Information processing device, training method, information processing method, and program

Abstract

An information processing device that trains a neural network model, said information processing device comprising a training unit that inputs, to the model, data in which a predetermined symbol is assigned to each output in one or more input / output examples, and that updates the parameters of the model on the basis of the outputs from the model and correct outputs.

Description

Information processing device, learning method, information processing method, and program

[0001] This invention relates to in-context learning, a technique that improves the performance of an unknown task by providing a language model with multiple examples.

[0002] The rapid development of language models has enabled them to perform a wider range of tasks that can be described in natural language with higher performance than before. In this context, a method called in-context learning exists to enable language models to solve unknown tasks more effectively.

[0003] In in-context learning, by providing multiple input / output examples for the task to be solved as sets of {task instruction, input, output}, the model can generate answers to the input more accurately according to those input / output examples.

[0004] Furthermore, a method called meta-training for in-context learning (MetaICL) has been proposed to bridge the gap between the use of in-context learning during language model inference and the actual learning process (Non-Patent Literature 1). This method involves fine-tuning language models using datasets that include multiple input and output examples. This method has shown that performing MetaICL can further improve the performance of in-context learning during inference.

[0005] S. Min, M. Lewis, L. Zettlemoyer, H. Hajishirzi. MetaICL: Learning to Learn In Context. In Proc. NAACL, pp. 2791-2809, 2022.

[0006] While MetaICL, disclosed in Non-Patent Document 1, can improve the performance of in-context learning during inference, it is unclear whether the language model is able to learn to solve the task according to the input and output examples.

[0007] In particular, in classification problems such as determining whether a text is positive or negative, it has been reported that reversing the input and output labels or assigning random labels does not significantly change performance compared to using the correct labels. This suggests that the system is not learning how to solve the task from multiple input and output examples, but rather predicting the output from the task instructions and their inputs.

[0008] This invention has been made in view of the above points, and aims to provide a technique for improving the performance of a learning model by using multiple input / output examples as inputs.

[0009] According to the disclosed technology, an information processing device is provided for learning a neural network model, which inputs data in which a predetermined symbol is assigned to each output in one or more input / output examples into the model, and updates the parameters of the model based on the output from the model and the correct output.

[0010] The disclosed technology provides a technique for improving the performance of a model that learns by using multiple input / output examples as input.

[0011] This is a diagram showing an example configuration of the learning device 100. This is a flowchart explaining the operation of the learning device 100. This is a diagram showing an example configuration of the generation unit 110. This is a flowchart explaining the operation of the generation unit 110. This is a diagram showing an example configuration of the inference device 200. This is a diagram showing the experimental results. This is a diagram showing an example of the hardware configuration of the device.

[0012] Hereinafter, embodiments of the present invention (this embodiment) will be described with reference to the drawings. The embodiments described below are merely examples, and the embodiments to which the present invention is applied are not limited to the embodiments described below.

[0013] In the following text of the specification, for convenience of description, the standard font is used to represent characters such as vectors, sets, matrices, etc. It is clear from the context that these characters represent vectors, sets, matrices, etc.

[0014] The following sections will provide a more detailed explanation of the prior art and its challenges, followed by a description of the technology related to this embodiment.

[0015] (Regarding conventional technologies and their challenges) As mentioned above, there is a method called in-context learning that allows language models to solve unknown tasks more efficiently. In this method, by providing multiple input and output examples of the task to be solved as sets of {task instruction, input, output}, the model can generate answers to the input more accurately according to those input and output examples.

[0016] For example, if we perform in-context learning with three examples in a translation task, the input would be {Instruction: Translate the following English into French. \n Input: sea otter \n Output: loutre de mer \n \n Instruction: Translate the following English into French. \n Input: peppermint \n Output: menthe poivre \n \n Instruction: Translate the following English into French. \n Input: plush giraffe \n Output: girafe peluche \n \n Instruction: Translate the following English into French. \n Input: cheese \n Output:} and the output would be {fromage}.

