Business demand book recommendation method and device

Abstract

The invention discloses a business demand book recommendation method and device, relates to the technical field of artificial intelligence, can be used in the financial science and technology field, and comprises the steps: employing a fine-grained semantic information extractor to respectively carry out fine-grained semantic extraction on a to-be-matched title and a corresponding label of a business demand book, and contents of stock and the corresponding label of the business demand book, obtaining fine-grained business semantics of the title and fine-grained business semantics of the content; a coarse-grained semantic information extractor is adopted to perform coarse-grained semantic extraction on a to-be-matched title and a corresponding label of the business demand book, and the content of the stock and the corresponding label of the business demand book, and coarse-grained business semantics of the title and coarse-grained business semantics of the content are obtained; according to all extracted business semantics, calculating a matching degree score of the title and the content of each stock; and recommending the content of the stock with the highest matching degree score to a business demand book writer. According to the method and the device, high-adaptability content can be recommended to the title of the target demand book.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology and can be used in the field of fintech. In particular, it relates to a method and device for recommending business requirement documents. Background Art

[0002] This section aims to provide background or context for the embodiments of the present invention described in the claims. The descriptions herein are not admitted to be prior art merely because they are included in this section.

[0003] Recommending business requirement documents means recommending appropriate requirement document content for the title of the target requirement document. Then, semantic extraction needs to be performed on the title and content of the requirement document to obtain their semantic representations respectively, and then a matching degree score is calculated for the two, and the most suitable content is recommended to the requirement document writer. Therefore, semantic extraction is the basis of the requirement document recommendation process, and extracting accurate and comprehensive features is one of the important factors affecting the recommendation effect of the requirement document recommendation system and is also a technical problem that urgently needs to be solved. Summary of the Invention

[0004] In a first aspect, an embodiment of the present invention provides a method for recommending business requirement documents for recommending content with a high degree of adaptation to the title of a target requirement document. The method includes:

[0005] Using a fine-grained semantic information extractor, perform fine-grained semantic extraction on the title to be matched and corresponding labels, and the existing content and corresponding labels of the business requirement document respectively, to obtain the fine-grained business semantics of the title and the fine-grained business semantics of the content. The fine-grained semantic information extractor performs fine-grained semantic extraction through a label-driven sentence-level attention mechanism and a label-driven paragraph-level attention mechanism;

[0006] Using a coarse-grained semantic information extractor, perform coarse-grained semantic extraction on the title to be matched and corresponding labels, and the existing content and corresponding labels of the business requirement document respectively, to obtain the coarse-grained business semantics of the title and the coarse-grained business semantics of the content. The coarse-grained semantic information extractor performs coarse-grained semantic extraction through a label-driven attention mechanism;

[0007] Calculate the matching degree score between the title and each existing content according to the fine-grained business semantics, coarse-grained business semantics of the title, and the fine-grained business semantics and coarse-grained business semantics of each existing content;

[0008] Recommend the existing content with the highest matching degree score to the business requirement document writer.

[0009] In a second aspect, an embodiment of the present invention provides a device for recommending business requirement documents for recommending content with a high degree of adaptation to the title of a target requirement document. The device includes:

[0010] A fine-grained semantic extraction module, which is used to adopt a fine-grained semantic information extractor to perform fine-grained semantic extraction on the title to be matched and the corresponding tags, as well as the existing content and the corresponding tags in the business requirement document, so as to obtain the fine-grained business semantics of the title and the fine-grained business semantics of the content. The fine-grained semantic information extractor performs fine-grained semantic extraction through a tag-driven sentence-level attention mechanism and a tag-driven paragraph-level attention mechanism;

[0011] A coarse-grained semantic extraction module, which is used to adopt a coarse-grained semantic information extractor to perform coarse-grained semantic extraction on the title to be matched and the corresponding tags, as well as the existing content and the corresponding tags in the business requirement document, so as to obtain the coarse-grained business semantics of the title and the coarse-grained business semantics of the content. The coarse-grained semantic information extractor performs coarse-grained semantic extraction through a tag-driven attention mechanism;

[0012] A matching module, which is used to calculate the matching degree scores between the title and the existing content according to the fine-grained business semantics and coarse-grained business semantics of the title, as well as the fine-grained business semantics and coarse-grained business semantics of each existing content;

[0013] A content recommendation module, which is used to recommend the content of the existing item with the highest matching degree score to the business requirement document writer.

[0014] In a third aspect, an embodiment of the present invention further provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the above-mentioned business requirement document recommendation method is implemented.

[0015] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned business requirement document recommendation method is implemented.

[0016] In a fifth aspect, an embodiment of the present invention proposes a computer program product, where the computer program product includes a computer program, and when the computer program is executed by a processor, the business requirement document recommendation method is implemented.

