Information processing apparatus, information processing method, and program
The apparatus addresses imprecise labeling of named entities by using document type-specific attributes to assign precise labels, enhancing the accuracy of named entity extraction and file naming.
Patent Information
- Application Number
- JP2023102556
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-06-22
- Publication Date
- 2025-01-09
AI Technical Summary
Existing technologies fail to accurately label named entities whose meaning changes depending on document attributes, such as document type, leading to imprecise labeling.
An information processing apparatus that assigns a first named entity label based on document image data, extracts named entities, and applies a second label associated with document type-specific attributes using conversion maps and priority rules to ensure precise labeling.
Enables accurate labeling of named entities by accounting for document type-specific meanings, improving precision in named entity extraction and file naming processes.
Smart Images
Figure 2025002396000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to an information processing technique for labeling named entities extracted from document data. [Background technology]
[0002] Conventionally, there is a technology that reads characters in a document image, converts it into an electronic file, and automatically assigns an appropriate file name. Elements that compose the automatically assigned file name include the name of an organization or person in the document, the document number, the document publication date, etc., in addition to the file creation date and time, and these can be extracted by a named entity extraction technology that is an applied technology in the field of natural language processing. Patent Document 1 discloses a named entity extraction device that uses a plurality of hierarchical recognition models to extract named entities that are assigned labels of extended named entities with a predefined hierarchical structure from input document data, and outputs document data in which the labels of each hierarchy are assigned to the named entities.
[0003] Incidentally, the label assigned to a named entity may change depending on the document attributes such as document type. For example, the label representing the "document issuer," which means the party that receives the issued document, may be the purchaser or the delivery address for an invoice, but the "document issuer" may be the order recipient or the delivery address for a purchase order. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2013-246795 A Summary of the Invention [Problem to be solved by the invention]
[0005] However, in the technology of Patent Document 1, the multiple hierarchical identification models label named entities using only the text information of the named entities, and therefore it is not possible to accurately assign appropriate labels to named entities whose meanings change depending on the attributes of the document. [Means for solving the problem]
[0006] The present invention is an information processing device comprising a first assignment means for assigning a first named entity label to a token obtained by dividing a character string included in input document image data, an extraction means for extracting a named entity based on the first named entity label assigned by the first assignment means, an acquisition means for acquiring attribute information including at least information indicating a document type of the document image data based on a predetermined character string included in the document image data, and a second named entity label associated with the first named entity label according to the attribute information, the second named entity label corresponding to the attribute information being assigned to the named entity to which the first named entity label corresponding to the second named entity label has been assigned, and the second named entity label is different from the first named entity label. Effect of the Invention
[0007] According to the present invention, it is possible to appropriately label named entities extracted from document data. [Brief description of the drawings]
[0008] [Figure 1] A block diagram showing the hardware configuration of an information processing device. [Diagram 2] A block diagram showing the main parts of the functional configuration of an information processing device. [Diagram 3] FIG. 1 is a block diagram showing an overall functional configuration of an information processing device according to a first embodiment. [Figure 4] Flowchart regarding processing of information processing device [Diagram 5] A flowchart of a named entity extraction process executed by a named entity extraction unit. [Figure 6] FIG. 13 is a schematic diagram of a specific example of a result of a named entity extraction process performed by a named entity extraction unit; [Figure 7] 1 is a flowchart of a second named entity labeling process executed by a second named entity labeling unit according to the first embodiment. [Figure 8] FIG. 1 is a schematic diagram of a specific example of a conversion map used by a second named entity labeling unit; [Figure 9] FIG. 1 is a schematic diagram of a specific example of a result of processing by a second named entity labeling unit using a conversion map; [Figure 10] FIG. 1 is a schematic diagram of a specific example of a priority map used by a second named entity labeling unit; [Figure 11] FIG. 1 is a schematic diagram of a specific example of a result of processing by a second named entity labeling unit using a priority map; [Figure 12] FIG. 13 is a schematic diagram of a specific example of a result of the second named entity labeling process executed by the second named entity labeling unit; [Figure 13] Specific examples of named entity extraction results displayed by the display means [Figure 14] A schematic diagram of a specific example of the hierarchical structure of files to be saved [Figure 15] FIG. 11 is a block diagram showing the overall functional configuration of an information processing device according to a second embodiment. [Figure 16] 1 is a flowchart of a second named entity labeling process executed by a second named entity labeling unit according to a second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Note that the configurations shown in the following embodiments are merely examples, and the present invention is not limited to the configurations shown in the drawings.
