Language processing device, and program

Abstract

To provide a language processing device and a program which can accurately determine whether there is a causal relationship between descriptions.SOLUTION: A language processing device 10 comprises a processor 14. The processor 40, when a sentence including a plurality of descriptions is inputted, analyzes the sentence, determines whether there is a causal relationship between a first description and a second description, and outputs a determination result. The processor 14: identifies items which each of the first and second descriptions belongs to when categorizing the sentence; gives item information indicating the item which each of the first and second descriptions belongs to, with respect to the sentence; and performs determination, by using the sentence to which the item information is given.SELECTED DRAWING: Figure 5

Description

[Technical field]

[0001] The present invention relates to a language processing device and a program. [Background technology]

[0002] In recent years, learning models (large-scale language models) capable of analyzing (learning) text by considering (learning) context, such as BERT (Bidirectional Encoder Representations from Transformers), have become widely known. In addition, Patent Documents 1 and 2 below describe configurations for extracting similar expressions by focusing on headings (items) of text when extracting similar expressions and important words. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2021-135782 A [Patent Document 2] Patent Publication No. 2022-047653 Summary of the Invention [Problem to be solved by the invention]

[0004] A text is formed from a collection of multiple sentences, and each sentence is formed from a collection of multiple tokens (morphemes and words). Furthermore, a text contains descriptions (phrases, named entities, entities) that are a collection of multiple tokens to form a single meaning. As such, a text contains multiple descriptions, and understanding the causal relationships between these multiple descriptions is important for efficiently utilizing the text. However, in the past, sufficient efforts were not made to understand such causal relationships, and there was a problem of low accuracy in understanding causal relationships (determining whether or not there is a causal relationship between descriptions).

[0005] The present invention has been made in consideration of the above background, and aims to provide a language processing device and program that are capable of accurately determining whether or not there is a causal relationship between statements contained in a sentence. [Means for solving the problem]

[0006] In order to solve the above problems, the language processing device of the present invention includes a processor, which accepts input of a sentence including a plurality of descriptions, and when the sentence is divided into categories, identifies the category to which each of a first description and a second description belongs, assigns category information to the sentence indicating the category to which each of the first and second descriptions belongs, and uses the sentence to which the category information has been assigned to determine whether or not there is a causal relationship between the first description and the second description.

[0007] The processor may separate a sentence into multiple tokens, convert the tokens that make up the sentence into features using a learning model capable of learning context, and make a judgment using the converted features.

[0008] Item information may be added by inserting a first item token as a special token, which is a type of token, at the start and end of the item to which the first description belongs, and by inserting a second item token as a special token at the start and end of the item to which the second description belongs.

[0009] The processor may insert a first description token as a special token at the start and end of the first description, and insert a second description token as a special token at the start and end of the second description, and make a judgment using a first description token feature which is a feature of the first description token and a second description token feature which is a feature of the second description token.

[0010] The processor may calculate a description token combination feature by combining the first description token feature and the second description token feature, and perform the determination using the description token combination feature.

[0011] The first description token feature and the second description token feature may be calculated as a multi-dimensional vector, and the description token combination feature may be calculated by adding the first description token feature and the second description token feature.

[0012] The processor may make a judgment using a first item token feature that is a feature of the first item token and a second item token feature that is a feature of the second item token in addition to the description token combination feature.

[0013] The processor may be configured to combine the first item token feature and the second item token feature to calculate an item token combined feature, combine the item token combined feature and the description token combined feature to calculate an integrated combined feature, and make a judgment using the integrated combined feature.

[0014] The item token combination feature and the description token combination feature may be calculated as a multi-dimensional vector, and the integrated combination feature may be calculated by adding the item token combination feature and the description token combination feature.

[0015] The first item token feature and the second item token feature may be calculated as a multi-dimensional vector, and the item token combination feature may be calculated by adding the first item token feature and the second item token feature.

[0016] The processor may create a transition model indicating the transition state of the strength of the causal relationship between multiple items including the items to which the first and second descriptions belong, and use the transition model to calculate an item token combination feature as a scalar value indicating the strength of the causal relationship between the first item token feature and the second item token feature, and calculate an integrated combination feature by multiplying the item token combination feature by the description token combination feature.

[0017] The processor may also provide an item in the document if one does not exist.

