CN116501837A - Retrieval method, system, device and storage medium based on twin-tower recall - Google Patents

Abstract

This application provides a retrieval method, system, device and storage medium based on twin-tower recall, which belongs to the field of data processing technology and the field of digital medical care, and searches the medical information required by users from a large number of electronic digital medical records based on an artificial intelligence model; According to the query text in the given medical field, determine the positively relevant articles and irrelevant articles of the given query text; input the trained pseudo-query model to get the first pseudo-query text of the positively relevant articles and the second of the irrelevant articles. Pseudo-query text; construct positive samples and negative samples; input positive samples and negative samples into the double-tower recall model for iterative training; input the query to be trained into the trained double-tower recall model to obtain retrieval results. The application can construct multiple strongly correlated negative samples for model learning through pseudo-queries, which greatly improves the retrieval efficiency and accuracy of the model.

Description

Retrieval method, system, device and storage medium based on twin-tower recall

technical field

The application belongs to the technical field of data processing and the field of digital medical care, and in particular relates to a retrieval method, system, equipment and storage medium based on twin-tower recall.

Background technique

Information query has become a channel for users to quickly obtain the required information in many scenarios. For example, in the medical field, the medical record information required by users can be queried from massive electronic medical records based on artificial intelligence models, which helps to provide medical record references for users. Traditional information retrieval is mainly based on keywords, such as semantic retrieval. The current semantic retrieval mainly adopts the neural network model, including the interactive single-tower model and the representational double-tower model.

The current mainstream dense vector recall method based on the two-tower architecture model requires fixed negative samples for training or a large degree of discrimination between negative samples and positive samples, resulting in too simple training; because of the increase in information interaction during model design The model increases the delay of the model, which greatly affects the final retrieval performance of the model.

Contents of the invention

The retrieval method, system, device and storage medium based on twin-tower recall proposed by the present invention can construct a plurality of strongly correlated negative samples through pseudo-query for model learning, greatly improving the retrieval efficiency and accuracy of the model.

According to the first aspect of the embodiment of the present application, a retrieval method based on twin-tower recall is provided, including:

According to the given query text, determine positively relevant articles and irrelevant articles of the given query text;

Input the positively related articles into the trained pseudo query model to obtain the first pseudo query text of the positively related articles; input the irrelevant articles into the trained pseudo query model to obtain the second pseudo query text of the irrelevant articles;

Construct a positive sample according to the given query text, positively related articles, and the first pseudo query text; construct a negative sample according to the given query text, positively related articles, unrelated articles, the first pseudo query text, and the second pseudo query text;

Input the positive samples and negative samples into the double-tower recall model for iterative training; input the query to the trained double-tower recall model to obtain the retrieval results.

In some embodiments of the present application, according to a given query text, positively related articles and irrelevant articles of a given query text are determined, including:

Select at least one public dataset;

Based on at least one public data set, match the given query text, obtain at least one positively related article related to the given query text, and obtain at least one irrelevant article not related to the given query text.

In some embodiments of the present application, positive samples and negative samples are input into the double-tower recall model for iterative training, including:

Based on the query text and the pseudo-query text based on a given article, determine a similarity score between the query text and the pseudo-query text;

Computes the total loss of the twin-tower recall model based on the vector similarity score such that the output distribution of the twin-tower recall model is fitted.

In some embodiments of the present application, according to the query text and the pseudo-query text based on a given article, determining the similarity score between the query text and the pseudo-query text includes:

Splice the query text object with special characters to get the final query text;

Inputting the final query text into the query encoder to obtain the first feature vector representations corresponding to all the characters in the query text;

Splicing the query article object with special characters, and inputting the spliced query article into the pseudo-query model to obtain at least one pseudo-query text;

Extract a pseudo-query text and stitch it again with the spliced query article to obtain the final article text;

Inputting the final article text into an article encoder to obtain a second feature vector representation corresponding to all characters in the final article text;

According to the first eigenvector representation and the second eigenvector, the similarity score of the query text and the extracted pseudo-query text is determined through vector inner product.

In some embodiments of the present application, the pseudo-query model is trained by using multiple article texts and multiple query texts corresponding to the articles as training samples; inputting a given article into the pseudo-query model to obtain at least one pseudo-query corresponding to the given article text.