[0017] In the embodiments of the present invention, a fine-grained semantic information extractor is adopted to perform fine-grained semantic extraction on the title to be matched and the corresponding tags, as well as the existing content and the corresponding tags in the business requirement document, so as to obtain the fine-grained business semantics of the title and the fine-grained business semantics of the content. The fine-grained semantic information extractor performs fine-grained semantic extraction through a tag-driven sentence-level attention mechanism and a tag-driven paragraph-level attention mechanism; a coarse-grained semantic information extractor is adopted to perform coarse-grained semantic extraction on the title to be matched and the corresponding tags, as well as the existing content and the corresponding tags in the business requirement document, so as to obtain the coarse-grained business semantics of the title and the coarse-grained business semantics of the content. The coarse-grained semantic information extractor performs coarse-grained semantic extraction through a tag-driven attention mechanism; according to the fine-grained business semantics and coarse-grained business semantics of the title, as well as the fine-grained business semantics and coarse-grained business semantics of each existing content, calculate the matching degree scores between the title and each existing content; recommend the content of the existing item with the highest matching degree score to the business requirement document writer. Through the above steps, two different semantic information extractors are used to extract coarse-grained and fine-grained semantic representations, which not only deeply excavate the structural semantics of the description information, but also use the tag-driven attention mechanism to guide the feature extraction process, and the extracted multi-granularity semantic representations have a higher correlation with the function; since both the coarse-grained semantic information extractor and the fine-grained semantic information extractor use the tag-driven attention mechanism to extract the business semantics of the description information when extracting the representation, the extracted representation has strong business properties; the effect of the fine-grained semantic information extractor is better than that of the coarse-grained semantic information extractor, because the fine-grained semantic information extractor further excavates the hierarchical structure of the description information, extracts the fine-grained structural semantics, and when performing multi-granularity semantic information matching, the coarse-grained semantics and the fine-grained semantics can complement each other, so as to achieve more accurate recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings. In the drawings:

[0019] Figure 1 It is a flowchart of the business requirement document recommendation method in the embodiments of the present invention;

[0020] Figure 2 It is the overall structure of the fine-grained semantic information extractor in the embodiments of the present invention;

[0021] Figure 3 It is the overall structure of the coarse-grained semantic information extractor in the embodiments of the present invention;

[0022] Figure 4 This is the overall structure of the multi-granularity semantic matching requirements book recommendation method in the embodiments of the present invention;

[0023] Figure 5 This is the structural schematic diagram of the business requirements book recommendation device in the embodiments of the present invention;

[0024] Figure 6 This is the schematic diagram of the computer device in the embodiments of the present invention. Detailed implementation manners

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer and more understandable, the following further describes the embodiments of the present invention in detail with reference to the accompanying drawings. Herein, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but not to limit the present invention.

[0026] In the technical solutions of this application, the acquisition, storage, use, processing, etc. of data all comply with the relevant regulations of national laws and regulations.

[0027] Figure 1 This is the flowchart of the business requirements book recommendation method in the embodiments of the present invention, including:

[0028] Step 101: Use a fine-grained semantic information extractor to respectively perform fine-grained semantic extraction on the title to be matched and the corresponding labels, and the existing content and the corresponding labels of the business requirements book, so as to obtain the fine-grained business semantics of the title and the fine-grained business semantics of the content. The fine-grained semantic information extractor performs fine-grained semantic extraction through a label-driven sentence-level attention mechanism and a label-driven paragraph-level attention mechanism;

[0029] Step 102: Use a coarse-grained semantic information extractor to respectively perform coarse-grained semantic extraction on the title to be matched and the corresponding labels, and the existing content and the corresponding labels of the business requirements book, so as to obtain the coarse-grained business semantics of the title and the coarse-grained business semantics of the content. The coarse-grained semantic information extractor performs coarse-grained semantic extraction through a label-driven attention mechanism;

[0030] Step 103: Calculate the matching degree scores between the title and the content of each existing item according to the fine-grained business semantics, coarse-grained business semantics of the title, and the fine-grained business semantics and coarse-grained business semantics of the content of each existing item;

[0031] Step 104: Recommend the content of the existing item with the highest matching degree score to the business requirements book writer.

[0032] In the embodiments of the present invention, two different semantic information extractors are adopted to extract coarse-grained and fine-grained semantic representations. It not only deeply mines the structural semantics of the description information, but also uses a label-driven attention mechanism to guide the feature extraction process. The multi-grained semantic representations extracted have a higher correlation with functions. Since both the coarse-grained semantic information extractor and the fine-grained semantic information extractor use a label-driven attention mechanism to extract the business semantics of the description information when extracting the representations, the extracted representations have strong business characteristics. The effect of the fine-grained semantic information extractor is better than that of the coarse-grained semantic information extractor because the fine-grained semantic information extractor further mines the hierarchical structure of the description information, extracts the fine-grained structural semantics, and when matching the multi-grained semantic information, the coarse-grained semantics and the fine-grained semantics can complement each other, thus achieving more accurate recommendations.