[0017] The method described above has the advantage of not requiring model parameter updates, as it only requires providing multiple input / output examples during inference. On the other hand, since the language model does not explicitly learn to generate answers to inputs using multiple tasks, there is a gap between the use of in-context learning during language model inference and the learning process itself.

[0018] Therefore, meta-training for in-context learning (MetaICL) has been proposed to bridge this gap (Non-Patent Literature 1). This method fine-tunes language models using datasets that include multiple examples as input, and it has been shown that performing MetaICL can further improve the performance of in-context learning during inference.

[0019] However, while MetaICL disclosed in Non-Patent Document 1 can improve the performance of in-context learning during inference, it is still unclear whether the language model is able to learn to solve the task according to the input and output examples. As mentioned above, it is suggested that the model does not learn how to solve the task from multiple input and output examples, but rather predicts the output from the task instruction and its input.

[0020] The following describes a technique for improving the performance of a model that learns by using multiple input / output examples as inputs, as part of this embodiment. In this embodiment, the learning device 100 learns the model, and the inference device 200 performs inference processing (predetermined processing) using the learned model learned by the learning device 100. However, this is just an example, and the learning device 100 may also perform inference processing together with learning.

[0021] The learning device 100 may also be called an information processing device. Similarly, the inference device 200 may also be called an information processing device. Furthermore, although this embodiment focuses on language models, the technology described in this embodiment is applicable to models other than language models.

[0022] (Overview of the Embodiment) When performing in-context learning, it is necessary to teach the model to learn the task from multiple input and output examples. Therefore, this embodiment proposes a MetaICL method for focusing on multiple examples during learning.

[0023] Specifically, during the training of the language model by the learning device 100 (specifically, for example, during fine-tuning), the same predetermined symbol is added to each output in the training data, which has multiple input / output examples, thereby giving the training data explicit regularity, which is then used for fine-tuning the language model. In addition to using normal training data for fine-tuning the language model, the training data with the explicit regularity described above may also be used for fine-tuning the language model.

[0024] For example, if we designate a specific symbol as ▲ and create training data for the translation task example mentioned above, the input will be {Instruction: Translate the following English into French. \n Input: sea otter \n Output: loutre de mer▲\n\n Instruction: Translate the following English into French. \n Input: peppermint \n Output: menthe poivre▲\n\n Instruction: Translate the following English into French. \n Input: plush giraffe \n Output: girafe peluche▲\n\n Instruction: Translate the following English into French. \n Input: cheese \n Output:}, and the output will be {fromage▲}.

[0025] Here, any symbol may be used as the designated symbol. For example, the designated symbol may be a sequence of numbers, a sequence of letters, etc. Any symbol is acceptable as long as it is consistently used in both input and output. Furthermore, the symbol may be inserted anywhere and any number of times, as long as it is consistently inserted in the same position in each output that constitutes the task. Possible positions for inserting (adding) the symbol include, for example, before the output, after the output, at both ends of the output, between each word of the text that constitutes the output, or among some of the words in the text that constitutes the output.

[0026] For example, in the example above, the output French text (let's call it abcd efg for explanation purposes) is followed by "abcd efg▲", and the symbol is added to the same position in the task. The way the symbol is added could be "abcd efg▲▲", "▲abcd efg▲", "▲▲abcd efg▲", "abcd▲ efg", or any other format.

[0027] As described above, in this embodiment, the training data to which the same symbol (e.g., ▲) has been added to each output in the training data is used to train the language model. During training, in order for the language model to predict the symbol (e.g., ▲), it is necessary to focus on multiple input / output examples included in the input and observe what kind of symbol it is and where that symbol should be output. Therefore, it is considered that the output will inevitably be predicted by focusing on multiple input / output examples included in the input.

[0028] Note that the number of input / output examples in the training data with symbols attached to the output may be just one.

[0029] The configuration and operation of the apparatus in this embodiment will be described in detail below.

[0030] (Configuration and Operation Overview of Learning Device 100) Figure 1 shows an example of the configuration of the learning device 100 in this embodiment. As shown in Figure 1, the learning device 100 comprises a generation unit 110, a learning unit 120, an input unit 130, and an output unit 140.

