Document creation device, program, and document creation method

Abstract

In document creation technology that generates documents according to a predetermined format based on input voice, it is possible to create documents based on unregistered words even when they are input via voice. [Solution] When the report creation and recording device 1 receives category information and audio information from the communication terminal 2, it selects a template that has been pre-registered and linked to this category information. If there is a word among the words extracted from the audio information that matches or has a similar meaning to a word that has been pre-registered as a replacement candidate for a label included in this template, the device replaces this label with the matching or similar word. If there is a word among the words extracted from the audio information that matches or has a similar pronunciation to a word registered as a replacement candidate for this label, the device modifies this extracted word to the matching or similar word and replaces this label with the modified extracted word. This creates the report.

Description

Technical Field

[0001] The present invention relates to a document creation technique for creating a document based on input speech, and more particularly to a document creation technique suitable for creating a report document according to a predetermined format.

Background Art

[0002] Patent Document 1 discloses a fixed phrase corpus creation device that creates a fixed phrase corpus used for generating synthetic speech and stores it in a speech database. This fixed phrase corpus creation device includes a fixed phrase input unit that receives fixed phrases and the positions of optional words to be inserted between the fixed phrases, an optional word input unit that receives the attributes of the optional words to be inserted and their insertion positions, an optional word selection unit that extracts words having the same attributes as the optional words received by the optional word input unit from the words registered in the word dictionary, a fixed sentence generation unit that inserts the words extracted by the optional word selection unit between the fixed phrases received by the fixed phrase input unit to generate a fixed sentence, and a fixed sentence output unit that outputs and stores the fixed sentence generated by the fixed sentence generation unit as a fixed phrase corpus in the speech database.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the field of caregiving, from the perspective of reducing the burden on caregivers, it is desired that a report document according to a predetermined format can be created based on voice input and recorded in a caregiving record database.

[0005] However, the fixed phrase corpus creation device described in Patent Document 1 does not assume creating a document based on voice input.

[0006] Furthermore, this fixed-phrase corpus creation device generates fixed-phrases by extracting words with the same attributes as the arbitrary word received by the arbitrary word input unit from the words registered in the word dictionary and inserting them between the fixed-phrases received by the fixed-phrase input unit. Therefore, it is not possible to create fixed-phrases using words not registered in the word dictionary.

[0007] This invention has been made in view of the above circumstances, and its purpose is to enable document creation based on unregistered words, even when unregistered words are input via voice, in a document creation technology that creates a document according to a predetermined format based on input voice. [Means for solving the problem]

[0008] To solve the above problems, the present invention pre-registers multiple templates containing labels to be replaced with words, each linked to category information, and for each of the multiple templates, at least one candidate word to replace the label included in that template is registered. When category information and audio information representing the report content are received, a word of a predetermined part of speech is extracted from the audio information, a template registered and linked to the received category information is selected, and the labels included in the selected template are replaced with words determined based on the words extracted from the audio information to create a document.

[0009] For example, if the extracted words from the audio information include an extracted word that matches a word registered as a replacement candidate for a label included in the selected template, a matching word replacement process is performed to replace this label with the extracted word. If the selected template contains an unreplaced label, and the extracted words from the audio information include an extracted word whose pronunciation matches or is similar to a word registered as a replacement candidate for this label, a pronunciation-matching / similar word replacement process is performed to replace this label with a replacement candidate whose pronunciation matches or is similar to the extracted word. If the extracted words from the audio information contain words that are semantically similar to words registered as replacement candidates for this label, a semantic similarity word replacement process is performed to replace this label with a replacement candidate that is semantically similar to this extracted word. If the selected template contains unreplaced labels, a semantically similar word to words registered as replacement candidates for this label is obtained from the trained model, and if the extracted words from the audio information contain words that have the same or similar pronunciation as the words obtained from the trained model, a semantic / pronunciation similarity word replacement process is performed to replace this label with the obtained word.

[0010] Of the matching word replacement process, pronunciation matching / similar word replacement process, semantic similar word replacement process, and semantic / pronunciation similar word replacement process, it is preferable that the matching word replacement process is performed first and the semantic / pronunciation similar word replacement process is performed last. The pronunciation matching / similar word replacement process and the semantic similar word replacement process may be performed in either order.

[0011] For example, the present invention is A document creation device that creates a document based on input speech, A template storage means in which multiple templates containing labels to be replaced with words are stored, each associated with category information, For each of the aforementioned multiple templates, a word dictionary means is provided in which at least one word is registered as a replacement candidate for each label included in the template. A means for receiving category information, A voice information receiving means that receives voice information representing the content of the report, A word extraction means for extracting words of a predetermined part of speech from the voice information received by the voice information receiving means, A template selection means that selects a template stored in the template storage means that is associated with the category information received by the category information receiving means, The document creation means includes a method for creating a document by replacing the labels included in the template selected by the template selection means with words determined based on the words extracted by the word extraction means, The document creation means is If a word that matches a word registered in the word dictionary means as a candidate for replacing a label included in the template selected by the template selection means is included in the words extracted by the word extraction means, then a matching word replacement process is performed to replace the label with that word. If the template selected by the template selection means contains an unreplaced label, and if a word whose pronunciation matches or is similar to a word registered in the word dictionary means as a replacement candidate for that label is included among the words extracted by the word extraction means, then a pronunciation-matching / similar word replacement process is performed to replace the label with the replacement candidate whose pronunciation matches or is similar to that word. If the template selected by the template selection means contains an unreplaced label, and if a word with a similar meaning to a word registered in the word dictionary means as a replacement candidate for that label is included among the words extracted by the word extraction means, then a semantic similarity word replacement process is performed to replace the label with the word with a similar meaning. If the template selected by the template selection means contains an unreplaced label, the system uses a trained model, which has been trained to acquire words with similar meanings to the words registered as replacement candidates for each label contained in each template stored in the template storage means, to acquire words with similar meanings to the words registered as replacement candidates for the unreplaced labels contained in the template selected by the template selection means. If a word whose pronunciation matches or is similar to the acquired word is included among the words extracted by the word extraction means, the system performs a semantic / pronunciation similar word replacement process to replace the label with the acquired word. [Effects of the Invention]

[0012] In this invention, if a word that matches a word pre-registered as a replacement candidate for a label included in a template registered in association with the received category information is found among the words extracted from the received audio information, this label is replaced with this extracted word. Furthermore, if the template contains an unreplaced label, if there is an extracted word among the words extracted from the received audio information whose pronunciation matches or is similar to a word pre-registered as a replacement candidate for this label, this label is replaced with this word. If there is a word among the words extracted from the received audio information whose meaning is similar to any of the words registered as replacement candidates for this label, this label is replaced with this word whose meaning is similar. In addition, for each unreplaced label included in the template, a word with a similar meaning is obtained from a trained model for each word registered as a replacement candidate for this label, and if there is a word among the words extracted from the received audio information whose pronunciation matches or is similar to any of the obtained words, this label is replaced with this word.

