JP2022049010A - Information search method, device, and computer readable storage medium - Google Patents

Abstract

To provide an information search method, a device, and a computer readable storage medium.SOLUTION: An information search method of the present invention includes: a step for acquiring a query command and first training data including a query result corresponding to the query command; a step for acquiring second training data by deleting noises in the first training data; a step for initializing an information search model by utilizing the second training data; and a step for searching information by utilizing the information search model. A technology proposal of the present invention enhances accuracy for information search results, which enables improving efficiency of the information search.SELECTED DRAWING: Figure 1

Description

The present invention relates to the field of information retrieval, and specifically relates to an information retrieval method, an apparatus, and a computer-readable storage medium.

Information retrieval technology is an important technology widely applied to search engines, question answering systems, recommendation systems and various other intelligent services. Better information retrieval technology allows manufacturers to accurately understand customer intent and provide appropriate products or services.

Currently, information retrieval determines the correlation between user queries and document semantics, primarily based on large neural network models. Training a large neural network model requires a large amount of labeling data, but artificial labeling is very costly. Related techniques have been presented to construct training labeling data in a generation-based manner. However, the generated data usually contains some noise, and the generated data lacks the correlation of the negative sample, which affects the effect of information retrieval.

It is an object of the present invention to provide an information retrieval method, an apparatus and a computer-readable storage medium, to improve the accuracy of information retrieval results, and to improve the efficiency of information retrieval.

According to one aspect of the embodiment of the present invention
The step of acquiring the first training data including the query command and the query result corresponding to the query command, and
The step of eliminating noise in the first training data and acquiring the second training data,
The step of initializing the information retrieval model using the second training data,
Steps to search for information using the information retrieval model,
Provides an information retrieval method including.

Further, according to at least one embodiment of the present invention, the method further comprises, after initializing the information retrieval model, the step of optimizing the information retrieval model by a hostile query.

Further, according to at least one embodiment of the present invention, the step of acquiring the first training data is
Acquiring open data including the query command and the query result corresponding to the query command, and
Using the open data, a query data generation model is generated by training, and the query data generation model generates a query command corresponding to the query result based on the input query result.
It includes inputting a document of a specific field into the query data generation model and generating the first training data.

Further, according to at least one embodiment of the present invention, the step of eliminating noise in the first training data is
Initializing the noise classification model using the first training data,
Training the noise classification model and
It includes eliminating noise in the first training data using a trained noise classification model.

Also, according to at least one embodiment of the invention, the step of training the noise classification model is
Obtained a noise classification model trained by performing N iterations, including that N is a positive integer.
Among them, in each iteration, the noise classification model is used to eliminate noise in the first training data, and the information retrieval model is trained and trained using the noise-erased data. The parameters of the noise classification model are updated using the loss function of the model.

Also, according to at least one embodiment of the invention, the step of optimizing the information retrieval model by the hostile query is.
The second training data is used to initialize the uncorrelated query generation model, the input of the uncorrelated query generation model is the query result and the first query command that correlates with the query result, and the output is uncorrelated with the query result. It is a second query command that is
This includes inputting the output result of the information retrieval model into the uncorrelated query generation model and training the information retrieval model using the output result of the uncorrelated query generation model.

Further, according to at least one embodiment of the present invention, the target function of the uncorrelated query generation model is
The correlation between the second query command generated by the uncorrelated query generation model and the query result,
The text similarity between the second query command and the first query command generated by the uncorrelated query generation model is included.

According to another aspect of the embodiment of the invention.
An acquisition unit that acquires the first training data including the query command and the query result corresponding to the query command, and the acquisition unit.
A noise elimination unit that eliminates noise in the first training data and acquires second training data,
An initialization unit that initializes the information retrieval model using the second training data,
An information retrieval unit that searches for information using the information retrieval model,
Provides an information retrieval device including.

Also included, according to at least one embodiment of the invention, is an optimization unit that optimizes the information retrieval model by hostile queries.

Further, according to at least one embodiment of the present invention, the acquisition unit is
A query command, an acquisition subunit that acquires open data containing the query results corresponding to the query command, and
The query data generation model is generated by training using the open data, and the query data generation model is a first processing subunit that generates a query command corresponding to the query result based on the input query result.
A second processing subunit, which inputs a document of a specific field into the query data generation model and generates the first training data, is included.