[0018] The processor It is also possible to calculate the error between the predicted probability distribution of the causal relationship between a plurality of items and the correct distribution, and to optimize the learning model for making the judgment using the calculated error.

[0019] In addition, in order to solve the above problem, the program of the present invention causes a language processing computer to execute the steps of accepting input of a sentence including multiple descriptions; identifying the items to which a first description and a second description belong when the sentence is divided into items; assigning item information to the sentence indicating the items to which the first and second descriptions belong; and determining whether or not there is a causal relationship between the first description and the second description using the sentence to which the item information has been assigned. Effect of the Invention

[0020] According to the present invention, it is possible to provide a language processing device and a program capable of accurately determining the presence or absence of a causal relationship between descriptions included in a sentence. [Brief description of the drawings]

[0021] [Figure 1] FIG. 1 is a schematic diagram illustrating a configuration of a language processing device. [Diagram 2] FIG. 11 is an explanatory diagram showing an example of an input sentence; [Diagram 3] FIG. 2 is a flowchart showing a flow of processing performed by the language processing device. [Figure 4] FIG. 2 is a block diagram showing the functions of a sentence analysis unit. [Diagram 5] FIG. 4 is an explanatory diagram showing the flow of processing performed by a text analysis unit. [Figure 6] FIG. 4 is an explanatory diagram showing the flow of processing performed by a text analysis unit. [Figure 7] FIG. 2 is an explanatory diagram of a transition model showing transition states of causal relationships between items. [Figure 8] FIG. 4 is an explanatory diagram showing the flow of processing performed by a text analysis unit. [Figure 9] FIG. 1 is a schematic diagram illustrating a configuration of a language processing device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0022] [First embodiment] 1, the language processing device 10 includes a memory 12 and a processor 14. The memory 12 stores a language analysis program for causing the processor 14 to function as the language processing device 10. When the language processing device 10 is started up, the processor 14 reads out and executes the language analysis program from the memory 12. When the language analysis program is executed, the processor 14 functions as an input sentence receiving unit 20, a sentence analysis unit 22, and a determination result output unit 24.

[0023] The input sentence receiving unit 20 receives sentences input to the language processing device 10. The sentences are input, for example, from an external terminal connected to the language processing device 10 via a network. The source of the sentence input is not limited to this. For example, the sentences may be input from an external terminal connected directly (without via a network) to the language processing device 10 wirelessly or via a wire. Also, for example, the sentences stored inside the language processing device 10 (for example, the memory 12) may be input via the input sentence receiving unit 20.

[0024] As the text, for example, a text 30 shown in FIG. 2 is input. The text 30 is formed from a set of multiple sentences, and each sentence is formed from a set of multiple tokens (morphemes and words). The text 30 includes descriptions (phrases, named entities, entities) in which multiple tokens are collected to form one meaning. The text 30 of this embodiment includes descriptions such as "decrease in sensitivity CD value," "shortage of polymer supply," and "limited use lot." The text 30 is further divided into multiple items (columns) based on the content, and each item is given an item name according to the content of the item. The text 30 of this embodiment is divided into multiple items with item names such as [Subject], [Content of abnormality], [Details of abnormality], [Investigation result of cause], and [Corrective action].

[0025] <Processing Overview> 3, the input sentence receiving unit 20 receives an input of a sentence 30, and then inputs the sentence 30 to the sentence analysis unit 22. The sentence analysis unit 22 performs an analysis process described later, and determines whether or not a causal relationship exists between a first description and a second description included in the input sentence 30. This determination result is output via the determination result output unit 24. The output determination result is displayed, for example, on a display (not shown) connected to the language processing device 10.

[0026] <Analysis process> The analysis process performed by the text analysis unit 22 will be described in detail below with reference to Figures 4 and 5. The analysis process by the text analysis unit 22 is performed using a learning model (large-scale language model) such as BERT (Bidirectional Encoder Representations from Transformers), which is constructed by learning the interpretation of text including context through machine learning using a large number of texts.

[0027] As shown in FIG. 4, the text analysis unit 22 functions as a description extraction unit 40, an information attachment unit 42, a feature conversion unit 44, a feature combination unit 46, and a causal relationship determination unit 48 when the analysis process starts.