[0031] Figure 1 also shows a symbol insertion unit 150 that inserts a predetermined symbol into each output in the training data. The symbol insertion unit 150 may be located inside the learning device 100 or outside the learning device 100. A symbol insertion unit 150 located outside the learning device 100 may be called a symbol insertion device. Alternatively, the insertion of symbols may be performed manually without a symbol insertion unit 150. The processing flow corresponding to the configuration in Figure 1 will be described later.

[0032] In this embodiment, the learning performed by the learning device 100 is assumed to be fine-tuning of a pre-trained language model, but is not limited to this. The learning performed by the learning device 100 may also be learning from scratch (learning from scratch).

[0033] When using a pre-trained language model as the training target, the structure, size, and language of the language model are not limited to any specific one and may be any. Furthermore, the language model is not limited to an openly accessible model; a language model created by the user may also be used. In addition, the training data used for training may be a dataset containing any type of task.

[0034] In this embodiment, as an example, we assume that a translation task is performed, and the source text of the translation task is referred to as the source text, and the translated text as the target text for explanation.

[0035] First, we will explain the operation of the generation unit 110. Note that the operation of the generation unit 110 is basically the same during training and inference.

[0036] The input to the generation unit 110 is a token sequence X of source text containing multiple input / output examples, and the output from the generation unit 110 is a token sequence Y of target text. Furthermore, the model parameter of the language model possessed by the generation unit 110 is θ. ｆｉｎｅ Let's assume that the initial values ​​of the model parameters are θ. ｐｒｅ You may also write it as follows.

[0037] The generation unit 110 takes a token sequence X of source text containing multiple input / output examples as input and generates a model parameter θ ｆｉｎｅ The token sequence Y of the target text is output by probability calculation according to the formula. Here, X = {x １ , ..., x ｍ}, Y = {y １ , , y ｎ It is expressed as}.

[0038] Regarding the token sequence, when using a pre-trained language model as the language model, the text is divided into a token sequence using the tokenizer provided in the model and then used as the input to the model. When creating the model from scratch, if the language to be handled is English, for example, expressions separated by spaces can be used as the token sequence. If the language is Japanese, for example, expressions automatically segmented by morphological analysis or expressions segmented by character units can be used as the token sequence. The tokenizer is assumed to be included in the generation unit 110, but is not limited to such an assumption. The divided token sequence may be input to the generation unit 110.

[0039] Next, the operation of the generation unit 110 will be described in more detail. The generation unit 110 receives "the token sequence X of the source text including a plurality of input-output examples", "the token sequences y １ ,..., y ｎ－１ " of the target text from the first to the (n - 1)-th that have already been obtained as the result of language generation, and "the model parameters θ ｆｉｎｅ ". Based on these inputs, the generation unit 110 outputs "the token yof the n-th target text ｎ ". In learning, "the model parameters θ ｆｉｎｅ " are the model parameters during learning. Note that during learning, the generation unit 110 may hold the model parameters θ ｆｉｎｅ , and the model parameters θ ｆｉｎｅ may be sequentially updated by the learning unit 1 ｆｉｎｅ <and>

[0040] In the language generation process executed by the generation unit 110 in the present embodiment, in order to perform in-context learning, when the token sequence of the source text including a plurality of input-output examples and the token sequences of the target text from the first to the (n - 1)-th obtained as the result of language generation are input to the generation unit 110, the generation unit 110 calculates the posterior probability distribution P(y ｎ |y １ ,..., y ｎ－１ , X; θ ｆｉｎｅ ) according to the probability calculation based on the model parameters, and obtains the token y of the n-th language generation result based on the maximum likelihood criterion.ｎ The determination is made according to the following formula.

[0041] The generation unit 110 performs this process recursively, thereby generating the token sequence Y = {y} of the target text. １ , , y ｎ The result is obtained. The structure of this generation unit 110 is the same as that of the pre-trained language model used. P(y ｎ | １ , , y ｎ－１ , X; θ ｆｉｎｅ The detailed formulation of this and its detailed calculation method will be described later.