[0010] [Embodiment 1] FIG. 1 is a block diagram showing the hardware configuration of information processing device 100. FIG. 2 is a block diagram showing the main parts of the functional configuration of information processing device 100. FIG. 3 is a block diagram showing the overall functional configuration of information processing device 100. FIG. 4 is a flowchart related to the processing of information processing device 100. The processing of information processing device 100 described below aims to extract a desired image area from characters in a document image and automatically transcribe it as an appropriate file name for document image data. Hereinafter, the operation of this embodiment will be described with reference to FIGS. 1 to 14.
[0011] The CPU 101 reads out programs and data stored in the ROM 103 into the RAM 102 and executes various processes, which will be described later.
[0012] In S1000, the image acquisition unit 200 acquires image data of a document to be processed by operating the input device 104 provided in the information processing device 100. The document image data is acquired by using a scanner 105 for a stack of paper documents. However, the document image data is not limited to this, and may be acquired from a storage device 106 such as a HDD, or may be acquired on a network 109 via a network interface 108.
[0013] In S1001, the character recognition unit 210 detects a character string area included in a document image, and acquires a character string written in the detected character string area. Furthermore, based on the coordinates of the character string area, acquires one character string by combining each character string in the input image as an input character string. Then, it creates document data D including the input character string, position information of each character string, position information of ruled lines, size of the document image, and other information.
[0014] In S1002, the named entity extraction unit 220 performs a named entity extraction process based on the document data D. Details of the process in S1002 will be described below with reference to FIG.
[0015] FIG. 5 is a flowchart of the named entity extraction process executed by the named entity extraction unit 220 in S1002.
[0016] In S1100, the division unit 221 in the named entity extraction unit 220 divides (separates) the input character string in the document data D into token units to generate an input token string. For example, if the input character string is "Publication date: June 11, 2017", it is divided into nine tokens, " / Publication / Day / : / 2017 / Year / June / Month / 11 / Day / ".
[0017] In S1101, the conversion unit 222 acquires the conversion dictionary 281 stored in the storage unit 280 of the storage device 106.
[0018] In S1102, the inference unit 223 loads the inference model 282 stored in the storage unit 280. Here, the inference model 282 is a parameter set that defines a machine learning model that realizes the inference unit 223. If the machine learning model has a neural network structure such as Transformer, the parameter set has various parameters that define the network configuration, the weights between networks, the bias, and the like.
[0019] In S1103, the conversion unit 222 uses the conversion dictionary 281 to map each token in the input token sequence to the feature space, thereby generating an input feature tensor.
[0020] In S1104, the inference unit 223 inputs the input feature tensor to the inference model 282, and the inference model 282 outputs the probability that each token becomes a named entity label for each named entity label belonging to the first named entity label group S1. The inference unit 223 outputs the probabilities output by the inference model 282 as a set of probabilities that each token becomes a named entity label.
[0021] In S1105, the first named entity label assignment unit 224 extracts the named entity label with the highest probability from the set of probabilities corresponding to each token, and assigns the extracted named entity label to the token corresponding to the set of probabilities from which the named entity label was extracted.
[0022] In S1106, the extraction unit 225 extracts a character string consisting of a plurality of tokens that is inferred to be a named entity based on the named entity label assigned in S1105, and generates a named entity extraction result.
