Information processing device, update method, and update program.

Abstract

This allows updating the transformation model that converts input queries into feature vectors based on arbitrary evaluation criteria. [Solution] The information processing device comprises an evaluation unit that evaluates related information detected by a search using feature vectors obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result by inputting data to be evaluated and evaluation criteria, and a learning unit that updates the transformation model based on the evaluation result. With this information processing device, it is also possible to optimize the transformation model according to the task and user.

Description

Technical Field

[0001] The present disclosure relates to an information processing apparatus, an update method, and an update program.

Background Art

[0002] When generating an answer to an input query using a large language model, in addition to the query, by inputting related information of the query detected by search into the large language model, a technique for generating an answer based on the related information is known. Such a technique is called Retrieval-Augmented Generation (RAG). As a prior art document disclosed about Retrieval-Augmented Generation, for example, Patent Document 1 below can be cited.

[0003] In the language model system described in Patent Document 1, a text database is searched using a feature vector calculated using an Embedding model from a question sentence input by a user terminal, and a text with a feature vector similar to the calculated feature vector is obtained. Then, in the language model system, a prompt obtained by adding the text obtained as described above to the input question sentence is input to a large language model (LLM) to generate an answer to the above question sentence.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] In the language model system described in Patent Document 1, the accuracy of the feature vector generated from the question significantly affects the accuracy of the generated answer. In other words, if an appropriate feature vector is generated that aligns with the user's intent, a sentence that matches that intent is detected from the text database, and an appropriate answer is generated based on that sentence. On the other hand, if a feature vector that does not align with the user's intent is generated, a sentence unrelated to the user's intent may be detected from the text database, potentially resulting in an irrelevant answer.

[0006] Therefore, it is important to properly train the Embedding model so that highly accurate feature vectors are generated. However, the criteria for judging the quality of text detected by search can differ depending on the task performed by the language model and the user. For example, for a user who prefers factual answers, text with high factual accuracy is "appropriate" relevant information, while for a user who prefers creative answers, text containing creative content is "appropriate" relevant information.

[0007] Therefore, in search extension generation, it is desirable to update the Embedding model, that is, the transformation model that converts the input query into a feature vector, based on arbitrary evaluation criteria. However, the language model system described in Patent Document 1 cannot perform such updates. An exemplary object of this disclosure is to provide a technique that enables updating the transformation model that converts the input query into a feature vector based on arbitrary evaluation criteria. [Means for solving the problem]

[0008] An information processing device relating to an exemplary aspect of this disclosure includes an evaluation means that evaluates related information of a query detected by a search using a feature vector obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated with the evaluation criteria by inputting the data to be evaluated and the evaluation criteria, and a learning means that updates the transformation model based on the evaluation result of the evaluation means.

[0009] An example of an update method relating to this disclosure includes an evaluation process in which at least one processor evaluates related information of a query detected by a search using feature vectors obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating data to be evaluated with evaluation criteria, and inputting data to be evaluated and evaluation criteria; and an update process that updates the transformation model based on the evaluation result of the evaluation process.

[0010] An example of an update program relating to this disclosure causes a computer to function as an evaluation means that evaluates related information of a query detected by a search using feature vectors obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated using the evaluation criteria, and a learning means that updates the transformation model based on the evaluation result of the evaluation means. [Effects of the Invention]

[0011] One illustrative effect of this disclosure is that it becomes possible to update the transformation model that converts an input query into a feature vector based on arbitrary evaluation criteria. [Brief explanation of the drawing]

[0012] [Figure 1] This is a block diagram showing the configuration of the information processing device related to this disclosure. [Figure 2] This flowchart shows the flow of the update procedure related to this disclosure. [Figure 3] This is a block diagram showing the configuration of other information processing devices related to this disclosure. [Figure 4] Figure 3 shows an example of generating responses to queries and updating transformation models using the information processing device shown. [Figure 5] Figure 3 is a flowchart showing the processing flow executed by the information processing device. [Figure 6] This is a block diagram showing the configuration of a computer that functions as an information processing device related to this disclosure. [Modes for carrying out the invention]

[0013] The following are examples of embodiments of the present invention. However, the present invention is not limited to the exemplary embodiments shown below, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining some or all of the technologies (things or methods) employed in each of the exemplary embodiments shown below may also be included in the scope of the present invention. Furthermore, embodiments obtained by appropriately omitting some of the technologies employed in each of the exemplary embodiments shown below may also be included in the scope of the present invention. In addition, the effects mentioned in each of the exemplary embodiments shown below are examples of effects that can be expected in that exemplary embodiment and do not define the scope of the present invention. That is, embodiments that do not produce the effects mentioned in each of the exemplary embodiments shown below may also be included in the scope of the present invention.

[0014] [First Exemplary Embodiment] A first exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. This exemplary embodiment is the basic form for each of the exemplary embodiments described later. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems occur. Furthermore, each technology shown in the drawings referenced to explain this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems occur.

[0015] (Configuration of Information Processing Device 1) The configuration of the information processing device 1 according to this exemplary embodiment will be described with reference to Figure 1. Figure 1 is a block diagram showing the configuration of the information processing device 1. As shown in Figure 1, the information processing device 1 includes an evaluation unit 101 and a learning unit 102.