[0028] <First and second description extraction> The description extraction unit 40 analyzes the sentence 30 and extracts a combination of the first and second descriptions. Specifically, the input sentence 30 is divided into token units and encoded using a learning model, that is, the learning model is used to learn (analyze) each token and convert it into a feature (in this embodiment, a multi-dimensional vector having learnable parameters). Next, the feature of each token is used to detect a description contained in the sentence 30, and a set (pair) of descriptions is extracted from the detected descriptions.

[0029] The text analysis unit 22 determines the causal relationship between the first and second descriptions, with one of a pair of descriptions extracted by the description extraction unit 40 as the first description and the other as the second description, through processing by an information addition unit 42, a feature conversion unit 44, a feature combination unit 46, and a causal relationship determination unit 48, which will be described later. Note that the description extraction unit 40 extracts all combinations while changing the combinations of descriptions to be extracted (combinations of the first and second descriptions), and the text analysis unit 22 determines the causal relationship between the descriptions for all combinations extracted in this manner.

[0030] <Tokenization> 5, the information providing unit 42 divides the inputted sentence 30 into token units. After that, the information providing unit 42 provides description information regarding the first and second descriptions and item information regarding the items to which the first and second descriptions belong to the sentence 30 divided into token units.

[0031] <Adding descriptive information> The description information is added by inserting [E1] as a special token (first description token) at the start of the first description and [ / E1] as a special token (first description token) at the end of the first description. The description information is added by inserting [E2] as a special token (second description token) at the start of the second description and [ / E2] as a special token (second description token) at the end of the second description.

[0032] The information addition unit 42 identifies the first description and the second description by referring to the results of the analysis (learning) of the sentence 30 performed by the description extraction unit 40 when extracting the first and second descriptions, and adds description information by inserting special tokens (first and second description tokens) at the above-mentioned positions.

[0033] FIG. 5 shows an example of adding descriptive information (an example of inserting special tokens (first and second description tokens)) when the description in FIG. 2, “Limitation of usage lot,” is the first description and the description in FIG. 2, “Decrease in sensitivity CD value,” is the second description.

[0034] <Adding item information> The item information is added by inserting a special token (first item token) at the start (start points of all tokens constituting the item to which the first description belongs) and at the end of the item to which the first description belongs. Also, the item information is added by inserting a special token (second item token) at the start and end of the item to which the second description belongs.

[0035] The information assignment unit 42 identifies the item to which the first description belongs and the item to which the second description belongs by referring to the results of the analysis (learning) of the sentence 30 performed by the description extraction unit 40 when extracting the first and second descriptions, and assigns item information by inserting special tokens (first and second item tokens) at the above-mentioned positions.

[0036] In this embodiment, as such special tokens (first and second item tokens), if the item to which the first or second description belongs is [Subject], [N] is inserted at the start and [ / N] at the end; if the item is [Abnormality content], [F] is inserted at the start and [ / F] at the end; if the item is [Abnormality content details], [Fd] is inserted at the start and [ / Fd] at the end; if the item is [Cause investigation results], [C] is inserted at the start and [ / C] at the end; and if the item is [Corrective action], [T] is inserted at the start and [ / T] at the end.

[0037] As described above, in this embodiment, the description "Limited to usage lot" in FIG. 2 is the first description, and the description "Decrease in sensitivity CD value" in FIG. 2 is the second description. The item to which the first description belongs is [Abnormality content], and the item to which the second description belongs is [Cause investigation result]. For this reason, in this embodiment, [F] is inserted as a special token (first item token) at the start of all tokens constituting the item [Abnormality content] to which the first description belongs, and [ / F] is inserted as a special token (first item token) at the end. In addition, [C] is inserted as a special token (second item token) at the start of all tokens constituting the item [Cause investigation result] to which the second description belongs, and [ / C] is inserted as a special token (second item token) at the end.

[0038] <Conversion to features> As described above, the feature conversion unit 44 encodes the sentence 30, which has been divided into tokens and has been assigned descriptive information and item information (sentence 30 into which special tokens have been inserted: the first descriptive token [E1], [ / E1], the second descriptive token [E2], [ / E2], the first item token [F], [ / F], and the second item token [C], [ / C]), using a learning model capable of learning context, such as the above-mentioned BERT, and converts each token, including the special tokens, into a feature.

