WO2020140635A1

WO2020140635A1 - Text matching method and apparatus, storage medium and computer device

Info

Publication number: WO2020140635A1
Application number: PCT/CN2019/118532
Authority: WO
Inventors: 于凤英; 王健宗
Original assignee: 平安科技（深圳）有限公司
Priority date: 2019-01-04
Filing date: 2019-11-14
Publication date: 2020-07-09
Also published as: CN109740126A; CN109740126B

Abstract

A text matching method and apparatus, a storage medium and a computer device. The method comprises: receiving an input target text (S100); obtaining a plurality of candidate sentences obtained by preliminary matching according to the target text (S200); inputting the target text and each candidate sentence into a text matching model formed by a convolutional neural network (CNN) and a GRU neural network, to obtain the semantic similarity between each candidate sentence and the target text, wherein the text matching model is used for representing the semantic similarities between the target text and the candidate sentences (S300); and recommending the candidate sentences to a user according to the semantic similarity corresponding to each candidate sentence (S400). According to the method above, the most semantically matched sentence in the candidate sentences for text matching can be obtained without manually defining a feature template, and the screening efficiency is improved.

Description

Text matching method, device, storage medium, and computer equipment

This application requires the priority of the Chinese patent application submitted to the China Patent Office on January 4, 2019, with the application number 201910008683.8 and the invention titled "text matching method, device and storage medium, computer equipment", the entire content of which is incorporated by reference In this application.

Technical field

The present application relates to the field of text processing technology. Specifically, the present application relates to a text matching method, device, storage medium, and computer equipment.

Background technique

Text matching is to measure the relevance or matching degree of the search term and the text on the text. The text matching method is an indispensable technology in the search system. In the retrieval question answering system, a very important step is to sort the retrieved answers to get the best answer. In other words, given a sentence and many candidate sentences in natural language form, according to the sentence, find the best matching sentence among the many candidate sentences.

The inventor realized that the traditional text matching method usually performs matching operation on the text according to the word frequency. For example, the TF-IDF (term frequency-inverse document frequency) algorithm. However, the text matching using word segmentation and word frequency has certain limitations in the accuracy of the matching results, and it is not possible to determine other texts with a high matching rate.

Summary of the invention

The present application proposes a text matching method, device, storage medium, and computer equipment, so as to achieve the sentence with the most matching semantics among text matching candidate sentences without manually defining a feature template.

In the first aspect, the present application discloses a text matching method, comprising: receiving input target text; obtaining a plurality of candidate sentences obtained by preliminary matching according to the target text; and inputting the target text with each of the candidate sentences A text matching model composed of a convolutional neural network CNN and a GRU neural network to obtain the semantic similarity between each candidate sentence and the target text; wherein the text matching model is used to characterize the target text and the candidate The semantic similarity of the sentences; according to the semantic similarity corresponding to each of the candidate sentences, recommend the candidate sentences to the user.

In a second aspect, the present application discloses a text matching device, including: a receiving module for acquiring target text input; a first acquiring module for acquiring a plurality of candidate sentences obtained by preliminary matching according to the target text; A second obtaining module, configured to input the target text and each candidate sentence into a text matching model formed by a convolutional neural network CNN and a GRU neural network, and obtain that each candidate sentence is semantically similar to the target text Wherein the text matching model is used to characterize the semantic similarity between the target text and the candidate sentence; the recommendation module is used to recommend the candidate to the user based on the semantic similarity corresponding to each candidate sentence Statement.

In a third aspect, the present application discloses a computer device, including: one or more processors; a memory; one or more computer programs, wherein the one or more computer programs are stored in the memory and configured as Executed by the one or more processors, the one or more computer programs are configured to perform a text matching method, the text matching method includes the following steps: receiving input target text; acquiring according to the target text Multiple candidate sentences obtained by preliminary matching; input the target text and each candidate sentence into a text matching model composed of a convolutional neural network CNN and a GRU neural network to obtain each candidate sentence and the target text Semantic similarity; wherein the text matching model is used to characterize the semantic similarity between the target text and the candidate sentence; according to the semantic similarity corresponding to each candidate sentence, recommend the candidate sentence to the user.