[0023] More specifically, steps S1105 and S1106 above are performed as follows. First, the named entity label uses the IOB2 format, which is commonly used in the field of named entity extraction technology. In the IOB2 format, the first token of a string that means a named entity is given the B-<arbitrary character string to identify the label> label, and the second and subsequent tokens are given the I-<arbitrary character string to identify the label> label. If the token is not one of the named entities, the O label is given. Then, the region of the same type of I label that follows the B label is considered to be a string of the same label. For example, for the token string " / order / day / : / 2017 / year / 6 / month / 11 / day / ", the B-DATEorder label is inferred for the token "2017", followed by the I-DATEorder label that continues up to the final "day". In this case, "June 11, 2017" is extracted as the string that corresponds to the named entity label "DATEorder".
[0024] The named entity labels belonging to the first named entity label group S1 assigned by the named entity label assignment process in S1105 include, in addition to the above-mentioned "DATEorder," the following: "DATEinvoice" assigned to the invoice date, "TITLE" assigned to the title, "IDquote" assigned to the quotation number, "IDinvoice" assigned to the invoice number, and "IDorder" assigned to the purchase order number. In addition, there are "ORGbuyer" assigned to the purchasing organization, "ORGseller" assigned to the selling organization, "ORGreceiver" assigned to the destination organization, and "ORGform" assigned to the organization that created the document template.
[0025] Fig. 6 shows a document image obtained by performing named entity extraction processing by the named entity extraction unit 220 using the first named entity label group S1, and a specific example of the processing result. Fig. 6(a) shows a document image obtained by reading a purchase order, and Fig. 6(b) shows the extraction result obtained by performing named entity extraction processing on the purchase order shown in Fig. 6(a). Fig. 6(c) shows a document image obtained by reading an invoice, and Fig. 6(d) shows the extraction result obtained by performing named entity extraction processing on the invoice shown in Fig. 6(c).
[0026] Returning now to FIG. 4, in S1003, the document type determination unit 230 performs document type determination for the document data D. Specifically, this is performed as follows. A character string corresponding to the named entity label of "TITLE" is obtained from the named entity extraction result, and it is determined whether or not the character string contains a reserved word that represents each document type. If a reserved word is contained, the document type associated with the reserved word is set as the document type of the document for which the named entity extraction process was performed. For example, if "invoice" has been extracted as the character string corresponding to the named entity label of "TITLE", the character string contains "claim", which is a reserved word for the document type "invoice", so in this case the document type can be determined to be "invoice".
[0027] Using the above method, the document type determination unit 230 determines the document type of the specific example data in Figures 6(a) and (b) to be "purchase order", and the document type of the specific example data in Figures 6(c) and (d) to be "invoice".
[0028] The method of document type determination is not limited to this, and other methods may be used. For example, an inference model for document type determination may be prepared separately, and an input token string obtained by dividing an input character string in document data D into token units may be input, and a document type determination result may be obtained as an output result. In addition, the inference model 282 may be an inference model capable of multitasking inference that combines named entity extraction and document type determination, so that named entity extraction results and document type determination results can be obtained.
[0029] In S1004, the second named entity label assignment unit 240 additionally assigns a named entity belonging to the second named entity label group S2 to the named entity extracted in the named entity extraction process in S1002. The named entity labels belonging to the second named entity label group S2 are presented to the user as the extraction result of the named entity by the display control unit 250 through a process described below. Examples of named entity labels belonging to the second named entity label group S2 include the following: "DATEpublish" assigned to the document publication date, "IDdoc" assigned to the document number, "TITLE" assigned to the title, "ORGfrom" assigned to the issuing organization, "ORGto" assigned to the issued organization, etc. Details of the process of S1004 will be described below with reference to FIG. 7.
[0030] FIG. 7 is a flowchart relating to the second named entity labeling process executed by the second named entity labeling unit 240 in S1004.
[0031] In S1200, the second named entity labeling unit 240 loads the conversion map 283 stored in the storage unit 280. FIG.
