Term detection device and program

Abstract

This function detects whether or not terms requiring special usage are permissible within the context of the text. [Solution] The pattern matching means 42 of the caution term detection means 21 provided in the teleprompter production support device 2 performs pattern matching between the text and each of the multiple caution terms registered in the caution term information 32 to detect caution terms. The generation AI model utilization means 43, if the caution term is a context-dependent term, uses the generation AI information 34 to estimate whether the caution term can be used in the context of the text, estimates an alternative if it cannot be used, and generates usability information. The detection result generation means 44 generates location information of the caution term, reads related information from the caution term information 32 for the caution term, and generates a detection result including the text, usability information, location information, related information, etc. This detection result is transmitted to the user terminal 3 and displayed on the screen.

Description

Technical Field

[0001] The present invention relates to a term detection device and a program that detect terms that require attention when used from input text.

Background Art

[0002] Conventionally, there is known a telop display technique that superimposes text related to broadcast content on an image of a television receiving a broadcast or displays text such as emergency news.

[0003] The telop display technique is not only used in the television broadcasting industry, but also includes those that display opinions arbitrarily input by users, such as on video sites on the Internet, as telops on the screen, and is widely used as a screen display technique.

[0004] As an example of such a telop display technique, a telop generation method that does not require a dedicated editing technique has been proposed for application to diversified video output techniques (see, for example, Patent Document 1).

[0005] This method generates a telop using various shared information such as information owned in advance in the television broadcasting industry, information that can be provided by public institutions such as the government and local governments, and information publicly disclosed by general companies. Specifically, this method extracts information necessary for creating a telop from shared information based on a telop creation request such as information regarding the distribution destination of the telop to be generated and information regarding the content of the telop display, and generates a telop using the extracted information.

[0006] However, there is a problem that the telop generated by such a method may include terms that require attention when used, and it is difficult to determine whether the generated telop can be used as it is.

[0007] To solve this problem, a method is envisioned that uses a database containing registered terms requiring attention to detect such terms from the generated on-screen text.

[0008] For example, although not specifically targeting on-screen text, a method has been proposed to determine whether or not text displayed on an electronic bulletin board contains terms that require attention (see, for example, Patent Document 2).

[0009] This method determines whether a post on an electronic bulletin board contains any terms requiring attention, and if so, sends the user web page data containing a list of the post and the URL of the electronic bulletin board, highlighting those terms.

[0010] On the other hand, methods have been proposed to provide easily readable text overlays for users who are presented with text and attempt to understand its content (see, for example, Patent Document 3).

[0011] This method replaces kanji characters in text with kana characters for elementary and junior high school students, according to their grade level.

[0012] Specifically, this method obtains the user's grade level, identifies kanji characters that the user cannot read from the text based on the obtained grade level and the assigned grade level table (a table defining kanji characters and the grade level in which those kanji characters are learned), and replaces the identified kanji characters with kana based on ruby ​​character data, which are then displayed as on-screen text. [Prior art documents] [Patent Documents]

[0013] [Patent Document 1] Japanese Patent Publication No. 2018-196076 [Patent Document 2] Patent No. 4937614 [Patent Document 3] Japanese Patent Publication No. 2022-101011 [Overview of the Initiative] [Problems that the invention aims to solve]

[0014] The method for detecting terms requiring attention using the aforementioned database involves comparing the terms registered in the database with the text of the generated caption to determine whether or not the text contains terms requiring attention.

[0015] However, this method had the problem of requiring a large database size because it necessitated registering various variations of terms that required attention in the database. Furthermore, comparing the terms registered in the database with the text resulted in a high processing load and a time-consuming process.

[0016] To solve this problem, a term detection device has been proposed, described in Japanese Patent Application No. 2024-001920, which was filed by the same applicant as the present patent application but was not published at the time of the present patent application.