In some implementations of the present application, according to the given query text, positively related articles, and the first pseudo query text, a positive sample is constructed, including:

Combine the given query text, a positively related article, and a first pseudo-query text to obtain a positive sample.

In some embodiments of the present application, according to the given query text, positively related articles, irrelevant articles, first pseudo query text and second pseudo query text, negative samples are constructed, including:

Combine the given query text, a positively related article, and a second pseudo-query text to obtain a negative sample;

And, combine the given query text, an irrelevant article and a second pseudo-query text to obtain a negative sample;

And, a negative sample is obtained by combining the given query text, an irrelevant article, and a first dummy query text.

According to the second aspect of the embodiment of the present application, a retrieval system based on twin-tower recall is provided, specifically including:

The article sample unit is used to determine positively relevant articles and irrelevant articles of a given query text according to the given query text;

The pseudo query unit is used to input the positively related articles into the trained pseudo query model to obtain the first pseudo query text of the positively related articles; it is used to input the irrelevant articles into the trained pseudo query model to obtain the second text of the irrelevant articles. dummy query text;

The training sample unit is used to construct positive samples according to the given query text, positively related articles and first pseudo query text; according to the given query text, positively related articles, unrelated articles, first pseudo query text and second pseudo query Text, construct negative samples;

The model retrieval unit is used for inputting positive samples and negative samples into the double-tower recall model for iterative training; inputting the query to be trained into the trained double-tower recall model to obtain retrieval results.

According to the third aspect of the embodiment of the present application, a retrieval device based on twin-tower recall is provided, including:

memory: for storing executable instructions; and

Processor: used to connect with the memory to execute executable instructions to complete the retrieval method based on the twin tower recall.

According to a fourth aspect of the embodiments of the present application, a computer-readable storage medium is provided, on which a computer program is stored; the computer program is executed by a processor to implement a retrieval method based on twin-tower recall.

Using the retrieval method, system, device and storage medium based on the recall of the twin towers of the present application, according to the given query text, determine the positively related articles and irrelevant articles of the given query text; input the positively related articles into the trained pseudo-query model to obtain the first pseudo-query text of positively related articles; input unrelated articles into the trained pseudo-query model to obtain the second pseudo-query text of unrelated articles; according to the given query text, positively related articles and the first pseudo-query text, construct positive samples; construct negative samples based on given query text, positively related articles, irrelevant articles, first pseudo query text, and second pseudo query text; input positive samples and negative samples into the twin-tower recall model for iterative training ; Input the query to the trained twin-tower recall model to obtain the retrieval result. The application can construct multiple strongly correlated negative samples for model learning through pseudo-queries, which greatly improves the retrieval efficiency and accuracy of the model.

Description of drawings

The drawings described here are used to provide a further understanding of the application and constitute a part of the application. The schematic embodiments and descriptions of the application are used to explain the application and do not constitute an improper limitation to the application. In the attached picture:

Fig. 1 shows a schematic diagram of the steps of a retrieval method based on twin tower recalls according to an embodiment of the present application;

Fig. 2 shows a schematic diagram of the steps of iterative training according to the twin-tower recall model in the embodiment of the present application;

Fig. 3 shows a schematic diagram of steps for calculating a similarity score according to an embodiment of the present application;

Figure 4 shows a flowchart of the principle of calculating the similarity score according to the embodiment of the present application;

Fig. 5 shows a schematic structural diagram of a retrieval system based on twin-tower recall according to an embodiment of the present application;

FIG. 6 shows a schematic structural diagram of a retrieval device based on twin-tower recall according to an embodiment of the present application.