Further, according to at least one embodiment of the present invention, the noise elimination unit is
The first initialization subunit that initializes the noise classification model using the first training data,
A training subunit that trains the noise classification model,
It includes an erasing subunit that eliminates noise in the first training data using a trained noise classification model.

Further, according to at least one embodiment of the present invention, the optimization unit is
The second training data is used to initialize the uncorrelated query generation model, the input of the uncorrelated query generation model is the query result and the first query command that correlates with the query result, and the output is uncorrelated with the query result. The second query command, the second initialization subunit,
It includes a hostile training subunit that inputs the output result of the information retrieval model into the uncorrelated query generation model and trains the information retrieval model using the output result of the uncorrelated query generation model.

The present invention
With the processor
Contains memory for storing computer program commands,
Provided is an information retrieval device that enables the processor to realize the steps of the information retrieval method when the computer program command is executed by the processor.

The present invention
A computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the computer-readable storage realizes the step of the information retrieval method. Provide a medium.

Compared with the prior art, the information retrieval method, apparatus and computer readable storage medium provided by the embodiments of the present invention acquire the first training data for training the information retrieval model and then directly use the first training data. The performance of the information retrieval model is optimized by first eliminating the noise in the first training data and then initializing the information retrieval model using the second training data that eliminates the noise without generating the information retrieval model. It is possible to improve the accuracy of information retrieval results and improve the efficiency of information retrieval.

In order to more clearly explain the technical plan of the embodiment of the present invention, the drawings required for the description of the embodiment of the present invention will be briefly introduced below. These drawings may be used by those skilled in the art to obtain other drawings on the premise that they do not require a high level of skill.
It is a flowchart of the information retrieval method of an Example of this invention. It is a flowchart which acquires the 1st training data of an Example of this invention. It is a flowchart which acquires the 1st training data of an Example of this invention. It is a schematic diagram which trains the noise classification model of the Example of this invention. It is a flowchart which optimizes the information retrieval model of the Example of this invention. It is a schematic diagram which generates the uncorrelated query of the Example of this invention. It is a schematic diagram for hostile training of the information retrieval model and the uncorrelated query generation model of the embodiment of the present invention. It is a structural schematic diagram of the information retrieval apparatus of an Example of this invention. It is another structural schematic diagram of another information retrieval apparatus of this invention embodiment. It is a structural schematic diagram of the acquisition unit of the Example of this invention. It is a structural schematic diagram of the noise elimination unit of the Example of this invention. It is a structural schematic diagram of the optimization unit of the Example of this invention. It is another structural schematic diagram of the information retrieval apparatus of this invention embodiment.

In order to clarify the technical problems, technical proposals, and excellent points to be solved by the present invention, the drawings and specific examples will be described in detail below. In the following description, the provision of specific arrangements and specific details of the components is solely to aid in a full understanding of the embodiments of the present invention. Therefore, it is obvious to those skilled in the art that various changes and improvements can be made to the embodiments described here without departing from the scope and gist of the present invention. Also, for the sake of clarity and conciseness, the description of known functions and structures is omitted.

As used herein, "one example" or "example" means that a particular feature, structure or property associated with an example is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" that appear throughout the specification does not necessarily refer to the same embodiment. It should be noted that these specific features, structures or properties can be arbitrarily combined into one or more embodiments in any suitable manner.

In each embodiment of the present invention, the magnitude of the serial number of each of the following processes does not mean the order before and after the execution procedure, and the execution procedure of each process is determined by its function and internal logic, and the execution of the present invention is carried out. No restrictions may be configured for the example implementation process.

To solve the problem of the large amount of labeling data required for information retrieval systems and the high cost of artificially labeling data, a query data generation model was used to generate a query on a document in the target field, which was then generated. Use queries to construct a "query-result" data pair to train an information retrieval model.

However, there are two problems with this method. One is that there is noise in the generated data. Second, in the above method, the generated query is a correlated query, the uncorrelated query is composed of queries from other randomly selected documents, and the number and quality of uncorrelated queries cannot meet the requirements. High-quality uncorrelated queries can effectively enhance the effectiveness of information retrieval systems. A high-quality uncorrelated query is one that is textually similar to a correlated query but does not correlate with the content of the query result. For example, in response to the query result "iphone X is produced by Apple Inc.", the correlated query is "What is the manufacturer of iphone X?", And the corresponding high-quality uncorrelated query is "What is the color of iphone X?", Which is low quality. The uncorrelated query for is "Who is Trump?".