[0032] In S1201, the second named entity labeling unit 240 derives a second named entity label or an indefinite label corresponding to the first named entity label extracted in the named entity extraction process in S1002, based on the document type obtained in S1003 and the conversion map 283. Fig. 9 is a schematic diagram of a specific example of a result of processing by the second named entity labeling unit 240 using the conversion map 283. Fig. 9(a) corresponds to the specific example shown in Figs. 6(a) and (b), and Fig. 9(b) corresponds to the specific example shown in Figs. 6(c) and (d).
[0033] In addition, in the conversion map shown in Figure 8, for the document type of "purchase order", the conversion destination of "ORGform" is "undefined" because, in the case of a purchase order, the organization that created the document template can be either the document issuer or the document recipient.
[0034] In S1202, the second named entity label assigning unit 240 replaces the indefinite label assigned in S1201 with a named entity label belonging to the second named entity label group S2 based on a predetermined rule. The conversion of the indefinite label to a named entity label belonging to the second named entity label group S2 can be performed in the following manner using the named entity labels assigned to the own and other named entities and text information indicating the character string extracted as a named entity. First, a character string extracted as a named entity whose conversion destination is an indefinite label is compared with a character string extracted as a named entity whose conversion destination is not an indefinite label. Then, if the similarity between the character strings extracted as the named entities is equal to or greater than a threshold, the indefinite label of the converted named entity whose similarity is equal to or greater than the threshold is replaced with the converted named entity label of the named entity whose conversion destination is equal to or greater than the threshold. In the example shown in FIG. 9(a), for the derived named entity whose second named entity label is "indefinite", there is a named entity whose converted named entity label is "ORGfrom" and whose character string extracted as a named entity is the same, "ABC Co., Ltd.". Therefore, by the above process, the named entity label derived by the conversion map 283 can be replaced from "undefined" to "ORGfrom".
[0035] If the above process fails to convert the indefinite label because the text similarity is below a threshold, the indefinite label is converted to another named entity label according to a specified rule. The specified rule is, for example, to replace a named entity label belonging to the first named entity label group S1 and a second named entity label with the same named entity label. Specifically, if a named entity label belonging to the first named entity label group S1 is "ORGform", a named entity label belonging to the second named entity label group S2 is replaced with "ORGfrom".
[0036] In S1203, the second named entity labeling unit 240 loads the priority order map 284 stored in the storage unit 280. FIG.
[0037] In S1204, the second named entity label assignment unit 240 acquires priority information of each named entity based on the document type, the named entity labels belonging to the first named entity label group S1, and the priority map 284.
[0038] In S1205, the second named entity labeling unit 240 converts some of the named entity labels belonging to the second named entity label group S2 into O labels based on the priority order among the named entity labels in the first named entity label group S1. Fig. 11 shows a specific example of a result of processing by the second named entity labeling unit 240 using the priority order map 284. Fig. 11(a) is a specific example of processing the example shown in Fig. 9(a), and Fig. 11(b) is a specific example of processing the example shown in Fig. 9(b).
[0039] The process of S1205 will be described with reference to FIG. 11(a). Based on the information that the document type is "purchase order", the information of the named entity labels belonging to the first named entity label group S1, and the priority map 284, the table shown in FIG. 11(a) is obtained. The priority of each named entity in the example shown in FIG. 9(a) is as shown in the priority column of the table shown in FIG. 11(a). The priority information is compared for each type of the converted named entity label at the time of S1202, and the named entity label other than the one with the top (highest) priority among the named entity labels of the same type is converted to an O label. In the example of FIG. 11(a), the highest priority of the named entity label with the converted named entity label "ORGfrom" is "1". Therefore, the converted named entity label of the named entity label "ORGreceiver" (the extracted character string is "ABC Sales Co., Ltd.") with the converted named entity label "ORGfrom" but with the priority "2" is converted to an O label. This makes it possible to extract only the most suitable named entities to be presented to the user in the final process described below, reducing the effort required for the user to confirm and correct the entities.