Detailed ways

In the process of implementing this application, the inventor found that in the medical field, the medical record information required by users can be queried from a large number of electronic medical records based on artificial intelligence models, which is helpful for providing medical record references for users. In order to meet the performance requirements, the mainstream dense vector recall method based on the twin-tower architecture generally explores from two aspects:

a) In terms of model training, the traditional BM25 method is usually used to select difficult negative samples for training. For a given query (query), the selected negative samples are always fixed and cannot be learned, which will lead to the final model training. The required negative samples are not representative, or the difference between negative samples and positive samples is too large, which is too simple, which will affect the final performance of the model.

b) In terms of model design, due to the use of the infrastructure of the twin towers, the information interaction in the later stage of the model is particularly important, and more complex modules are usually introduced to enhance the information building between the query (query) and the article (passage). On the one hand, the introduction of these modules increases the delay of the model and brings a great burden to the deployment and inference of the model. On the other hand, due to the introduction of complex modules, it cannot adapt to the existing large-scale dense vector retrieval framework, such as Faiss, Milvus, etc. have weakened the original advantages of the twin-tower architecture.

Summarize the problems encountered in the existing technical solutions: the negative samples required for training are fixed or the degree of discrimination between negative samples and positive samples is very large, which makes the training too simple; in the model design, because of the information modeling of information interaction, the increase of The delay of the model greatly affects the final retrieval performance of the model.

When obtaining training samples, such as in medical application scenarios, the sample images are medical images, and the types of objects included in the sample images are lesions, that is, the parts of the body where lesions occur. Medical imaging refers to internal tissues obtained in a non-invasive manner for medical treatment or medical research, such as images of the stomach, abdomen, heart, knees, and brain, such as CT (Computed Tomography, computerized tomography), MRI (Magnetic Resonance Imaging, magnetic resonance imaging), US (ultrasonic, ultrasound), X-ray images, electroencephalograms, and images generated by medical instruments with optical photography lights.

The data used in this application is medical data, such as personal health records, prescriptions, inspection reports and other data.

Based on this, this application proposes a pseudo-query-based multi-view double-tower dense vector recall method, in which: a) through pseudo-query to dynamically construct multiple strongly correlated negative samples for model learning; b) for a given An article of , generates multiple different perspectives in advance to describe it, and no longer depends on the complex modules introduced by the information interaction in the later stage of the model.

Specifically, the retrieval method, system, device, and storage medium based on the recall of the twin towers of the present application determine the positively related articles and irrelevant articles of the given query text according to the given query text; input the positively related articles into the trained Pseudo-query model, get the first pseudo-query text of positively related articles; input unrelated articles into the trained pseudo-query model, get the second pseudo-query text of unrelated articles; according to the given query text, positively related articles and the first Pseudo-query text to construct positive samples; according to given query text, positively related articles, irrelevant articles, first pseudo-query text and second pseudo-query text, negative samples are constructed; positive samples and negative samples are input into the twin-tower recall model for Iterative training; input the query to the trained twin-tower recall model to obtain the retrieval result. The application can construct multiple strongly correlated negative samples for model learning through pseudo-queries, which greatly improves the retrieval efficiency and accuracy of the model.

The application provides multiple descriptions from different perspectives for the article presentation by sampling different pseudo-queries, so that the core important information in the article presentation can be enhanced and highlighted.

At the same time, by migrating the complex information interaction in the later stage of the model to the early stage of the model, the pseudo-query of all articles can be pre-calculated offline, which greatly reduces the reasoning delay of the model while ensuring performance, and retains the original advantages of the two-tower architecture.

In this application, due to the randomness introduced by the pseudo-query, the composition of negative samples is no longer static, so that the model can continuously learn new information. In addition, in terms of negative sample design, the innovative introduction of the negative sample that is "closest" to the given query greatly strengthens the model's ability to distinguish between positive and negative samples.

In order to make the technical solutions and advantages in the embodiments of the present application clearer, the exemplary embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings. Apparently, the described embodiments are only part of the embodiments of the present application, and Not an exhaustive list of all embodiments. It should be noted that, in the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

Example 1

FIG. 1 shows a schematic diagram of steps of a retrieval method based on twin-tower recall according to an embodiment of the present application. As shown in Figure 1, the retrieval method based on the twin-tower recall of the embodiment of the present application includes the following steps:

S1: According to the given query text, determine positively relevant articles and irrelevant articles of the given query text.

S2: Input positively related articles into the trained pseudo-query model to obtain the first pseudo-query text of positively related articles; input irrelevant articles into the trained pseudo-query model to obtain the second pseudo-query text of irrelevant articles.

