting the topic words in the target dataset based on the topic word portrait, the execution entity can match each word in the target dataset with each label of the topic word portrait, and extract the words in the target dataset that match the labels of the topic word portrait as topic words, thereby improving the accuracy of topic word extraction. After extracting the topic words, the execution entity can convert the extracted topic words into topic word semantic vectors by means of word embedding, and further obtain a topic word semantic vector list.

[0106] Step S203: Calculate the similarity between each topic word semantic vector in the topic word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and further obtain a word importance ratio list.

[0107] In some embodiments, the execution entity may perform cosine similarity calculation on each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list, so as to determine the calculated cosine similarity as the importance ratio of each word in the word list, thereby accurately obtaining a word importance ratio list.

[0108] In some embodiments, the execution entity may also calculate the similarity between each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list. Specifically, the Euclidean distance (Euclidean distance: a measure of the straight-line distance between two points in space, applicable to numerical data) between each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list may be calculated to obtain the corresponding similarity, and the calculated Euclidean distance may be determined as the importance ratio of each word in the word list, thereby accurately obtaining a word importance ratio list, such as word_weight1, word_weight2, word_weight3, ..., word_weightl.

[0109] Step S204, performing convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data.

[0110] Convolution calculation is performed based on a word importance ratio list (e.g., word_weight1, word_weight2, word_weight3, ..., word_weightl) and a semantic vector list (e.g., word1, word2, word3, ..., wordl) to obtain convolution calculation result data.

[0111] Step S205, determining a summary vector according to the convolution calculation result data.

[0112] The convolution calculation result data is decoded through the decoding network to obtain the summary vectors corresponding to the single text data.

[0113] Step S206: convolve each summary vector with each subject word semantic vector in the subject word semantic vector list to obtain each similarity.

[0114] The personalized attention mechanism (IN-ATT) that incorporates the idea of ​​attention is called to convolve each summary vector with each subject word semantic vector in the subject word semantic vector list to obtain each similarity.

[0115] In order to focus on important information, the summary vector output by the decoding network is convolved with the topic words, and then the proportion of key information can be obtained. It can also be seen from this that for a data set, the overall data set background is considered, and each text is processed to focus on the more important topic words of each single text. This approach takes into account both the field of the overall data set and the individual differences. With the parameters such as the topic words in the overall target data set, personalized information extraction is performed for individual text data.

[0116] Step S207 determines the words in the word list corresponding to the summary vector corresponding to the similarity greater than the preset threshold as the target extraction data.

[0117] According to each similarity and a preset threshold, the target extraction data is determined. By comparing each similarity with a preset threshold, the target extraction data is determined according to the comparison result. Specifically, each similarity can be sorted in descending order, and the words in the word list corresponding to the summary vector corresponding to the first 5 (or 10, the embodiment of the present application does not specifically limit this number) similarities greater than the preset threshold are determined as the target extraction data. Thereby improving the accuracy of the determined target extraction data.

[0118] Figure 3 FIG. 1 is a schematic diagram of the main flow of a data extraction method according to an embodiment of the present application. Figure 3 As shown, at the topic coding layer, the acquired single text data is segmented to obtain a word list, such as word_text1, word_text2, word_text3, ..., word_text1, where l can be the number of word segmentations for a single text data, and the embodiment of the present application does not specifically limit what letters are used to represent this number. The obtained word list is input into the text conversion coding layer (such as the BERT text conversion coding layer) of the deep learning model (such as the BERT model, the GPT model or the Word2Vec model) to obtain a semantic vector list: word1, word2, word3, ..., word1.

[0119] The execution subject can extract the topic words in the target data set, for example, the target data set can be input into a topic word extraction model (such as an LDA model, a QDA model, a PLSA model or an LSA model) to obtain the topic words: topic_text1, topic_text2, topic_text3, ..., topic_text1. The obtained topic words: topic_text1, topic_text2, topic_text3, ..., topic_text1 are input into a text conversion encoding layer (such as a BERT model, a GPT model or a Word2Vec model) of a deep learning model to obtain a topic word semantic vector list: topic_word1, topic_word2, topic_word3, ..., topic_word1.