[0013] Therefore, according to the present invention, even when an unregistered word is input by voice, a document can be created based on this unregistered word according to a predetermined format. Further, according to the present invention, even when a word that is registered as a replacement candidate for a label and has a similar meaning is erroneously recognized by voice as a word with a different meaning, the document can be created by performing a phrase correction on the correct word.

Brief Description of the Drawings

[0014] [Figure 1] FIG. 1 is a schematic configuration diagram of a report document creation / recording system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a schematic functional configuration diagram of the report document creation / recording apparatus 1. [Figure 3] FIG. 3 is a diagram schematically showing an example of the registered content of the template storage unit 101. [Figure 4] FIG. 4 is a diagram schematically showing an example of the registered content of the word dictionary unit 102. [Figure 5] FIG. 5 is a diagram schematically showing an example of the registered content of the report document storage unit 103. [Figure 6] FIG. 6 is a diagram showing an example of a word acquired by the similar word acquisition unit 110. [Figure 7] FIG. 7 is a flowchart for explaining the operation of the report document creation / recording apparatus 1. [Figure 8] FIG. 8 is a flowchart for explaining the matching word replacement process S13 shown in FIG. 7. [Figure 9] FIG. 9 is a flowchart for explaining the title matching / similar word replacement process S15 shown in FIG. 7. [Figure 10] FIG. 10 is a flowchart for explaining the semantically similar word replacement process S17 shown in FIG. 7. [Figure 11] FIG. 11 is a flowchart for explaining the semantically similar word replacement process S17 shown in FIG. 7 and is a continuation of FIG. 10. [Figure 12] FIG. 12 is a flowchart for explaining the semantically and title-similar word replacement process S19 shown in FIG. 7. [Modes for carrying out the invention]

[0015] An embodiment of the present invention will be described below.

[0016] Figure 1 is a schematic diagram of the report creation and recording system according to this embodiment.

[0017] As shown in the figure, the report creation and recording system according to this embodiment is used in nursing care settings and the like, and is configured by connecting a report creation and recording device 1 and at least one communication terminal 2 to a network 3. In Figure 1, a wired network such as a LAN (Local Area Network) or WAN (Wide Area Network) is shown as an example of network 3, but a wireless network such as a wireless LAN or mobile phone network may be used instead of or in conjunction with a wired network.

[0018] The communication terminal 2 is equipped with voice input / output functionality and transmits category information received from the reporter, such as a caregiver, and voice information of the report content entered via voice input to the report creation / recording device 1 via the network 3. It also receives and displays the report containing the report text from the report creation / recording device 1. Furthermore, it accepts corrections to the report text from the reporter as needed and transmits these corrections to the report creation / recording device 1.

[0019] Communication terminal 2 may be a dedicated communication terminal such as an intercom, or a general-purpose communication terminal such as a mobile phone, smartphone, or tablet PC (Personal Computer) equipped with communication functions.

[0020] The report creation and recording device 1 creates and records a report in a predetermined format (template) based on category information and voice information representing the report content received from the communication terminal 2 via the network 3. It also transmits the created report to the communication terminal 2 and accepts corrections to the report from the communication terminal 2.

[0021] Figure 2 is a schematic diagram of the functional configuration of the report creation and recording device 1.

[0022] As shown in the figure, the report creation and recording device 1 includes a network interface unit 100, a template storage unit 101, a word dictionary unit 102, a report storage unit 103, a category information receiving unit 104, a voice information receiving unit 105, a word extraction unit 106, a template selection unit 107, a report creation unit 108, a similarity measurement unit 109, a similar word acquisition unit 110, a correction receiving unit 111, and a dictionary update unit 112.

[0023] The network interface unit 100 is an interface for connecting to network 3.

[0024] The template storage unit 101 stores templates that serve as the basis for report texts in a predetermined format, linked to category information and the verbs (in this case, in their base form) used within those texts.

[0025] Figure 3 is a schematic diagram showing an example of the registered contents of the template storage unit 101.

[0026] As shown in the figure, the template storage unit 101 stores a template record 1010 for each template. The template record 1010 has a field 1011 in which a template ID for identifying the template is registered, a field 1012 in which category information related to this template is registered, a field 1013 in which verbs used in the sentences contained in this template are registered, and a field 1014 in which this template is registered. Field 1014 may also contain the storage location of the file in which the template is described.

[0027] The template contains a sentence in which one or more labels (strings enclosed in <> in Figure 3) 1015, each associated with a word's attribute, are placed. The report is created by replacing the labels 1015 in this sentence with the words whose attributes are associated with them.

[0028] The word dictionary section 102 contains, for each template, a list of words that can be used as replacements for each of the labels 1015 included in that template.

[0029] Figure 4 is a schematic diagram showing an example of the registered contents of the word dictionary unit 102.

[0030] As shown in the diagram, the word dictionary unit 102 stores a word dictionary table 1020 associated with the template ID of each template stored in the template storage unit 101.

[0031] The word dictionary table 1020 contains replacement candidate records 1021 for each label 1015 included in the template identified by the template ID associated with this table 1020. Each replacement candidate record 1021 has a field 1022 in which the attributes of label 1015 are registered, and a field 1023 in which at least one word that is a replacement candidate for label 1015 is registered.

[0032] The report storage unit 103 stores the report created by the report creation unit 108.

[0033] Figure 5 is a schematic diagram showing an example of the registered contents in the report storage unit 103.

[0034] As shown in the figure, the report storage unit 103 stores a report record 1030 for each report. The report record 1030 has a field 1031 in which the date and time of the report is registered, a field 1032 in which the terminal ID, which is the identification information of the reporter's (e.g., caregiver's) communication terminal 2, is registered, and a field 1033 in which the report for the person the reporter is in charge of (e.g., the person receiving care) is registered.

[0035] The category information receiving unit 104 receives category information from the communication terminal 2 via the network interface unit 100.

[0036] The voice information receiving unit 105 receives voice information representing the report content from the communication terminal 2 via the network interface unit 100. It is also possible for the voice information receiving unit 105 to receive voice information of category information from the communication terminal 2 via the network interface unit 100. In this case, the category information receiving unit 104 may acquire the category information by voice recognition of the voice information of category information received by the voice information receiving unit 105.