[0017] This term detection device, when detecting terms requiring attention from text (hereinafter referred to as "terms requiring attention"), not only detects terms at the word level but also performs morphological analysis of the text and replaces the conjugated forms of verbs contained in the text with their base forms, thereby detecting terms at the verb level. Since the conjugated forms of verbs contained in the text are replaced with their base forms and terms requiring attention are detected for the base forms of verbs, it is not necessary to register all conjugated forms of verbs in the database. As a result, it is sufficient to register only the base forms of verbs in the database, which reduces the database capacity and the processing load of term detection.

[0018] However, while this term detection device can detect terms that require caution when used, it cannot detect terms that are unsuitable in the context of the text (terms that pose a contextual risk (require caution) when used).

[0019] For example, when the term "Mecca" is included in the text, "Mecca" in the context of using this term as a place name is a term for usage caution, but it can be used in the context (there is no context risk when using it).

[0020] On the other hand, "Mecca of XX" in the context of use in religions other than Islam or in a figurative context such as "Mecca of XX" is a term for usage caution and cannot be used in the context (there is a context risk when using it), and its use should be refrained from. Also, for "flung" and "peculiar", they are similarly terms for usage caution and cannot be used in the context, and their use should be refrained from.

[0021] More specifically, when the text is "The road for pilgrims to access Mecca has been improved", the term "Mecca" is used as a place name, so it is a term for usage caution, but it can be used in the context. On the other hand, when the text is "Paris is the Mecca of the fashion world", the term "Mecca" is used figuratively, so it is a term for usage caution and cannot be used in the context.

[0022] Thus, for the input text, it has been desired to detect whether the terms for usage caution included in the text can be used in the context.

[0023] Therefore, the present invention has been made to solve the above problems, and its object is to provide a term detection device and a program capable of detecting whether terms for usage caution included in a text can be used in the context of the text.

Means for Solving the Problems

[0024] To solve the above problems, the term detection device according to claim 1 receives text input by a user from a user terminal, detects a predetermined term from the text, and transmits a detection result to the user terminal. In the term detection device, a term that requires attention when used is defined as a term that requires attention, and a storage unit stores a plurality of terms that require attention as information on terms that require attention. A morphological analysis unit performs morphological analysis on the text and generates text after morphological analysis. Based on the text after morphological analysis generated by the morphological analysis unit and the information on terms that require attention stored in the storage unit, a term that requires attention included in the text is detected as a detected term that requires attention, and using a large language model, it is estimated whether the detected term that requires attention can be used in the context of the text, and information on whether the detected term that requires attention can be used in the context of the text is generated as the detection result. It is characterized by comprising a detection unit.

[0025] Further, the term detection device according to claim 2 is the term detection device according to claim 1, wherein the detection unit has a means for detecting a term that requires attention. The means for detecting a term that requires attention uses information on an explanation of the detected term that requires attention included in the text, the detected term that requires attention, and the information on terms that require attention stored in the storage unit to generate an instruction text including an instruction content for estimating whether the detected term that requires attention can be used in the context of the text, and an output format for obtaining, as an answer, an alternative that serves as a substitute for the detected term that requires attention when it is estimated that it cannot be used, together with the information on whether it can be used. The instruction text generated by the instruction text generation means is output to the large language model, and a model interface means obtains the answer from the large language model. A position information generation unit generates position information of the detected term that requires attention included in the text, and a detection result generation unit generates the detection result including the information on whether the detected term that requires attention included in the answer obtained by the model interface means can be used and the alternative in the case where it cannot be used, and the position information. It is characterized by comprising a detection result generation unit.

[0026] Furthermore, the term detection device of claim 3 is a term detection device of claim 2, wherein the detection unit further generates a kana-attached text by adding phonetic readings to the text, and has a phonetic reading estimation means that uses the large-scale language model to estimate whether the phonetic readings included in the kana-attached text are correct in the context of the text, and the phonetic reading estimation means generates an instruction sentence that includes an instruction content for estimating whether the phonetic readings of the kana-attached text are correct in the context of the text, and an output format for obtaining a response that returns the kana-attached text as is if it is estimated to be correct, and returns a kana-attached text corrected to the correct phonetic readings if it is estimated to be incorrect, and