S3: Construct a positive sample according to the given query text, positively related articles and the first pseudo query text; construct a negative sample according to the given query text, positively related articles, irrelevant articles, the first pseudo query text and the second pseudo query text sample.

S4: Input the positive samples and negative samples into the double-tower recall model for iterative training; input the query to the trained double-tower recall model to obtain the retrieval results.

The application can construct multiple strongly correlated negative samples for model learning through pseudo-queries, which greatly improves the retrieval efficiency and accuracy of the model.

During specific implementation, according to the given query text in S1, positively related articles and irrelevant articles of the given query text are determined, including:

First, select at least one public dataset; for example, use the public dataset TREC 2022 Passageranking dataset.

Then, based on at least one public data set, the given query text is matched, at least one positively related article related to the given query text is obtained, and at least one irrelevant article not related to the given query text is obtained.

In a specific digital medical field scenario, the data set is medical data, and the query text is medical text. Medical data such as personal health records, prescriptions, inspection reports and other data. Medical texts can be electronic medical records (Electronic Healthcare Records), electronic personal health records, including medical records, electrocardiograms, medical images, and a series of electronic records that are valuable for future reference.

FIG. 2 shows a schematic diagram of the steps of iterative training according to the two-tower recall model in the embodiment of the present application.

As shown in Figure 2, in step S4, positive samples and negative samples are input into the double-tower recall model for iterative training, including:

S41: According to the query text and the pseudo-query text based on the given article, determine the similarity score between the query text and the pseudo-query text; S42: Calculate the total loss of the two-tower recall model based on the vector similarity score, so that the two-tower recall model Output distribution fit.

FIG. 3 shows a schematic diagram of steps for calculating a similarity score according to an embodiment of the present application. Fig. 4 shows a flow chart of the principle of calculating the similarity score according to the embodiment of the present application.

As shown in Figures 3 and 4, specifically, in S41, according to the query text and the pseudo-query text based on a given article, determine the similarity score between the query text and the pseudo-query text, including:

S411: concatenate the query text object with special characters to obtain the final query text;

S412: Input the final query text into the query encoder to obtain the first feature vector representations corresponding to all the characters in the query text;

S413: Splice the query article object with special characters, and input the spliced query article into the pseudo-query model to obtain at least one pseudo-query text;

S414: extracting a pseudo-query text and splicing it again with the spliced query article to obtain the final article text;

S415: Input the final article text into the article encoder to obtain a second feature vector representation corresponding to all characters in the final article text;

S416: According to the first eigenvector representation and the second eigenvector, determine the similarity score of the query text and the extracted pseudo-query text through vector inner product.

Among them, the pseudo-query model of S2 is trained by using multiple article texts and multiple query texts corresponding to the articles as training samples; a given article is input into the pseudo-query model to obtain at least one pseudo-query text corresponding to the given article.

Specifically, in S3, a positive sample is constructed according to the given query text, positively related articles, and the first pseudo query text, including: combining the given query text, a positively related article, and a first pseudo query text to obtain a positive sample .

Specifically, in S3, negative samples are constructed according to the given query text, positively related articles, irrelevant articles, the first pseudo-query text and the second pseudo-query text, including: combining the given query text, a positively related article and a first pseudo-query text Two pseudo-query texts are combined to obtain a negative sample; and, a given query text, an irrelevant article and a second pseudo-query text are combined to obtain a negative sample; and, a given query text, a Irrelevant articles and a first pseudo-query text are combined to obtain a negative sample.

Referring to FIG. 4 , in order to further illustrate the retrieval solution based on the twin-tower recall of the embodiment of the present application, the following will be described through specific implementation scenarios.

Step 1) Use q to represent the query text object, q1, q2, q3, q4, q5,..., qi represent each character in the query text object q sequence, and i is the length of the query text object.

Step 2) splicing special characters [CLS] before the query text object to mark the beginning of the query text, and splicing special characters [SEP] to mark the end of the query text after the query text object;

The final query text is of the form [CLS],q1,q2,q3,q4,q5,...,ql,[SEP].