[0042] (Operation of the learning device 100) The language model used by the inference device 200 is learned by the learning device 100. The operation of the learning device 100 is described below.

[0043] <Overall operation of the learning device 100> The learning process involves inputting a token sequence of source text containing multiple input / output examples into a pre-trained language model, and then adjusting the model parameters θ ｆｉｎｅ This is achieved by learning. Regarding the model to be used, if a pre-trained language model is used, it will follow its structure; if a model is created from scratch, any language model will suffice, such as the encoder-decoder type or decoder-only type of Transformer. For more details, see, for example, "A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, AN Gomez, I. Polosukhin. Attention is all you need. In Proc. NIPS, pp. 5998-6008, 2017."

[0044] The operation of the learning device 100, which has the configuration shown in Figure 1, will now be described. The learning device 100 has a set of D, which is a set of "a set of token sequences of source text containing multiple input / output examples and a set of token sequences of target text". ｎｏｒｍａｌ = {(X １ , Y １ ), ..., (X｜Ｄ＿ｎｏｒｍａｌ｜ , Y ｜Ｄ＿ｎｏｒｍａｌ｜ )} and D is a set of "a set of token sequences of source text and a set of token sequences of target text that include a predetermined symbol in each output of multiple input / output examples". ｓｙｍｂｏｌ = {(X １ , Y １ ), ..., (X ｜Ｄ＿ｓｙｍｂｏｌ｜ , Y ｜Ｄ＿ｓｙｍｂｏｌ｜ )} is entered.

[0045] Furthermore, as assumed in this embodiment, when fine-tuning a pre-trained model, the pre-trained model parameters θ ｐｒｅ This is input to the learning device 100. Here, θ is used as the initial value for the parameters of the pre-trained model. ｐｒｅ The learning device 100 uses the learned model parameters θ. ｆｉｎｅ Outputs.

[0046] The training data D160 in Figure 1 is the same as the above D ｎｏｒｍａｌ Corresponding to this, the training data D'170 is the above D ｓｙｍｂｏｌ This corresponds to the following. In Figure 1, in order to clearly show the features of the technology according to the present invention, as an example of operation, the symbol insertion unit 150 is used to D ｎｏｒｍａｌ From D ｓｙｍｂｏｌ This indicates that a signal is generated and input to the learning device 100.

[0047] The learning device 100 is a set D of "a pair of token sequences from source text containing multiple input / output examples and token sequences from target text". ｎｏｒｍａｌ and a set D of "a set of token sequences of source text and target text containing a predetermined symbol in each output of multiple input / output examples". ｓｙｍｂｏｌ From the model parameter θ ｆｉｎｅ Learn about it.

[0048] In this case, if a pre-trained language model is used, θ ｐｒｅ Initial value of the model parameter θ ｆｉｎｅ Learn about it.

[0049] Here, D ｓｙｍｂｏｌ D ｎｏｒｍａｌ (X ｔ, Y ｔ ) can be created by inserting a predetermined symbol into the part corresponding to the "output" of ) or D ｎｏｒｍａｌ It is also possible to create it using a supervised dataset that is completely unrelated to the above. Figure 1 is D ｓｙｍｂｏｌ D ｎｏｒｍａｌ This shows an example of how it can be created from.

[0050] In this case, the same symbol is inserted in the same position for the t-th sample (which may also be called a task), but D ｓｙｍｂｏｌ It is not necessary to use the same symbol throughout. Also, regarding the insertion method, D ｓｙｍｂｏｌ There is no need to maintain uniformity across the entire dataset; different insertion methods may be used to create the data for each sample.

[0051] D = {D ｎｏｒｍａｌ , D ｓｙｍｂｏｌ The language model's model parameters θ, optimized by} ｆｉｎｅ The following formula applies. Note that any format can be used for the prompt (which may also be called an instruction) used for D. For example, multiple types may be mixed and used. For optimization, a common optimization method in neural network training, such as RAdam, can be used.

[0052] The trained model parameters θ learned by the learning device 100 ｆｉｎｅ However, the trained model parameters θ in the inference device 200 ｆｉｎｅ It is used as such.