[0027] Furthermore, the term detection device of claim 4 is a term detection device according to claim 2, wherein the detection unit further has input error detection means for estimating whether or not there are user input errors in the text using the large-scale language model, and the input error detection means generates an instruction statement that uses the text and the morphologically analyzed text generated by the morphological analysis unit to estimate whether or not there are user input errors in the text, and if an input error is estimated to exist, returns the term with the input error, the correct term obtained by correcting the term, and the position information of the term with the input error included in the text, and if no input error is estimated, returns the text as is, in an output format for obtaining a response, the instruction statement generated by the instruction statement generation means outputs the instruction statement to the large-scale language model and obtains the response from the large-scale language model, and the detection result generation means generates the detection result including the correct term obtained by the model interface means and the position information, or the text included in the response.

[0028] Furthermore, the program of claim 5 is characterized in that it causes a computer to function as a term detection device according to any one of claims 1 to 4. [Effects of the Invention]

[0029] As described above, according to the present invention, it is possible to detect whether or not a term requiring caution in use included in a text is usable within the context of the text. [Brief explanation of the drawing]

[0030] [Figure 1] This is a schematic diagram showing an example of the overall configuration of a text overlay production support system. [Figure 2] This is a block diagram showing an example configuration of a text overlay production support device according to an embodiment of the present invention. [Figure 3] This flowchart shows an example of processing by a text overlay production support device. [Figure 4] This block shows an example of the configuration of a means for detecting terms requiring caution in use. [Figure 5] This is a flowchart showing an example of processing by the term detection means for terms requiring caution (step S303). [Figure 6] This figure shows an example of the data structure for information on terms requiring caution in use. [Figure 7] This is a block diagram showing an example configuration of a means for utilizing a generated AI model, which is included in a means for detecting terms requiring caution in use. [Figure 8] This figure shows an example of instruction text used in the process of detecting terms requiring caution. [Figure 9] This figure shows an example of a response to an instruction in the process of detecting terms requiring caution. [Figure 10] This figure shows an example of the display screen on a user terminal (processing for detecting terms requiring caution). [Figure 11] This is a block diagram showing an example configuration of a reading analogy tool. [Figure 12] Flowchart showing an example of the processing method for inferring phonetic readings (step S304). [Figure 13] This block diagram shows an example configuration of a generative AI model utilization method provided for a reading inference method. [Figure 14] This figure shows an example of an instruction sentence used in phonetic transcription processing. [Figure 15] This figure shows an example of a response to an instruction in the phonetic transcription process. [Figure 16] This figure shows an example of the display screen on a user terminal (for phonetic transcription). [Figure 17] This is a block diagram showing an example configuration of an input error detection means. [Figure 18] This is a flowchart showing an example of the processing by the input error detection means (step S305). [Figure 19] This block diagram shows an example configuration of a means for utilizing a generated AI model, which is included in an input error detection means. [Figure 20] This figure shows an example of an instruction statement used in input error detection processing. [Figure 21]This figure shows an example of a response to an instruction in the input error detection process. [Figure 22] This figure shows an example of the display screen on a user terminal (input error detection process). [Modes for carrying out the invention]

[0031] The embodiments for carrying out the present invention will be described in detail below with reference to the drawings. The present invention will be described below using a system that supports the production of teleprompters (teleprompter production support system) as an example, and the processing of terms requiring caution, the processing of phonetic transcription, and the processing of input errors will be described.

[0032] The process for detecting terms requiring caution in use involves detecting terms requiring caution from the text used for on-screen text, using a generative AI model (large-scale language model) to estimate whether or not those terms can be used in the context of the text, estimating alternatives if they cannot be used, and generating the results as detection results for on-screen display.

[0033] The reading inference process involves adding readings to kanji characters in the text used for on-screen text, and then using a generative AI model to estimate whether the readings are correct in the context of the text. If the readings are incorrect, the correct readings are estimated, and the results are generated as detection results for on-screen display.

[0034] The input error detection process uses a generation AI model to estimate whether there are input errors in the text used for the on-screen text, and if there are input errors, it estimates the correct sentence and generates the result as a detection result for on-screen display.