[0039] <Combining features> The feature combining unit 46 calculates a description token combined feature 54 by combining a feature 50 of the first description token [E1] and a feature 52 of the second description token [E2]. As described above, in this embodiment, the feature of each token is a vector. In this embodiment, the description token combined feature 54 is calculated by combining a vector that is a feature of the first description token [E1] and a vector that is a feature of the second description token [E2]. In this combination, a combination method such as combining or concatenating vectors, or average pooling, max pooling, or element-by-element addition can be used, and the description token combined feature 54 is calculated by using (implementing) any one or a combination of a plurality of these.

[0040] The description token combination feature calculated as described above includes information indicating the presence or absence and strength of a causal relationship between the first description and the second description.

[0041] That is, in this embodiment, item information (first and second item tokens) regarding the items to which the first and second descriptions belong is added to the input sentence 30, and the sentence 30 to which the item information is added is encoded by a learning model capable of learning context. Therefore, the feature amount of each token obtained by encoding is influenced by the items to which the first and second descriptions belong (the item information is reflected by learning (analyzing) the sentence to which the item information is added).

[0042] In addition, it is known that the features of special tokens inserted at the beginning of a sentence, such as BERT's [CLS] token, summarize the meaning of the entire sentence, and the features of the first description token [E1] also summarize the meaning of the first description, and the features of the second description token [E2] also summarize the meaning of the second description.

[0043] In this way, the features 50 and 52 of the first and second description tokens [E1] and [E2] are summaries of the meanings of the first and second description tokens, and furthermore, as described above, the feature reflects the item information by encoding the sentence 30 with the item information (first and second item tokens) added. That is, the features 50 and 52 of the first and second description tokens [E1] and [E2] are summaries of the meanings of the first and second description tokens, and also reflect the item information. Furthermore, the feature 54 of the description token combination obtained by combining the features 50 and 52 of the first and second description tokens [E1] and [E2] represents the relevance between the first description and the second description. Furthermore, since the items are a division of a sentence by content, the relevance between a pair of descriptions (presence or absence and strength of causal relationship) tends to change depending on the item to which one of the pair of descriptions belongs and the item to which the other belongs. From the above, the description token combination feature 54 can be considered to include information indicating the presence or absence and strength of a causal relationship between the first description and the second description.

[0044] <Determining causal relationships> The text analysis unit 22 uses the above-mentioned property to determine whether or not there is a causal relationship between the first and second descriptions, using the description token combination feature 54. Specifically, the text analysis unit 22 inputs the description token combination feature to the causal relationship determination unit 48. The causal relationship determination unit 48 uses a learning model that has learned how the description token combination feature 54 changes depending on the causal relationship between a pair of descriptions included in a text and the relationship with the item to which each of the pair of descriptions belongs, for example, and compares the input description token combination feature 54 with the learning result to determine whether or not the input description token combination feature 54 indicates the presence of a causal relationship between the first and second descriptions, and outputs the determination result.

[0045] As described above, according to this embodiment, a sentence 30 to which item information has been added is learned (analyzed) using a learning model capable of learning context, so that it is possible to accurately determine whether or not a causal relationship exists between the descriptions contained in the sentence 30.

[0046] [Second embodiment] Hereinafter, the second embodiment of the present invention will be described with reference to Fig. 6. In the description using Fig. 6 and subsequent drawings, the same members and components as those in the above-described embodiment will be denoted by the same reference numerals and description thereof will be omitted.

[0047] As shown in Figure 6, in the second embodiment, in addition to the description token combination feature 54 used in the first embodiment described above (which combines the features 50, 52 of the first and second description tokens [E1] and [E2]), an item token combination feature 64 which combines the features 60, 62 of the first and second item tokens [F] and [C], and an integrated combination feature 70 which combines the description token combination feature 54 and the item token combination feature 64 are used to determine whether or not a causal relationship exists.

[0048] Specifically, in the second embodiment, the feature combining unit 46 (see FIG. 4) combines the feature 60 of the first item token [F] with the feature 62 of the second item token [C] to calculate the item token combined feature 64. As described above, the feature of each token is a vector, and the item token combined feature 64 is calculated by combining the vector that is the feature of the first item token [F] with the vector that is the feature of the second item token [C]. Note that, in this combination, as in the case of the description token combined feature 54 described above, any one or a combination of a combination method such as combining or concatenating vectors, or average pooling, max pooling, or element-by-element addition, can be used (implemented).