[0040] In S1206, the second named entity label assignment unit 240 assigns, to each named entity, information on the named entity label of the second named entity label group S2, which is a converted named entity label used for presenting to the user.
[0041] Fig. 12 is a schematic diagram showing a specific example of a result obtained by the second named entity labeling process executed by the second named entity labeling unit 240. Fig. 12(a) is an example of document data of a purchase order, and Fig. 12(b) is a result obtained by performing the second named entity labeling process on the document data shown in Fig. 12(a). Fig. 12(c) is an example of document data of an invoice, and Fig. 12(d) is a result obtained by performing the second named entity labeling process on the document data shown in Fig. 12(c).
[0042] As shown in FIG. 12(b) and (d), the named entity label belonging to the first named entity label group S1 assigned by the named entity extraction unit 220 is a named entity label having a meaning that the named entity to which the label is assigned does not change regardless of the document type, which is an attribute of the document. For example, the character string "orderer" extracted as a named entity is assigned a named entity label "ORGbuyer" having an appropriate meaning whether it is a purchase order or an invoice. Therefore, the named entity extraction unit 220 can assign a named entity label belonging to the first named entity label group S1 handled by itself with high accuracy regardless of the document type. On the other hand, the second named entity label assignment unit 240 assigns a named entity label having a meaning that the named entity to which the label is assigned changes for each document type. The second named entity label assignment unit 240 assigns a named entity label belonging to the second named entity label group S2 to a named entity to which a named entity label belonging to the first named entity label group S1 is assigned, based on the assigned named entity label and the document type. This allows for appropriate named entity label information to be assigned for final output, taking into account differences in the meaning of named entities according to document type. As a result, this technology allows for accurate named entity extraction and label assignment even for documents with different attributes, such as document type.
[0043] As a method of assigning a named entity label belonging to the second named entity label group S2 to a character string extracted as a named entity, there is also a method of extracting named entities taking into account the document type, using an inference model that performs multitask inference for both named entity extraction and document type determination. However, this method requires an increased number of model parameters so that the model can learn the difference in meaning of the character string extracted as a named entity according to the document type. Therefore, this method of performing named entity extraction and document type determination using a single inference model requires a larger inference model size than the method of this embodiment, and requires more training data and processing resources during inference.
[0044] Returning now to Fig. 4. In S1005, the display control unit 250 displays the named entity extraction results on the display device 107, such as a touch panel. Fig. 13 shows a specific example of a user interface (UI) screen 251 on which the display control unit 250 displays the named entity extraction results. Fig. 13(a) displays the results shown in Fig. 12(a), and Fig. 13(b) displays the results of Fig. 12(b).
[0045] The UI screen 251 has a preview pane 252 which is an area for displaying a preview of a document image obtained using the scanner 105, and a property pane 253 which is an area for displaying the extraction results of named entities. In the property pane 253, a character string 254 indicating a category and an extracted character string 255 are displayed for each extracted named entity. There is a one-to-one correspondence between the named entity labels belonging to the second named entity label group S2 assigned by the second named entity label assignment unit 240 and the character string 254 indicating a category displayed on the UI screen 251, and the correspondence is as follows: If the named entity label is a character string indicating a category, then TITLE is the title, DATEpublish is the document publication date, IDdoc is the document number, ORGfrom is the organization publisher, and ORGto is the organization destination.
[0046] If multiple named entities with the same named entity label are detected, the one with the character string extracted as the named entity located at the top left of the document image is selected and displayed. Also, in the preview pane 252, the portion of the document image that corresponds to the named entity is highlighted. In Fig. 13, the highlight is represented by a dashed rectangle.