Step 3) take the spliced query text obtained in step 2) as input, through the query encoder, where the query encoder selects the current mainstream BERT model, and obtains the encoding vector representation Eq_ori corresponding to all positions;

Eq_ori=[E[CLS],Eq1,Eq2,Eq3,...,Eql,E[SEP]];

Among them, each position in Eq_ori is a 768-dimensional vector.

Step 4) Take out the first vector E[CLS] in Eq_ori obtained in step 3), standardize it through L2, make its L2 norm equal to 1, and record the obtained result as Eq, which is represented as the feature vector of the query.

Step 5) Use d to represent the article text object, d1, d2, d3, d4, d5, ..., dn represent each character in the sequence d of the article text object, and n is the length of the article text object.

Step 6) Splicing special characters [cls] to mark the beginning of the article text before the article text object, and splicing special characters [SEP] to mark the end of the article text after the article text object;

The form of the article text at this stage is A=[cls],d1,d2,d3,d4,d5,...,dn,[SEP].

Step 7) Use the result obtained in step 6) as the input of the docTTTTTquery model to obtain the pseudo query list Pseudo_query_list:

Pseudo_query_list = docTTTTTquery(A);

Pseudo_query_list=[qq1,qq2,qq3,...,qq10];

Among them, the docTTTTTquery model uses a large number of passages and corresponding queries as training samples, and the trained model can predict the possible query list corresponding to the passages by using the given passages as input. For the details of the docTTTTTquery model, please refer to the existing model, so I won't go into details here.

Here, through the top-k sampling strategy of the docTTTTTquery model itself, set k=10, and finally get 10 different pseudo-query descriptions corresponding to a given article, where each pseudo-query description qqi can represent a given input article A A summary, a perspective.

Step 8) Extract each qqi in the pseudo-query list Pseudo_query_list, and splicing with the article object in step 5), splicing special characters [cls] at the beginning and splicing special characters [SEP] at the end; the form of the final article text is: [cls], qqi, d1, d2, d3, d4, d5, ..., dn, [SEP].

Step 9) take the spliced article text obtained in step 8) as input, and obtain the encoding vector representation Ed_ori corresponding to all positions through the article encoder, where the article encoder also selects the BERT model;

Ed_ori=[E[cls], Eqqi, Ed1, Ed2, Ed3, . . . , Edn].

Step 10) Take out the first vector E[CLS] in Ed_ori obtained in step 9), standardize it by L2, make its L2 norm equal to 1, and denote the obtained result as Ed.

Step 11) Calculate the similarity between the vector representation of the query obtained in step 4) and the vector representation of the article obtained in step 10); here the similarity is measured using a simple vector inner product, and the final score is the given The query sums the similarity score for the combination of a given article and one of the pseudo-queries.

Next, the training phase of the twin-tower recall model will be described in detail.

First, use the public dataset TREC 2022 Passage ranking dataset, where for a given query q, there is a positively related article d+ and an irrelevant article d-, forming a combination (q, d+, d-). Using steps 5)-7), we can obtain the pseudo-query list qq+ for the positively related article d+, and the pseudo-query list qq- for the irrelevant article d-.

For a given q, the construction form of the positive sample is: (q,[qqi+,d+]); that is, the combination of each pseudo-query in the pseudo-query list qq+ of the positively related article d+ and the positively related article d+, the combination form is is step 8), for the convenience of subsequent formula derivation, it is abbreviated as (q,d ₊ ) here.

The construction form of the negative sample is: (q, [qqi-, d+]) or (q, [qqi+, d-]) or (q, [qqi-, d-]) three forms. That is, the combination of each pseudo-query qqi- of irrelevant articles and the combination of positively relevant articles d+, or the combination of each pseudo-query qqi+ of positively relevant articles and d- of irrelevant articles; or the combination of each pseudo-query of irrelevant articles The combination of qqi- and irrelevant articles d-; finally, the sum of all negative samples is abbreviated as . Among them, the first two negative samples actually contain some positive samples, which further increases the difficulty of learning the model.

When the model is iteratively trained, try to make the model score as high as possible for positive samples and as low as possible for negative samples, so the following loss function is designed:

Wherein, s() is the similarity score calculated in the above step 11).