[0016] The evaluation unit 101 evaluates the information related to the query. Here, "query" means an inquiry or request. This query may be entered by a person (for example, a user of the information processing device 1) or generated by the information processing device 1 or other devices. Typically, the above query is text data that expresses the content of the inquiry or request in natural language. However, the data format of the query is arbitrary. For example, a query that mixes multiple data formats, such as a combination of image data (which may be moving image data or still image data) and text data, may be applied.

[0017] Furthermore, the "related information" mentioned above is information related to the above query. More specifically, the related information is information detected by searching using feature vectors obtained by transforming the above query using a transformation model. Therefore, the above query and the related information can be said to be related in that the feature vectors that represent their characteristics are similar. The data format of the related information is also arbitrary.

[0018] The above "conversion model" is a model that converts the input data into a feature vector representing the characteristics of the data. For example, the Embedding model described in the background art can also be used as the above conversion model. The above conversion model can be generated by machine learning the relationship between the input data and the feature vector representing the characteristics of the data. The above conversion model may be selected according to the data format of the query to be converted. For example, if the query is text data described in natural language, a conversion model that can convert text data described in natural language into a feature vector may be used. Also, when the query includes both text data and image data, a conversion model that can convert both text data and image data into feature vectors may be used. Further, a conversion model that can convert text data into a feature vector and a conversion model that can convert image data into a feature vector may be used in combination.

[0019] Note that the above conversion model may be incorporated into a generation model that generates an answer to the above query. This generation model is a model that generates an answer to the above query using the feature vector generated by the above conversion model. As such a generation model, for example, a language model that has learned the order of its components (such as words) in a natural language sentence and the order of sentences in a text can be applied. Examples of such language models include, for example, BERT (Bidirectional Encoder Representations from Transformers), RoBERTa (Robustly optimized BERT approach), ELECTRA (Efficiently Learning an Encoder that Classifies Token Replacements Accurately), etc.

[0020] The evaluation unit 101 evaluates the above-mentioned related information using a predetermined evaluation model. This "evaluation model" is a model that outputs an evaluation result obtained by evaluating the data to be evaluated against the evaluation criteria by inputting the data to be evaluated and the evaluation criteria. Therefore, the evaluation unit 101 can output an evaluation result obtained by evaluating the related information against the evaluation criteria by inputting the related information and the evaluation criteria applicable to the evaluation of the related information into the above evaluation model. Note that the evaluation unit 101 may use the output of the evaluation model as the evaluation result as it is, or may generate an evaluation result using the output of the evaluation model. The evaluation result may be any indicator that indicates the goodness or badness or validity of the related information. The evaluation model will be described in more detail in Exemplary Embodiment 2.

[0021] The learning unit 102 updates the above conversion model based on the evaluation result of the evaluation unit 101. Here, updating the conversion model based on the evaluation result means updating the parameters of the conversion model (parameters to be updated by learning) so that a feature vector that makes it easier to detect related information with a good evaluation result when evaluated against the above evaluation criteria is generated. Note that the feature vector that makes it easier to detect related information with a good evaluation result when evaluated against the above evaluation criteria is a feature vector that has a high similarity to the feature vector of related information with a good evaluation result when evaluated against the above evaluation criteria. Therefore, it can also be said that the learning unit 102 updates the parameters of the conversion model so that a feature vector that has a high similarity to the feature vector of related information with a good evaluation result when evaluated against the above evaluation criteria is generated. Note that the update of the conversion model will be described in more detail in Exemplary Embodiment 2.

[0022] As described above, the information processing device 1 according to this exemplary embodiment employs a configuration that includes an evaluation unit 101 that evaluates the related information of the above query detected by a search using feature vectors obtained by transforming the input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated with the evaluation criteria when the data to be evaluated and the evaluation criteria are input, and a learning unit 102 that updates the transformation model based on the evaluation result of the evaluation unit 101.

[0023] According to the above configuration, the query's related information is evaluated using an evaluation model that outputs an evaluation result of evaluating the data against the evaluation criteria by inputting the data to be evaluated and the evaluation criteria. Since this evaluation model outputs an evaluation result of evaluating against the evaluation criteria by inputting the evaluation criteria, it is possible to evaluate the related information using any evaluation criteria by using this evaluation model. Then, according to the above configuration, the conversion model is updated based on the evaluation result of the evaluation unit 101.

[0024] Therefore, the above configuration has the effect of making it possible to update the transformation model that converts the input query into a feature vector based on arbitrary evaluation criteria. As a result, the information processing device 1 can also optimize the above transformation model according to the task and user. Furthermore, in this update, the user does not need to prepare the training data (the input data of the model with the correct answer data labeled), which is generally required when updating a trained model. Therefore, the above configuration also has the advantage of being able to update the transformation model while keeping human costs down.

[0025] (Wheat flour) The functions of the information processing device 1 described above can also be implemented by a program. The update program according to this exemplary embodiment is an update program for a transformation model, which causes a computer to function as an evaluation means that evaluates the information related to the query detected by a search using feature vectors obtained by transforming the input query with the transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated using the evaluation criteria, and a learning means that updates the transformation model based on the evaluation result of the evaluation means. This update program makes it possible to update the transformation model that transforms the input query into feature vectors based on arbitrary evaluation criteria.

[0026] (Update process) The flow of the update method according to this exemplary embodiment will be explained with reference to Figure 2. Figure 2 is a flowchart showing the flow of the update method. Note that the entity executing each step in this update method may be a processor provided in the information processing device 1, a processor provided in another device, or the entity executing each step may be a processor provided in a different device.