It is an object of the present invention to provide an information retrieval method, an apparatus and a computer-readable storage medium, to improve the accuracy of information retrieval results, and to improve the efficiency of information retrieval.

An embodiment of the present invention provides an anomalous data detection method comprising the following steps, as shown in FIG.

Step 101: Acquire the first training data including the query command and the query result corresponding to the query command.

For example, in response to the query command "Who is the manufacturer of iphone X?", The corresponding query result is "iphone X is manufactured by Apple Inc.". Among them, the first training data is a data pair of "query command-query result" of a specific target field.

As shown in FIG. 2, the step of acquiring the first training data includes the following steps.

Step 1011: Acquire open data including the query command and the query result corresponding to the query command.

Of these, open data can be datasets published to the network and can also be collected and retrieved from the network. For example, the "question-answer" data on the Q & A site can be viewed as the query command and the query result corresponding to the query command, and the "question-answer" data can be the training data used to generate the query data generation model. ..

The difference from the first training data is that the acquired open data is not necessarily the data of the specific target field, for example, if the specific target field is in the medical field, the acquired open data is other such as the mechanical field. It may be field data.

Step 1012: The query data generation model is generated by training using the open data, and the query data generation model generates a query command corresponding to the query result based on the input query result.

Among them, the open data generation model is a neural network model, which uses the acquired open data to generate a query data generation model by training, and the query data generation model generates a query command corresponding to the query result based on the input query result. can do. For example, if the input query result is "Trump is the President of the United States", the output generated is the query command "Who is Trump?".

Step 1013: A document of a specific field is input to the query data generation model to generate the first training data.

Among them, documents in a specific field can be input to the query data generation model as needed, and the specific field includes, but is not limited to, the medical field, the machine manufacturing field, and the like. By inputting a specific field document into the query data generation model, it is possible to generate a data set of "query command-query result", which utilizes the query data generation model to generate a specific field "query command-query result". It is possible to solve the problem that a large amount of data is generated, the amount of labeling data required for an information retrieval system is large, and the cost of artificially labeling the data is high.

Step 102: Eliminate the noise in the first training data and acquire the second training data.

The information retrieval model is initially initialized because the presence of noise (ie, inaccurate data) in the data set of the "query command-query result" of the specific field generated in step 101 adversely affects the accuracy of the information retrieval model. Before training, data noise must first be eliminated. In this embodiment, the noise classification model can be used to eliminate noise in the first training data, the noise classification model can be any text classification model, and whether one data is noise or not. Can be classified.

As shown in FIG. 3, in this embodiment, the step of eliminating noise in the first training data includes the following steps.

Step 1021: Initialize the noise classification model using the first training data.

Step 1022: Train the noise classification model.

In this embodiment, a noise classification model trained by performing N iterations can be obtained, where N is a positive integer.

As shown in FIG. 4, in each iteration, the noise classification model is used to eliminate noise in the first training data, and the noise-removed data is used to train an information retrieval model and trained information retrieval. Optimize the noise classification model by updating the parameters of the noise classification model using the loss function of the model.

In this embodiment, the noise classification model can predict the probability p _j that the data is noise, as shown below.

Of these, π is a function of the noise classification model, a = 1 indicates that the data is not noise, and a = 0 indicates that the data is noise.

The parameter θ of the noise classification model can be updated by the following loss function.

Of these, U _i is the data remaining after eliminating the noise in the first training data using the noise classification model when iterating at the i-th time, and U _i-1 is the noise when it is iterated at the i-1st time. F is a performance index for evaluating the information retrieval model, which is the data remaining after eliminating the noise in the first training data using the classification model.

Of these, N can be a default value, for example 50 or 100, and can also be determined based on the performance of the information retrieval model after iteration. After the Nth iteration, if the performance of the information retrieval model is limited to the performance improvement from the N-1th iteration, the iteration is terminated.

Step 1023: The trained noise classification model is used to eliminate the noise in the first training data.

The trained noise classification model can be used to eliminate noise in the first training data and then eliminate erroneous data in the first training data to improve the accuracy of the information retrieval model.

Step 103: Initialize the information retrieval model using the second training data.

In this embodiment, by initializing the information retrieval model using the second training data in which noise is eliminated, the accuracy of the information retrieval results can be improved and the efficiency of the information retrieval can be improved.