[0053] <Processing Flow of Learning Device 100> The operation of the learning device 100 (and symbol insertion unit 150) shown in Figure 1 will be explained according to the procedure in the flowchart of Figure 2. In the example above, learning is D = {D ｎｏｒｍａｌ , D ｓｙｍｂｏｌ This is done using}. However, in the example shown in Figure 1, in order to clearly show the features of the technology according to the present invention, as an example of operation, the symbol insertion unit 150 is used to D ｎｏｒｍａｌ From D ｓｙｍｂｏｌ This indicates that a value has been generated and will be used for learning by the learning device 100.

[0054] First, the training data D160 (D ｎｏｒｍａｌ The corresponding () is input to the symbol insertion unit 150. The learning data D160 is a set of "pairs of token sequences from source text containing multiple input / output examples and token sequences from target text".

[0055] In S100, the symbol insertion unit 150 inserts a predetermined symbol into each output of multiple input / output examples in each source text of the training data D160, and outputs training data D'170, which is the training data in which the predetermined symbol has been inserted. As mentioned above, the symbols do not need to be consistent throughout the entire training data D160.

[0056] In S101, learning data D'170 is input from the input unit 130. The "token sequence of source text including multiple input / output examples" in the learning data D'170 is input to the generation unit 110, and the corresponding "token sequence of target text" is input to the learning unit 120. In both the "token sequence of source text including multiple input / output examples" and the "token sequence of target text," a predetermined symbol is assigned to the "output."

[0057] In S102, the generation unit 110 generates the model parameter θ ｆｉｎｅ Using a language model, a token sequence of the target text is generated from the input token sequence of the source text, which includes multiple input and output examples, and output. The generated token sequence of the target text is input to the learning unit 120.

[0058] In S103, the learning unit 120 adjusts the model parameter θ so that the error between the correct "target text token sequence" and the generated "target text token sequence" is reduced. ｆｉｎｅ Update the data. For example, backpropagation can be used for the update.

[0059] If learning is complete (Yes in S104), proceed to S105; otherwise, return to S102. If return to S102, execute S102 and S103 using the next training data. In S105, the output unit 140 outputs the trained model parameters θ. ｆｉｎｅ The output is displayed. Note that the determination of whether training is complete can be made, for example, by checking whether the number of training iterations has reached a predetermined number, by checking whether the amount of parameter updates has fallen below a threshold, or by using any other method.

[0060] In this embodiment, the output trained model parameter θ ｆｉｎｅ It is assumed that this will be used in the inference device 200. When the learning device 100 performs inference along with learning, the learning device 100 will use the learned model parameters θ ｆｉｎｅ It is possible to perform inferences using this directly.

[0061] (Details of the generation unit 110) Next, the generation unit 110 will be described in more detail. The description here applies to both the generation unit 110 in the learning device 100 and the generation unit 110 in the inference device 200.

[0062] As described above, the generating unit 110 is P(y ｎ | １ , , y ｎ－１ , X; θ ｆｉｎｅ The calculation is performed. The detailed formulation for this calculation is shown in equation (1) above. Figure 3 shows the detailed configuration of the generation unit 110 for performing this calculation.

[0063] As shown in Figure 3, the generation unit 110 includes a token embedding vector calculation unit 111, a context vector calculation unit 112, a posterior probability calculation unit 113, and a text generation unit 114. Below, the posterior probability P(y) for the nth token of the target text is calculated. ｎ | １ , , y ｎ－１ , X; θ ｆｉｎｅ The detailed processing of each part for calculating ) is explained. Here, the model parameter is θ ｆｉｎｅ = {θ} ｈ , θ ｕ, θ ｐ shall be represented as {.

[0064] <Token Embedding Vector Calculation Unit 111> The token embedding vector calculation unit 111 receives a token sequence X of source text including a plurality of input-output examples, and token sequences y of the target text from the first to the (n - 1)-th that have already been obtained as the result of language generation １ , ···, y ｎ－１ and the model parameter θ ｈ as inputs, and outputs a token embedding vector sequence h.