In a fourth aspect, the present application discloses a storage medium on which a computer program is stored; the computer program is adapted to be loaded by a processor and execute a text matching method, the text matching method includes the following steps: receiving an input target Text; obtain multiple candidate sentences obtained by preliminary matching according to the target text; input the target text and each candidate sentence into a text matching model composed of a convolutional neural network CNN and a GRU neural network to obtain each The semantic similarity between the candidate sentence and the target text; wherein the text matching model is used to characterize the semantic similarity between the target text and the candidate sentence; according to the semantic similarity corresponding to each candidate sentence, Recommend the candidate sentence to the user.

Additional aspects and advantages of the present application will be partially given in the following description, which will become apparent from the following description, or be learned through the practice of the present application.

BRIEF DESCRIPTION

The above-mentioned and/or additional aspects and advantages of this application will become apparent and easy to understand from the following description of the embodiments in conjunction with the accompanying drawings, in which:

1 is a flowchart of a method in an embodiment of a text matching method provided by this application;

FIG. 2 is a flowchart of the method in an embodiment of step S300 provided by the present application;

3 is a flowchart of a training method in an embodiment of a text matching model provided by this application;

4 is a structural block diagram of an embodiment of a text matching device provided by this application;

FIG. 5 is a schematic structural diagram of an embodiment of a computer device provided by this application.

detailed description

Those skilled in the art can understand that the term “terminal” and “terminal device” used here include not only devices with wireless signal receivers, but only devices with wireless signal receivers that do not have transmitting capabilities, but also devices that receive and transmit hardware. A device having a device capable of performing receiving and transmitting hardware for bidirectional communication on a bidirectional communication link. Such devices may include: cellular or other communication devices with single-line displays or multi-line displays or cellular or other communication devices without multi-line displays; PCS (Personal Communications Services), which can combine voice and data Processing, fax and/or data communication capabilities; PDA (Personal Digital Assistant), which can include radio frequency receivers, pagers, Internet/Intranet access, web browsers, notepads, calendars and/or GPS (Global Positioning System (Global Positioning System) receiver; conventional laptop and/or palmtop computer or other device that has and/or includes a conventional radio frequency receiver and/or palmtop computer or other device. As used herein, "terminal" and "terminal equipment" may be portable, transportable, installed in a vehicle (aeronautical, maritime, and/or terrestrial), or suitable and/or configured to operate locally, and/or In a distributed form, it operates on any other location on the earth and/or in space. The "terminal" and "terminal device" used herein may also be a communication terminal, an Internet terminal, a music/video playback terminal, for example, may be a PDA, MID (Mobi Internet) Device, and/or have music/video The mobile phone with playback function can also be a smart TV, set-top box and other devices.

This application provides a text matching method. In an embodiment, as shown in FIG. 1, the text matching method includes the following steps:

S100. Receive the input target text.

In this embodiment, the system receives the target text input by the user. The target text may be a paragraph of characters. Specifically, it may be a searchable question and answer system in which the user inputs a target text.

S200. Acquire a plurality of candidate sentences obtained by preliminary matching according to the target text.

In this embodiment, after receiving the target text input by the user, the system may initially filter to obtain a plurality of matching candidate sentences according to the target text. The initial screening method here may be a general text matching method. For example, by cutting the target text, and analyzing the semantics of each word segmentation after the word cutting, the overall semantics of the target text are determined to obtain multiple candidate sentences that match the target text semantically from the database. Or, according to a conventional technical method in the technical field, multiple candidate sentences that semantically match the target text are obtained from the system database.

S300. Input the target text and each candidate sentence into a text matching model composed of a convolutional neural network CNN and a GRU neural network to obtain a semantic similarity between each candidate sentence and the target text.

In this embodiment, it should be noted that the text matching model is used to characterize the semantic similarity between the target text and the candidate sentence. The system matches each candidate sentence in the multiple candidate sentences with the target text input text matching model to obtain the semantic similarity between each candidate sentence and the target text. Among them, the text matching model is a model composed of CNN and GRU neural network.

In this scheme, convolutional neural network CNN is more suitable for the learning of sentence vectors, but recurrent neural network is more suitable for the learning of temporal relationships. Therefore, in the text matching model used in this embodiment, the convolutional neural network is used to convolution the sentence, and then the results of the CNN are input into the GRU neural network, and the recurrent neural network is used to condense the words. The product representation learns its sequence relationship, so that the resulting sentence vector representation can characterize more sentence information and sentence features. The following explains the CNN and GRU neural network:

Convolutional neural network (Convolutional Neural Network, CNN) is one of the most common network structures in the deep learning network model, and its recognition effect in the image field has been greatly improved, which makes the convolutional neural network famous. Vibrate. The two most important layers of the convolutional neural network are the convolutional layer and the pooling layer, which are also the core steps of the convolutional neural network. The following figure shows a typical CNN structure, called LeNet-5 network, which includes multiple convolutional layers and pooling layers. The convolutional layer (Convolutional Layer) uses a convolution kernel (parameter matrix) to slide the windows one by one in the entire picture matrix to perform inner product to obtain another intermediate matrix. The size of the intermediate matrix depends on the dimension of the convolution kernel. Pooling layer (Pooling Layer), also called subsampling layer (Subsampling Layer), after the convolution layer is convolved, the corresponding features can be obtained, and the obtained features can be used for classifier training. However, this method still has a huge amount of calculation, which is prone to overfitting problems. To further reduce the fitting degree of the model, the convolutional layer needs to be pooled.

Recurrent Neural Network (RNN) is widely used to process variable-length text sequence input, which can learn the word order features of sentences. The key structure is a memory unit (Memory Unit). The memory unit can memorize the information of a certain period of time, and can selectively memorize the information of the words of the previous moment for a sentence. At present, there are mainly two different variants of LSTM and GRU in recurrent neural networks. They can solve the problem of long-term distance dependence and gradient disappearance in traditional RNN. Compared with the LSTM network structure, the hidden unit inside the GRU has one less control gate, fewer parameters, and faster convergence. While ensuring the model effect, the model structure has been effectively simplified.

In this embodiment, the CNN and the GRU neural network are combined to obtain the text matching model. Among them, the text matching model is used to characterize the semantic similarity of two input texts. Through the text matching model, the similarity value of the two input texts can be obtained to judge the matching degree of the two input texts.

In an embodiment, as shown in FIG. 2, step S300 includes:

S310, input the target text into the convolutional neural network CNN for convolution processing to obtain a first convolution vector, and input the candidate sentence into the convolutional neural network CNN for convolution processing to obtain a second convolution vector.

S320: Input the first convolution vector into the GRU neural network, and then enter the first convolution vector into the GRU neural network to obtain a second neural network vector.

S330: Obtain a semantic similarity between the candidate sentence and the target text according to the cosine similarity between the first neural network vector and the second neural network vector.

In this embodiment, the system first uses a convolutional neural network to convolve the target text and candidate sentences, and then enters the results of the convolutional neural network CNN into the GRU neural network, and uses the recurrent neural network to target the text and candidate sentences. To learn the time sequence of the words in the two to get more sentence information and sentence characteristics, and then determine the similarity between the target text and the candidate sentence according to the sentence information and sentence characteristics of the target text and the candidate sentence .

In an embodiment, as shown in FIG. 3, the text matching model composed of the CNN and GRU neural network is trained according to the following manner:

S10. Acquire a target training sentence, a first training sentence semantically similar to the target training sentence, and a second training sentence not semantically similar to the target training sentence.

S20. Use the convolutional neural network CNN to perform convolution processing on the target training sentence, the first training sentence, and the second training sentence, respectively, to obtain a first vector corresponding to the target training sentence, and A second vector corresponding to the first training sentence, and a third vector corresponding to the second training sentence.

In an implementation manner of this embodiment, step S20 includes: setting a convolution window of the convolutional neural network CNN to preset N words; using the set convolutional neural network CNN to respectively target the target The training sentence, the first training sentence and the second training sentence are subjected to convolution processing.

S30: Input the first vector, the second vector, and the third vector into the GRU neural network, respectively, to obtain a fourth vector corresponding to the first vector, and a third vector corresponding to the second vector Five vectors and a sixth vector corresponding to the third vector.

In an implementation manner of this embodiment, after step S30, the method further includes: respectively passing the fourth vector, the fifth vector, and the sixth vector through a pooling layer to analyze the fourth vector, the The fifth vector and the sixth vector undergo dimension change processing. At this time, step S30 includes: according to the cosine similarity between the fourth vector and the fifth vector after the dimensional change processing, and the cosine similarity between the fourth vector and the sixth vector after the dimensional change processing, to obtain A first score for the target training sentence and the first training sentence, and a second score for the target training sentence and the second training sentence.

S40. According to the cosine similarity between the fourth vector and the fifth vector, and the cosine similarity between the fourth vector and the sixth vector, obtain the first score of the target training sentence and the first training sentence Value, and the second score of the target training sentence and the second training sentence.

S50. Determine related parameters in the text matching model according to the first score and the second score.