[0033] See Figure 2 This is the overall structure of the fine-grained semantic information extractor in the embodiments of the present invention, which extracts fine-grained business information from information at different levels (sentences, paragraphs) in the description, that is: at the sentence level, words related to the business have higher weights; at the paragraph level, sentences more related to the business have higher weights, so as to extract key fine-grained business information from the description. The core idea of the fine-grained semantic information extractor is to decompose the input object (the title and corresponding label of the business requirement document, the content and corresponding label) to obtain text, then encode the text to obtain vectors, and then encode each vector to represent its paragraph-level description information, reflect the hierarchical structure in the description information, and obtain fine-grained structural semantics; and introduce a label-driven attention mechanism in each encoding to make the fine-grained semantic information extractor pay more attention to features more related to the business and obtain fine-grained functional semantics.

[0034] The fine-grained semantic information extractor includes a first input layer, a first embedding layer, a sentence-level extraction layer, a label-driven sentence-level attention mechanism layer, a first paragraph-level extraction layer, a label-driven paragraph-level attention mechanism layer, and a first output layer; among them,

[0035] The first input layer is used to decompose the input object to obtain multiple texts. When the object is a title or the corresponding label of the title, the text is a word; when the object is content or the corresponding label of the content, the text is a sentence.

[0036] The first embedding layer is used to embed multiple texts into a vector space to obtain the text vector of each text.

[0037] The sentence-level extraction layer is used to obtain the sentence matrix representation of the object according to all text vectors.

[0038] A label-driven sentence-level attention mechanism layer, which is used to obtain the attention mechanism sentence matrix representation of an object by adopting a label-driven sentence-level attention mechanism according to the sentence matrix representation of the object;

[0039] A first paragraph-level extraction layer, which is used to obtain the paragraph matrix representation of an object according to the attention mechanism sentence matrix representation of the object;

[0040] A label-driven paragraph-level attention mechanism layer, which is used to obtain the attention mechanism paragraph matrix representation of an object by adopting a label-driven paragraph-level attention mechanism according to the paragraph matrix representation of the object;

[0041] A first output layer, which is used to output the attention mechanism paragraph matrix representation of an object.

[0042] In one embodiment, the label-driven sentence-level attention mechanism layer is specifically used for: according to the sentence matrix representation of an object, adopting a label-driven sentence-level attention mechanism to activate the weights of each text in the sentence matrix representation, so as to obtain the attention mechanism sentence matrix representation of the object;

[0043] The label-driven paragraph-level attention mechanism layer is specifically used for: according to the paragraph matrix representation of an object, adopting a label-driven paragraph-level attention mechanism to activate the weights of each sentence in the paragraph matrix representation, so as to obtain the attention mechanism paragraph matrix representation of the object.

[0044] Specifically in implementation, the purpose of the first input layer is to decompose the input object to obtain multiple texts. The object is a title or a label corresponding to the title, and one title corresponds to one label. The object can also be content or a content object label, and the content includes two cases of sentences and paragraphs. Each sentence has a corresponding label, and each paragraph has one label.

[0045] The purpose of the first embedding layer is to obtain the vectorized representation of the text. Currently, the Word2vec tool is used to embed all texts into a low-dimensional space as the first embedding layer. Through the first embedding layer, all texts are uniformly embedded into a k-dimensional vector space. Taking the text as a word as an example, each word can be represented by a k-dimensional vector. The finally obtained text vectors (the text vectors of words can be called word vectors, and the text vectors of sentences can be called sentence vectors) are uniformly represented by k. The word vector representation of word x is The word vector representation of word c is The word vector corresponding to label t is

[0046] Sentence-level extraction layer: To better understand the business information of the requirements document, it is not enough to understand each text in isolation. It is necessary to understand the entire sequence connected together. The embodiment of the present invention adopts the BiLSTM network structure (a type of LSTM network structure) as the extraction model of the sentence-level extraction layer. For the title s, the text is a word, and s can be expressed as a set of words, that is, s = {x1, x2, ..., x |s|}, after the first embedding layer, the word is represented by the word vector, x j The word vector is The BiLSTM memory cell at the jth word position in s consists of an input gate A forget gate An output gate and a state vector The update of these doors is based on the hidden state of the previous position And the word vector input at the current position Calculated. The update formula is:

[0047]