[0037] The word extraction unit 106 extracts verbs and nouns from the audio information received by the audio information receiving unit 105 that are included in the report content represented by the audio information. Specifically, it performs speech recognition processing on the audio information to convert it into text data, and then performs morphological analysis processing on this text data to decompose it into multiple words. Then, it extracts verbs and nouns from among the multiple words obtained by decomposing this text data.

[0038] The template selection unit 107 selects a template stored in the template storage unit 101 that is associated with the category information received by the category information receiving unit 104 and the verbs extracted by the word extraction unit 106.

[0039] The report creation unit 108 works in conjunction with the similarity measurement unit 109 to create a report using the template selected by the template selection unit 107, the word dictionary table 1020 stored in the word dictionary unit 102 and linked to the template ID of this template, the nouns extracted by the word extraction unit 106, and the words obtained by the similar word acquisition unit 110, which will be described later.

[0040] The similarity measurement unit 109 measures the semantic similarity between words according to the instructions of the report creation unit 108. Specifically, it measures the semantic similarity (score value) between words input from the report creation unit 108 using a pre-trained model such as Word2Vec, which has been learned from general documents. Here, the pre-trained model does not necessarily have to be provided by the report creation and recording device 1 itself; it may also be provided by another AI (Artificial Intelligence) server connected to the network 3.

[0041] Furthermore, the similarity measurement unit 109 measures the similarity of pronunciations between words according to the instructions of the report creation unit 108. For example, it measures the Levenshtein distance between words input from the report creation unit 108 (the smaller the distance, the higher the similarity of pronunciations between words).

[0042] The similar word acquisition unit 110, following instructions from the report creation unit 108, uses a pre-trained model such as Word2Vec, learned from general documents, to acquire words whose meaning is more or less similar to the word specified by the report creation unit 108, along with a score value representing their similarity. Here, the pre-trained model does not necessarily have to be provided by the report creation / recording device 1 itself; it may also be provided by another AI server connected to the network 3. Furthermore, the pre-trained model may be shared with the pre-trained model used by the similarity measurement unit 109.

[0043] Figure 6 shows an example of a word obtained by the similar word acquisition unit 110.

[0044] This example shows the case where, for the word "breakfast" 1100 specified by the report creation unit 108, words 1101 with a semantic similarity score of "0.84" or higher are obtained along with their respective score values ​​1102.

[0045] The correction reception unit 111 receives correction instructions for the report created by the report creation unit 108 from the communication terminal 2, which is the source of the voice information used to create the report.

[0046] The dictionary update unit 112, following instructions from the report creation unit 108, registers the unregistered words (nouns) used in creating the report in the word dictionary table 1020 used to create the report.

[0047] The functional configuration of the report creation and recording device 1 shown in Figure 2 can be implemented in hardware using integrated logic ICs such as ASICs (Application Specific Integrated Circuits) and FPGAs (Field Programmable Gate Arrays), or in software using a computer such as a DSP (Digital Signal Processor). Alternatively, it can be implemented as a process in a computer system such as a PC, which includes a CPU, memory, auxiliary storage devices such as HDDs and DVD-ROMs, and communication interfaces such as modems, NICs, and wireless LAN adapters, by the CPU loading a predetermined program from the auxiliary storage device into memory and executing it.

[0048] Figure 7 is a flowchart illustrating the operation of the report creation and recording device 1.

[0049] This flow begins when the category information receiving unit 104 receives category information, and the voice information receiving unit 105 receives voice information representing the report content, both from the same communication terminal 2 via the network interface unit 100.

[0050] First, the category information receiving unit 104 outputs the category information to the template selection unit 107, linking it to the communication terminal 2 that transmitted it. Also, the voice information receiving unit 105 outputs the voice information to the word extraction unit 106, linking it to the communication terminal 2 that transmitted it.

[0051] Next, the word extraction unit 106 performs speech recognition processing on the speech information received by the speech information receiving unit 105 to convert the speech information into text data, and then performs morphological analysis processing on this text data to break it down into multiple words. Then, it extracts verbs and nouns from these words, and outputs the extracted verbs (hereinafter referred to as extracted verbs) to the template selection unit 107, linked to the communication terminal 2 associated with this speech information, and outputs the extracted nouns (hereinafter referred to as extracted nouns) to the report creation unit 108, linked to this communication terminal 2 (S10).

[0052] Next, the template selection unit 107 receives category information and extracted verbs associated with the same communication terminal 2 from the category information receiving unit 104 and the word extraction unit 106, respectively, and selects a template stored in the template storage unit 101 that is associated with this category information and extracted verbs. Then, it outputs the selected template (hereinafter referred to as the selected template) and its template ID to the report creation unit 108, associated with this communication terminal 2 (S11).

[0053] Next, the report creation unit 108 receives the extracted nouns, selected templates, and their template IDs, which are linked to the same communication terminal 2, from the word extraction unit 106 and the template selection unit 107, respectively, and selects the word dictionary table (hereinafter referred to as the selection dictionary table) 1020 stored in the word dictionary unit 102, which is linked to the template ID of the selected template (S12).

[0054] Subsequently, the report creation unit 108 performs the matching word replacement process S13 described later, replacing the label 1015 included in the selection template with an extracted noun registered in the selection dictionary table 1020 as a replacement candidate for this label.

[0055] As a result, if all labels 1015 included in the selection template have been replaced with extracted nouns (YES in S14), proceed to S21. If there are any unreplaced labels 1015 that have not been replaced with extracted nouns (hereinafter referred to as unreplaced labels) (NO in S14), perform the pronunciation matching / similar word replacement process S15 described below. Modify the extracted nouns registered in the selection dictionary table 1020 that have a pronunciation matching or similarity to the unreplaced label replacement candidate, and replace these unreplaced labels with the modified extracted nouns.

[0056] As a result, if all labels 1015 included in the selection template have been replaced with extracted nouns (YES in S16), the process proceeds to S21. If there are any unreplaced labels that have not been replaced with extracted nouns (NO in S16), the semantic similar word replacement process S17 described below is performed to replace the unreplaced labels with extracted nouns that are semantically similar to the replacement candidates for the unreplaced labels registered in the selection dictionary table 1020.

[0057] As a result, if all labels 1015 included in the selection template have been replaced with extracted nouns (YES in S18), the process proceeds to S21. If there are any unreplaced labels that have not been replaced with extracted nouns (NO in S18), the semantic / pronunciation-similar word replacement process S19 described below is performed to modify extracted nouns that have a similar pronunciation to or match the semantic candidate of the unreplaced label registered in the selection dictionary table 1020, and then replace these unreplaced labels with the modified extracted nouns.