[0027] In S1 (evaluation process), at least one processor evaluates the relevant information of the above query, which is detected by searching using feature vectors obtained by transforming the input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated and the evaluation criteria, given as input data and evaluation criteria.

[0028] In S2 (update process), at least one processor updates the above conversion model based on the evaluation results of S1.

[0029] As described above, the update method according to this exemplary embodiment is a method for updating a transformation model, and employs a configuration that includes: an evaluation process in which at least one processor evaluates the information related to the query detected by a search using feature vectors obtained by transforming the input query with the transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated using the evaluation criteria, and an update process that updates the transformation model based on the evaluation result of the evaluation process. Therefore, according to the update method according to this embodiment, it becomes possible to update the transformation model that transforms the input query into feature vectors based on arbitrary evaluation criteria.

[0030] [Second exemplary embodiment] A second exemplary embodiment, which is an example of an embodiment of the present invention, will be described in detail with reference to the drawings. Components having the same function as those described in the above-described exemplary embodiment are denoted by the same reference numerals, and their descriptions are omitted as appropriate. The scope of application of each technology adopted in this exemplary embodiment is not limited to this exemplary embodiment. That is, each technology adopted in this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems arise. Furthermore, each technology shown in the drawings referenced to describe this exemplary embodiment can also be adopted in other exemplary embodiments included in this disclosure, to the extent that no particular technical problems arise.

[0031] (Configuration of Information Processing Device 1A) The configuration of the information processing device 1A according to this exemplary embodiment will be described with reference to Figure 3. Figure 3 is a block diagram showing the configuration of the information processing device 1A. The information processing device 1A is a device that has the function of receiving query input and outputting an answer to that query. The information processing device 1A may be a local device used by individual users, or it may be a server that provides a service of generating answers to queries for multiple users.

[0032] As shown in the figure, the information processing device 1A includes a control unit 10A that controls all parts of the information processing device 1A, and a storage unit 11A that stores various data used by the information processing device 1A. The information processing device 1A also includes a communication unit 12A for the information processing device 1A to communicate with other devices, an input unit 13A that receives input to the information processing device 1A, and an output unit 14A for the information processing device 1A to output data. The control unit 10A includes an evaluation unit 101A, a learning unit 102A, a reception unit 103A, a search unit 104A, a generation control unit 105A, a presentation control unit 106A, and a DB (Data Base) update unit 107A. The storage unit 11A stores a conversion model 111A, a generation model 112A, and an evaluation model 113A. The DB update unit 107A will be explained later in the section "Database Update Method".

[0033] Similar to the evaluation unit 101 of the exemplary embodiment 1, the evaluation unit 101A evaluates the information related to the query detected by a search using the feature vector obtained by transforming the input query with the transformation model 111A, using the evaluation model 113A.

[0034] The conversion model 111A, like the model of the same name described in Exemplary Embodiment 1, is a model that converts input data into feature vectors that represent the characteristics of that data.

[0035] Evaluation model 113A, like the model of the same name described in Exemplary Embodiment 1, is a model that takes the data to be evaluated and the evaluation criteria as input and outputs an evaluation result in which the data has been evaluated using the evaluation criteria. Evaluation model 113A can be generated by machine learning.

[0036] For example, the evaluation model 113A may be a language model trained on natural language. This provides the added benefit of being able to input evaluation criteria written in natural language directly into the evaluation model 113A and output evaluation results, in addition to the effects of the information processing device 1. More specifically, training on natural language means training on the arrangement of constituent elements (such as words) in natural language sentences, and the arrangement of sentences in texts.

[0037] Alternatively, a model such as a VLM (Vision and Language Model), which is a type of language model that accepts image data and text data as input, may be used as the evaluation model 113A. This makes it possible to directly input related information, which is image data, into the evaluation model 113A and evaluate it.

[0038] The learning unit 102A updates the conversion model 111A based on the evaluation results of the evaluation unit 101A, similar to the learning unit 102 in Exemplary Embodiment 1. Details of the update method will be explained later in the section "Method for updating the conversion model".

[0039] The reception unit 103A receives various inputs from the user of the information processing device 1A. For example, the reception unit 103A receives the input of the query mentioned above. Also, for example, the reception unit 103A receives the specification of evaluation criteria for evaluating the information related to the query. Such inputs and specifications may be made via the input unit 13A or via the communication unit 12A. In the latter case, the user will specify the query input and evaluation criteria using the terminal device they are using.

[0040] The search unit 104A searches for relevant information related to the input query. More specifically, the search unit 104A inputs the query received by the reception unit 103A into the transformation model 111A and converts it into a feature vector. Then, the search unit 104A performs a vector search using the feature vector obtained by the transformation and detects relevant information related to the query.

[0041] The generation control unit 105A inputs the query received by the reception unit 103A and the related information detected by the search unit 104A into the generation model 112A to generate a response to the query. The data format of the response is arbitrary. For example, the response may be text data, image data (still image data or moving image data), or a combination thereof. Alternatively, instead of the query received by the reception unit 103A, the generation control unit 105A may input a feature vector generated by the search unit 104A by inputting the query into the transformation model 111A into the generation model 112A.