[0065] The token embedding vector calculation unit 111 converts a token sequence X of source text including a plurality of input-output examples and token sequences y of the target text from the first to the (n - 1)-th that have already been obtained as the result of language generation １ , ···, y ｎ－１ and the model parameter θ ｈ into a token embedding vector sequence h by a conversion function based on the model parameter θ.

[0066] The embedding vector represents a vector obtained by embedding the information possessed by each token, and includes the semantic information possessed by the word in the text, which is necessary for predicting the n-th token.

[0067] Also, the larger the dimension number of the vector, the more information can be embedded. For example, the dimension number is determined manually as a 512-dimensional vector. At this time, any conversion function can be used as the conversion function as long as it is a function that converts a variable-length symbol sequence into a variable-length vector. For example, as the conversion function, a function that constructs a frequency vector of tokens in the text, a neural network such as a fully connected layer, etc. can be used.

[0068] <Context Vector Calculation Unit 112> The context vector calculation unit 112 receives a token embedding vector sequence h and the model parameter θ ｕ as inputs, and outputs a context vector u ｎ .

[0069] That is, the context vector calculation unit 112 uses the model parameter θ to process the token embedding vector sequence hｕ Convert it to the context vector u by a conversion function based on ｎ This context vector embeds the semantic information necessary to output the n-th token of the target text, taking into account the information of the input token sequence.

[0070] At this time, any conversion function can be used as long as it converts a variable-length vector into a single vector. If it is the simplest one, a function that constructs a vector by averaging the embedding vectors of tokens can be used.

[0071] Also, a function that introduces a self-attention mechanism into the vector sequence and expresses it as a single context vector by adding an attention mechanism can be used.

[0072] <Posterior probability calculation unit 113> The posterior probability calculation unit 113 takes the context vector u ｎ and the model parameter θ ｐ as inputs, and outputs the posterior probability P(y ｎ | y １ , ···, y ｎ－１ , X; θ ｆｉｎｅ ) for the n-th token of the target text.

[0073] That is, the posterior probability calculation unit 113 converts the context vector u ｎ into the posterior probability P(y ｐ | y ｎ , ···, y １ , X; θ ｎ－１ ) for the n-th token of the target text by a conversion function based on the model parameter θ ｆｉｎｅ .

[0074] The posterior probability can be represented as a vector with each token as an element, and it is possible to represent the posterior probability distribution by a vector conversion function. At this time, any conversion function can be used as long as it converts a vector into a posterior probability distribution. For example, it can be realized by a function that performs conversion using a softmax function. Otherwise, a function that can convert the sum of the elements of the output vector corresponding to the posterior probability distribution to 1.0 is applicable.

[0075] <Text generation unit 114> The text generation unit 114 generates the posterior probability of the first to nth tokens of the target text, "P(y1│X;θ) ｆｉｎｅ ), ..., P(y ｎ | １ , , y ｎ－１ , X; θ ｆｉｎｅ The input is ) and the target text token sequence Y is output.

[0076] In other words, the text generation unit 114 calculates a posterior probability P(y1│X;θ) for the first to nth tokens in the target text. ｆｉｎｅ ), ..., P(y ｎ | １ , , y ｎ－１ , X; θ ｆｉｎｅ P(y) such that the product of ) is maximized ｎ | １ , , y ｎ－１ , X; θ ｆｉｎｅ By selecting one element from the above, the token sequence Y of the target text is output.

[0077] At this time, P(y ｎ | １ , , y ｎ－１ , X; θ ｆｉｎｅ The selection of elements in ) can be any method that maximizes the probability, but this can be achieved by using approximation algorithms such as greedy algorithms.

[0078] <Processing Flow of Generation Unit 110> The above process is repeated recursively until all words of the target text are generated. The flow of the above process in the generation unit 110 will be explained with reference to Figure 4. The initial value of n is 1. In S201, a token sequence X of the source text, which includes multiple input / output examples, is input to the generation unit 110. If n > 1, proceed to S206 (if YES in S202), otherwise proceed to S203 (if NO in S202).