Make the model predict the score of the positive sample s(q,d ₊ ) as high as possible, close to 1; for the negative sample s(q,d) the score prediction is as low as possible, close to 0.

Finally, the stochastic gradient descent algorithm (SGD) and the pytorch framework are used to model the model and update the parameters.

Finally, inference stage: For a given query q and article d, calculate the final similarity score of each pseudo-query in the pseudo-query list according to steps 1)-11), and take the highest score as the final similarity between query q and article d sex score.

Finally, specifically, the retrieval method based on the recall of the twin towers of this application determines the positively related articles and irrelevant articles of the given query text according to the given query text; input the positively related articles into the trained pseudo-query model to obtain The first pseudo-query text of positively related articles; input the unrelated articles into the trained pseudo-query model to obtain the second pseudo-query text of unrelated articles; according to the given query text, positively related articles and the first pseudo-query text, construct Positive samples; construct negative samples based on given query text, positively related articles, irrelevant articles, first pseudo query text, and second pseudo query text; input positive samples and negative samples into the double-tower recall model for iterative training; Query and input the trained twin-tower recall model to get the retrieval results. The application can construct multiple strongly correlated negative samples for model learning through pseudo-queries, which greatly improves the retrieval efficiency and accuracy of the model.

The application provides multiple descriptions from different perspectives for the article presentation by sampling different pseudo-queries, so that the core important information in the article presentation can be enhanced and highlighted.

At the same time, by migrating the complex information interaction in the later stage of the model to the early stage of the model, the pseudo-query of all articles can be pre-calculated offline, which greatly reduces the reasoning delay of the model while ensuring performance, and retains the original advantages of the two-tower architecture.

In this application, due to the randomness introduced by the pseudo-query, the composition of negative samples is no longer static, so that the model can continuously learn new information. In addition, in terms of negative sample design, the innovative introduction of the negative sample that is "closest" to the given query greatly strengthens the model's ability to distinguish between positive and negative samples.

Example 2

This embodiment provides a retrieval system based on twin-tower recall. For the undisclosed details in the retrieval system based on twin-tower recall in this embodiment, please refer to the specific implementation of the retrieval method based on twin-tower recall in other embodiments. content.

FIG. 5 shows a schematic structural diagram of a retrieval system based on twin-tower recall according to an embodiment of the present application.

As shown in FIG. 5 , the retrieval system based on twin-tower recall in the embodiment of the present application specifically includes an article sample unit 10 , a pseudo-query unit 20 , a training sample unit 30 and a model retrieval unit 40 .

specific,

The article sample unit 10 is used to determine positively related articles and irrelevant articles of the given query text according to the given query text;

Pseudo-query unit 20, for inputting positively related articles into the trained pseudo-query model to obtain the first pseudo-query text of positively related articles; for inputting irrelevant articles into the trained pseudo-query model to obtain the first pseudo-query text of irrelevant articles Two dummy query text;

The training sample unit 30 is used to construct a positive sample according to the given query text, positively related articles and first pseudo query text; according to the given query text, positively related articles, unrelated articles, first pseudo query text and second pseudo Query text and construct negative samples;

The model retrieval unit 40 is configured to input positive samples and negative samples into the double-tower recall model for iterative training; input the query to be trained into the trained double-tower recall model to obtain retrieval results.

By adopting the retrieval system based on twin-tower recall of the present application, multiple strongly correlated negative samples can be constructed through pseudo-queries for model learning, which greatly improves the retrieval efficiency and accuracy of the model.

The application provides multiple descriptions from different perspectives for the article presentation by sampling different pseudo-queries, so that the core important information in the article presentation can be enhanced and highlighted.

At the same time, by migrating the complex information interaction in the later stage of the model to the early stage of the model, the pseudo-query of all articles can be pre-calculated offline, which greatly reduces the reasoning delay of the model while ensuring performance, and retains the original advantages of the two-tower architecture.

Example 3

This embodiment provides a retrieval device based on twin-tower recall. For details not disclosed in the retrieval device based on twin-tower recall in this embodiment, please refer to the retrieval method or system based on twin-tower recall in other embodiments for details. implementation content.

FIG. 6 shows a schematic structural diagram of a retrieval device 400 based on twin-tower recall according to an embodiment of the present application.