[0042] The generative model 112A is a model that generates answers to queries. Various models can be applied as generative model 112A depending on the answer to be generated. For example, a language model such as BERT that has been trained on natural language can be applied as generative model 112A, or a model that generates images according to the input query can be applied as generative model 112A.

[0043] The transformation model 111A and the generation model 112A may be configured as a single model. In this case, the feature vector of the query generated by the transformation model 111A is used for vector searching by the search unit 104A, and together with the related information detected by the search, is used to generate an answer to the query.

[0044] The presentation control unit 106A controls the presentation of various types of information to the user. For example, the presentation control unit 106A presents the user with answers to queries. Alternatively, for example, the presentation control unit 106A may present the user with evaluation results from the evaluation unit 101A. The manner of presentation is arbitrary, as long as the presented content is recognizable to the target person. For example, the presentation control unit 106A may transmit the information to be presented to the user's terminal device via the communication unit 12A and present the information by displaying or outputting audio from the terminal device. Alternatively, for example, the presentation control unit 106A may present the information by outputting it to the output unit 14A.

[0045] (Example of processing) An example of the processing performed by the information processing device 1A will be explained based on Figure 4. Figure 4 shows an example of the generation of a response to a query and the updating of the transformation model 111A by the information processing device 1A. The query input in the example in Figure 4 is query A1 shown in the figure. Specifically, query A1 is natural language text data that says, "I would like some groundbreaking ideas to solve the shortage of medical personnel."

[0046] When the reception unit 103A receives the input of query A1, it outputs the received query A1 to the search unit 104A. The reception unit 103A may accept either text data of query A1 or audio data of query A1. In the latter case, the reception unit 103A only needs to obtain text data obtained by speech recognition of the input audio data. Speech recognition may be performed by the information processing device 1A or by another device.

[0047] Next, the search unit 104A inputs query A1 into the transformation model 111A and converts it into a feature vector A2. Then, the search unit 104A uses the feature vector A2 to perform a vector search targeting database D1 and detects related information A3. In this way, the search target may be limited to a predetermined database. This makes it possible to generate answers based on information recorded in the predetermined database.

[0048] Furthermore, database D1 stores each piece of information that is a candidate for related information, along with its corresponding feature vector. This allows the search unit 104A to detect information in database D1 that has a feature vector with a high similarity to the feature vector A2 of query A1, and to acquire this as related information. From the information stored in database D1, the search unit 104A can acquire a document related to query A1, such as "Report on efforts to alleviate labor shortages at X Hospital," as related information A3.

[0049] The method for calculating similarity is arbitrary. For example, the search unit 104A may calculate the cosine similarity between each feature vector recorded in the database D1 and feature vector A2. Then, the search unit 104A can obtain as related information A3 information that is associated with feature vectors whose cosine similarity to feature vector A2 is above a predetermined threshold, or information that is associated with a predetermined number of feature vectors with high similarity to feature vector A2.

[0050] Next, the generation control unit 105A inputs the acquired related information A3 and feature vector A2 into the generation model 112A. As a result, the generation model 112A outputs the answer A4. The generated answer A4 is presented to the user by the presentation control unit 106A. Answer A4 is the answer to query A1 and takes into account the content of related information A3.

[0051] Furthermore, the evaluation unit 101A evaluates the related information A3 using the evaluation model 113A. In the example in Figure 4, the evaluation unit 101A generates prompt A5 using the evaluation criteria to be applied to the evaluation of related information A3, the related information A3 to be evaluated, and query A1, and inputs the generated prompt A5 to the evaluation model 113A. Prompt A5 shows three evaluation criteria, 1 to 3. Prompt A5 also instructs the evaluation model 113A to output the evaluation result for each of these evaluation criteria as a numerical value from 1 to 5. As a result, the evaluation model 113A outputs an evaluation result A6 for each of the three evaluation criteria, showing the evaluation result for that evaluation criterion as a numerical value in the range of 1 to 5. Prompt A5 can be generated, for example, by inputting the evaluation criteria to be applied, the related information A3 to be evaluated, and query A1 into a pre-prepared template.

[0052] Here, if the evaluation unit 101A has the evaluation results for each of the multiple evaluation criteria output to the evaluation model 113A, it may combine these evaluation results to generate an overall evaluation result. For example, in the example in Figure 4, the average value of the evaluation results for each evaluation criterion is calculated as the overall evaluation result A7. The overall evaluation result only needs to be generated based on each evaluation result, and the generation method is arbitrary. For example, if each evaluation result is expressed numerically, the overall evaluation result may be a statistical value such as the median, mode, maximum, or minimum of those numerical values, or the sum of each evaluation result may be used as the overall evaluation result. Alternatively, for example, the user may be allowed to specify the weights for each evaluation criterion, and the weighted average value calculated using the specified weights may be used as the overall evaluation result. This allows the evaluation results of the evaluation criteria that the user wants to emphasize to be strongly reflected in the update of the conversion model 111A.

[0053] Here, the presentation control unit 106A may present the evaluation results of the evaluation unit 101 to the user. This provides the user with information that serves as a basis for judging the validity of the answer to the query, in addition to the effects performed by the information processing device 1, by providing the evaluation results of related information. The presentation control unit 106A may present both the evaluation result A6 and the overall evaluation result A7, or it may present either one of them.

[0054] Next, the learning unit 102A updates the transformation model 111A based on the overall evaluation result A7. Note that the evaluation of related information and the updating of the transformation model 111A may be performed each time a query is entered, or each time a predetermined number of queries are received, or at predetermined intervals.