[0058] Next, the report creation unit 108 determines whether the created report contains unreplaced labels that have not been replaced with extracted nouns (S20). If all labels 1015 included in the selection template have been replaced with extracted nouns (NO in S20), the process proceeds to S21. If unreplaced labels are included (YES in S20), the unit performs predetermined error processing, such as sending a message via the network interface unit 100 to the communication terminal 2 associated with the extracted noun and selection template, prompting the terminal to re-enter category information and voice information of the report content because the report creation failed (S27).

[0059] Based on the above, if all labels 1015 included in the selection template have been replaced and the operator (reporter) of the communication terminal 2 that is the source of the category information and voice information has completed creating the report (if the answer is YES in S14, YES in S16, YES in S18, or NO in S20), then in S21, the report creation unit 108 instructs the correction reception unit 111 to send the report to the communication terminal 2 that is the source of the category information and voice information. The correction reception unit 111 sends the report created by the report creation unit 108 to the communication terminal 2 that is the source of the category information and voice information via the network interface unit 100 for display and prompts the reporter to confirm the report. Then, it receives the reporter's recording instruction from this communication terminal 2.

[0060] Next, if the correction request received by the correction request unit 111 from the communication terminal 2 includes the correction details for the report (YES in S22), the report creation unit 108 corrects the report according to these correction details. Then, the report containing the corrected report is recorded in the report storage unit 103, linked to the date and time the recording request was received (date and time of reporting) and the terminal ID of the communication terminal 2 that sent the recording request (S23).

[0061] On the other hand, if the correction request received by the correction request unit 111 from the communication terminal 2 does not include any corrections to the report (NO in S22), the report creation unit 108 records the report, including the report, in the report storage unit 103, linking it to the date and time of the report and the terminal ID of the communication terminal 2 that sent the recording request (S24).

[0062] Then, in creating the report, if the extracted noun obtained by replacing the label 1015 included in the selection template contains a word not registered in the selection dictionary table 1020 (YES in S25), the dictionary update unit 112 adds this unregistered word to the selection dictionary table 1020 as a replacement candidate for this label and updates the word dictionary unit 102 (S26).

[0063] Figure 8 is a flowchart illustrating the matching word replacement process S13 shown in Figure 7.

[0064] First, the report creation unit 108 selects an unselected label 1015 from among the labels 1015 included in the selection template (S130), and identifies replacement candidates associated with this label (hereinafter referred to as the selected label in the matching word replacement process S13) 1015 from the selection dictionary table 1020 (S131).

[0065] Next, the report generation unit 108 passes all the identified replacement candidates and all the extracted nouns that have not been flagged as replaced (as described later) to the similarity measurement unit 109 and instructs it to measure the similarity between them. In response, the similarity measurement unit 109 measures the similarity between each replacement candidate and each extracted noun, and passes the measurement results to the report generation unit 108 (S132).

[0066] Then, the report creation unit 108 refers to the measurement results received from the similarity measurement unit 109 and determines whether or not there is an extracted noun that matches any of the replacement candidates (S133). If there is an extracted noun that matches any of the replacement candidates (YES in S133), it replaces the selection label 1015 in the selection template with this extracted noun (S134) and assigns a replaced flag to this extracted noun (S135).

[0067] Next, if there is an unselected label 1015 in the selection template (YES in S136), the report creation unit 108 returns to S130 and continues processing. On the other hand, if there is no unselected label 1015 in the selection template (NO in S136), this flow is terminated and the process proceeds to S14 in Figure 7.

[0068] Figure 9 is a flowchart illustrating the name matching and similar word replacement process S15 shown in Figure 7.

[0069] First, the report creation unit 108 selects an unselected label 1015 from among the unreplaced labels 1015 remaining in the selection template after the matching word replacement process S13 has been performed (S150), and identifies the replacement candidate associated with this label (hereinafter referred to as the selected label in the pronunciation matching / similar word replacement process S15) 1015 from the selection dictionary table 1020 (S151).

[0070] Next, the report generation unit 108 passes all the identified replacement candidates and all extracted nouns that have not been flagged as replaced to the similarity measurement unit 109 and requests that it measure the similarity of their pronunciations. In response, the similarity measurement unit 109 measures the similarity of each replacement candidate's pronunciation with each extracted noun and passes the measurement results back to the report generation unit 108 (S152).

[0071] Next, the report generation unit 108 refers to the measurement results received from the similarity measurement unit 109 and determines whether or not there is an extracted noun whose pronunciation matches any of the replacement candidates (S153). If an extracted noun whose pronunciation matches any of the replacement candidates exists (YES in S153), this extracted noun is modified to a replacement candidate whose pronunciation matches this extracted noun (S154). On the other hand, if no extracted noun whose pronunciation matches any of the replacement candidates does not exist (NO in S153), the unit further determines whether or not there is an extracted noun whose pronunciation similarity to any of the replacement candidates is higher than a predetermined value (for example, an extracted noun whose Levenshtein distance is less than or equal to a predetermined value) (S155). If an extracted noun whose pronunciation similarity to any of the replacement candidates is higher than a predetermined value exists (YES in S155), this extracted noun is modified to a replacement candidate whose pronunciation similarity to this extracted noun is higher than a predetermined value (S156).

[0072] Then, the report creation unit 108 replaces the selection label 1015 in the selection template with the corrected extracted noun (S157), and assigns a replaced flag to this corrected extracted noun (S158).

[0073] Subsequently, if there is an unselected label 1015 in the selection template (YES in S159), the report creation unit 108 returns to S150 and continues processing. On the other hand, if there is no unselected label 1015 in the selection template (NO in S159), this flow is terminated and the process proceeds to S16 in Figure 7.

[0074] Figures 10 and 11 are flowcharts illustrating the semantic similar word replacement process S17 shown in Figure 7.

[0075] First, the report generation unit 108 compares the number of extracted nouns that have not been flagged as replaced with the number of unreplaced labels 1015 that have not been replaced by the extracted nouns (S170). If the number of extracted nouns that have not been flagged as replaced is smaller than the number of unreplaced labels 1015 (NO in S170), the process proceeds to S18 in Figure 7.

[0076] On the other hand, if the number of extracted nouns that have not been flagged as replaced is greater than or equal to the number of unreplaced labels 1015 (YES in S170), the report generation unit 108 identifies the replacement candidate words associated with each unreplaced label 1015 from the selection dictionary table 1020 (S171).