[0047] In S1006, the display control unit 250 requests the user to approve or correct the named entity extraction result. When correcting the named entity extraction result, the user selects the character string 254 indicating the category that needs to be corrected on the property pane 253 through the display device 107 such as a touch panel, and selects the area in the preview pane 252 where the corrected character string is displayed. Note that, when multiple named entities corresponding to the category for which correction has been requested are detected, the portion of the named entity that is not displayed in the property pane 253 is presented to the user.
[0048] In S1007, the named entity output unit 260 outputs the named entity extraction result and the document type determination result to the file storage unit 270.
[0049] In S1008, the file storage unit 270 assigns a file name to the document image data acquired by the image acquisition unit 200 and stores it in a predetermined folder based on the output result of the named entity output unit 260. The document image data is converted into a PDF format and stored in the storage device 106 or in an external server via the network 109.
[0050] FIG. 14 shows an example of the hierarchical structure of files to be saved. The destination folder name and the file name of the document image data are determined according to rules previously set by the user. In the example shown in FIG. 14, the folder name and file name are determined as follows: First, the destination folder name is determined based on the document type determination result. If the document type determination result is "invoice", the document image file is saved under the "invoice" folder. The document image file name is created by connecting the organization issuer, document number, and document issue date with underscores. Note that the dots between the year, month, and date of the document issue date have been removed to accommodate differences in file name assignment depending on the display format.
[0051] [Embodiment 2] In the first embodiment, the second named entity labeling process is performed using the conversion map 283 and the priority map 284, but this is merely one embodiment of the technology of the present disclosure. In the present embodiment, an example will be described in which the second named entity labeling process is performed using a machine learning model that has been trained in advance.
[0052] Fig. 15 is a block diagram showing the overall functional configuration of the information processing device 100 in this embodiment. Fig. 16 is a flowchart relating to the second named entity labeling process executed by the second named entity labeling unit 240 in this embodiment. This flowchart is executed in S1004 after the information processing device 100 extracts named entities (S1002) and determines the document type (S1003). Note that S1000 to S1003 and S1005 to S1008 are the same as in the first embodiment, and therefore will not be described here.
[0053] In S2000, the second named entity labeling unit 240 loads the document attribute information 285 and the second named entity labeling model 286 stored in the storage unit 280. The document attribute information 285 includes at least page information of the document, and may further include information such as the country of transaction, the language used, whether the document was published after a specific legal amendment, whether it is a cover, etc. The second named entity labeling model 286 is a learning model of a neural network that has been trained in advance. However, without being limited thereto, the second named entity labeling model 286 may be, for example, a machine learning model of a pre-trained SVM or decision tree.
[0054] In S2001, the second named entity label assignment unit 240 generates an input feature tensor based on the document type, the document attribute information 285, the character string extracted as a named entity by the named entity extraction unit 220, and the named entity label belonging to the first named entity label group S1.
[0055] The document type used here is the determination result in S1003 by the document type determination unit 230. In this embodiment, the document type is treated as information separate from the document attribute information 285, but the document type may be included in the document attribute information 285.
[0056] In S2002, the second named entity labeling unit 240 performs inference by inputting the generated input feature tensor to the second named entity labeling model 286. The inference result is an output of the probability that each named entity label belonging to the second named entity label group S2 becomes a named entity label for each named entity.
[0057] In S2003, the second named entity label assignment unit 240 determines, for each named entity, the one with the highest probability (inference score) of being a named entity label belonging to the second named entity label group S2 as the named entity label for each named entity.
[0058] In S2004, label information of the determined second named entity label group is assigned to each named entity.
[0059] As described above, by using a machine learning model for the second named entity label assignment, the named entity labels of the second named entity label group to be presented to the user can be determined using information on a wider range of document attributes. This can provide the following effects. That is, the named entity labels belonging to the first named entity label group S1 assigned by the named entity extraction unit 220 have the same meaning of the named entity to which the label is assigned, regardless of the document type, which is an attribute of the document, so that the named entities can be extracted with high accuracy. In addition, the second named entity label assignment unit 240 assigns additional named entity labels set in consideration of the difference in meaning of the named entity to which the label is assigned according to the document attributes, so that the information of the named entity label set according to the document attributes can be assigned to the named entity. This makes it possible to perform named entity extraction and named entity label assignment with high accuracy even for documents with different attributes.