[0048] Where W and U represent the weight matrices, and the subscripts of the weight matrices indicate their meaning. b represents the offset vector parameter, which needs to be learned during training. The initial values c0 = 0, h0 = 0, and ° represents element-wise multiplication. After the BiLSTM state update, the output at position j is:

[0049]

[0050] Therefore, the sentence matrix representation extracted from the title s by the sentence-level extraction layer can be defined as:

[0051]

[0052] in, It is the output of the corresponding position after passing through BiLSTM at position j, Indicates that the outputs of all positions are concatenated to form a sentence matrix representation. Define sentence matrix representation H s,v as follows:

[0053]

[0054] Where k is the dimension of the word vector, H s,v It is the representation of the title s of a business requirements document. This representation is extracted from the word and represents the meaning of s.

[0055] For the content m of the business requirements document, m can be expressed as a set of sentences, and for a certain sentence s={x1,x2,...,x|s|}, through the same state update as in Formulas (1) to (5), it can be obtained that the output of the content through the sentence-level extraction layer is consistent with Formula (7).

[0056] Define the overall output sentence matrix of the sentence-level extraction layer as H s,m , that is For the body content, the sentence matrix representation of the sentence-level extraction layer is also the representation of each sentence in the content.

[0057] Label-driven sentence-level attention mechanism layer: To extract the fine-grained functional semantics of the described object, according to the different importance of different texts in the description, the label of the object is used as the functional guidance for feature extraction. In the fine-grained semantic information extractor, the label-driven sentence-level attention mechanism is used to activate the weights of each text, so as to pay attention to more important texts.

[0058] For example, for the sentence s (the entire title) of the title v, the attention function Attention is used to guide its function. Among them, first aggregate the label Tag v of the title v to obtain the query of the attention function as G(·) represents the aggregation function. The output sentence matrix representation H s,v of the sentence-level extraction layer is used as the key and values of the attention function. The sentence matrix representation of the attention mechanism of the title is:

[0059]

[0060] Among them, is the parameter matrix, represents the transpose of H s,v . Through the calculation of the attention mechanism, the final representation of the title is

[0061] Similarly, for the body content, the label Tag m is aggregated to obtain r Tag,m as the query, and the output sentence matrix representation H s,m of the sentence-level extraction layer is used as the key and values. The sentence matrix representation of the attention mechanism of the final content is

[0062]

[0063] The first paragraph-level extraction layer: For example, for the body content, after passing through the sentence-level attention layer, each sentence can be represented by , but the description information is a natural language paragraph, which will contain multiple sentences. Therefore, multiple sentences in the content will be feature-extracted by the first paragraph-level extraction layer.

[0064] The first paragraph-level extraction layer also uses the BiLSTM network structure as the basic extraction network. The input at the j-th sentence position is the output attention mechanism sentence matrix representation of the previous layer's attention mechanism. Through the same state update as in formulas (1) to (6), the output of the first paragraph-level extraction layer of the content is defined as the paragraph matrix representation H. a,m , and its value is:

[0065]

[0066] Label-driven paragraph-level attention mechanism layer: Similar to the label-driven sentence-level attention mechanism layer, when obtaining the paragraph matrix representation of the entire paragraph, more attention should be paid to the sentences with stronger business relevance. Therefore, the fine-grained semantic information extractor guides the process of paragraph-level feature extraction through labels again to extract fine-grained functional semantics and focus on the sentences most relevant to the function. For example, for the content, the query of the paragraph-level attention mechanism is the aggregation r of the labels. Tag,m , and the key and values are the paragraph matrix representation H output by the first paragraph-level extraction layer. a,m . After passing through the label-driven attention mechanism, the attention mechanism paragraph matrix representation of the content (which is the final fine-grained semantic representation at this time) is:

[0067]

[0068] Among them, is the parameter matrix, represents the transpose of H. a,m .

[0069] Through the fine-grained semantic information extractor, the fine-grained semantic representation of the title is obtained as r. fine-grained,v , and the fine-grained semantic representation of the content is obtained as r. fine-grained,m , and these representations can all be used for the calculation of the matching degree score of the subsequent recommendation system. The output of the fine-grained semantic information extractor, that is, the attention mechanism paragraph matrix representation, can be uniformly represented as r. fine-grained .