To improve performance, the information retrieval model is initialized and then the method further comprises optimizing the information retrieval model by hostile queries. As shown in FIG. 5, the step of optimizing the information retrieval model by a hostile query further comprises the following steps:

Step 1051: Initialize the uncorrelated query generation model using the second training data, the input of the uncorrelated query generation model is the query result and the first query command that correlates with the query result, and the output is the query result. This is a second query command that does not correlate with.

Of these, the second query command must be a high quality uncorrelated query, the second query command must be uncorrelated with the query results, and have text similarities to the first query command. Among them, the target function of the uncorrelated query generation model is the correlation between the second query command generated by the uncorrelated query generation model and the query result, and the second query command and the first one generated by the uncorrelated query generation model. Includes textual similarities in query commands. The smaller the objective function, the more desirable it is.

As shown in Figure 6, the input of the uncorrelated query generation model correlates with a query result such as "iphone X is produced by Apple Inc." and a query result such as "Who is the manufacturer of iphone X?" In a query command, the output is another query command that is uncorrelated with the query result, for example "What is the color of iphone X?", And that uncorrelated query command has a text similarity of "Who is the manufacturer of iphone X?" Have, i.e., a high quality uncorrelated query.

In the initialization process, the following objective function is used to make the generated query command and the query result uncorrelated, but the correlated query command is similar in text.
(Outside 1)

の部分は生成済みの無相関クエリーとクエリー結果の相関性を示し、上記初期化トレーニングした情報検索モデルによって取得することができる。
（外３）

は編集距離など、テキストの類似性を判断する指標であることができる。この二つの部分はいずれも小さいほど望ましい。λは第二部分の重要度を調整する重み係数で、必要に応じてλの値を調整することができる。 Of which
(Outside 2)

The part of shows the correlation between the generated uncorrelated query and the query result, and can be obtained by the above-mentioned initialization-trained information retrieval model.
(Outside 3)

Can be an index for judging the similarity of texts, such as the editing distance. The smaller these two parts are, the more desirable. λ is a weighting factor that adjusts the importance of the second part, and the value of λ can be adjusted as needed.

Step 1052: The output result of the information retrieval model is input to the uncorrelated query generation model, and the information retrieval model is trained by using the output result of the uncorrelated query generation model.

As shown in Fig. 7, the information retrieval model is trained using the output result of the uncorrelated query generation model as training data, and the result output by the information retrieval model is input to the uncorrelated query generation model as feedback for hostile training. Optimize the information retrieval model.

If the input of the information retrieval model is "query command-query result" and the output is probability, it indicates that the query result is the probability of the exact query result corresponding to the query command, and if "correlated query command and correspondence". Since the accuracy of the information retrieval model is limited if the information retrieval model is trained using only the "query result to be performed", in this embodiment, the "uncorrelated query command-query result" data output by the uncorrelated query generation model. Use to train an information retrieval model, hostile training an information retrieval model and an uncorrelated query generation model, and allow the two models to optimize each other's effects by iteration.

Step 104: Search for information using the information retrieval model.

In this embodiment, the information retrieval model can be used to accurately determine the correlation between the user's query command and the meaning of the document.

In this embodiment, after the first training data is acquired, the information retrieval model is not generated by directly using the first training data, but the noise in the first training data is first eliminated and the noise is eliminated. By initializing the information retrieval model using the training data, it is possible to optimize the performance of the information retrieval model, improve the accuracy of the information retrieval results, and improve the efficiency of the information retrieval.

The embodiments of the present invention also provide an information retrieval apparatus including the following units, as shown in FIG.

Acquisition unit 21: Acquires the first training data including the query command and the query result corresponding to the query command.

For example, in response to the query command "Who is the manufacturer of iphone X?", The corresponding query result is "iphone X is manufactured by Apple Inc.". Among them, the first training data is a data pair of "query command-query result" of a specific target field.

Noise elimination unit 22: The noise in the first training data is erased, and the second training data is acquired.

The information retrieval model is initialized and trained because the presence of noise (that is, inaccurate data) in the generated "query command-query result" dataset adversely affects the accuracy of the information retrieval model. Before we do, we must first eliminate the noise in the data. In this embodiment, the noise classification model can be used to eliminate noise in the first training data, the noise classification model can be any text classification model, and whether one data is noise or not. Can be classified.