In this embodiment, the system acquires the target training sentence, the first training sentence semantically similar to the target training sentence, and the second training sentence not semantically similar to the target training sentence as the training corpus. The text matching model formed by the network is used for model training. Specifically, the target training sentence, the first training sentence, and the second training sentence are respectively input into the convolutional neural network CNN for preliminary convolution training, and the first vector corresponding to the target training sentence and the first training sentence are respectively obtained The second vector of, and the third vector corresponding to the second training sentence. Further, the first vector, the second vector, and the third vector are respectively input into the GRU neural network to obtain a fourth vector corresponding to the first vector, a fifth vector corresponding to the second vector, and a sixth vector corresponding to the third vector vector. Finally, the cosine similarity between the fourth vector and the fifth vector and the cosine similarity between the fourth vector and the sixth vector are obtained to obtain the corresponding cosine similarity score, and the text matching model is determined according to the corresponding cosine similarity score The parameters associated in.

In one implementation of this embodiment, step S50 includes: determining the associated parameters of the cost function corresponding to the text matching model according to the first score and the second score; the cost function includes a hinge The loss function Hinge loss.

For the training method of the text matching model composed of the CNN and the GRU neural network described in the above embodiment, a specific embodiment is provided below for detailed description:

Specifically, in the text matching model training process, the text matching model inputs three sentences, the current input sentence (target training sentence, Sentence), a sentence similar to the current input sentence (Similar, semantic, sentence), and the current input sentence Different semantic sentences (Different semantic). First, use CNN's convolutional layer to convolve three sentences, and set the convolution window to two words, and learn the corresponding sentence vector; then use the recurrent neural network to further obtain the timing information of the sentence; and then pass through a layer of pools The mean/max pooling (mean/max pooling, mean value pooling or maximum pooling), do a dimensional change of the vector to prevent the model from overfitting; then use this vector to calculate the similarity (cosine), the similarity uses cosine Similarity, get the score of the currently input sentence and the semantic similarity and dissimilarity. Among them, the cost function uses hinge loss, the purpose is to make the score of semantically similar sentences higher than the score of semantically dissimilar sentences, not just close to 1.

Therefore, before the pooling layer, two deep learning network structures, CNN and GRU, are used to learn sentence information. In the specific implementation process, CNN's convolution window is two words, and the recurrent neural network uses the basic model GRU. This embodiment does not use a bidirectional GRU, and does not require more complicated improvements to optimize, so as to avoid the problem of overfitting and slow training due to too many parameters. Combined with the construction of the model graph, it can be seen that this model uses an additional layer of recurrent neural network to learn sentences compared to the CNN model, and can learn more information and features.

S400: Recommend the candidate sentence to the user according to the semantic similarity corresponding to each candidate sentence.

In this embodiment, each candidate sentence in the plurality of candidate sentences and the target text are input into the above text matching model, respectively, to obtain the semantic similarity between each candidate sentence and the target text, thereby according to each candidate sentence The corresponding semantic similarity recommends to the user candidate sentences that are semantically similar to the target text. In one embodiment, step S400 includes: filtering out candidate sentences with the highest semantic similarity from the plurality of candidate sentences according to the semantic similarity corresponding to each of the candidate sentences, and recommending the semantic similarity to the user The candidate with the highest degree.

The following provides a specific implementation scenario to further illustrate the above text matching method:

In this specific implementation scenario, the text matching method can be used in a retrieval question answering system. In the retrieval question answering system, the user inputs a target text, and the system reads a plurality of candidate sentences for the semantic matching of the target text from the corresponding database by analyzing the target text. Generally, in this step, the system obtains multiple candidate sentences through preliminary matching, but the semantic similarity between each candidate sentence and the target sentence is not determined. Therefore, it is difficult to filter out the candidate sentences that best match the semantics of the target text input by the user. In this implementation scenario, a text matching model composed of CNN and GRU neural network is provided. Each candidate sentence and target text are input into the text matching model to obtain the semantic meaning of each candidate sentence and target text. Similarity, so that the candidate sentences with the highest similarity can be selected and recommended to the user. Therefore, there is no need to manually define the feature template, and finally obtain the sentence with the best semantic match among the candidate sentences.

The present application also provides a text matching device. As shown in FIG. 4, in an embodiment, the text matching device includes a receiving module 100, a first acquiring module 200, a second acquiring module 300, and a recommendation module 400.