As shown in Fig. 6, the retrieval device 400 based on the recall of the twin towers includes:

Memory 402: for storing executable instructions; and

Processor 401: used to connect with memory 402 to execute executable instructions so as to recall the retrieval method based on the twin towers.

Those skilled in the art can understand that the schematic diagram 6 is only an example of the retrieval device 400 based on the recall of the twin towers, and does not constitute a limitation to the retrieval device 400 based on the recall of the twin towers, and may include more or less components than those shown in the figure. Or combine certain components, or different components, for example, the retrieval device 400 based on the recall of the twin towers may also include input and output devices, network access devices, buses, and the like.

The processor 401 (Central Processing Unit, CPU) may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field programmable gate array ( Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor 401 can also be any conventional processor, etc. The processor 401 is the control center of the retrieval device 400 recalled based on the twin towers, and utilizes various interfaces and lines to connect the entire twin tower-based Various parts of Retrieval Device 400 are recalled.

The memory 402 can be used to store computer-readable instructions, and the processor 401 realizes the retrieval device 400 based on the twin-tower recall by running or executing the computer-readable instructions or modules stored in the memory 402, and calling the data stored in the memory 402. various functions. The memory 402 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playback function, an image playback function, etc.); The retrieval device 400 based on the twin-tower recall uses the created data and the like. In addition, the memory 402 may include a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash memory card (Flash Card), at least one magnetic disk storage device, a flash memory device, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), or other non-volatile/volatile storage devices.

If the integrated modules of the twin-tower recall-based retrieval device 400 are realized in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the present invention realizes all or part of the processes in the methods of the above embodiments, and can also use computer-readable instructions to instruct related hardware to complete, and the computer-readable instructions can be stored in a computer-readable storage medium. When the computer-readable instructions are executed by the processor, the steps of the above-mentioned various method embodiments can be realized.

Example 4

This embodiment provides a computer-readable storage medium, on which a computer program is stored; the computer program is executed by a processor to implement the retrieval method based on twin-tower recall in other embodiments.

The retrieval device and storage medium based on the recall of the twin towers in the embodiment of the present application determine the positively related articles and irrelevant articles of the given query text according to the given query text; input the positively related articles into the trained pseudo-query model to obtain The first pseudo-query text of positively related articles; input the unrelated articles into the trained pseudo-query model to obtain the second pseudo-query text of unrelated articles; according to the given query text, positively related articles and the first pseudo-query text, construct Positive samples; construct negative samples based on given query text, positively related articles, irrelevant articles, first pseudo query text, and second pseudo query text; input positive samples and negative samples into the double-tower recall model for iterative training; Query and input the trained twin-tower recall model to get the retrieval results. The application can construct multiple strongly correlated negative samples for model learning through pseudo-queries, which greatly improves the retrieval efficiency and accuracy of the model.

The application provides multiple descriptions from different perspectives for the article presentation by sampling different pseudo-queries, so that the core important information in the article presentation can be enhanced and highlighted.

Those skilled in the art should understand that the embodiments of the present application may be provided as methods, systems, or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present application is described with reference to flowcharts and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in the flow chart or blocks of the flowchart and/or the block or blocks of the block diagrams.

The terminology used in the present invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used herein and in the appended claims, the singular forms "a", "the", and "the" are intended to include the plural forms as well, unless the context clearly dictates otherwise. It should also be understood that the term "and/or" as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, the information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of the present invention, first information may also be called second information, and similarly, second information may also be called first information. Depending on the context, the word "if" as used herein may be interpreted as "at" or "when" or "in response to a determination."

While preferred embodiments of the present application have been described, additional changes and modifications to these embodiments can be made by those skilled in the art once the basic inventive concept is appreciated. Therefore, the appended claims are intended to be construed to cover the preferred embodiment and all changes and modifications which fall within the scope of the application.