[0070] Figure 3 This is the overall structure of the coarse-grained semantic information extractor in the embodiment of the present invention. In one embodiment, the coarse-grained semantic information extractor includes a second input layer, a second embedding layer, a second paragraph-level extraction layer, a label-driven attention mechanism layer, and a second output layer; among them,

[0071] The second input layer is used to decompose the input object to obtain multiple words, and the object is the title, the label corresponding to the title, the content, or the label corresponding to the content;

[0072] The second embedding layer is used to embed multiple words into a vector space to obtain word vectors for each word;

[0073] The second paragraph-level extraction layer is used to obtain a paragraph matrix representation of the object based on the word vectors of all words of the object;

[0074] The label-driven attention mechanism layer is used to obtain a paragraph matrix representation of the object with the label-driven attention mechanism based on the paragraph matrix representation of the object;

[0075] The second output layer is used to output the paragraph matrix representation of the object with the attention mechanism.

[0076] In one embodiment, the label-driven attention mechanism layer is specifically used to: based on the paragraph matrix representation of the object, adopt the label-driven attention mechanism to activate the weights of each sentence in the paragraph matrix representation to obtain the paragraph matrix representation of the object with the attention mechanism.

[0077] Among them, the purpose of the coarse-grained semantic information extractor is to capture the coarse-grained semantic features of the object in the requirement document and generate an overall semantic representation at the paragraph level for matching calculation.

[0078] For example, for content m, the second input layer decomposes the input object to obtain multiple words The second embedding layer embeds multiple words into the vector space to obtain word vectors for each word. For example, the word vector of word c j is

[0079] The second paragraph-level extraction layer: The second paragraph-level extraction layer uses the BiLSTM network structure as the basic network unit. For the j-th word position in the content, the input to the BiLSTM memory unit is The update methods of its input gate, forget gate, output gate, and state vector are the same as the update formulas of the fine-grained semantic information extractor. Through BiLSTM, the output of the second paragraph-level extraction layer of content m can be defined as the paragraph matrix representation, and the formula is:

[0080]

[0081] The label-driven attention mechanism layer: After obtaining the paragraph matrix representation, this output directly represents the paragraph information of the content. In order to obtain more accurate description features, the label-driven attention mechanism is used to guide the output of the second paragraph-level extraction layer, so that the model pays more attention to the words related to the business. The key, query, and values of the attention mechanism are H b,m , r Tag,m and H b,m. After passing through the label-driven attention layer, the final overall coarse-grained representation of the content is the attention mechanism paragraph matrix representation, and the formula is:

[0082]

[0083] For the title v, the entire title can be regarded as a paragraph, and the input is Q v , after being calculated by BiLSTM and guided by the label-driven attention layer for the H b,v extracted by the paragraph-level extraction layer, the coarse-grained representation of the title is obtained as:

[0084]

[0085] Through the coarse-grained semantic information extractor, the coarse-grained semantic representation r coarse-grained,v of the title and the coarse-grained semantic representation r coarse-grained,m of the content can be obtained. The coarse-grained semantic representation can also be used for the matching degree scoring calculation of the subsequent recommendation system. The output of the coarse-grained semantic information extractor is uniformly represented as r coarse-grained .

[0086] The overall structure of the multi-granularity semantic matching requirement book recommendation method is as Figure 4 shown. For the target requirement book content and all candidate requirement book titles to be recommended, their main information and labels respectively generate two semantic representations from different granularities through the fine-grained semantic information extractor and the coarse-grained semantic information extractor, and then combine these two semantic representations for multi-granularity semantic matching recommendation.

[0087] In one embodiment, according to the fine-grained business semantics, coarse-grained business semantics of the title, and the fine-grained business semantics and coarse-grained business semantics of the content of each inventory, calculate the matching degree score between the title and the content of each inventory, including:

[0088] Connect the fine-grained business semantics and coarse-grained business semantics of the title to obtain the title representation;

[0089] Connect the fine-grained business semantics and coarse-grained business semantics of the content of each inventory to obtain the content representation of each inventory;

[0090] Calculate the matching degree score between the title representation and the content representation of each inventory.

[0091] Among them, connect the fine-grained business semantics and coarse-grained business semantics of the title to obtain the title representation r v , r v can be expressed as:

[0092] r v = [r fine-grained,v ||rcoarse-grained,v (16)

[0093] Among them, the fine-grained business semantics and the coarse-grained business semantics of the content of each stock are connected to obtain the content representation r of each stock m , r m can be expressed as:

[0094] r m = [r fine-grained,m || r coarse-grained,m (17)

[0095] Among them, [r fine-grained,v || r coarse-grained,v means directly connecting the vectors r fine-grained,v and r coarse-grained,v . Therefore, for the title, each content representation is used to calculate the matching degree with it to evaluate the content candidate closest to the target title.

[0096] Among them, after connecting the title representation and the content representation of each stock, the matching degree score between the title representation and the content representation of each stock is calculated. That is to say, the representation r v of the target title and the content r m are connected as [r v || r m .[[]END]

[0097] In one embodiment, calculating the matching degree score between the title representation and the content representation of each stock includes:

[0098] Connect the title representation and the content representation of each stock to obtain a connection representation;

[0099] Use a multi-layer perceptron to calculate the matching degree of the connection representation to obtain a matching degree score.