[0055] (How to update the conversion model) The learning unit 102A only needs to update the transformation model 111A so that the more favorable the evaluation result of the relevant information by the evaluation unit 101A, the higher the similarity between the feature vector of that relevant information and the feature vector of the query. This makes it possible to generate feature vectors that make it easier to detect relevant information that has a favorable evaluation result according to the evaluation criteria applied by the evaluation unit 101A when generating answers to subsequent queries.

[0056] For example, the learning unit 102A may update the transformation model 111A using a loss function that represents the difference between the probability distribution of the similarity between the feature vector of the query and the feature vector of the related information, and the probability distribution of the numerical value representing the evaluation result of the related information. This makes it possible to generate feature vectors that make it easier to detect related information that has a good evaluation result according to the evaluation criteria applied by the evaluation unit 101A when generating answers to subsequent queries.

[0057] Specifically, the learning unit 102A first converts the similarity between the query's feature vector and the related information's feature vector into a probability value. This process is performed for each piece of related information for which similarity is calculated. This yields the probability distribution of the similarity. Similarly, the learning unit 102A also obtains the probability distribution of numerical values ​​representing the evaluation results of the related information.

[0058] Furthermore, when converting similarity to probability values, high similarity should be assigned a high probability, and when converting evaluation results to probability values, high evaluation results should be assigned a high probability. For example, the learning unit 102A may use a softmax function, which converts input values ​​to probability values, to convert the numerical values ​​representing similarity and evaluation results into probability values.

[0059] The learning unit 102A updates the parameters of the transformation model 111A so that the two probability distributions generated as described above become closer. Specifically, the learning unit 102A generates a loss function that shows the difference between the generated probability distributions. Then, the learning unit 102A updates the parameters of the transformation model 111A using the generated loss function. For example, the learning unit 102A may generate the Kullback-Leibler divergence, which shows the difference between the obtained probability distributions, as the loss function. Known methods can be applied as specific methods for updating using the Kullback-Leibler divergence as the loss function.

[0060] Alternatively, the learning unit 102A may update the transformation model 111A based on optimal transport theory instead of using Kullback-Leibler divergence. In this case, the learning unit 102A can use a loss function that expresses the difference between the probability distribution of the similarity between the feature vector of the query and the feature vector of the related information, and the probability distribution of the numerical value representing the evaluation result of the related information, based on optimal transport theory. This makes it possible to appropriately update the transformation model 111A by considering the geometric structure of each probability distribution.

[0061] Furthermore, the learning unit 102A may update the transformation model 111A using reinforcement learning, which uses the evaluation results from the evaluation unit 101A as a reward. In this case, the transformation model 111A is updated so that the evaluation results from the evaluation unit 101A improve. Even when reinforcement learning is applied to the update, it becomes possible to generate feature vectors that make it easier to detect relevant information with good evaluation results according to the evaluation criteria applied by the evaluation unit 101A when generating answers to subsequent queries.

[0062] (How to update the database) As described above, the search by the search unit 104A may be performed on a database (for example, database D1 shown in Figure 4) that records each piece of information that is a candidate for related information and its feature vector in association with each piece of information. In this case, the DB update unit 107A updates each feature vector recorded in the database using the feature vector obtained by transforming each piece of information recorded in the database with the transformation model 111A after the learning unit 102A has updated.

[0063] With the information processing device 1A equipped with the DB update unit 107A, in addition to the effects of the information processing device 1, it becomes possible to make it easier to detect relevant information from the database that has good evaluation results according to the evaluation criteria applied by the evaluation unit 101A.

[0064] When updating database D1, the DB update unit 107A first retrieves each piece of information recorded in database D1 and converts that information into feature vectors using the conversion model 111A. Then, the DB update unit 107A updates database D1 by overwriting each feature vector (a feature vector corresponding to the same information) recorded in database D1 with each feature vector obtained through the conversion.

[0065] (Process flow) The processing flow performed by the information processing device 1A will be explained with reference to Figure 5. Figure 5 is a flowchart showing the processing flow performed by the information processing device 1A. Figure 5 includes each step of the update method of this exemplary embodiment.

[0066] In S11, the reception unit 103A receives the query input. Next, in S12, the search unit 104A inputs the query received in S11 into the transformation model 111A and converts it into a feature vector.

[0067] In S13, the search unit 104A performs a search using the feature vector generated by the transformation in S12 to detect related information related to the query received in S11. This search may be performed on a predetermined database, such as the database D1 shown in Figure 4. In S13, the search unit 104A may detect multiple pieces of related information. For example, the search unit 104A may detect a predetermined number of pieces of information recorded in the database D1 that have a high degree of similarity between the feature vector associated with the information in question and the feature vector generated by the transformation in S12, as related information.

[0068] In S14, the generation control unit 105A inputs the feature vector generated by the transformation in S12 (or the query received in S11) and the related information detected in S13 to the generation model 112A to generate an answer to the query received in S11. Then, in S15, the presentation control unit 106A presents the answer generated in S14 to the user.