[0060] (Other Examples) The present invention can also be realized by a process in which a program for implementing one or more of the functions of the above-described embodiments is supplied to a system or device via a network or a storage medium, and one or more processors in a computer of the system or device read and execute the program. The present invention can also be realized by a circuit (e.g., ASIC) that implements one or more of the functions.
[0061] The present disclosure includes the following configurations and methods. [Configuration 1] a first assigning means for assigning a first named entity label to a token obtained by dividing a character string included in document image data obtained by scanning a document; an extracting means for extracting a named entity based on the first named entity label assigned by the first assigning means; an acquisition means for acquiring attribute information including at least information indicating a type of the document based on a predetermined character string included in the document image data; a second assigning means for assigning a second named entity label corresponding to the attribute information to the named entity to which the first named entity label corresponding to the second named entity label has been assigned, the second named entity label being associated with the first named entity label according to the attribute information; Equipped with the second named entity label is different from the first named entity label; 23. An information processing apparatus comprising: [Configuration 2] The predetermined character string is included in the named entity to which a predetermined first named entity label is assigned. 2. The information processing device according to configuration 1. [Configuration 3] The predetermined first named entity label is a first named entity label assigned to a title in the document image data. 3. The information processing device according to configuration 2. [Configuration 4] the acquiring means acquires priority information assigned to the first named entity label in accordance with the attribute information; when the second assigning means assigns the same second named entity label to named entities to which a plurality of different first named entity labels have been assigned, the second assigning means replaces the second named entity labels to be assigned to named entities to which the other first named entity labels have been assigned, except for the first named entity label having the highest priority in the priority information, among the named entities to which the plurality of different first named entity labels to which the same second named entity label has been assigned, with an indefinite label; 4. The information processing device according to claim 1, wherein: [Configuration 5] the second assigning means, when there is a named entity similar to the named entity to which the indefinite label is assigned among the named entities to which the second named entity label other than the indefinite label is assigned, replaces the indefinite label with the second named entity label assigned to the similar named entity; 5. The information processing device according to configuration 4. [Configuration 6] The second assigning means replaces the indefinite label assigned to the named entity having no similar named entity with a named entity label other than the indefinite label in accordance with a predetermined rule. 6. The information processing device according to configuration 5. [Configuration 7] the second assigning means replaces the indefinite label assigned to a named entity having no similar named entity with the first named entity label assigned to the named entity to which the indefinite label has been assigned; 7. The information processing device according to configuration 6. [Configuration 8] a display control means for displaying, on a display device, a character string corresponding to the second named entity label and the named entity to which the second named entity label has been assigned; 8. The information processing device according to any one of configurations 1 to 7, further comprising: [Configuration 9] the second assigning means derives the second named entity label associated with the first named entity label in accordance with the attribute information, based on a conversion map that associates the first named entity label with the second named entity label; 9. The information processing device according to any one of configurations 1 to 8. [Configuration 10] the second assigning means derives the second named entity label associated with the first named entity label according to the attribute information by using an inference model having the named entity, the first named entity label, and the attribute information as inputs; 9. The information processing device according to any one of configurations 1 to 8. [Configuration 11] The attribute information further includes at least one of page information of the document, a country of commerce, a language used, whether or not a specific legal amendment has been made, and whether or not the document has a cover page; 11. The information processing device according to any one of configurations 1 to 10. [Configuration 12] a step of assigning a first named entity label to tokens obtained by dividing a character string included in document image data obtained by scanning a document; extracting a named entity based on the first named entity label assigned by the step of assigning the first named entity label; acquiring attribute information including at least information indicating a type of the document based on a predetermined character string included in the document image data; a step of assigning a second named entity label that is associated with the first named entity label according to the attribute information, the second named entity label corresponding to the attribute information, to the named entity to which the first named entity label corresponding to the second named entity label has been assigned; Equipped with the second named entity label is different from the first named entity label; 23. An information processing method comprising: [Configuration 13] A program for causing a computer to function as the information processing device according to any one of configurations 1 to 10.