[0100] In one embodiment, using a multi-layer perceptron to calculate the matching degree of the connection representation to obtain a matching degree score includes:

[0101] Use the following formula to calculate the matching degree of the connection representation:

[0102]

[0103] Among them, is the matching degree score, MLP is the multi-layer perceptron, sigmoid is the activation function, r v and r m are the title representation and the content representation respectively, is the objective function of the multi-layer perceptron, and y is the matching degree label.

[0104] Figure 5A schematic structural diagram of the business requirement book recommendation device in an embodiment of the present invention, including:

[0105] A fine-grained semantic extraction module 501, configured to use a fine-grained semantic information extractor to perform fine-grained semantic extraction on the titles and corresponding labels to be matched, and the existing content and corresponding labels of the business requirement book respectively, to obtain the fine-grained business semantics of the titles and the fine-grained business semantics of the content. The fine-grained semantic information extractor performs fine-grained semantic extraction through a label-driven sentence-level attention mechanism and a label-driven paragraph-level attention mechanism;

[0106] A coarse-grained semantic extraction module 502, configured to use a coarse-grained semantic information extractor to perform coarse-grained semantic extraction on the titles and corresponding labels to be matched, and the existing content and corresponding labels of the business requirement book respectively, to obtain the coarse-grained business semantics of the titles and the coarse-grained business semantics of the content. The coarse-grained semantic information extractor performs coarse-grained semantic extraction through a label-driven attention mechanism;

[0107] A matching module 503, configured to calculate the matching degree scores between the titles and the existing content respectively according to the fine-grained business semantics and coarse-grained business semantics of the titles, and the fine-grained business semantics and coarse-grained business semantics of each existing content;

[0108] A content recommendation module 504, configured to recommend the content of the existing item with the highest matching degree score to the business requirement book writer.

[0109] In one embodiment, the fine-grained semantic information extractor includes a first input layer, a first embedding layer, a sentence-level extraction layer, a label-driven sentence-level attention mechanism layer, a first paragraph-level extraction layer, a label-driven paragraph-level attention mechanism layer, and a first output layer; wherein,

[0110] The first input layer is configured to decompose the input object to obtain multiple texts. When the object is a title or a title corresponding label, the text is a word; when the object is content or a content corresponding label, the text is a sentence;

[0111] The first embedding layer is configured to embed the multiple texts into a vector space to obtain the text vectors of each text;

[0112] The sentence-level extraction layer is configured to obtain the sentence matrix representation of the object according to all the text vectors;

[0113] The label-driven sentence-level attention mechanism layer is configured to obtain the attention mechanism sentence matrix representation of the object according to the sentence matrix representation of the object by using the label-driven sentence-level attention mechanism;

[0114] The first paragraph-level extraction layer is configured to obtain the paragraph matrix representation of the object according to the attention mechanism sentence matrix representation of the object;

[0115] A label-driven paragraph-level attention mechanism layer, which is used to obtain the attention mechanism paragraph matrix representation of an object by adopting a label-driven paragraph-level attention mechanism according to the paragraph matrix representation of the object;

[0116] The first output layer is used to output the attention mechanism paragraph matrix representation of the object.

[0117] In one embodiment, the coarse-grained semantic information extractor includes a second input layer, a second embedding layer, a second paragraph-level extraction layer, a label-driven attention mechanism layer, and a second output layer; wherein,

[0118] The second input layer is used to decompose the input object to obtain multiple words, and the object is a title, a label corresponding to the title, content, or a label corresponding to the content;

[0119] The second embedding layer is used to embed multiple words into a vector space to obtain the word vectors of each word;

[0120] The second paragraph-level extraction layer is used to obtain the paragraph matrix representation of the object according to the word vectors of all words of the object;

[0121] A label-driven paragraph-level attention mechanism layer, which is used to obtain the attention mechanism paragraph matrix representation of the object by adopting a label-driven paragraph-level attention mechanism according to the paragraph matrix representation of the object;

[0122] The second output layer is used to output the attention mechanism paragraph matrix representation of the object.

[0123] In one embodiment, the label-driven sentence-level attention mechanism layer is specifically used for: activating the weights of each text in the sentence matrix representation by adopting a label-driven sentence-level attention mechanism according to the sentence matrix representation of the object to obtain the attention mechanism sentence matrix representation of the object;

[0124] The label-driven paragraph-level attention mechanism layer is specifically used for: activating the weights of each sentence in the paragraph matrix representation by adopting a label-driven paragraph-level attention mechanism according to the paragraph matrix representation of the object to obtain the attention mechanism paragraph matrix representation of the object;

[0125] The label-driven attention mechanism layer is specifically used for: activating the weights of each sentence in the paragraph matrix representation by adopting a label-driven attention mechanism according to the paragraph matrix representation of the object to obtain the attention mechanism paragraph matrix representation of the object.