[0077] Next, the report creation unit 108 creates replacement patterns for all combinations of unreplaced labels 1015 and extracted nouns that have not been flagged as replaced, assigning any extracted noun that has not been flagged as replaced to each of the unreplaced labels 1015 in the selected template, ensuring that duplicate extracted nouns are not used between labels (S172).

[0078] Next, the report creation unit 108 selects an unselected replacement pattern from the created replacement patterns (S173). Then, for each unreplaced label 1015 included in the selected replacement pattern, it passes the extracted noun assigned to this label 1015 and each replacement candidate for this label 1015 to the similarity measurement unit 109 and instructs the similarity measurement unit 109 to measure the similarity between the two. In response, the similarity measurement unit 109 measures the semantic similarity between each unreplaced label 1015, the extracted noun assigned to this label 1015, and each replacement candidate for this label 1015, and passes the measurement results to the report creation unit 108 (S174).

[0079] Next, the report generation unit 108 refers to the measurement results received from the similarity measurement unit 109 and determines whether there is a label 1015 in the selected replacement pattern in which the maximum similarity between the extracted noun assigned to this label 1015 and each of the replacement candidates for this label 1015 is less than or equal to a predetermined value (S175). If such an unreplaced label 1015 exists (YES in S175), the similarity of the selected replacement pattern is set to "zero" (S176). On the other hand, if such an unreplaced label 1015 does not exist (NO in S175), the maximum similarity between the extracted noun assigned to this label 1015 and each of the replacement candidates for this label 1015 is identified for each unreplaced label 1015 in the selected replacement pattern, the sum of these values ​​is calculated, and the similarity of the selected replacement pattern is set to this sum (S177).

[0080] Next, if the report creation unit 108 finds that there are any unselected replacement patterns among the created replacement patterns (NO in S178), it returns to S173 and continues processing.

[0081] On the other hand, if all created substitution patterns have been selected (YES in S178), it is determined whether there are any substitution patterns among these that have a similarity equal to or greater than a predetermined threshold (S179). If there are no substitution patterns with a similarity equal to or greater than a predetermined threshold (NO in S179), it is determined that there are no extracted nouns among the extracted nouns that have not been flagged as replaced that can be replaced with the unsubstituted label 1015, and the process proceeds to S18 in Figure 7.

[0082] On the other hand, if there are substitution patterns with a similarity equal to or greater than a predetermined threshold (YES in S179), the report creation unit 108 selects the substitution pattern with the greatest similarity from among such substitution patterns as the adopted pattern (S180). Then, according to this adopted pattern, each of the unsubstituted labels 1015 included in the selection template is replaced with an extracted noun (a word with a similar meaning to the replacement candidate for this label 1015) assigned to that label 1015 (S181), and a replaced flag is assigned to this extracted noun (S182). After that, the process proceeds to S18 in Figure 7.

[0083] Figure 12 is a flowchart illustrating the semantic / pronunciation-similar word replacement process S19 shown in Figure 7.

[0084] First, the report creation unit 108 selects an unselected label 1015 from among the unreplaced labels 1015 remaining in the selection template after the semantic similar word replacement process S17 has been performed (S190), and identifies the replacement candidate associated with this label (hereinafter referred to as the selected label in the semantic / pronunciation similar word replacement process S19) 1015 from the selection dictionary table 1020 (S191).

[0085] Next, the report creation unit 108 instructs the similar word acquisition unit 110 to acquire words that are similar in meaning to the replacement candidates for the selected label (hereinafter referred to as similar words in the semantic / pronunciation similar word replacement process S19) by specifying the replacement candidates for the selected label. In response, as shown in Figure 6, the similar word acquisition unit 110 uses a trained model learned from general documents to acquire words whose similarity to the replacement candidates for the selected label specified by the report creation unit 108 is equal to or greater than a predetermined score value, along with their similarity score value (S192). The acquired similar words, along with their similarity score value, are then passed to the report creation unit 108.

[0086] Next, if the report generation unit 108 receives similar words from the similar word acquisition unit 110 and there are multiple words similar to replacement candidates (YES in S193), it sums the similarity scores of these similar words and these replacement candidates, and changes the sum to the similarity score of the similar word (S194). Then, the report generation unit 108 sorts the similar words received from the similar word acquisition unit 110 in descending order of similarity score (S195).

[0087] Next, the report generation unit 108 selects one unselected similar word from the similar word acquisition unit 110, in descending order of similarity score (S196). Then, it passes the selected similar word and all extracted nouns that have not been flagged as replaced to the similarity measurement unit 109 and instructs it to measure the similarity of their pronunciations. In response, the similarity measurement unit 109 measures the similarity of the pronunciations between the selected similar word and each extracted noun, and passes the measurement results to the report generation unit 108 (S197).

[0088] Next, the report creation unit 108 refers to the measurement results received from the similarity measurement unit 109 and determines whether there are any extracted nouns that have a pronunciation similarity of a predetermined level or higher to the selected similar words (S198). If none of the extracted nouns have a pronunciation similarity of a predetermined level or higher to the selected similar words (NO in S198), and there are unselected similar words among the similar words received from the similar word acquisition unit 110 (YES in S202), the process returns to S196 and continues. If there are no unselected similar words (NO in S202), the process proceeds to S203.

[0089] On the other hand, if there is an extracted noun that has a similarity in pronunciation to the selected similar word above a predetermined level (YES in S198), the report generation unit 108 corrects this extracted noun to the selected similar word (S199). Then, the report generation unit 108 replaces the selection label 1015 in the selection template with the corrected extracted noun (S200) and assigns a replaced flag to this corrected extracted noun (S201). After that, the process proceeds to S203.

[0090] In S203, if there is an unselected label 1015 in the selected template (YES in S203), the report creation unit 108 returns to S190 and continues processing. On the other hand, if there is no unselected label 1015 in the selected template (NO in S203), this flow is terminated and the process proceeds to S20 in Figure 7.

[0091] For example, suppose the selection template is "I ate <food> in <amount> at <location>" and includes labels 1015 for the attributes <location>, <food>, and <amount>, and the replacement candidates for label 1015 of attribute <location> registered in the selection dictionary table 1020 are "private room", "room", and "dining hall", the replacement candidates for label 1015 of attribute <food> registered in the selection dictionary table 1020 are "breakfast", "lunch", "dinner", and "snack", and the replacement candidates for label 1015 of attribute <amount> registered in the selection dictionary table 1020 are "full amount", "half amount", and "small amount".