[0049] Next, the feature combination unit 46 calculates an integrated combined feature 70 by combining the description token combined feature 54 and the item token combined feature 64. The integrated combined feature 70 is calculated by adding the description token combined feature 54 and the item token combined feature 64 (adding vectors together).

[0050] Then, this integrated bond feature 70 is input to the causal relationship determination unit 48. In the second embodiment, the causal relationship determination unit 48 uses a learning model that has learned how the integrated bond feature 70 changes depending on the causal relationship between a pair of descriptions included in a sentence and the relationship between each of the pair of descriptions and the item to which the pair of descriptions belongs, and compares the input integrated bond feature 70 with the learning result to determine whether the input integrated bond feature 70 indicates the presence of a causal relationship between the first and second descriptions, and outputs the determination result.

[0051] [Third embodiment] Hereinafter, a third embodiment of the present invention will be described with reference to Figs. 7 and 8. As shown in Fig. 7, in the third embodiment, for example, a transition model 80 is stored in the memory 12 (see Fig. 1). The transition model 80 is a model in the form of a lookup table that predicts the probability of a causal relationship existing between an item to which a first description belongs and an item to which a second description belongs, and stores items to which the first and second descriptions belong (feature amounts 60, 62 of the first and second item tokens [F] and [C]) and a prediction score (in this embodiment, S 1,1 -S 5,5 In this embodiment, the predicted score is a scalar value that increases as the probability of the existence of a causal relationship increases.

[0052] The transition model 80 is created, for example, by learning causal relationships between items. The transition model 80 may be a trained model created in advance (before the input of the sentence 30) that is continuously used without updating, or a trained model created in advance that is appropriately updated and used by learning (optimizing, fine-tuning) the sentence 30 and other sentences. However, the latter is preferable in order to obtain a more accurate determination result of the causal relationship.

[0053] As shown in FIG. 8, in the third embodiment, the feature combination unit 46 (see FIG. 4) obtains a prediction score 84 indicating the strength of the causal relationship between the items to which the first and second descriptions belong by collating the items to which the first and second descriptions belong (the features 60, 62 of the first and second item tokens [F] and [C]) with a transition model 80. Then, the prediction score 84 is treated as an item token combination feature combining the first and second item tokens, and this (prediction score 84 (item token combination feature)) is combined with the description token combination feature 54 described above to calculate an integrated combination feature 90. In the third embodiment, the integrated combination feature 90 is calculated by multiplying the prediction score 84 (item token combination feature), which is a scalar value, by the description token combination feature 54, which is a vector (multiplication of a scalar value and a vector).

[0054] Then, this integrated bond feature 90 is input to the causal relationship determination unit 48. In the third embodiment, the causal relationship determination unit 48 uses a learning model that has learned how the integrated bond feature 90 changes depending on the causal relationship between a pair of descriptions included in a sentence and the relationship between each of the pair of descriptions and the item to which the pair of descriptions belongs, and compares the input integrated bond feature 90 with the learning result to determine whether or not the input integrated bond feature 90 indicates the presence of a causal relationship between the first and second descriptions, and outputs the determination result.

[0055] [Fourth embodiment] <Adding pseudo items> In the above embodiment, an example was described in which a sentence 30 (see FIG. 2) to which an item is added is input, but some sentences do not have an item added. In the fourth embodiment, when a sentence to which no item is added is input, a pseudo item is added to the sentence. Specifically, as in the language processing device 100 shown in FIG. 9, the processor 14 is made to function as an item adding unit 102 in addition to the above-mentioned units. For example, the item adding unit 102 uses a learning model constructed by learning (analyzing) a plurality of sentences to which an item is added, and when a sentence to which no item is added is input, it adds a pseudo item to the sentence and outputs it. In the language processing device 100, the sentence to which an item (pseudo item) is added in this way is analyzed by the above-mentioned sentence analysis unit 22. In this way, even for a sentence to which no item is added, it is possible to accurately determine a causal relationship taking into account the item to which the description belongs.