[0126] In one embodiment, the matching module is specifically used for:

[0127] Connect the fine-grained business semantics and the coarse-grained business semantics of the title to obtain the title representation;

[0128] Connect the fine-grained business semantics and the coarse-grained business semantics of the content of each stock to obtain the content representation of each stock;

[0129] Calculate the matching degree score between the title representation and the content representation of each stock.

[0130] In one embodiment, the matching module is specifically used for:

[0131] Connect the title representation and the content representation of each stock to obtain a connection representation;

[0132] Use a multi-layer perceptron to calculate the matching degree of the connection representation to obtain the matching degree score.

[0133] In one embodiment, the matching module is specifically used for:

[0134] Use the following formula to calculate the matching degree of the connection representation:

[0135]

[0136] Wherein, is the matching degree score, MLP is the multi-layer perceptron, sigmoid is the activation function, r v and r m are the title representation and the content representation respectively, is the objective function of the multi-layer perceptron, and y is the matching degree label.

[0137] In summary, in the method and device proposed in the embodiments of the present invention, two different semantic information extractors are used to extract the coarse-grained and fine-grained semantic representations, which not only deeply excavate the structural semantics of the description information, but also use the label-driven attention mechanism to guide the feature extraction process, and the extracted multi-granularity semantic representations have a higher correlation with the function; because both the coarse-grained semantic information extractor and the fine-grained semantic information extractor use the label-driven attention mechanism to extract the business semantics of the description information when extracting the representation, the extracted representation has strong business nature; the effect of the fine-grained semantic information extractor is better than that of the coarse-grained semantic information extractor, because the fine-grained semantic information extractor further excavates the hierarchical structure of the description information, extracts the fine-grained structural semantics, and when matching the multi-granularity semantic information, the coarse-grained semantics and the fine-grained semantics can complement each other, so as to achieve more accurate recommendation.

[0138] The embodiments of the present invention also provide a computer device, Figure 6Schematic diagram of a computer device in an embodiment of the present invention. The computer device 600 includes a memory 610, a processor 620, and a computer program 630 stored in the memory 610 and executable on the processor 620. When the processor 620 executes the computer program 630, the above-mentioned business requirement document recommendation method is implemented.

[0139] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above-mentioned business requirement document recommendation method is implemented.

[0140] An embodiment of the present invention further provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the above-mentioned business requirement document recommendation method is implemented.

[0141] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0142] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0143] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0144] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, thereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one process or a plurality of processes and / or blocks Figure 1 one process or a plurality of processes and / or blocks Figure 1 steps for implementing the functions specified in one block or a plurality of blocks.

[0145] The specific embodiments described above further elaborate the object, technical solution and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A business requirement document recommendation method, characterized in that, Including: Using a fine-grained semantic information extractor, respectively perform fine-grained semantic extraction on the title to be matched and its corresponding label, and the existing content and its corresponding label in the business requirement document, to obtain the fine-grained business semantics of the title and the fine-grained business semantics of the content. The fine-grained semantic information extractor performs fine-grained semantic extraction through a label-driven sentence-level attention mechanism and a label-driven paragraph-level attention mechanism; Using a coarse-grained semantic information extractor, respectively perform coarse-grained semantic extraction on the title to be matched and its corresponding label, and the existing content and its corresponding label in the business requirement document, to obtain the coarse-grained business semantics of the title and the coarse-grained business semantics of the content. The coarse-grained semantic information extractor performs coarse-grained semantic extraction through a label-driven attention mechanism; According to the fine-grained business semantics and coarse-grained business semantics of the title, as well as the fine-grained business semantics and coarse-grained business semantics of the content of each existing item, calculate the matching degree score between the title and the content of each existing item; Recommend the content of the existing item with the highest matching degree score to the business requirement document writer.