Claims
1. a first labeling means for labeling a first named entity label to a token obtained by dividing a character string included in the input document image data; an extracting means for extracting a named entity based on the first named entity label assigned by the first assigning means; an acquiring means for acquiring attribute information including at least information indicating a document type of the document image data based on a predetermined character string included in the document image data; a second assigning means for assigning a second named entity label corresponding to the attribute information to the named entity to which the first named entity label corresponding to the second named entity label has been assigned, the second named entity label being associated with the first named entity label according to the attribute information; Equipped with the second named entity label is different from the first named entity label; 23. An information processing apparatus comprising:
2. The predetermined character string is included in the named entity to which a predetermined first named entity label is assigned.
2. The information processing apparatus according to claim 1,
3. the predetermined first named entity label is a first named entity label assigned to a title in the document image data; 3. The information processing apparatus according to claim 2.
4. the acquiring means acquires priority information assigned to the first named entity label in accordance with the attribute information; when the second assigning means assigns the same second named entity label to named entities to which a plurality of different first named entity labels have been assigned, the second assigning means replaces the second named entity labels to be assigned to named entities to which the other first named entity labels have been assigned, except for the first named entity label having the highest priority in the priority information, among the named entities to which the plurality of different first named entity labels have been assigned and to which the same second named entity label has been assigned, with an indefinite label; 2. The information processing apparatus according to claim 1,
5. the second assigning means, when there is a named entity similar to the named entity to which the indefinite label is assigned among the named entities to which the second named entity label other than the indefinite label is assigned, replaces the indefinite label with the second named entity label assigned to the similar named entity; 5. The information processing apparatus according to claim 4.
6. The second assigning means replaces the indefinite label assigned to the named entity having no similar named entity with a named entity label other than the indefinite label in accordance with a predetermined rule.
6. The information processing apparatus according to claim 5,
7. the second assigning means replaces the indefinite label assigned to a named entity having no similar named entity with the first named entity label assigned to the named entity to which the indefinite label has been assigned; 7. The information processing apparatus according to claim 6,
8. a display control means for displaying, on a display device, a character string corresponding to the second named entity label and the named entity to which the second named entity label has been assigned; 2. The information processing apparatus according to claim 1, further comprising:
9. the second assigning means derives the second named entity label associated with the first named entity label in accordance with the attribute information, based on a conversion map that associates the first named entity label with the second named entity label; 2. The information processing apparatus according to claim 1,
10. the second assigning means derives the second named entity label associated with the first named entity label according to the attribute information by using an inference model having the named entity, the first named entity label, and the attribute information as inputs; 2. The information processing apparatus according to claim 1,
11. The attribute information further includes at least one of page information of the document, a country of commerce, a language used, whether or not a specific legal amendment has been made, and whether or not a cover page has been made.
11. The information processing apparatus according to claim 1,
12. a step of assigning a first named entity label to tokens obtained by dividing a character string included in the input document image data; extracting a named entity based on the first named entity label assigned by the step of assigning the first named entity label; acquiring attribute information including at least information indicating a document type of the document image data based on a predetermined character string included in the document image data; a step of assigning a second named entity label that is associated with the first named entity label according to the attribute information, the second named entity label corresponding to the attribute information, to the named entity to which the first named entity label corresponding to the second named entity label has been assigned; Equipped with the second named entity label is different from the first named entity label; 23. An information processing method comprising:
13. A program for causing a computer to function as the information processing device according to any one of claims 1 to 10.
Citation Information
Patent Citations
Proper expression extraction device, method and program
JP2013246795A