[0069] In S16, the reception unit 103A accepts the specification of evaluation criteria for the related information detected in S13. The acceptance of the specification of evaluation criteria may be carried out, for example, by presenting candidate evaluation criteria to the user and having the user select the criteria to be used from among the candidates. Alternatively, if the evaluation model 113A is a language model, the evaluation criteria may be freely entered in natural language. The acceptance of the specification of evaluation criteria for related information may be carried out at any time that is earlier than the processing in S17, which evaluates the related information. For example, the processing in S16 may be carried out before S11, or the specification of evaluation criteria may be accepted along with the query in S11. Alternatively, multiple evaluation criteria may be predetermined, and the user may be allowed to specify the weight of each evaluation criterion.

[0070] In S17 (evaluation process), the evaluation unit 101A evaluates the relevant information detected in S13 using the evaluation model 113A. More specifically, the evaluation unit 101A inputs the relevant information detected in S13 and the evaluation criteria specified in S16 into the evaluation model 113A and outputs the evaluation result of the relevant information. In this way, by accepting the specification of evaluation criteria and inputting those evaluation criteria into the evaluation model 113A, in addition to the effects performed by the information processing device 1, the effect of being able to perform evaluation according to the evaluation criteria desired by the user is obtained.

[0071] In S18, the presentation control unit 106A presents the evaluation results from S17 to the user. At this time, the reception unit 103A may accept modifications to the presented evaluation results from the user. This makes it possible to update the conversion model 111A based on evaluation results that align with the user's intentions.

[0072] In S19 (update process), the learning unit 102A updates the transformation model 111A based on the evaluation results from S17. As described above, the learning unit 102A should update the transformation model 111A so that the more favorable the evaluation result by the evaluation unit 101A for the related information, the higher the similarity between the feature vector of that related information and the feature vector of the query.

[0073] In S20, the DB update unit 107A determines whether or not to update the database targeted for the search of related information (for example, database D1 shown in Figure 4). If the result in S20 is NO, the process in Figure 5 ends. On the other hand, if the result in S20 is YES, the process proceeds to S21. The database update conditions can be predetermined. For example, the database update conditions may be that the cumulative number of updates of the conversion model 111A since the last database update has reached a predetermined number, or that a predetermined period of time has elapsed since the last database update.

[0074] In S21, the DB update unit 107A updates the database targeted for the search of related information (for example, database D1 shown in Figure 4). This completes the process shown in Figure 5. Alternatively, after S15, the process may return to S11 to accept input for a new query.

[0075] [Variation] The entities executing each process described in the exemplary embodiments above are arbitrary and not limited to the examples given. For example, a system having the same functions as the information processing devices 1 and 1A can be constructed using multiple devices that can communicate with each other. Furthermore, the entities executing each process shown in the flowchart in Figure 5 may be a single device (which can also be called a processor) or multiple devices (which can also be called processors).

[0076] [Examples of implementation using software] Some or all of the functions of the information processing devices 1,1A (hereinafter also referred to as "each of the above devices") may be implemented by hardware such as integrated circuits (IC chips) or by software.

[0077] In the latter case, each of the above devices is implemented, for example, by a computer that executes instructions for a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as Computer C) is shown in Figure 6. Figure 6 is a block diagram showing the hardware configuration of Computer C, which functions as each of the above devices.

[0078] Computer C comprises at least one processor C1 and at least one memory C2. Memory C2 stores a program (update program) P that causes computer C to operate as each of the above-mentioned devices. In computer C, processor C1 reads program P from memory C2 and executes it, thereby realizing each of the above-mentioned devices.

[0079] For processor C1, for example, a CPU (Central Processing Unit), GPU (Graphic Processing Unit), DSP (Digital Signal Processor), MPU (Micro Processing Unit), FPU (Floating Point Number Processing Unit), PPU (Physics Processing Unit), TPU (Tensor Processing Unit), quantum processor, microcontroller, or a combination thereof can be used. For memory C2, for example, flash memory, HDD (Hard Disk Drive), SSD (Solid State Drive), or a combination thereof can be used.

[0080] Computer C may also be equipped with RAM (Random Access Memory) for loading program P at runtime and for temporarily storing various data. Furthermore, computer C may be equipped with communication interfaces for sending and receiving data with other devices. Additionally, computer C may be equipped with input / output interfaces for connecting input / output devices such as keyboards, mice, displays, and printers.

[0081] Furthermore, program P can be recorded on a non-temporary, tangible recording medium M that is readable by computer C. Such a recording medium M could be, for example, tape, disk, card, semiconductor memory, or programmable logic circuitry. Computer C can acquire program P via such a recording medium M. Program P can also be transmitted via a transmission medium. Such a transmission medium could be, for example, a communication network or broadcast waves. Computer C can also acquire program P via such a transmission medium.

[0082] Furthermore, each of the above functions of each of the above devices may be implemented by a single processor in a single computer, by multiple processors in a single computer working together, or by multiple processors in each of multiple computers working together. In addition, the programs for implementing each of the above functions in each of the above devices may be stored in a single memory in a single computer, distributed and stored in multiple memories in a single computer, or distributed and stored in multiple memories in each of multiple computers.

[0083] [Additional Notes] This disclosure includes the technologies described in the following appendices. However, the present invention is not limited to the technologies described in the following appendices, and various modifications are possible within the scope of the claims.

[0084] (Note A1) An information processing device comprising: an evaluation means that evaluates related information of a query detected by a search using feature vectors obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated with the evaluation criteria by inputting the data to be evaluated and the evaluation criteria; and a learning means that updates the transformation model based on the evaluation result of the evaluation means.