[0120] Calculate the similarity between the product copy words and the topic words, that is, calculate the similarity between each word vector word_text1, word_text2, word_text3, ..., word_textl in the semantic vector list and each topic word semantic vector topic_word1, topic_word2, topic_word3, ..., topic_wordl in the topic word semantic vector list to determine the importance ratio of each word in the word list, and then obtain a word importance ratio list: word_weight1, word_weight2, word_weight3, ..., word_weightl.

[0121] Input the word importance ratio list into a unidirectional encoding layer (e.g., an LSTM encoding layer) to obtain an encoded word importance ratio list: encoder_WIRL1, encoder_WIRL2, encoder_WIRL3, ..., encoder_WIRL1; input the semantic vector list into a bidirectional encoding layer (e.g., a BiLSTM encoding layer) to obtain an encoded semantic vector list: ESL1, ESL2, ESL3, ..., ESL1; convolve the encoded semantic vectors ESL1, ESL2, ESL3, ..., ESL1 in the encoded semantic vector list with the encoded importance ratios encoder_WIRL1, encoder_WIRL2, encoder_WIRL3, ..., encoder_WIRL1 of the corresponding words in the encoded word importance ratio list through a convolution layer to obtain the convolution calculation result data content_vector t :

[0122]

[0123] Among them, the meaning of i can be the i-th word obtained by segmenting the acquired single text data, the range of i can be 1 to m, m is the maximum value that i can reach, m is an integer, t can be any value from 1 to m, and can represent the t-th word obtained by segmenting the acquired single text data, m can represent the m-th word obtained by segmenting the acquired single text data, 1≤t≤m, and t is an integer.

[0124] The convolution calculation result data content_vector or the specific content_vector with the subscript t (t can represent the tth word obtained by segmenting the single text data obtained) t Input the Encoder decoding network to obtain the summary vector y, and convolve the summary vector y with the topic word semantic vector corresponding to the topic word of the extracted target data set through the Attention convolution layer, so as to more accurately determine the target extracted data.

[0125] The embodiment of the present application adds a keyword extraction model to extract important "keywords" from the target data set, and then performs similarity calculations to match the words in the single text data. A deep learning model is added so that the semantic vector can better represent the actual meanings of different words. The idea of ​​attention is incorporated, and the summary vector output by the decoding layer is convolved with the keyword to obtain the similarity between the keyword and the summary word. The keyword extraction model is an extractive model that extracts important words from the article, and the extracted words are some keyword words by default.

[0126] The embodiments of the present application can be applied to the e-commerce industry to recommend products to different users. When recommending products, it is necessary to show the characteristics of the corresponding products to customers. Therefore, browsing from the details page of the product will make it more difficult for users to choose products. In addition, since the description is too redundant and there is a lot of text information, users cannot capture the recommended points of the product in a short time. Based on the need to summarize and generalize information, the automatic text summarization technology will be used in this regard. Taking into account the diversity of data sets and describing different product themes, it is considered to use a keyword extraction model to annotate key information, so that the model pays more attention to the key information of a single text data.

[0127] The embodiment of the present application uses a subject information attention layer to allow more attention to important information. A deep learning model is used for encoding conversion. The extracted information is then compared with the subject word information, so that more attention is paid to the key information of a single text data when extracting data.