[0056] In addition, when an item (pseudo item) is assigned to a sentence without an item as described above, the item to be assigned to the sentence may be a preset one. That is, items such as [symptom], [cause], and [measure] may be prepared in advance as items to be assigned to the sentence, and when a sentence without an item is input, the sentence may be classified (an item is assigned) into the previously prepared items based on the learning result of learning (analyzing) the sentence. Of course, the item to be assigned to the sentence may be any item other than the preset items. In this case, the type of item (item name) to be assigned to the sentence and the position (range) to which the item is assigned may be determined by comparing the learning result (learning result of learning a plurality of sentences with items) obtained in conjunction with the construction of a learning model (a learning model that assigns items to sentences without items when the sentence without items is input and outputs the sentence).

[0057] In the above example, an item is added to the actual text, but it is not necessary to add an item to the actual text. In this case, the type of item (item name) and the position (range) to which the item should be added can be stored as data, and special tokens (first and second item tokens) can be inserted based on the stored data when adding item information.

[0058] In the above-mentioned first to fourth embodiments, an example is described in which it is determined whether or not a causal relationship exists between the first and second descriptions, but the present invention is not limited to this. It is also possible to calculate the probability that a causal relationship exists between the first and second descriptions, or the strength of the causal relationship between the first and second descriptions, and output these as a determination result of the causal relationship.

[0059] Furthermore, the present invention only requires that a learning result (determination result of the presence or absence of a causal relationship between the first and second descriptions) that takes into account the items to which the first and second descriptions belong is obtained by learning (analyzing) a sentence with item information added using a learning model capable of learning context. Therefore, the detailed configurations, such as the method of extracting the first and second descriptions, the method of tokenization, whether description information is added or used, the method of converting to features, whether feature quantities are combined and a specific combination method, and a specific method of determining the causal relationship, are not limited to the above embodiment and can be modified as appropriate.

[0060] The learning model that encodes the sentence into which the special token is inserted may be constructed by learning sentences into which no special token is inserted as pre-learning. In other words, general pre-learning may be performed without considering the presence or absence and strength of the causal relationship due to the item to which the description belongs. Conversely, the learning model may be constructed by learning sentences into which the special token is inserted as pre-learning. In other words, pre-learning may be performed taking into account the presence or absence and strength of the causal relationship due to the item to which the description belongs. However, the latter is preferable in order to convert the special token into a feature that more appropriately reflects the causal relationship between the descriptions. In the former case, it is preferable to optimize (fine-tune) the learning model by learning using sentences into which the special token is inserted. Of course, optimization may be performed even in the latter case.

[0061] In the above embodiment, the example was described in which learning and optimization were performed by inserting special tokens (first and second item tokens) only into the items to which the first and second descriptions belong, but learning and optimization may be performed by inserting special tokens into all items. In this way, it becomes possible to efficiently learn and optimize the causal relationships between items.

[0062] For example, cross-entropy loss can be used to optimize the various models (learning models) described above. Specifically, the cross-entropy error is calculated (the error between the predicted probability distribution of the model and the correct distribution) and the model can be optimized by minimizing the error.

[0063] In the above embodiment, the hardware structure of the processing units that execute various processes, such as the processor 14, the input text receiving unit 20, the text analyzing unit 22, the judgment result output unit 24, the description extracting unit 40, the information adding unit 42, the feature converting unit 44, the feature combining unit 46, the causal relationship judging unit 48, and the item adding unit 102, is various processors as shown below. The various processors include a CPU (Central Processing Unit), which is a general-purpose processor that executes software (programs) and functions as various processing units, a programmable logic device (PLD), which is a processor whose circuit configuration can be changed after manufacture, such as an FPGA (Field Programmable Gate Array), and a dedicated electric circuit, which is a processor having a circuit configuration designed specifically for executing various processes.

[0064] One processing unit may be configured with one of these various processors, or may be configured with a combination of two or more processors of the same type or different types (for example, multiple FPGAs, or a combination of a CPU and an FPGA). Also, multiple processing units may be configured with one processor. As an example of configuring multiple processing units with one processor, first, as represented by a computer such as a client or server, one processor is configured with a combination of one or more CPUs and software, and this processor functions as multiple processing units. Second, as represented by a system on chip (SoC), there is a form in which a processor is used that realizes the functions of the entire system including multiple processing units with one IC (Integrated Circuit) chip. In this way, the various processing units are configured using one or more of the above various processors as a hardware structure.