[0092] Here, let's assume that, for example, the extracted nouns "private room," "breakfast," and "entire amount" are extracted from the audio information. In this case, the replacement candidate "private room" for attribute <place> label 1015 matches the extracted noun "private room," the replacement candidate "breakfast" for attribute <food> label 1015 matches the extracted noun "breakfast," and the replacement candidate "entire amount" for attribute <amount> label 1015 matches the extracted noun "entire amount." Therefore, the matching word replacement process S13 shown in Figure 8 replaces the labels 1015 of attributes <place>, <food>, and <amount> included in the selection template with the extracted nouns "private room," "breakfast," and "entire amount," respectively. This creates the report sentence "I ate the entire amount of breakfast in a private room."

[0093] Furthermore, suppose the extracted nouns "cafeteria," "championship," and "good-natured" are extracted from the audio information. In this case, the replacement candidate "cafeteria" for the attribute <place> label 1015 matches the extracted noun "cafeteria." Therefore, the matching word replacement process S13 shown in Figure 8 replaces the attribute <place> label 1015 included in the selected template with the extracted noun "cafeteria." Also, none of the replacement candidates for the attribute <food> label 1015 match the extracted nouns "championship" or "good-natured," but the replacement candidate "dinner" differs from the extracted noun "championship" by only one character in its pronunciation, and the two are similar in pronunciation. Therefore, the pronunciation matching / similar word replacement process S15 shown in Figure 9 corrects the extracted noun "championship" to "dinner," and the unreplaced label 1015 for the attribute <food> included in the selected template is replaced with the corrected extracted noun "dinner." Furthermore, none of the replacement candidates for the attribute <quantity> label 1015 match the extracted noun "good-natured," but the replacement candidate "total amount" phonetically matches the extracted noun "good-natured." Therefore, the phonetic matching / similar word replacement process S15 shown in Figure 9 corrects the extracted noun "good-natured" to "total amount," and the unreplaced label 1015 of the attribute <quantity> included in the selected template is replaced with the corrected extracted noun "total amount." As a result, the report sentence "I ate the entire amount of dinner at the cafeteria." is created.

[0094] Furthermore, suppose the extracted nouns "dining," "lunch," and "total amount" are extracted from the audio information. In this case, the replacement candidate "total amount" for the attribute <quantity> label 1015 matches the extracted noun "total amount." Therefore, the matching word replacement process S13 shown in Figure 8 replaces the attribute <quantity> label 1015 included in the selected template with the extracted noun "total amount." Also, the replacement candidates for the attributes <place> and <food> label 1015 do not match the extracted nouns "dining" and "lunch," and their pronunciations do not match or are similar. Therefore, the semantic similarity word replacement process S17 shown in Figures 10 and 11 is performed to create replacement patterns in which the extracted nouns "dining" and "lunch" are assigned to the unreplaced labels 1015 of the attributes <place> and <food> in the selected template, respectively, and replacement patterns in which the extracted nouns "lunch" and "dining" are assigned to respectively, and the semantic similarity is calculated for each replacement pattern. Here, the extracted noun "dining" has a high semantic similarity to the replacement candidate "cafeteria" for label 1015 of attribute <place>, and the extracted noun "lunch" has a high semantic similarity to the replacement candidate "lunchtime" for label 1015 of attribute <food>. Therefore, one of the replacement patterns that assigns the extracted nouns "dining" and "lunch" to labels 1015 of attribute <place> and <food> respectively is determined to be the adopted pattern, and the unreplaced labels 1015 of attribute <place> and <food> included in the selected template are replaced with the extracted nouns "dining" and "lunch," respectively. As a result, the report sentence "I ate the entire lunch at the dining room." is created.

[0095] Furthermore, suppose the extracted nouns "cafeteria," "dice," and "total amount" are extracted from the audio information. In this case, the replacement candidate "cafeteria" for label 1015 of attribute <place> matches the extracted noun "cafeteria," and the replacement candidate "total amount" for label 1015 of attribute <amount> matches the extracted noun "total amount." Therefore, the matching word replacement process S13 shown in Figure 8 replaces the labels 1015 of attributes <place> and <amount> included in the selection template with the extracted nouns "cafeteria" and "total amount," respectively.

[0096] Furthermore, the replacement candidates for label 1015 of attribute <food> do not match the extracted noun "dice," nor do their pronunciations match or are similar. Moreover, their meanings are not similar. For this reason, the semantic and pronunciation similar word replacement process S19 shown in Figure 12 is performed, and similar words whose meanings are more or less similar to each of the replacement candidates for label 1015 of attribute <food>—"breakfast," "lunch," "dinner," and "snack"—are obtained along with their similarity scores.

[0097] Here, for example, the similar words obtained for the replacement candidate "breakfast" are "morning meal" (similarity score: 0.92), "morning" (similarity score: 0.90), "rice" (similarity score: 0.88), "rice" (similarity score: 0.86), and "bread" (similarity score: 0.84). The similar words obtained for the replacement candidate "lunch" are "lunch" (similarity score: 0.92), "lunch" (similarity score: 0.90), "rice" (similarity score: 0.88), "rice" (similarity score: 0.86), and "bread" (similarity score: 0.84). Assume that 0.84) is obtained, and the following similar words for the replacement candidate "dinner" are obtained: "evening meal" (similarity score: 0.92), "supper" (similarity score: 0.91), "dinner" (similarity score: 0.90), "rice" (similarity score: 0.88), "rice" (similarity score: 0.86), and "bread" (similarity score: 0.84). And, the following similar words for the replacement candidate "snack" are obtained: "sweets" (similarity score: 0.95), "ice cream" (similarity score: 0.92), and "tea" (similarity score: 0.84).

[0098] The similar words "gohan" (rice), "rice" (bread), and "pan" (bread) are similar to the three replacement candidates "breakfast," "lunch," and "dinner." Therefore, their similarity scores are changed to the sum of their similarity scores with the three replacement candidates "breakfast," "lunch," and "dinner," which are "2.64," "2.58," and "2.52," respectively. Then, sorting the retrieved similar words in descending order of their similarity score, the order is as follows: "gohan" (similarity score: 2.64), "rice" (similarity score: 2.58), "pan" (similarity score: 2.52), "kashi" (similarity score: 0.95), "asagohan" (similarity score: 0.92), "hirugohan" (similarity score: 0.92), "yugohan" (similarity score: 0.92), "aisui" (similarity score: 0.92), "supgohan" (similarity score: 0.91), "moring" (similarity score: 0.90), "lunch" (similarity score: 0.90), "dinner" (similarity score: 0.90), and "ocha" (similarity score: 0.84).