[0128] For example, in an embodiment of the present application, taking the e-commerce industry as an example, the target data set is input into the topic word extraction model to extract the topic words in the product copy corresponding to the target data set, namely topic_text1, topic_text2, topic_text3, ..., topic_text1, and a topic word semantic vector list is generated based on the extracted topic words, namely topic_word1, topic_word2, topic_word3, ..., topic_word1, and these words topic_word1, topic_word2, topic_word3, ..., topic_word1 are input into the deep learning model for conversion between text and numerical vectors to obtain a topic word semantic vector list (topic_word_semanteme_list), namely topic_word1, topic_word2, topic_word3, ..., topic_word1;

[0129] The product copy corresponding to a single text data is segmented, and the "stop words" without actual meaning are removed to obtain a word list (product_word_list), namely word_text1, word_text2, word_text3, ..., word_text1. The word list (product_word_list) is passed into the deep learning model, and the semantic vector of the corresponding word conversion, namely the word vector, is output to obtain a semantic vector list (word_semanteme_list), namely word1, word2, word3, ..., word1.

[0130] Calculate the similarity between each word vector word1, word2, word3, ..., word1 in the semantic vector list corresponding to the word list (product_word_list) and each topic word semantic vector topic_word1, topic_word2, topic_word3, ..., topic_word1 in the topic word semantic vector list (topic_word_semanteme_list), and calculate the sum. The specific formula is as follows:

[0131]

[0132] Among them, j is the jth subject word extracted from the target data set, i can mean the i-th word obtained by segmenting the acquired single text data, n is the number of subject words extracted from the target data set, and when j=n, it represents the nth subject word extracted from the target data set.

[0133] This can be used to obtain the importance (i.e., importance ratio) of the words in the word list corresponding to each single text data (e.g., product copy words). i The vector represents the importance ratio of the i-th subject word to the theme of this single text data (for example, product copy). The larger the importance ratio, the more important the word is; otherwise, the probability that it is considered to be an "irrelevant word" is higher. Therefore, the importance ratio of each word (for example, product copy word) constitutes the word importance ratio list (word_importance_ratio_list, hereinafter referred to as WIRL) of the words corresponding to the entire single text data (for example, product copy), for example, word_weight1, word_weight2, word_weight3, ..., word_weightl (in the embodiment of the present application, it can also be word_weight1, word_weight2, word_weight3, ..., word_weight l ). Input the semantic vector list (word_semanteme_list) into the encoding layer of a bidirectional deep learning model (e.g., LSTM) to obtain an encoded semantic vector list (encoder_semanteme_list, hereinafter referred to as ESL), for example, ESL1, ESL2, ESL3, ..., ESL1. Input the word importance ratio list (WIRL) of the entire product copy: word_weight1, word_weight2, word_weight3, ..., word_weight1 into a unidirectional deep learning model (e.g., LSTM) to obtain an encoded word importance ratio list (encoder_WIRL), for example, encoder_WIRL1, encoder_WIRL2, encoder_WIRL3, ..., encoder_WIRL1. Convolve the encoded semantic vector list (encoder_semanteme_list, hereinafter referred to as ESL), i.e., the encoded semantic vector (ESL) in ESL1, ESL2, ESL3, ..., ESL1, with the encoded word importance ratio list (encoder_WIRL), i.e., the important ratio of the word in encoder_WIRL1, encoder_WIRL2, encoder_WIRL3, ..., encoder_WIRL1, as shown below, and then perform sum processing to obtain the input content_vector of the decoding network t .

[0134]

[0135] Among them, the meaning of i can be the i-th word obtained by segmenting the acquired single text data, the range of i can be 1 to m, m is the maximum value that i can reach, m is an integer, t can be any value from 1 to m, and can represent the t-th word obtained by segmenting the acquired single text data, m can represent the m-th word obtained by segmenting the acquired single text data, 1≤t≤m, and t is an integer.

[0136] Decoding network: The extracted information can be obtained by decoding using the decoding network. The decoding network is built using a single-layer LSTM network.