The receiving module 100 is used to obtain the input target text. In this embodiment, the system receives the target text input by the user. The target text may be a paragraph of characters. Specifically, it may be a searchable question and answer system in which the user inputs a target text.

The first obtaining module 200 is used to obtain a plurality of candidate sentences obtained by preliminary matching according to the target text. In this embodiment, after receiving the target text input by the user, the system may initially filter to obtain a plurality of matching candidate sentences according to the target text. The initial screening method here may be a general text matching method. For example, by cutting the target text, and analyzing the semantics of each word segmentation after the word cutting, the overall semantics of the target text are determined to obtain multiple candidate sentences that match the target text semantically from the database. Or, according to a conventional technical method in the technical field, multiple candidate sentences that semantically match the target text are obtained from the system database.

The second obtaining module 300 is used to input the target text and each of the candidate sentences into a text matching model composed of a convolutional neural network CNN and a GRU neural network, and obtain that each candidate sentence is semantically similar to the target text Degree; wherein, the text matching model is used to characterize the semantic similarity of the target text and the candidate sentence. In this embodiment, it should be noted that the text matching model is used to characterize the semantic similarity between the target text and the candidate sentence. The system matches each candidate sentence in the multiple candidate sentences with the target text input text matching model to obtain the semantic similarity between each candidate sentence and the target text. Among them, the text matching model is a model composed of CNN and GRU neural network.

The recommendation module 400 is used to recommend the candidate sentence to the user according to the semantic similarity corresponding to each of the candidate sentences. In this embodiment, each candidate sentence in the plurality of candidate sentences and the target text are input into the above text matching model, respectively, to obtain the semantic similarity between each candidate sentence and the target text, thereby according to each candidate sentence The corresponding semantic similarity recommends to the user candidate sentences that are semantically similar to the target text.

In other embodiments, each module in the text matching device provided by this application is also used to perform the operations performed in accordance with each step in the text matching method described in this application, and no detailed description will be given here.

The application also provides a storage medium. A computer program is stored on the storage medium; when the computer program is executed by the processor, the text matching method described in any of the above embodiments is implemented. The storage medium may be a memory. The storage medium in this embodiment is a volatile storage medium or a non-volatile storage medium. For example, internal memory or external memory, or include both internal memory and external memory. The internal memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory. The external memory may include hard disks, floppy disks, ZIP disks, U disks, magnetic tapes, etc. The storage media disclosed in this application include but are not limited to these types of memories. The memory disclosed in this application is only an example and not a limitation.

This application also provides a computer device. A computer device includes: one or more processors; memory; one or more application programs. Wherein the one or more application programs are stored in the memory and are configured to be executed by the one or more processors, and the one or more application programs are configured to perform the operations described in any of the foregoing embodiments Text matching method.

5 is a schematic structural diagram of a computer device in an embodiment of the present application. The computer device in this embodiment may be a server, a personal computer, and a network device. As shown in FIG. 5, the device includes a processor 703, a memory 705, an input unit 707, a display unit 709, and other devices. Those skilled in the art may understand that the device structure device shown in FIG. 5 does not constitute a limitation on all devices, and may include more or less components than those illustrated, or combine certain components. The memory 705 may be used to store application programs 701 and various functional modules. The processor 703 runs the application programs 701 stored in the memory 705 to execute various functional applications and data processing of the device. The memory may be internal memory or external memory, or include both internal memory and external memory. The internal memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory. The external memory may include hard disks, floppy disks, ZIP disks, U disks, magnetic tapes, etc. The memories disclosed in this application include but are not limited to these types of memories. The memory disclosed in this application is only an example and not a limitation.

The input unit 707 is used to receive an input of a signal and a keyword input by a user. The input unit 707 may include a touch panel and other input devices. The touch panel can collect the user's touch operations on or near it (such as the user's operation with any suitable objects or accessories such as fingers, stylus, etc. on or near the touch panel), and according to the preset The program drives the corresponding connection device; other input devices may include but are not limited to one or more of a physical keyboard, function keys (such as playback control keys, switch keys, etc.), trackball, mouse, joystick, etc. The display unit 709 can be used to display information input by the user or information provided to the user and various menus of the computer device. The display unit 709 may take the form of a liquid crystal display, an organic light emitting diode, or the like. The processor 703 is the control center of the computer equipment. It uses various interfaces and lines to connect the various parts of the entire computer. By running or executing the software programs and/or modules stored in the memory 705, and calling the data stored in the memory, the Various functions and processing data.