[0065] Furthermore, the hardware structure of these various processors is, more specifically, an electric circuit in the form of a combination of circuit elements such as semiconductor elements. [Explanation of symbols]

[0066] 10, 100 Language Processing Device 12. Memory 14 processors 20 Input text reception section 22 Text analysis department 24 Judgment result output section 30 sentences 40 Description Extraction Unit 42 Information provision section 44 Feature conversion unit 46 Feature Combination Unit 48 Causal Relation Determination Section 50 Features (Features of the first token [E1]) 52 Features (Features of the second description token [E2]) 54 Features (descriptive token combination features) 60 features (features of the first token [F]) 62 Features (Features of the second item token [C]) 64 Features (item token combination features) 70 features (integrated combined features) 80 Transition Model 84 Prediction score (item token combination feature) 90 Features (Integrated Combined Features) 102 Item assignment section

Claims

1. a processor; The processor, Accepts input of sentences containing multiple descriptions, When the text is classified into categories, the categories to which the first and second descriptions belong are identified; adding item information indicating an item to which each of the first and second descriptions belongs to the sentence; determining whether or not there is a causal relationship between the first description and the second description by using the sentence to which the item information is added; Language processing device.

2. The processor, Separating the sentence into a number of tokens; Converting tokens constituting the sentence into features using a learning model capable of learning context; The determination is made using the transformed feature amount. The language processing device according to claim 1 .

3. the item information is assigned by inserting a first item token as a special token, which is a type of the token, at the start and end of the item to which the first description belongs, and by inserting a second item token as the special token at the start and end of the item to which the second description belongs; The language processing device according to claim 2 .

4. The processor, inserting a first description token as the special token at the start and end of the first description, and inserting a second description token as the special token at the start and end of the second description; performing the determination using a first description token feature that is the feature of the first description token and a second description token feature that is the feature of the second description token; The language processing device according to claim 3 .

5. The processor, calculating a description token combined feature by combining the first description token feature and the second description token feature; The determination is made using the description token combination feature. The language processing device according to claim 4.

6. the first description token feature amount and the second description token feature amount are calculated as multi-dimensional vectors; the description token combination feature is calculated by adding the first description token feature and the second description token feature. The language processing device according to claim 5 .

7. The processor, In addition to the token combination features, performing the determination using a first item token feature that is the feature of the first item token and a second item token feature that is the feature of the second item token; The language processing device according to claim 6.

8. The processor, calculating an item token combined feature by combining the first item token feature and the second item token feature; Calculating an integrated combined feature by combining the item token combined feature and the description token combined feature; The determination is performed using the integrated binding feature. The language processing device according to claim 7.

9. The item token combination feature and the description token combination feature are calculated as a multi-dimensional vector, the integrated combined feature is calculated by adding the item token combined feature and the description token combined feature. The language processing device according to claim 8.

10. the first item token feature and the second item token feature are calculated as multi-dimensional vectors; the item token combined feature is calculated by adding the first item token feature and the second item token feature. The language processing device according to claim 9.

11. The processor, creating a transition model showing a transition state of the strength of the causal relationship between a plurality of items including the items to which the first and second descriptions belong; calculating the item token combination feature as a scalar value indicating a strength of a causal relationship between the first item token feature and the second item token feature using the transition model; calculating the integrated combined feature by multiplying the item token combined feature by the description token combined feature; The language processing device according to claim 9.

12. The processor, Adding an item to the text if the item does not exist; The language processing device according to any one of claims 1 to 11.

13. The processor, Calculating an error between a predicted probability distribution of the causal relationship between a plurality of items and a correct distribution, and optimizing a learning model for making the judgment using the calculated error. The language processing device according to any one of claims 1 to 11.

14. The language processing device computer A step of accepting an input of a sentence including a plurality of descriptions; a step of identifying a category to which each of a first description and a second description of the descriptions belongs when the sentence is classified into categories; adding item information indicating an item to which each of the first and second descriptions belongs to the sentence; determining whether or not there is a causal relationship between the first description and the second description by using the sentence to which the item information is added; A program for executing the above.

Images