Initialization unit 23: The information retrieval model is initialized by using the second training data.

In this embodiment, by initializing the information retrieval model using the second training data in which noise is eliminated, the accuracy of the information retrieval results can be improved and the efficiency of the information retrieval can be improved.

Information retrieval unit 24: Search for information using the information retrieval model.

In this embodiment, the information retrieval model can be used to accurately determine the correlation between the user's query command and the meaning of the document.

In this embodiment, after the first training data is acquired, the information retrieval model is not generated by directly using the first training data, but the noise in the first training data is first eliminated and the noise is eliminated. By initializing the information retrieval model using the training data, it is possible to optimize the performance of the information retrieval model, improve the accuracy of the information retrieval results, and improve the efficiency of the information retrieval.

In some embodiments, the apparatus further comprises an optimization unit 25 that optimizes the information retrieval model by hostile queries, as shown in FIG.

In some embodiments, the acquisition unit 21 includes the following units, as shown in FIG.

Acquisition subunit 211: Acquires open data including the query command and the query result corresponding to the query command.

First processing subunit 212: The open data is used to generate a query data generation model by training, and the query data generation model generates a query command corresponding to the query result based on the input query result.

Of these, open data can be datasets published to the network and can also be collected and retrieved from the network. For example, the "question-answer" data on the Q & A site can be viewed as the query command and the query result corresponding to the query command, and the "question-answer" data can be the training data used to generate the query data generation model. ..

The difference from the first training data is that the acquired open data is not necessarily the data of the specific target field, for example, if the specific target field is in the medical field, the acquired open data is other such as the mechanical field. It may be field data.

Among them, the open data generation model is a neural network model, which uses the acquired open data to generate a query data generation model by training, and the query data generation model generates a query command corresponding to the query result based on the input query result. can do. For example, if the input query result is "Trump is the President of the United States", the output generated is the query command "Who is Trump?".

Second processing subunit 213: A document of a specific field is input to the query data generation model to generate the first training data.

Among them, documents in a specific field can be input to the query data generation model as needed, and the specific field includes, but is not limited to, the medical field, the machine manufacturing field, and the like. By inputting a specific field document into the query data generation model, it is possible to generate a data set of "query command-query result", which utilizes the query data generation model to generate a specific field "query command-query result". It is possible to solve the problem that a large amount of data is generated, the amount of labeling data required for an information retrieval system is large, and the cost of artificially labeling the data is high.

In some embodiments, as shown in FIG. 11, the noise elimination unit 22 includes the following units.

First initialization subunit 221: Initialize the noise classification model using the first training data.

Training subunit 222: Trains the noise classification model.

In this embodiment, a noise classification model trained by performing N iterations can be obtained, where N is a positive integer.

As shown in FIG. 4, in each iteration, the noise classification model is used to eliminate noise in the first training data, and the noise-removed data is used to train an information retrieval model and trained information retrieval. Optimize the noise classification model by updating the parameters of the noise classification model using the loss function of the model.

In this embodiment, the noise classification model can predict the probability p _j that the data is noise, as shown below.

Of these, π is a function of the noise classification model, a = 1 indicates that the data is not noise, and a = 0 indicates that the data is noise.

The parameter θ of the noise classification model can be updated by the following loss function.

Of these, U _i is the data remaining after eliminating the noise in the first training data using the noise classification model when iterating at the i-th time, and U _i-1 is the noise when it is iterated at the i-1st time. F is a performance index for evaluating the information retrieval model, which is the data remaining after eliminating the noise in the first training data using the classification model.

Of these, N can be a default value, for example 50 or 100, and can also be determined based on the performance of the information retrieval model after iteration. After the Nth iteration, if the performance of the information retrieval model is limited to the performance improvement from the N-1th iteration, the iteration is terminated.

Elimination subunit 223: Eliminates noise in the first training data using the trained noise classification model.

The trained noise classification model can be used to eliminate noise in the first training data and then eliminate erroneous data in the first training data to improve the accuracy of the information retrieval model.

In some embodiments, as shown in FIG. 12, the optimization unit 25 includes the following units.

Second initialization subunit 251: Initialize the uncorrelated query generation model using the second training data, and the input of the uncorrelated query generation model is the query result and the first query command that correlates with the query result. The output is a second query command that does not correlate with the query result.