In one embodiment, the device includes one or more processors 703, and one or more memories 705, and one or more applications 701. The one or more application programs 701 are stored in the memory 705 and are configured to be executed by the one or more processors 703, and the one or more application programs 701 are configured to perform the operations described in the above embodiments Text matching method.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units are integrated into one module. The above integrated modules can be implemented in the form of hardware or software function modules. If the integrated module is implemented in the form of a software function module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

A person of ordinary skill in the art may understand that all or part of the steps for implementing the above-mentioned embodiments may be completed by hardware, or by a program instructing related hardware. The program may be stored in a computer-readable storage medium, and the storage medium may include Memory, magnetic disk or optical disk, etc.

The above is only part of the implementation of this application. It should be pointed out that for those of ordinary skill in the art, without departing from the principles of this application, a number of improvements and retouches can be made. These improvements and retouches also It should be regarded as the scope of protection of this application.

It should be understood that the functional units in the embodiments of the present application may be integrated into one processing module, or each unit may exist alone physically, or two or more units may be integrated into one module. The above integrated modules can be implemented in the form of hardware or software function modules.

Claims

A text matching method, including:

Receive the input target text;

Acquiring multiple candidate sentences obtained by preliminary matching according to the target text;

Input the target text and each candidate sentence into a text matching model composed of a convolutional neural network CNN and a GRU neural network to obtain a semantic similarity between each candidate sentence and the target text; wherein, the text The matching model is used to characterize the semantic similarity between the target text and the candidate sentence;

According to the semantic similarity corresponding to each candidate sentence, the candidate sentence is recommended to the user.
The method according to claim 1, the inputting the target text and each candidate sentence into a text matching model formed by a convolutional neural network CNN and a GRU neural network to obtain each candidate sentence and the target The semantic similarity of text, including:

Input the target text into the convolutional neural network CNN to perform convolution processing to obtain a first convolution vector, and input the candidate sentence into the convolutional neural network CNN to perform convolution processing to obtain a second convolution vector;

Input the first convolution vector into the GRU neural network to the first neural network vector, and input the second convolution vector into the GRU neural network to obtain a second neural network vector;

According to the cosine similarity between the first neural network vector and the second neural network vector, a semantic similarity between the candidate sentence and the target text is obtained.
The method according to claim 1, the text matching model composed of the CNN and the GRU neural network is trained according to the following manner:

Acquiring a target training sentence, a first training sentence semantically similar to the target training sentence, and a second training sentence not semantically similar to the target training sentence;

Use the convolutional neural network CNN to perform convolution processing on the target training sentence, the first training sentence, and the second training sentence, respectively, to obtain a first vector corresponding to the target training sentence, and the A second vector corresponding to the first training sentence, and a third vector corresponding to the second training sentence;

Input the first vector, the second vector and the third vector into the GRU neural network respectively to obtain a fourth vector corresponding to the first vector and a fifth vector corresponding to the second vector And a sixth vector corresponding to the third vector;

Obtain the first score of the target training sentence and the first training sentence based on the cosine similarity between the fourth vector and the fifth vector, and the cosine similarity between the fourth vector and the sixth vector, And the second score of the target training sentence and the second training sentence;

The associated parameters in the text matching model are determined according to the first score and the second score.
According to the method of claim 3, the convolutional neural network CNN respectively performs convolution processing on the target training sentence, the first training sentence, and the second training sentence, including:

Setting the convolution window of the convolutional neural network CNN to preset N words;

The convolutional neural network CNN is used to perform convolution processing on the target training sentence, the first training sentence, and the second training sentence, respectively.
The method according to claim 3, the inputting the first vector, the second vector and the third vector to the GRU neural network respectively to obtain a fourth vector corresponding to the first vector, After the fifth vector corresponding to the second vector and the sixth vector corresponding to the third vector, the method further includes: pooling the fourth vector, the fifth vector, and the sixth vector, respectively Layer to perform dimension change processing on the fourth vector, the fifth vector, and the sixth vector;

Obtaining the first score of the target training sentence and the first training sentence according to the cosine similarity of the fourth vector and the fifth vector, and the cosine similarity of the fourth vector and the sixth vector, respectively Value, and the second score of the target training sentence and the second training sentence, including: the cosine similarity between the fourth vector and the fifth vector after the dimensional change processing and the dimensional change processing The cosine similarity of the fourth vector and the sixth vector to obtain the first score of the target training sentence and the first training sentence, and the second score of the target training sentence and the second training sentence value.
The method according to claim 3, the determining the associated parameters in the text matching model according to the first score and the second score includes:

The related parameters of the cost function corresponding to the text matching model are determined according to the first score and the second score; the cost function includes a hinge loss function Hinge loss.
According to the method of claim 1, the recommending the candidate sentence to the user according to the semantic similarity corresponding to each of the candidate sentences includes:

According to the semantic similarity corresponding to each candidate sentence, the candidate sentence with the highest semantic similarity is selected from the plurality of candidate sentences, and the candidate sentence with the highest semantic similarity is recommended to the user.
A text matching device, including:

The receiving module is used to obtain the target text of the received input;

A first obtaining module, configured to obtain a plurality of candidate sentences obtained by preliminary matching according to the target text;

A second obtaining module, configured to input the target text and each candidate sentence into a text matching model formed by a convolutional neural network CNN and a GRU neural network, and obtain that each candidate sentence is semantically similar to the target text Degree; wherein, the text matching model is used to characterize the semantic similarity between the target text and the candidate sentence;

The recommendation module is configured to recommend the candidate sentence to the user according to the semantic similarity corresponding to each candidate sentence.
A computer device, including:

One or more processors;

Memory

One or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, the one or more computer programs are configured to execute A text matching method, the text matching method includes the following steps:

Receive the input target text;

Acquiring multiple candidate sentences obtained by preliminary matching according to the target text;

Input the target text and each candidate sentence into a text matching model composed of a convolutional neural network CNN and a GRU neural network to obtain a semantic similarity between each candidate sentence and the target text; wherein, the text The matching model is used to characterize the semantic similarity between the target text and the candidate sentence;

According to the semantic similarity corresponding to each candidate sentence, the candidate sentence is recommended to the user.
The computer device according to claim 9, wherein the target text and each candidate sentence are input into a text matching model formed by a convolutional neural network CNN and a GRU neural network, and each candidate sentence and the The semantic similarity of the target text, including:

Input the target text into the convolutional neural network CNN to perform convolution processing to obtain a first convolution vector, and input the candidate sentence into the convolutional neural network CNN to perform convolution processing to obtain a second convolution vector;

Input the first convolution vector into the GRU neural network to the first neural network vector, and input the second convolution vector into the GRU neural network to obtain a second neural network vector;

According to the cosine similarity between the first neural network vector and the second neural network vector, a semantic similarity between the candidate sentence and the target text is obtained.
According to the computer device of claim 9, the text matching model composed of the CNN and GRU neural network is trained according to the following manner:

Acquiring a target training sentence, a first training sentence semantically similar to the target training sentence, and a second training sentence not semantically similar to the target training sentence;

Use the convolutional neural network CNN to perform convolution processing on the target training sentence, the first training sentence, and the second training sentence, respectively, to obtain a first vector corresponding to the target training sentence, and the A second vector corresponding to the first training sentence, and a third vector corresponding to the second training sentence;

Input the first vector, the second vector and the third vector into the GRU neural network respectively to obtain a fourth vector corresponding to the first vector and a fifth vector corresponding to the second vector And a sixth vector corresponding to the third vector;

Obtain the first score of the target training sentence and the first training sentence based on the cosine similarity between the fourth vector and the fifth vector, and the cosine similarity between the fourth vector and the sixth vector, And the second score of the target training sentence and the second training sentence;

The associated parameters in the text matching model are determined according to the first score and the second score.
The computer device according to claim 11, wherein the convolutional neural network CNN respectively performs convolution processing on the target training sentence, the first training sentence, and the second training sentence, including:

Setting the convolution window of the convolutional neural network CNN to preset N words;

The convolutional neural network CNN is used to perform convolution processing on the target training sentence, the first training sentence, and the second training sentence, respectively.
The computer device according to claim 11, the inputting the first vector, the second vector and the third vector into the GRU neural network respectively to obtain a fourth vector corresponding to the first vector After the fifth vector corresponding to the second vector and the sixth vector corresponding to the third vector, the method further includes: passing the fourth vector, the fifth vector, and the sixth vector through the pool, respectively A layer to perform dimension change processing on the fourth vector, the fifth vector, and the sixth vector;