[0079] In S203, the generation unit 110 generates X to y １ This generates the following. In S204, let n = n + 1.

[0080] If n = N, the process ends (if YES is answered in S205), otherwise the process returns to S202. In S206, which proceeds if n > 1, the generation unit 110 generates X and y １ , , y ｎ－１ Generate and proceed to S204.

[0081] (Regarding the inference device 200) Figure 5 shows an example of the configuration of the inference device 200 in this embodiment. As shown in Figure 5, the inference device 200 has a generation unit 110, an input unit 210, and an output unit 220. The function and operation of the generation unit 110 are as described above.

[0082] However, the generation unit 110 of the inference device 200 uses the trained model parameters θ learned by the learning device 100. ｆｉｎｅ Inference is performed using the language model with the settings configured. Note that the generation unit 110 itself can be considered a language model. Figure 5 shows the trained model parameters θ. ｆｉｎｅ The model parameter is shown as 230.

[0083] The operation of the inference device 200 will be explained with reference to Figure 5. First, a token sequence of source text, including multiple input / output examples, is input from the input unit 210. This token sequence is input to the generation unit 110. The multiple input / output examples used for inference may or may not be assigned predetermined symbols.

[0084] The generation unit 110 generates a token sequence of the target text from a token sequence of the source text, which includes multiple input / output examples, and outputs it. The token sequence of the target text is output externally from the output unit 220.

[0085] (Regarding experimental results) An experiment was conducted to confirm the effectiveness of the technology according to this embodiment, and the experimental results are shown in Figure 6. The technology according to this embodiment is referred to as the "proposed method".

[0086] In Figure 6, "Baseline" shows the results when using a pre-trained language model. "Conventional" shows the results when using a model fine-tuned with the MetaICL method disclosed in Non-Patent Literature 1. "Proposed" shows the results when using a model fine-tuned with the proposed method. Figure 6 shows the results when the number of training tasks is 1,500 and when the number of training tasks is 3,231.

[0087] The training data used for both "Conventional" and "Proposed" training includes 2 to 4 input / output examples for each task.

[0088] "Original test data in context learning" means that no symbols are assigned to the test data. "Shot" indicates the number of input / output examples. For example, "{Instruction: Translate the following English into French. \n Input: sea otter \n Output: loutre de mer \n \n Instruction: Translate the following English into French. \n Input: peppermint \n Output: menthe poivre \n \n Instruction: Translate the following English into French. \n Input: plush giraffe \n Output: girafe peluche \n \n Instruction: Translate the following English into French. \n Input: cheese \n Output:}" has three input / output examples, so it is a "3-shot" example.

[0089] "Style-unified test data 4-shot in context learning" indicates that the test data has four input / output examples, and each output is assigned a symbol.

[0090] In "Style-unified test data 4-shot in context learning," a is a pattern in which symbols are randomly inserted before or after the output, and such patterns have been trained in the "Proposed" model. b is a pattern in which symbols are inserted before or after the output, c is a pattern in which symbols are inserted to the right of each token in the output, and d is a pattern in which symbols are inserted to the left of each token in the output. Patterns b, c, and d are all untrained.

[0091] The numerical values ​​shown in Figure 6 represent the accuracy of the model output, with higher values ​​indicating greater accuracy. As shown in Figure 6, the proposed method is superior to conventional methods in "Original test data in context learning."

[0092] Furthermore, in the "Style-unified test data 4-shot in context learning" test, the proposed method outperforms the baseline in all patterns a-d. On the other hand, the accuracy of the conventional method is significantly lower than the baseline.

[0093] (Example Hardware Configuration) Any of the devices described in this embodiment (learning device 100, inference device 200, information processing device) can be realized, for example, by having a computer execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0094] In other words, the device can be realized by using hardware resources such as the CPU and memory built into a computer to execute a program corresponding to the processing performed by the device. The program can be recorded on a computer-readable recording medium (such as portable memory), saved, and distributed. It can also be provided via a network, such as the Internet or email.