[0099] When the similarity of the extracted noun "dice" is calculated in descending order of similarity score, the similar word "gohan" (rice) has a low similarity of pronunciation (its similarity of pronunciation is below a predetermined value), while the similar word "rice" has a high similarity of pronunciation (its similarity of pronunciation is above a predetermined value). Therefore, the semantic / pronunciation similar word replacement process S19 shown in Figure 12 corrects the extracted noun "dice" to "rice," and the unreplaced label 1015 of the attribute <food> included in the selected template is replaced with the corrected extracted noun "rice." As a result, the report sentence "I ate all of the rice at the cafeteria." is created.

[0100] One embodiment of the present invention has been described above.

[0101] In this embodiment, the report creation and recording device 1 replaces a label with a matching extracted noun if such a noun exists among the extracted nouns extracted from the audio information. On the other hand, if no such extracted noun exists, the device replaces the label with an extracted noun that has a similar meaning to any of the replacement candidates for this label, if such an extracted noun exists among the extracted nouns extracted from the audio information. Therefore, according to this embodiment, even if an unregistered word is input via voice, a report can be created in a predetermined format based on this unregistered word.

[0102] Furthermore, in this embodiment, if the report creation and recording device 1 contains one or more unreplaced labels in the selection template for which there are no extracted nouns that match the replacement candidates, it selects a combination of an unreplaced label and an abstract noun that maximizes the sum of the maximum semantic similarity between the words registered as replacement candidates for the unreplaced labels and the extracted nouns (similarity of replacement patterns), and replaces the unreplaced labels with the extracted nouns according to this combination. Therefore, according to this embodiment, even if the audio information contains one or more words that are not registered as replacement candidates, a more accurate report can be created based on the unregistered words.

[0103] Furthermore, in this embodiment, the report creation and recording device 1 modifies extracted nouns whose pronunciation matches or is similar to a predetermined number of words with similar meanings to replacement candidates for labels in the selected template, or words with similar meanings to replacement candidates, by changing them to replacement candidates whose pronunciation matches or is similar to a predetermined number of words with similar meanings to replacement candidates. Therefore, according to this embodiment, even if the extracted nouns cannot be correctly extracted from the audio information due to misrecognition in the speech recognition process, a more accurate report can be created.

[0104] Furthermore, in this embodiment, the report creation and recording device 1 stores one or more templates linked to category information and verbs used in the text. It receives category information and audio information of the report content from the communication terminal 2 and selects the template linked to the extracted verbs in this category information and audio information as the selected template. Therefore, according to this embodiment, it is possible to appropriately select a template to be used as the basis for the report from among the one or more stored templates.

[0105] Furthermore, in this embodiment, if the label of the selected template is converted to an extracted noun that does not match any of the replacement candidates for this label, the report creation and recording device 1 adds this extracted noun to the list of replacement candidates for the selected template label. Therefore, according to this embodiment, whenever a word not yet registered as a replacement candidate is input via voice, this word is automatically added to the word dictionary as a replacement candidate, thus improving convenience.

[0106] It should be noted that the present invention is not limited to the embodiments described above, and numerous modifications are possible within the scope of its essence.

[0107] For example, in the above embodiment, if there are unreplaced labels on the selection plate after the pronunciation matching and similar word replacement process S15 is performed, the semantic similar word replacement process S17 is performed (see Figure 7). However, the present invention is not limited thereto. The pronunciation matching and similar word replacement process S15 may be performed if there are unreplaced labels on the selection plate after the semantic similar word replacement process S17 is performed. Even in this case, it is preferable to perform the matching word replacement process S13 first and the semantic and pronunciation similar word replacement process S19 last.

[0108] Furthermore, in the above embodiment, the report creation and recording device 1 stores templates linked to category information and verbs used in the text, receives category information and audio information of the report content from the communication terminal 2, extracts verbs from the audio information, and determines the template linked to the extracted verbs in the category information and audio information as the selected template. For this reason, if verbs cannot be extracted from the audio information by speech recognition processing or morphological analysis processing, or if the audio information does not contain verbs at all, it may not be possible to determine a single selected template using only category information, and multiple selected templates may be determined. In such cases, a report may be created for each selected template, as shown below, and the most suitable report may be selected from among the created report.

[0109] In other words, if the report creation and recording device 1 fails to extract verbs from the audio information, it selects one of several templates associated with the category information as the template and creates a report for each template according to the procedure described above. For each created report, it measures the similarity between the extracted noun (for which the label has been replaced) and the replacement candidate for that label for each label included in the base template of the report (if the extracted noun and the replacement candidate match, the similarity between the two is taken as the highest value), and calculates the sum of the maximum similarity values ​​measured for each label. Then, it sends the report with the highest total similarity value to the communication terminal 2 as the optimal report and displays it, allowing the reporter to confirm this report.

[0110] Furthermore, in the above embodiment, the report creation and recording device 1 stores templates linked to category information and verbs used in the text, receives category information and audio information of the report content from the communication terminal 2, extracts verbs from the audio information, and determines the template linked to this category information and the extracted verbs from the audio information as the selected template. However, the present invention is not limited to this. Templates may be stored linked to category information, category information may be received from the communication terminal 2, and the template linked to this category information may be determined as the selected template. In this case, the category information may be subdivided so that one template is linked to one piece of category information.

[0111] Furthermore, in the above embodiment, the report creation and recording device 1 extracts nouns from the voice information received from the communication terminal 2 and replaces each label included in the selection template with the extracted noun from the voice information. However, the present invention is not limited to this. The present invention only needs to extract words of a predetermined part of speech from voice information and replace each label included in the selection template with the extracted word from the voice information. For example, if adjectives are also extracted in addition to nouns from the voice information, and there is an extracted word (here, an extracted noun or extracted adjective) that matches any of the label replacement candidates, this label is replaced with this matching extracted word. On the other hand, if there is no such extracted word, and there is an extracted word that is similar in meaning to any of the label replacement candidates, this label is replaced with this extracted word. Also, if there is an extracted word whose pronunciation matches or is similar to a predetermined amount or more to a label replacement candidate or a word that is similar in meaning to a replacement candidate, this label is replaced with the replacement candidate or a word that is similar in meaning to a replacement candidate. [Explanation of symbols]

[0112] 1: Report creation and recording device 2: Communication terminal 3: Network 100: Network interface unit 101: Template storage unit 102: Vocabulary Dictionary Section 103: Report Sentence Memory Section 104: Category Information Reception Department 105: Voice Information Reception Department 106: Word extraction unit 107: Template selection unit 108: Report Writing Department 109: Similarity Measurement Department 110: Similar word acquisition unit 111: Correction acceptance unit 112: Dictionary update unit