[0137] Individualized attention mechanism (IN-ATT) that incorporates the idea of ​​attention: In order to make "important information get attention", the vector output by the decoding network is processed with the theme words as a convolution, and then the proportion of key information can be obtained. For a data set, the overall background of the target data set is considered, and the single text data is processed to focus on the more important theme words of the single text data. Therefore, the embodiment of the present application takes into account both the field situation of the overall data set and the differences of the single text data. With the parameters such as the theme words in the overall target data set, personalized important information extraction is performed for the single text data.

[0138] The embodiment of the present application uses a topic word extraction model (such as an LDA model, a QDA model, a PLSA model, or an LSA model) to add the text domain of the overall data set to the encoding, and then uses a deep learning model (such as a BERT model, a GPT model, or a Word2Vec model) to obtain a better vector conversion. Then, the idea of ​​attention is incorporated, and the importance of the decoded information for a single text is compared with the topic words, so that when the information is extracted, the importance of the text is more clearly defined. Therefore, the model not only takes into account the overall background, but also extracts personalized important information for a single text, so that important information can be extracted more efficiently.

[0139] The embodiment of the present application uses a subject word extraction model (such as an LDA model, a QDA model, a PLSA model, or an LSA model) to first identify important subject words based on the entire data set, and then use the current deep learning model with good performance (such as a BERT model, a GPT model, or a Word2Vec model) to semantically encode these subject words. After the single text data (such as a product copy) is semantically encoded by a deep learning model (such as a BERT model, a GPT model, or a Word2Vec model), it is compared with the semantics of the subject words. Thus, the words that are more relevant to the subject words will be given higher permissions in the encoding network, so that the execution subject can automatically pay attention to these more important information in the single text data to improve the extraction quality of the entire single text data summary.

[0140] Since the important information of a single text in the overall data set may be different, the embodiment of the present application adds an individual attention mechanism (individuation-attention, referred to as IN-ATT) that incorporates the idea of ​​attention, so that more attention is paid to the individual important information of the current single text data during data extraction. Therefore, based on the data set in the overall text field, the embodiment of the present application also considers the need to pay attention to the important information of a single text during the information extraction process.

[0141] Figure 4 Schematic diagram of the main units of the data extraction device according to an embodiment of the present application. Figure 4 As shown, the data extraction device 400 includes a word segmentation unit 401, an extraction unit 402, a similarity calculation unit 403, a convolution calculation unit 404 and a target extraction data determination unit 405.

[0142] The word segmentation unit 401 is configured to perform word segmentation on the acquired single text data to obtain a word list, and convert the word list into a semantic vector list.

[0143] The extraction unit 402 is configured to extract the subject words in the target data set, and generate a subject word semantic vector list based on the subject words.

[0144] The similarity calculation unit 403 is configured to perform similarity calculation on each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtain a word importance ratio list.

[0145] The convolution calculation unit 404 is configured to perform convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data.

[0146] The target extraction data determination unit 405 is configured to determine a summary vector according to the convolution calculation result data, and call the topic information attention layer to determine the target extraction data based on the summary vector.

[0147] In some embodiments, the convolution calculation unit 404 is further configured to: input the word importance ratio list into the unidirectional encoding layer to obtain the encoded word importance ratio list; input the semantic vector list into the bidirectional encoding layer to obtain the encoded semantic vector list; convolve the encoded semantic vector in the encoded semantic vector list with the encoded importance ratio of the corresponding word in the encoded word importance ratio list to obtain the convolution calculation result data.

[0148] In some embodiments, the target extraction data determination unit 405 is further configured to: convolve each summary vector with each subject word semantic vector in the subject word semantic vector list to obtain each similarity; and determine the target extraction data according to each similarity and a preset threshold.

[0149] In some embodiments, the target extraction data determination unit 405 is further configured to: determine the words in the word list corresponding to the summary vector corresponding to the similarity greater than a preset threshold as the target extraction data.

[0150] In some embodiments, the word segmentation unit 401 is further configured to: delete stop words from the single text data, and split the remaining single text data after the stop words are deleted based on the placeholders of the stop words to obtain a word list.