[0095] Figure 7 shows an example of the hardware configuration of the computer described above. The computer in Figure 7 has a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, etc., all of which are interconnected by bus B. The computer may also be equipped with a GPU.

[0096] The program that enables processing on the computer is provided on a recording medium 1001, such as a CD-ROM or memory card. When the recording medium 1001 containing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001; it may also be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files and data.

[0097] The memory device 1003 reads and stores a program from the auxiliary storage device 1002 when a program startup command is received. The CPU 1004 implements the functions related to the memory device 1003 according to the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network, etc. The display device 1006 displays a GUI (Graphical User Interface) etc., based on a program. The input device 1007 consists of a keyboard and mouse, buttons, or a touch panel, etc., and is used to input various operation commands. The output device 1008 outputs the calculation results.

[0098] (Effects of the Embodiment) As described above, the technology described in this embodiment makes it possible to improve the performance of a model that learns by using multiple input / output examples as inputs.

[0099] The following additional information is disclosed regarding the embodiments described above.

[0100] <Notes> (Note 1) An information processing device for learning a neural network model, comprising: a memory; and at least one processor connected to the memory, wherein the processor inputs data in which a predetermined symbol is assigned to each output in one or more input / output examples to the model, and updates the parameters of the model based on the output from the model and the correct output. (Note 2) The information processing device according to Note 1, wherein the predetermined symbol is assigned before the output, after the output, at both ends of the output, or to a word in the text constituting the output. (Note 3) The information processing device according to Note 1, wherein the processor assigns the predetermined symbol to each output in data having one or more input / output examples. (Note 4) An information processing device for performing predetermined processing using a neural network model, comprising: a memory; and at least one processor connected to the memory, wherein the processor performs the predetermined processing using the model learned using data in which a predetermined symbol is assigned to each output in one or more input / output examples as input. (Appendix 5) A learning method executed by an information processing device for learning a neural network model, comprising the steps of inputting data in which a predetermined symbol is assigned to each output in one or more input / output examples into the model, and updating the parameters of the model based on the output from the model and the correct output. (Appendix 6) A non-temporary storage medium storing a program for causing a computer to function as an information processing device described in any one of Appendix 1 to 4.

[0101] Although this embodiment has been described above, the present invention is not limited to this specific embodiment, and various modifications and changes are possible within the scope of the gist of the invention as described in the claims.

[0102] 100 Learning device 110 Generation unit 111 Token embedding vector calculation unit 112 Context vector calculation unit 113 Posterior probability calculation unit 114 Text generation unit 120 Learning unit 130 Input unit 140 Output unit 150 Symbol insertion unit 200 Inference device 210 Input unit 220 Output unit 1000 Drive device 1001 Recording medium 1002 Auxiliary storage device 1003 Memory device 1004 CPU 1005 Interface device 1006 Display device 1007 Input device 1008 Output device

Claims

1. An information processing device for learning a neural network model, comprising a learning unit that inputs data in which a predetermined symbol is assigned to each output in one or more input / output examples into the model, and updates the parameters of the model based on the output from the model and the correct output.

2. The information processing apparatus according to claim 1, wherein the predetermined symbol is attached before the output, after the output, at both ends of the output, or to a word in the text constituting the output.

3. The information processing apparatus according to claim 1, further comprising a symbol insertion unit for assigning the predetermined symbol to each output in data having one or more input / output examples.

4. An information processing device that performs predetermined processing using a neural network model, comprising a generation unit that performs the predetermined processing using the model learned using data in which predetermined symbols are assigned to each output in one or more input / output examples as input.

5. A learning method executed by an information processing device for learning a neural network model, comprising the steps of inputting data into the model in which a predetermined symbol is assigned to each output in one or more input / output examples, and updating the parameters of the model based on the output from the model and the correct output.

6. An information processing method performed by an information processing device that performs predetermined processing using a neural network model, comprising the step of performing the predetermined processing using the model that has been trained using data in which predetermined symbols are assigned to each output in one or more input / output examples as input.

7. A program for causing a computer to function as a component of the information processing apparatus described in any one of claims 1 to 4.

Images