Claims

1. A document creation device that creates a document based on input speech, A template storage means in which multiple templates containing labels to be replaced with words are stored, each associated with category information, For each of the aforementioned multiple templates, a word dictionary means is provided in which at least one word is registered as a replacement candidate for each label included in the template. A means for receiving category information, A voice information receiving means that receives voice information representing the content of the report, A word extraction means for extracting words of a predetermined part of speech from the voice information received by the voice information receiving means, A template selection means that selects a template stored in the template storage means that is associated with the category information received by the category information receiving means, The document creation means includes a method for creating a document by replacing the labels included in the template selected by the template selection means with words determined based on the words extracted by the word extraction means, The document creation means is If a word that matches a word registered in the word dictionary means as a replacement candidate for a label included in the template selected by the template selection means is included in the words extracted by the word extraction means, then a matching word replacement process is performed to replace the label with that word. If the template selected by the template selection means contains an unreplaced label, and if a word whose pronunciation matches or is similar to a word registered in the word dictionary means as a replacement candidate for that label is included among the words extracted by the word extraction means, then a pronunciation-matching / similar word replacement process is performed to replace the label with the replacement candidate whose pronunciation matches or is similar to that word. If the template selected by the template selection means contains an unreplaced label, and if a word with a similar meaning to a word registered in the word dictionary means as a replacement candidate for that label is included among the words extracted by the word extraction means, then a semantic similarity word replacement process is performed to replace the label with the replacement candidate with a similar meaning to that word. If the template selected by the template selection means contains an unreplaced label, the template storage means uses a trained model that has been trained to acquire words with similar meanings to the words registered as replacement candidates for the unreplaced labels in each template stored in the template storage means. The trained model then acquires words with similar meanings to the words registered as replacement candidates for the unreplaced labels in the template selected by the template selection means. If a word whose pronunciation matches or is similar to the acquired word is included among the words extracted by the word extraction means, the label is replaced with the acquired word in a semantic / pronunciation similarity word replacement process. A document creation device characterized by the following features.

2. A document creation device according to claim 1, The document creation means is In the aforementioned semantic and pronunciation-similar word replacement process, if there are multiple words obtained from the trained model, the words obtained from the trained model are selected in order of decreasing semantic similarity to the replacement candidate for the label. The system then determines whether a word with a matching or similar pronunciation is included among the words extracted by the word extraction means. If such a word is included, the label is replaced with the word obtained from the trained model. A document creation device characterized by the following features.

3. A document creation apparatus according to claim 2, The document creation means is In the aforementioned semantic and pronunciation-similar word replacement process, if any of the multiple words obtained from the trained model have a semantic similarity to any of the multiple words registered as replacement candidates for the label, the semantic similarity of that word is changed to the sum of the semantic similarities of each of the multiple words registered as replacement candidates for the label. A document creation device characterized by the following features.

4. A document creation apparatus according to any one of claims 1 to 3, The document creation means further includes a dictionary update means that, if the label of the template selected by the template selection means is replaced by a word other than a word registered in the word dictionary means as a replacement candidate for the label, registers the replaced word in association with the label in the word dictionary means. A document creation device characterized by the following features.

5. A program that causes a computer to function as a document creation device that creates documents based on input voice, A template storage means in which multiple templates containing labels to be replaced with words are stored, each associated with category information. For each of the aforementioned multiple templates, a word dictionary means is provided in which at least one word is registered as a replacement candidate for each label included in the template. Category information receiving means for receiving the aforementioned category information, A voice information receiving means that receives voice information representing the content of the report. A word extraction means for extracting words of a predetermined part of speech from the voice information received by the voice information receiving means. A template selection means for selecting a template stored in the template storage means that is associated with the category information received by the category information receiving means, and The computer functions as a document creation means, which replaces the labels included in the template selected by the template selection means with words determined based on the words extracted by the word extraction means to create the document. The document creation means is If a word that matches a word registered in the word dictionary means as a replacement candidate for a label included in the template selected by the template selection means is included in the words extracted by the word extraction means, then a matching word replacement process is performed to replace the label with that word. If the template selected by the template selection means contains an unreplaced label, and if a word whose pronunciation matches or is similar to a word registered in the word dictionary means as a replacement candidate for that label is included among the words extracted by the word extraction means, then a pronunciation-matching / similar word replacement process is performed to replace the label with the replacement candidate whose pronunciation matches or is similar to that word. If the template selected by the template selection means contains an unreplaced label, and if a word with a similar meaning to a word registered in the word dictionary means as a replacement candidate for that label is included among the words extracted by the word extraction means, then a semantic similarity word replacement process is performed to replace the label with the replacement candidate with a similar meaning to that word. If the template selected by the template selection means contains an unreplaced label, the template storage means uses a trained model that has been trained to acquire words with similar meanings to the words registered as replacement candidates for the unreplaced labels in each template stored in the template storage means. The trained model then acquires words with similar meanings to the words registered as replacement candidates for the unreplaced labels in the template selected by the template selection means. If a word whose pronunciation matches or is similar to the acquired word is included among the words extracted by the word extraction means, the label is replaced with the acquired word in a semantic / pronunciation similarity word replacement process. A program characterized by the following features.

6. A document creation method in which a document creation device creates a document based on input speech, The document creation device is The system receives audio information representing the report content, extracts words of a specified part of speech from the audio information, The system accepts category information and selects a template that includes labels in the text that are pre-stored and associated with that category information, which will be replaced with words. If a word that matches a word registered as a replacement candidate for a label included in the selected template exists among the words extracted from the audio information, then the label is replaced with that word. If the selected template contains an unreplaced label, and if there is a word among the words extracted from the audio information that matches or is similar in pronunciation to a word registered as a replacement candidate for that label, the label is replaced with the replacement candidate that matches or is similar in pronunciation to that word. Furthermore, if there is a word among the words extracted from the audio information that is similar in meaning to a word registered as a replacement candidate for that label, the label is replaced with the replacement candidate that is similar in meaning to that word. If the selected template contains an unreplaced label, a pre-trained model, which has been trained to learn words similar in meaning to the replacement candidate words for each label contained in each of the pre-stored templates, is used to obtain words similar in meaning to the words registered as replacement candidates for the unreplaced labels in the selected template. If a word whose pronunciation matches or is similar to the obtained word is included among the words extracted from the speech information, the label is replaced with the obtained word. A document creation method characterized by the following:

Images