Of these, the second query command must be a high quality uncorrelated query, the second query command must be uncorrelated with the query results, and have text similarities to the first query command. Among them, the target function of the uncorrelated query generation model is the correlation between the second query command generated by the uncorrelated query generation model and the query result, and the second query command and the first one generated by the uncorrelated query generation model. Includes textual similarities in query commands. The smaller the objective function, the more desirable it is.

As shown in Figure 6, the input of the uncorrelated query generation model correlates with a query result such as "iphone X is produced by Apple Inc." and a query result such as "Who is the manufacturer of iphone X?" In a query command, the output is another query command that is uncorrelated with the query result, for example "What is the color of iphone X?", And that uncorrelated query command has a text similarity of "Who is the manufacturer of iphone X?" Have, i.e., a high quality uncorrelated query.

In the initialization process, the following objective function is used to make the generated query command and the query result uncorrelated, but the correlated query command is similar in text.
(Outside 4)

の部分は生成済みの無相関クエリーとクエリー結果の相関性を示し、上記初期化トレーニングした情報検索モデルによって取得することができる。
（外６）

は編集距離など、テキストの類似性を判断する指標であることができる。この二つの部分はいずれも小さいほど望ましい。λは第二部分の重要度を調整する重み係数で、必要に応じてλの値を調整することができる。 Of which
(Outside 5)

The part of shows the correlation between the generated uncorrelated query and the query result, and can be obtained by the above-mentioned initialization-trained information retrieval model.
(Outside 6)

Can be an index for judging the similarity of texts, such as the editing distance. The smaller these two parts are, the more desirable. λ is a weighting factor that adjusts the importance of the second part, and the value of λ can be adjusted as needed.

Hostile training subunit 252: The output result of the information retrieval model is input to the uncorrelated query generation model, and the information retrieval model is trained using the output result of the uncorrelated query generation model.

As shown in Fig. 7, the information retrieval model is trained using the output result of the uncorrelated query generation model as training data, and the result output by the information retrieval model is input to the uncorrelated query generation model as feedback for hostile training. Optimize the information retrieval model.

If the input of the information retrieval model is "query command-query result" and the output is probability, it indicates that the query result is the probability of the exact query result corresponding to the query command, and if "correlated query command and correspondence". Since the accuracy of the information retrieval model is limited if the information retrieval model is trained using only the "query result to be performed", in this embodiment, the "uncorrelated query command-query result" data output by the uncorrelated query generation model. Use to train an information retrieval model, hostile training an information retrieval model and an uncorrelated query generation model, and allow the two models to optimize each other's effects by iteration.

Examples of the present invention are also shown in FIG.
Processor 32 and
It provides a memory 34 for storing computer program commands and an information retrieval device 30 including the memory 34.

In it, when the computer program command is executed by the processor, the processor 32 is given the next step, i.e.
The step of acquiring the first training data including the query command and the query result corresponding to the query command, and
The step of eliminating noise in the first training data and acquiring the second training data,
The step of initializing the information retrieval model using the second training data,
The step of searching for information using the information retrieval model and the step of searching for information are executed.

Further, as shown in FIG. 13, the information retrieval device 30 includes a network interface 31, an input device 33, a hard disk 35 and a display device 36.

Each of the above interfaces and devices can be interconnected through a bus architecture. The bus architecture can include any number of interconnected buses and bridges. Specifically, one or more central processing units (CPUs) represented by the processor 32 and various electric circuits of one or more memories represented by the memory 34 are connected to each other. Bus architectures can, for example, connect peripherals, various other electrical circuits such as voltage regulators and power management circuits to each other. It should be understood that the bus architecture is used to achieve connection communication between these components. The bus architecture includes outside the data bus, power bus, control bus and status signal bus, which are well known in the art and are not described in detail herein.

The network interface 31 can be connected to a network (for example, the Internet, Ethernet, etc.), and can acquire related data such as public data from the network and store it in the hard disk 35.

The input device 33 can receive various commands input by the operator and transmit them to the processor 32 for execution. The input device 33 may include a keyboard or click device (eg, mouse, trackball, touch pad, touch panel, etc.).

The display device 36 can display the result obtained by executing the command by the processor 32.

The memory 34 stores data such as programs and data necessary for executing the operating system and intermediate results of the calculation process of the processor 32.