[0151] In some embodiments, the extraction unit 402 is further configured to: input the target data set into a topic word extraction model, so as to identify and extract topic words in the target data set through the topic word extraction model.

[0152] In some embodiments, the extraction unit 402 is further configured to: input each topic word into the deep learning model, so as to convert each topic word into each topic word semantic vector through the text conversion coding layer in the deep learning model, and then obtain a topic word semantic vector list.

[0153] In some embodiments, the similarity calculation unit 403 is further configured to: for each word in the word list, calculate the vector difference between the corresponding subject word semantic vector in the subject word semantic vector list and the corresponding word vector in the semantic vector list; determine the importance ratio of the word based on the vector difference and the vector module length of the corresponding subject word semantic vector.

[0154] It should be noted that the data extraction method and the data extraction device of the present application have a corresponding relationship in terms of specific implementation contents, so the repeated contents will not be described again.

[0155] Figure 5 An exemplary system architecture 500 is shown to which the data extraction method or data extraction device according to the embodiment of the present application can be applied.

[0156] like Figure 5 As shown, system architecture 500 may include terminal devices 501, 502, 503, a network 504 and a server 505. Network 504 is used to provide a medium for communication links between terminal devices 501, 502, 503 and server 505. Network 504 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0157] Users can use terminal devices 501, 502, 503 to interact with server 505 through network 504 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 501, 502, 503, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only as examples).

[0158] The terminal devices 501 , 502 , and 503 may be various electronic devices having a data extraction processing screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0159] Server 505 can be a server that provides various services, such as a background management server that provides support for single text data obtained by users using terminal devices 501, 502, and 503 (only as an example). The background management server can segment the obtained single text data to obtain a word list, and convert the word list into a semantic vector list; extract the subject words in the target data set, and generate a subject word semantic vector list based on the subject words; perform similarity calculations on each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtain a word importance ratio list; perform convolution calculations based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data; determine the summary vector based on the convolution calculation result data, and call the subject information attention layer to determine the target extraction data based on the summary vector. Thereby improving the accuracy of the determined target extraction data.

[0160] It should be noted that the data extraction method provided in the embodiment of the present application is generally executed by the server 505 , and accordingly, the data extraction device is generally arranged in the server 505 .

[0161] It should be understood that Figure 5The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0162] Reference below Figure 6 , which shows a schematic diagram of the structure of a computer system 600 of a terminal device suitable for implementing an embodiment of the present application. Figure 6 The terminal device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0163] like Figure 6 As shown, the computer system 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage part 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the computer system 600 are also stored. The CPU 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0164] The following components are connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 610 as needed, so that a computer program read therefrom is installed into the storage section 608 as needed.

[0165] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 609, and / or installed from the removable medium 611. When the computer program is executed by the central processing unit (CPU) 601, the above-mentioned functions defined in the system of the present application are executed.

[0166] It should be noted that the computer-readable medium shown in the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. Computer-readable storage media may include, for example, but are not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0167] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the above-mentioned module, program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0168] The units involved in the embodiments described in the present application may be implemented by software or hardware. The units described may also be set in a processor, for example, it may be described as: a processor includes a word segmentation unit, an extraction unit, a similarity calculation unit, a convolution calculation unit, and a target extraction data determination unit. The names of these units do not constitute limitations on the units themselves in some cases.

[0169] As another aspect, the present application also provides a computer-readable medium, which may be included in the device described in the above embodiment; or it may exist independently and not be assembled into the device. The above computer-readable medium carries one or more programs. When the above one or more programs are executed by a device, the device performs word segmentation on the acquired single text data to obtain a word list, and converts the word list into a semantic vector list; extracts the subject words in the target data set, and generates a subject word semantic vector list based on the subject words; performs similarity calculation on each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the important proportion of each word in the word list, and then obtains a word important proportion list; performs convolution calculation based on the word important proportion list and the semantic vector list to obtain convolution calculation result data; determines the summary vector according to the convolution calculation result data, and calls the subject information attention layer to determine the target extraction data based on the summary vector. Thereby improving the accuracy of the determined target extraction data.