Obviously, those skilled in the art can make various changes and modifications to the application without departing from the spirit and scope of the application. In this way, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims (10)

1. A retrieval method based on Twin Towers recall, characterized in that it comprises: Determine positively relevant articles and irrelevant articles of the given query text according to the given query text; Input the positively related articles into the trained pseudo query model to obtain the first pseudo query text of the positively related articles; input the unrelated articles into the trained pseudo query model to obtain the second text of the positively related articles dummy query text; Construct a positive sample according to the given query text, positively related articles and first pseudo query text; according to the given query text, positively related articles, irrelevant articles, first pseudo query text and second pseudo query text, Build a negative sample; Input the positive samples and negative samples into the double-tower recall model for iterative training; input the query into the trained double-tower recall model to obtain retrieval results. 2. the retrieval method based on twin tower recall according to claim 1, is characterized in that, described according to given query text, determines the positively relevant article of described given query text and irrelevant article, comprises: Select at least one public dataset; said dataset is medical data; Based on the at least one public data set, match the given query text, obtain at least one positively related article related to the given query text, and obtain at least one irrelevant article not related to the given query text ; The query text is a medical text. 3. the retrieval method based on twin-tower recall according to claim 1, is characterized in that, described positive sample and negative sample input twin-tower recall model carry out iterative training, comprising: Based on the query text and the pseudo-query text based on a given article, determine a similarity score between the query text and the pseudo-query text; A total loss of the two-tower recall model is calculated based on the vector similarity score such that the output distribution of the two-tower recall model is fitted. 4. the retrieval method based on twin tower recall according to claim 3, is characterized in that, described according to query text and the pseudo-query text based on given article, determine the similarity score between query text and pseudo-query text, include: Splice the query text object with special characters to get the final query text; Inputting the final query text into a query encoder to obtain a first feature vector representation corresponding to all characters in the query text; Splicing the query article object with special characters, and inputting the spliced query article into the pseudo-query model to obtain at least one pseudo-query text; Extract a pseudo-query text and stitch it again with the spliced query article to obtain the final article text; Inputting the final article text into an article encoder to obtain a second feature vector representation corresponding to all characters in the final article text; According to the first eigenvector representation and the second eigenvector, the similarity score of the query text and the extracted pseudo-query text is determined through a vector inner product. 5. the retrieval method based on twin-tower recall according to claim 1, is characterized in that, described pseudo-query model is trained as training samples by multiple query texts corresponding to multiple article texts and articles; The pseudo-query model is input to obtain at least one pseudo-query text corresponding to a given article. 6. The retrieval method based on twin-tower recall according to claim 1, wherein said constructing a positive sample according to said given query text, positively related articles and first pseudo-query text includes: Combine the given query text, a positively related article, and a first pseudo-query text to obtain a positive sample. 7. the retrieval method based on twin towers recall according to claim 1, is characterized in that, according to said given query text, positively related articles, irrelevant articles, the first pseudo-query text and the second pseudo-query text , construct negative samples, including: Combine the given query text, a positively related article, and a second pseudo-query text to obtain a negative sample; And, combine the given query text, an irrelevant article and a second pseudo-query text to obtain a negative sample; And, a negative sample is obtained by combining the given query text, an irrelevant article, and a first dummy query text. 8. A retrieval system based on Twin Towers recall, characterized in that it specifically includes: An article sample unit, configured to determine positively relevant articles and irrelevant articles of the given query text according to the given query text; A pseudo-query unit, configured to input the positively related articles into the trained pseudo-query model to obtain the first pseudo-query text of the positively related articles; to input the unrelated articles into the trained pseudo-query model to obtain the second pseudo-query text of the positively related articles; The training sample unit is used to construct a positive sample according to the given query text, positively related articles and first pseudo query text; according to the given query text, positively related articles, unrelated articles, first pseudo query text and The second pseudo-query text, constructing a negative sample; The model retrieval unit is used to input the positive samples and negative samples into the double-tower recall model for iterative training; input the query to be entered into the trained double-tower recall model to obtain the retrieval result. 9. A retrieval device based on twin-tower recall, comprising: memory: for storing executable instructions; and Processor: be used for being connected with storer to carry out executable instruction so as to complete the retrieval method based on twin-tower recall as described in any one of claims 1-7. 10. A computer-readable storage medium, characterized in that a computer program is stored thereon; the computer program is executed by a processor to implement the retrieval method based on twin-tower recall according to any one of claims 1-7.