It should be understood that in the embodiments of the present invention, the memory 34 can be volatile or non-volatile memory, or can include both volatile and non-volatile memory. Among them, non-volatile memory is read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. be able to. Volatile memory can be random access memory (RM), which is used for external cache. The memory 34 of the devices and methods described herein is intended to include, but is not limited to, these and any other suitable type of memory.

In some embodiments, the memory 34 can store the following elements and execute modules or data structures, or a subset thereof, or an extension set thereof, operating systems 341 and applications 342. ..

Among them, the operating system 341 includes various system programs such as a frame layer, a core library layer, and a drive layer, and is used for realizing various basic tasks and processing hardware-based tasks. The application 342 includes various applications such as a browser and is used for realizing various application operations. A program that implements the methods of the embodiments of the present invention can be included in application 342.

When the processor 32 calls and executes an application and data stored in the memory 34, specifically, when the program or command is stored in the application 342, the query command and the query result corresponding to the query command are used. The first training data including the above is acquired, the noise in the training data is eliminated, the second training data is acquired, the information retrieval model is initialized by using the second training data, and the information retrieval model is used. Search for information.

Further, when the processor 32 calls and executes an application and data stored in the memory 34, specifically, when it is a program or a command stored in the application 342, the information retrieval model is executed by a hostile query. Optimize.

Further, the processor 32 corresponds to a query command and the query command when calling and executing an application and data stored in the memory 34, specifically, when the program or command is stored in the application 342. The open data including the query result is acquired, the query data generation model is generated by training using the open data, and the query data generation model generates a query command corresponding to the query result based on the input query result. The document of a specific field is input to the query data generation model to generate the first training data.

Further, when the processor 32 calls and executes the application and data stored in the memory 34, specifically, when the program or command is stored in the application 342, the processor 32 utilizes the first training data. The noise classification model is initialized, the noise classification model is trained, and the trained noise classification model is used to eliminate the noise in the first training data.

Further, the processor 32 is trained by performing N iterations when calling and executing the application and data stored in the memory 34, specifically, when the program or command is stored in the application 342. A noise classification model is acquired, and N is a positive integer, of which, in each iteration, the noise classification model is used to eliminate the noise in the first training data, and the noise-removed data is used. The information retrieval model is trained, and the parameters of the noise classification model are updated by using the loss function of the trained information retrieval model.

Further, when the processor 32 calls and executes the application and data stored in the memory 34, specifically, when the program or command is stored in the application 342, the second training data is used. Initialize the uncorrelated query generation model, the input of the uncorrelated query generation model is the query result and the first query command that correlates with the query result, and the output is the second query command that does not correlate with the query result. The output result of the search model is input to the uncorrelated query generation model, and the information retrieval model is trained using the output result of the uncorrelated query generation model.

The goal function of the uncorrelated query generation model is
The correlation between the second query command generated by the uncorrelated query generation model and the query result,
Includes text similarities between the second query command and the first query command generated by the uncorrelated query generation model.

The method shown in the above embodiment of the present invention can be applied to the processor 32 or can be realized by the processor 32. The processor 32 can be an integrated circuit chip and has signal processing capability. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit of the processor 32 or a software type command. The methods, steps and logic blocks disclosed in the embodiments of the present invention include general purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices. It can be realized or implemented by the processor 32 described above, which can be a discrete gate or transistor logic device, a discrete hardware component. The general purpose processor can be a microprocessor, or the processor can be any conventional processor or the like. The steps fused with the methods disclosed in the embodiments of the present invention can complete the execution directly on the hardware decode processor, or the hardware and software modules of the decode processor can be combined to complete the execution. Software modules can be located in storage media mature in the art such as random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable read-only memory, registers. The memory is located in memory 34, and the processor 32 reads the information stored in memory 34 and combines it with the hardware to complete the steps of the above method.

It should be understood that these embodiments described herein can be implemented by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays ( It can be implemented by FPGAs), general purpose processors, controllers, microcontrollers, microcontrollers, and other electronic units or combinations thereof for performing the functions described in this claim.

For software implementation, the techniques described herein can be implemented by executing modules of functionality described herein (eg, processes, functions, etc.). The software code is stored in memory and executed by the processor. Memory can be implemented on the processor or outside the processor.