[0085] (Appendix A2) The information processing device described in Appendix A1, comprising a receiving means for receiving the specification of evaluation criteria for evaluating the related information of the query, wherein the evaluation means inputs the related information of the query and the specified evaluation criteria into the evaluation model and outputs the evaluation result of the related information.

[0086] (Note A3) The information processing apparatus according to Appendix A1 or A2, comprising a presentation control means for presenting the evaluation results obtained by the evaluation means to a user.

[0087] (Note A4) The information processing apparatus described in any of Appendix A1 to A3, wherein the search is performed on a database in which each piece of information that is a candidate for the related information is recorded in association with the feature vector of each piece of information, and the database update means updates each feature vector recorded in the database by the feature vector obtained by transforming each piece of information recorded in the database with the transformation model after it has been updated by the learning means.

[0088] (Note A5) The aforementioned evaluation model is an information processing device described in any of the appendices A1 to A4, which is a language model trained on natural language.

[0089] (Note A6) The learning means updates the transformation model so that the similarity between the feature vector of the related information and the feature vector of the query increases for related information for which the evaluation result by the evaluation means is good. This is an information processing device as described in any of Appendix A1 to A5.

[0090] (Note A7) The information processing apparatus described in Appendix A6, wherein the learning means updates the transformation model using a loss function that shows the difference between the probability distribution of the similarity between the feature vector of the query and the feature vector of the related information and the probability distribution of a numerical value indicating the evaluation result of the related information.

[0091] (Note A8) The information processing device described in Appendix A6, wherein the learning means updates the conversion model by reinforcement learning, with the evaluation result by the evaluation means as a reward.

[0092] (Note B1) A method for updating a transformation model, comprising: an evaluation process in which at least one processor evaluates related information of a query detected by a search using feature vectors obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating data to be evaluated using said evaluation criteria, and an update process that updates the transformation model based on the evaluation result of the evaluation process.

[0093] (Note B2) The update method described in Appendix B1, wherein the at least one processor includes a receiving process for receiving the specification of evaluation criteria for evaluating the related information of the query, and in the evaluation process, the at least one processor inputs the related information of the query and the specified evaluation criteria into the evaluation model and outputs the evaluation result of the related information.

[0094] (Note B3) The update method according to Appendix B1 or B2, which includes a presentation control process in which at least one processor presents the evaluation results obtained from the evaluation process to the user.

[0095] (Note B4) The search is performed on a database in which each piece of information that is a candidate for the related information and the feature vector of each piece of information are recorded in association, and the update method described in any of Appendix B1 to B3 includes a DB update process in which at least one processor updates each feature vector recorded in the database by the feature vector obtained by transforming each piece of information recorded in the database with the transformation model after the learning process has been updated.

[0096] (Note B5) The aforementioned evaluation model is a language model trained on natural language, and the update method is one of those described in Appendix B1 to B4.

[0097] (Note B6) In the learning process, the update method described in any of Appendix B1 to B5, wherein the at least one processor updates the transformation model such that the similarity between the feature vector of the related information and the feature vector of the query increases for related information for which the evaluation result of the evaluation process is favorable.

[0098] (Note B7) The update method described in Appendix B6, wherein in the learning process, the at least one processor updates the transformation model using a loss function that shows the difference between the probability distribution of the similarity between the feature vector of the query and the feature vector of the related information and the probability distribution of a numerical value indicating the evaluation result of the related information.

[0099] (Note B8) The update method described in Appendix B6, wherein in the learning process, at least one processor updates the transformation model by reinforcement learning using the evaluation result from the evaluation process as a reward.

[0100] (Note C1) A programmable program for a transformation model that enables a computer to function as an evaluation means for evaluating related information of a query detected by a search using feature vectors obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated using the evaluation criteria, and a learning means for updating the transformation model based on the evaluation result of the evaluation means.

[0101] (Note C2) The update program described in Appendix C1, wherein the computer functions as a receiving means for receiving the specification of evaluation criteria for evaluating the related information of the query, and the evaluation means inputs the related information of the query and the specified evaluation criteria into the evaluation model and outputs the evaluation result of the related information.

[0102] (Note C3) The update program described in Appendix C1 or C2, which causes the computer to function as a presentation control means for presenting evaluation results obtained by the evaluation means to a user.

[0103] (Note C4) The aforementioned search is performed on a database in which each piece of information that is a candidate for the related information is associated with the feature vector of each piece of information, and the computer is made to function as a DB update means that updates each feature vector recorded in the database by the feature vector obtained by transforming each piece of information recorded in the database with the transformation model after it has been updated by the learning means, as described in any of the appendices C1 to C3.

[0104] (Note C5) The aforementioned evaluation model is a language model trained on natural language, and is an update program described in any of the appendices C1 to C4.

[0105] (Appendix C6) The learning means updates the transformation model so that the similarity between the feature vector of the related information and the feature vector of the query increases for related information for which the evaluation result by the evaluation means is good, as described in any of the appendices C1 to C5.

[0106] (Note C7) The learning means updates the transformation model using a loss function that shows the difference between the probability distribution of the similarity between the feature vector of the query and the feature vector of the related information and the probability distribution of the numerical value representing the evaluation result of the related information, as described in Appendix C6.

[0107] (Note C8) The learning means updates the conversion model by reinforcement learning, which uses the evaluation result from the evaluation means as a reward, as described in Appendix C6.