Obtaining the first score of the target training sentence and the first training sentence according to the cosine similarity of the fourth vector and the fifth vector, and the cosine similarity of the fourth vector and the sixth vector, respectively Value, and the second score of the target training sentence and the second training sentence, including: the cosine similarity between the fourth vector and the fifth vector after the dimensional change processing and the dimensional change processing The cosine similarity of the fourth vector and the sixth vector to obtain the first score of the target training sentence and the first training sentence, and the second score of the target training sentence and the second training sentence value.
The computer device according to claim 11, said determining the associated parameters in the text matching model according to the first score and the second score includes:

The associated parameters of the cost function corresponding to the text matching model are determined according to the first score and the second score; the cost function includes a hinge loss function Hinge loss.
According to the computer device of claim 9, the recommendation of the candidate sentence to the user according to the semantic similarity corresponding to each of the candidate sentences includes:

According to the semantic similarity corresponding to each candidate sentence, the candidate sentence with the highest semantic similarity is selected from the plurality of candidate sentences, and the candidate sentence with the highest semantic similarity is recommended to the user.
A storage medium on which a computer program is stored; the computer program is adapted to be loaded by a processor and execute a text matching method, including:

Receive the input target text;

Acquiring multiple candidate sentences obtained by preliminary matching according to the target text;

Input the target text and each candidate sentence into a text matching model composed of a convolutional neural network CNN and a GRU neural network to obtain a semantic similarity between each candidate sentence and the target text; wherein, the text The matching model is used to characterize the semantic similarity between the target text and the candidate sentence;

According to the semantic similarity corresponding to each candidate sentence, the candidate sentence is recommended to the user.
The storage medium according to claim 16, wherein the target text and each candidate sentence are input into a text matching model composed of a convolutional neural network CNN and a GRU neural network, and each candidate sentence and the The semantic similarity of the target text, including:

Input the target text into the convolutional neural network CNN to perform convolution processing to obtain a first convolution vector, and input the candidate sentence into the convolutional neural network CNN to perform convolution processing to obtain a second convolution vector;

Input the first convolution vector into the GRU neural network to the first neural network vector, and input the second convolution vector into the GRU neural network to obtain a second neural network vector;

According to the cosine similarity between the first neural network vector and the second neural network vector, a semantic similarity between the candidate sentence and the target text is obtained.
The storage medium according to claim 16, the text matching model composed of the convolutional neural network CNN and the GRU neural network is trained according to the following manner:

Acquiring a target training sentence, a first training sentence semantically similar to the target training sentence, and a second training sentence not semantically similar to the target training sentence;

Use the convolutional neural network CNN to perform convolution processing on the target training sentence, the first training sentence, and the second training sentence, respectively, to obtain a first vector corresponding to the target training sentence, and the A second vector corresponding to the first training sentence, and a third vector corresponding to the second training sentence;

Input the first vector, the second vector and the third vector into the GRU neural network respectively to obtain a fourth vector corresponding to the first vector and a fifth vector corresponding to the second vector And a sixth vector corresponding to the third vector;

Obtain the first score of the target training sentence and the first training sentence based on the cosine similarity between the fourth vector and the fifth vector, and the cosine similarity between the fourth vector and the sixth vector, And the second score of the target training sentence and the second training sentence;

The associated parameters in the text matching model are determined according to the first score and the second score.
The storage medium according to claim 18, wherein the convolutional neural network CNN performs convolution processing on the target training sentence, the first training sentence, and the second training sentence, respectively, including:

Setting the convolution window of the convolutional neural network CNN to preset N words;

The convolutional neural network CNN is used to perform convolution processing on the target training sentence, the first training sentence, and the second training sentence, respectively.
The storage medium according to claim 18, wherein the first vector, the second vector, and the third vector are input to the GRU neural network to obtain a fourth vector corresponding to the first vector After the fifth vector corresponding to the second vector and the sixth vector corresponding to the third vector, the method further includes: passing the fourth vector, the fifth vector, and the sixth vector through the pool, respectively A layer to perform dimension change processing on the fourth vector, the fifth vector, and the sixth vector;

Obtaining the first score of the target training sentence and the first training sentence according to the cosine similarity of the fourth vector and the fifth vector, and the cosine similarity of the fourth vector and the sixth vector, respectively Value, and the second score of the target training sentence and the second training sentence, including: the cosine similarity between the fourth vector and the fifth vector after the dimensional change processing and the dimensional change processing The cosine similarity of the fourth vector and the sixth vector to obtain the first score of the target training sentence and the first training sentence, and the second score of the target training sentence and the second training sentence value.