The embodiments of the present invention also provide a computer-readable storage medium. If the computer-readable storage medium stores a computer program and the computer program is executed by the processor, the next step, ie, to the processor.
The step of acquiring the first training data including the query command and the query result corresponding to the query command, and
The step of eliminating noise in the first training data and acquiring the second training data,
The step of initializing the information retrieval model using the second training data,
The step of searching for information using the information retrieval model and the step of searching for information are executed.

The above is a preferred embodiment of the present invention, and a person skilled in the art can make some improvements and modifications on the premise that the principle of the present invention is not deviated, and these improvements and modifications are also regarded as the scope of protection of the present invention. Will be.

Claims (14)

The step of acquiring the first training data including the query command and the query result corresponding to the query command, and
The step of eliminating noise in the first training data and acquiring the second training data,
The step of initializing the information retrieval model using the second training data,
Steps to search for information using the information retrieval model,
An information retrieval method characterized by including. After initializing the information retrieval model
The information retrieval method according to claim 1, further comprising a step of optimizing the information retrieval model by a hostile query. The step of acquiring the first training data is
Acquiring open data including the query command and the query result corresponding to the query command, and
Using the open data, a query data generation model is generated by training, and the query data generation model generates a query command corresponding to the query result based on the input query result.
By inputting a document of a specific field into the query data generation model and generating the first training data,
The information retrieval method according to claim 1, wherein the information retrieval method comprises. The step of eliminating noise in the first training data is
Initializing the noise classification model using the first training data,
Training the noise classification model and
Using the trained noise classification model, eliminating the noise in the first training data and
The information retrieval method according to claim 1, wherein the information retrieval method comprises. The step of training the noise classification model is
Obtained a noise classification model trained by performing N iterations, including that N is a positive integer.
Among them, in each iteration, the noise classification model is used to eliminate noise in the first training data, and the information retrieval model is trained and trained using the noise-erased data. Update the parameters of the noise classification model using the loss function of the model,
The information retrieval method according to claim 4, wherein the information is retrieved. The step of optimizing the information retrieval model by the hostile query is
The second training data is used to initialize the uncorrelated query generation model, the input of the uncorrelated query generation model is the query result and the first query command that correlates with the query result, and the output does not correlate with the query result. Being a second query command
Input the output result of the information retrieval model into the uncorrelated query generation model, and train the information retrieval model using the output result of the uncorrelated query generation model.
The information retrieval method according to claim 2, wherein the information is retrieved. The goal function of the uncorrelated query generation model is
The correlation between the second query command generated by the uncorrelated query generation model and the query result,
The text similarity between the second query command and the first query command generated by the uncorrelated query generation model,
The information retrieval method according to claim 6, wherein the information retrieval method comprises. An acquisition unit that acquires the first training data including the query command and the query result corresponding to the query command, and the acquisition unit.
A noise elimination unit that eliminates noise in the first training data and acquires second training data,
An initialization unit that initializes the information retrieval model using the second training data,
An information retrieval unit that searches for information using the information retrieval model, and
An information retrieval device characterized by including. The information retrieval device according to claim 8, further comprising an optimization unit that optimizes the information retrieval model by a hostile query. The acquisition unit
A query command, an acquisition subunit that acquires open data containing the query results corresponding to the query command, and
The query data generation model is generated by training using the open data, and the query data generation model is a first processing subunit that generates a query command corresponding to the query result based on the input query result.
A second processing subunit that inputs a document of a specific field into the query data generation model and generates the first training data,
The information retrieval apparatus according to claim 8, wherein the information retrieval apparatus comprises. The noise elimination unit
The first initialization subunit that initializes the noise classification model using the first training data,
A training subunit that trains the noise classification model,
Using the trained noise classification model, an erasing subunit that eliminates noise in the first training data,
The information retrieval apparatus according to claim 8, wherein the information retrieval apparatus comprises. The optimization unit
The second training data is used to initialize the uncorrelated query generation model, the input of the uncorrelated query generation model is the query result and the first query command that correlates with the query result, and the output is uncorrelated with the query result. The second query command, the second initialization subunit,
A hostile training subunit that inputs the output result of the information retrieval model into the uncorrelated query generation model and trains the information retrieval model using the output result of the uncorrelated query generation model.
The information retrieval apparatus according to claim 9, wherein the information retrieval apparatus comprises. A program for causing a computer to execute the information retrieval method according to any one of claims 1 to 7. A computer-readable storage medium that stores the program of claim 13.

Images