[0170] The computer program product of the present application includes a computer program, and when the computer program is executed by a processor, the data extraction method in the embodiment of the present application is implemented.

[0171] According to the technical solution of the embodiment of the present application, the accuracy of the determined target extraction data can be improved.

[0172] The above specific implementations do not constitute a limitation on the protection scope of this application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of this application should be included in the protection scope of this application.

Claims

1. A data extraction method, characterized in that: include: Segmenting the acquired single text data to obtain a word list, and converting the word list into a semantic vector list; Extracting the subject words in the target data set, and generating a subject word semantic vector list based on the subject words; Calculating the similarity between each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtaining a word importance ratio list; Performing convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data; A summary vector is determined according to the convolution calculation result data, and a topic information attention layer is called to determine target extraction data based on the summary vector.

2. The method according to claim 1, characterized in that The convolution calculation is performed based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data, including: Inputting the word importance ratio list into a one-way encoding layer to obtain an encoded word importance ratio list; Inputting the semantic vector list into a bidirectional encoding layer to obtain an encoded semantic vector list; The encoded semantic vectors in the encoded semantic vector list are convolved with the encoded importance ratios of the corresponding words in the encoded word importance ratio list to obtain convolution calculation result data.

3. The method according to claim 1, characterized in that: The calling of the subject information attention layer to determine target extraction data based on the summary vector includes: Convolving each of the abstract vectors with each of the subject word semantic vectors in the subject word semantic vector list to obtain each similarity; According to the respective similarities and preset thresholds, target data is determined for extraction.

4. The method according to claim 3, characterized in that The step of determining target extraction data according to the respective similarities and preset thresholds includes: The words in the word list corresponding to the summary vector corresponding to the similarity greater than a preset threshold are determined as target extraction data.

5. The method according to claim 1, characterized in that: The acquired single text data is segmented to obtain a word list, including: Stop words are deleted from the single text data, and the remaining single text data after the stop words are deleted are split based on the placeholders of the stop words to obtain a word list.

6. The method according to claim 1, characterized in that The extracting of subject words in the target data set includes: The target data set is input into a subject word extraction model so as to identify and extract subject words in the target data set through the subject word extraction model.

7. The method according to claim 1, characterized in that The step of generating a list of semantic vectors of subject words based on the subject words includes: Each of the subject words is input into a deep learning model, so as to convert each of the subject words into a subject word semantic vector through a text conversion coding layer in the deep learning model, thereby obtaining a subject word semantic vector list.

8. The method according to any one of claims 1 to 7, characterized in that: Determining the importance ratio of each word in the word list includes: For each word in the word list, calculating the vector difference between the corresponding subject word semantic vector in the subject word semantic vector list and the corresponding word vector in the semantic vector list; The importance ratio of the word is determined according to the vector difference and the vector modulus of the corresponding subject word semantic vector.

9. A data extraction device, characterized in that: include: A word segmentation unit is configured to perform word segmentation on the acquired single text data to obtain a word list, and convert the word list into a semantic vector list; An extraction unit is configured to extract subject words in the target data set, and generate a subject word semantic vector list based on the subject words; a similarity calculation unit configured to perform similarity calculation on each subject word semantic vector in the subject word semantic vector list and each word vector in the semantic vector list to determine the importance ratio of each word in the word list, and then obtain a word importance ratio list; A convolution calculation unit, configured to perform convolution calculation based on the word importance ratio list and the semantic vector list to obtain convolution calculation result data; The target extraction data determination unit is configured to determine a summary vector according to the convolution calculation result data, and call the topic information attention layer to determine the target extraction data based on the summary vector.

10. A data extraction electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 8.

11. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

12. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images