[0035] The term "words requiring caution" includes not only prohibited terms, terms requiring caution, terms to be voluntarily restricted, and problematic terms in the television broadcasting industry, but also terms that are prohibited from use in other industries or terms that require caution when used. "Words requiring caution" is a general term encompassing all of these.

[0036] Furthermore, the information set for each term requiring caution is defined as information containing information related to that term. This information set consists of multiple terms requiring caution, and each term includes related information such as furigana (pronunciation guide), risk type, explanation, and whether or not it is a context-dependent term, and is pre-registered by the user (see Figure 6 below).

[0037] Furthermore, terms that are usable within the context of the text indicate that they pose no risk when used within that context. In other words, terms that are usable within the context of the text can be used as is because they pose no risk in that context.

[0038] On the other hand, terms that are unsuitable within the context of the text indicate terms that pose a risk when used in that context. In other words, terms that are unsuitable within the context of the text should be avoided from being used as is in that text because they pose a contextual risk.

[0039] [Subtitle Production Support System] First, a system including a text overlay production support device according to an embodiment of the present invention will be described. Figure 1 is a schematic diagram showing an example of the overall configuration of the text overlay production support system.

[0040] This caption production support system 1 is a system that supports the work of producing captions for educational programs, and consists of a caption production support device (terminology detection device) 2 and a user terminal 3. The caption production support device 2 and the user terminal 3 are connected by a network 4 such as an intranet.

[0041] The user accesses a web page (a program that processes the text overlays in

[0042] The caption creation support device 2 receives text entered from the user terminal 3 according to user operations and performs processing for detecting words requiring caution, inferring phonetic readings, and detecting input errors. The caption creation support device 2 then sends the generated detection results to the user terminal 3, displaying the results on the user terminal 3's screen. Details of the caption creation support device 2 will be described later.

[0043] User terminal 3 inputs text according to user instructions and sends the text to the caption creation support device 2. User terminal 3 then receives detection results for the text from the caption creation support device 2 and displays the detection results on the screen along with the text. In addition, user terminal 3 inputs predetermined conditions according to user instructions, and if there are kanji characters that match the predetermined conditions along with the input text, it changes the color of the characters in the corresponding sections, for example. Details of user terminal 3 and the predetermined conditions will be described later.

[0044] [Subtitle Production Support Device 2] Next, we will explain in detail the text overlay production support device 2 shown in Figure 1.

[0045] (composition) Figure 2 is a block diagram showing an example configuration of the teleprompter production support device 2 according to an embodiment of the present invention. As described above, the teleprompter production support device 2 is a term detection device that receives text entered by the user terminal 3 according to user operations, performs a term detection process for terms requiring caution, a reading inference process, and an input error detection process, and transmits the generated detection results to the user terminal 3.

[0046] This caption production support device 2 comprises a communication unit 11, a morphological analysis unit 12, a detection unit 13, and a storage unit 14. The detection unit 13 includes a means for detecting words to be used with caution 21, a means for inferring readings 23, and an input error detection means 24. The storage unit 14 also has a morphological analysis dictionary 31 and information on words to be used with caution 32 pre-stored (registered).

[0047] Furthermore, the detection unit 13 may include a kanji detection means 22 in addition to the usage caution term detection means 21, the reading kana guessing means 23, and the input error detection means 24. Also, the memory unit 14 may have a morphological analysis dictionary 31 and usage caution term information 32, as well as grade-level kanji information 33 pre-stored in it. The user terminal 3 may also be equipped with a kanji detection means 22, or the grade-level kanji information 33 may be pre-stored in the memory unit of the user terminal 3. In other words, the processing of the kanji detection means 22 using the grade-level kanji information 33 is performed by either the teleprompter production support device 2 or the user terminal 3.

[0048] Since the morphological analysis dictionary 31 stored in the memory unit 14 is known, a detailed explanation is omitted here. Details of the usage caution term information 32 will be described later. In addition, the grade-level kanji information 33 is a list of kanji in accordance with the curriculum guidelines, and the kanji to be taught (kanji to be learned) are defined for each grade level. Since the grade-level kanji information 33 is known, a detailed explanation is omitted here.