[0108] (Note D1) An information processing device comprising at least one processor, wherein the at least one processor performs an evaluation process that evaluates related information of a query detected by a search using feature vectors obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating data to be evaluated using said evaluation criteria, and a learning process that updates the transformation model based on the evaluation result of the evaluation process.

[0109] The information processing device may also include memory. Furthermore, the memory may store a program that causes at least one processor to execute each of the aforementioned processes.

[0110] (Note D2) The information processing apparatus according to Appendix D1, wherein the at least one processor performs an acceptance process to receive the specification of evaluation criteria for evaluating the related information of the query, and in the evaluation process, the at least one processor inputs the related information of the query and the specified evaluation criteria into the evaluation model and outputs the evaluation result of the related information.

[0111] (Note D3) The information processing apparatus according to Appendix D1 or D2, wherein the at least one processor performs a presentation control process for presenting the evaluation results obtained from the evaluation process to the user.

[0112] (Note D4) The information processing apparatus according to any one of the appendices D1 to D3, wherein the search is performed on a database in which each piece of information that is a candidate for the related information is recorded in association with the feature vector of each piece of information, and at least one processor performs a DB update process to update each feature vector recorded in the database by the feature vector obtained by transforming each piece of information recorded in the database with the transformation model after the learning process has been updated.

[0113] (Note D5) The aforementioned evaluation model is an information processing device described in any of the appendices D1 to D4, which is a language model trained on natural language.

[0114] (Note D6) An information processing apparatus according to any one of the appendices D1 to D5, wherein in the learning process, the at least one processor updates the transformation model so that the similarity between the feature vector of the related information and the feature vector of the query increases for related information for which the evaluation result of the evaluation process is good.

[0115] (Note D7) The information processing apparatus according to Appendix D6, wherein in the learning process, the at least one processor updates the transformation model using a loss function that shows the difference between the probability distribution of the similarity between the feature vector of the query and the feature vector of the related information and the probability distribution of a numerical value indicating the evaluation result of the related information.

[0116] (Note D8) The information processing apparatus according to Appendix D6, wherein in the learning process, at least one processor updates the transformation model by reinforcement learning using the evaluation result from the evaluation process as a reward.

[0117] (Note E) A non-temporary recording medium that records an update program that causes a computer to execute an evaluation process, which evaluates related information of a query detected by a search using feature vectors obtained by transforming an input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated using the evaluation criteria, and a learning process that updates the transformation model based on the evaluation result of the evaluation process. [Explanation of Symbols]

[0118] 1. Information Processing Device 101 Evaluation Unit (Evaluation Means) 102 Learning Section (Learning Methods) 1A Information Processing Device 101A Evaluation Unit (Evaluation Means) 102A Learning Section (Learning Methods) 103A Reception area (reception method) 106A Presentation Control Unit (Presentation Control Means) 107A DB update unit (DB update means) 111A conversion model 113A Evaluation Model A1 Query A2 Feature vector A3 Related Information A4 answer D1 Database

Claims

1. An evaluation means that evaluates the related information of the query detected by a search using feature vectors obtained by transforming the input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated and the evaluation criteria, by inputting the data to be evaluated and the evaluation criteria, An information processing apparatus comprising a learning means for updating the conversion model based on the evaluation results of the evaluation means.

2. The system includes a means for receiving the specification of evaluation criteria for evaluating the relevant information of the aforementioned query, The information processing apparatus according to claim 1, wherein the evaluation means inputs the relevant information of the query and the specified evaluation criteria into the evaluation model and outputs the evaluation result of the relevant information.

3. The information processing apparatus according to claim 1 or 2, further comprising a presentation control means for presenting the evaluation results obtained by the evaluation means to a user.

4. The aforementioned search is performed on a database that records the correspondence between each candidate piece of information related to the aforementioned information and the feature vector of each piece of information. The information processing apparatus according to claim 1 or 2, further comprising a DB update means for updating each feature vector recorded in the database by converting each piece of information recorded in the database with the feature vector obtained by converting it using the conversion model after it has been updated by the learning means.

5. The information processing apparatus according to claim 1 or 2, wherein the evaluation model is a language model trained on natural language.

6. The information processing apparatus according to claim 1 or 2, wherein the learning means updates the transformation model so that the similarity between the feature vector of the related information and the feature vector of the query increases for related information for which the evaluation result by the evaluation means is good.

7. The information processing apparatus according to claim 6, wherein the learning means updates the transformation model using a loss function that shows the difference between the probability distribution of the similarity between the feature vector of the query and the feature vector of the related information and the probability distribution of a numerical value indicating the evaluation result of the related information.

8. The information processing apparatus according to claim 6, wherein the learning means updates the conversion model by reinforcement learning, with the evaluation result by the evaluation means as a reward.

9. At least one processor, An evaluation process that evaluates the relevant information of the query detected by a search using feature vectors obtained by transforming the input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated and the evaluation criteria, by inputting the data to be evaluated and the evaluation criteria, A method for updating a conversion model, which includes an update process that updates the conversion model based on the evaluation results of the evaluation process.

10. Computers, An evaluation means that evaluates the related information of the query detected by a search using feature vectors obtained by transforming the input query with a transformation model, using an evaluation model that outputs an evaluation result obtained by evaluating the data to be evaluated and the evaluation criteria, and A programmable program for a transformation model that functions as a learning means for updating the transformation model based on the evaluation results of the evaluation means.

Images