2. The method according to claim 1, wherein The fine-grained semantic information extractor includes a first input layer, a first embedding layer, a sentence-level extraction layer, a label-driven sentence-level attention mechanism layer, a first paragraph-level extraction layer, a label-driven paragraph-level attention mechanism layer, and a first output layer; among them, The first input layer is used to decompose the input object to obtain multiple texts. When the object is a title or a title corresponding label, the text is a word. When the object is content or a content corresponding label, the text is a sentence; The first embedding layer is used to embed multiple texts into a vector space to obtain the text vector of each text; The sentence-level extraction layer is used to obtain the sentence matrix representation of the object according to all text vectors; The label-driven sentence-level attention mechanism layer is used to obtain the attention mechanism sentence matrix representation of the object by using the label-driven sentence-level attention mechanism according to the sentence matrix representation of the object; The first paragraph-level extraction layer is used to obtain the paragraph matrix representation of the object according to the attention mechanism sentence matrix representation of the object; The label-driven paragraph-level attention mechanism layer is used to obtain the attention mechanism paragraph matrix representation of the object by using the label-driven paragraph-level attention mechanism according to the paragraph matrix representation of the object; The first output layer is used to output the attention mechanism paragraph matrix representation of the object.

3. The method according to claim 2, wherein The coarse-grained semantic information extractor includes a second input layer, a second embedding layer, a second paragraph-level extraction layer, a label-driven attention mechanism layer, and a second output layer; among them, The second input layer is used to decompose the input object to obtain multiple words, and the object is a title, a title corresponding label, content, or a content corresponding label; The second embedding layer is used to embed multiple words into a vector space to obtain the word vector of each word; The second paragraph-level extraction layer is used to obtain the paragraph matrix representation of the object according to the word vectors of all words of the object; The label-driven attention mechanism layer is used to obtain the attention mechanism paragraph matrix representation of the object by using the label-driven attention mechanism according to the paragraph matrix representation of the object; The second output layer is used to output the attention mechanism paragraph matrix representation of the object.

4. The method according to claim 3, characterized in that, The label-driven sentence-level attention mechanism layer is specifically used for: according to the sentence matrix representation of the object, activating the weights of each text in the sentence matrix representation by using the label-driven sentence-level attention mechanism to obtain the attention mechanism sentence matrix representation of the object; The label-driven paragraph-level attention mechanism layer is specifically used for: according to the paragraph matrix representation of the object, activating the weights of each sentence in the paragraph matrix representation by using the label-driven paragraph-level attention mechanism to obtain the attention mechanism paragraph matrix representation of the object; The label-driven attention mechanism layer is specifically used for: according to the paragraph matrix representation of the object, activating the weights of each sentence in the paragraph matrix representation by using the label-driven attention mechanism to obtain the attention mechanism paragraph matrix representation of the object.

5. The method according to claim 1, characterized in that Calculate the matching degree scores between the title and the content of each stock according to the fine-grained business semantics, coarse-grained business semantics of the title, and the fine-grained business semantics and coarse-grained business semantics of the content of each stock, including: Connect the fine-grained business semantics and coarse-grained business semantics of the title to obtain the title representation; Connect the fine-grained business semantics and coarse-grained business semantics of the content of each stock to obtain the content representation of each stock; Calculate the matching degree scores between the title representation and the content representation of each stock.

6. The method according to claim 5, wherein Calculate the matching degree scores between the title representation and the content representation of each stock, including: Connect the title representation and the content representation of each stock to obtain a connection representation; Use a multi-layer perceptron to calculate the matching degree of the connection representation to obtain the matching degree score.

7. The method according to claim 6, wherein Use a multi-layer perceptron to calculate the matching degree of the connection representation to obtain the matching degree score, including: Use the following formula to calculate the matching degree of the connection representation: Among them, is the matching degree score, MLP is the multi-layer perceptron, sigmoid is the activation function, r v and r m are the title representation and the content representation respectively, is the objective function of the multi-layer perceptron, and y is the matching degree label.

8. A business requirement document recommendation device, characterized in that, Including: The fine-grained semantic extraction module is used to respectively perform fine-grained semantic extraction on the title to be matched and its corresponding label, and the content and its corresponding label of the stock in the business requirement document by using a fine-grained semantic information extractor to obtain the fine-grained business semantics of the title and the fine-grained business semantics of the content. The fine-grained semantic information extractor performs fine-grained semantic extraction through a label-driven sentence-level attention mechanism and a label-driven paragraph-level attention mechanism; The coarse-grained semantic extraction module is used to respectively perform coarse-grained semantic extraction on the title to be matched and its corresponding label, and the content and its corresponding label of the stock in the business requirement document by using a coarse-grained semantic information extractor to obtain the coarse-grained business semantics of the title and the coarse-grained business semantics of the content. The coarse-grained semantic information extractor performs coarse-grained semantic extraction through a label-driven attention mechanism; The matching module is used to calculate the matching degree scores between the title and the content of each stock according to the fine-grained business semantics, coarse-grained business semantics of the title, and the fine-grained business semantics and coarse-grained business semantics of the content of each stock; The content recommendation module is used to recommend the content of the stock with the highest matching degree score to the business requirement document writer.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the method according to any one of claims 1 to 7.

11. A computer program product, characterized in that, The computer program product includes a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images