[0049] Furthermore, the generated AI information 34 is a generated AI model (large-scale language model) stored on an external server. Examples of generated AI models used include Claude® and ChatGPT®. The instruction sentences input to the generated AI model are generally called prompts (instruction sentences).

[0050] This generated AI information 34 is accessed by the model interface means 52, 72, and 92 shown in Figures 7, 13, and 19, which will be described later, via the detection unit 13's warning term detection means 21, reading comprehension means 23, and input error detection means 24.

[0051] The communication unit 11 receives text input from the user terminal 3 according to user operations, performs predetermined reception processing, and outputs the processed text to the morphological analysis unit 12. The communication unit 11 also receives detection results from the detection unit 13, performs predetermined transmission processing, and transmits the processed detection results to the user terminal 3.

[0052] The morphological analysis unit 12 receives text from the communication unit 11 and performs morphological analysis on the text using the morphological analysis dictionary 31 stored in the memory unit 14. The morphological analysis unit 12 then generates text consisting of multiple morphemes distinguished by part of speech, with the number of characters and position of each term clearly indicated, and this is used as the morphologically analyzed text. The morphological analysis unit 12 outputs the input text and the generated morphologically analyzed text to the detection unit 13.

[0053] The detection unit 13 receives text and the text after morphological analysis from the morphological analysis unit 12, performs a usage caution term detection process using the usage caution term detection means 21, a reading kana analogy process using the reading kana analogy means 23, and an input error detection process using the input error detection means 24, and generates detection results according to each process. The detection unit 13 then outputs the detection results to the communication unit 11.

[0054] If the detection unit 13 is equipped with a kanji detection means 22, the kanji detection means 22 receives grade-specific detection conditions (not shown) entered by the user in accordance with user operations on the user terminal 3 via the communication unit 11. Details of the grade-specific detection conditions will be described later. The kanji detection means 22 then extracts one or more kanji from the text and, for each extracted kanji, uses the grade-specific kanji information 33 stored in the memory unit 14 to identify the grade in which students will receive instruction on that kanji in accordance with the curriculum guidelines (the grade in which the kanji is studied).

[0055] The kanji detection means 22 determines whether each extracted kanji matches the grade-level detection conditions, generates a detection result that includes kanji that match the grade-level detection conditions, and outputs the detection result to the communication unit 11.

[0056] (process) Next, we will explain the processing of the text overlay production support device 2 shown in Figure 2. Figure 3 is a flowchart showing an example of the processing of the text overlay production support device 2.

[0057] The communication unit 11 receives text input from the user terminal 3 according to user operations (step S301). The morphological analysis unit 12 then performs morphological analysis on the text using the morphological analysis dictionary 31 and generates the morphologically analyzed text (step S302), and proceeds to steps S303 to S305.

[0058] For example, if the text is "Paris is the mecca of the fashion world," the text after morphological analysis will be "Paris" (noun) / "is" (particle) / "fashion" (noun) / "world" (noun) / of" (particle) / "mecca" (noun) / is" (auxiliary verb). This morphologically analyzed text will include the number of characters and position of each of the multiple morphemes (for "Paris," "is," etc.).

[0059] The detection unit 13's caution term detection means 21 moves from step S302 to perform a caution term detection process to generate a detection result (step S303), and then moves to step S306.

[0060] Specifically, the cautionary term detection means 21 uses cautionary term information 32 to detect cautionary terms from the text. If the cautionary term detection means 21 determines that the cautionary term is a context-dependent term, it uses generated AI information 34 to estimate whether the cautionary term can be used in the context of the text, and if it cannot be used, it estimates an alternative. The cautionary term detection means 21 generates the estimation result as a detection result for display on the user terminal 3. Details of the cautionary term detection means 21 will be described later.

[0061] For example, suppose the text is "Paris is the mecca of the fashion world." The cautionary term detection means 21 uses cautionary term information 32 to detect the cautionary term "mecca of" and determines that it is a context-dependent term. The cautionary term detection means 21 uses generated AI information 34 to estimate that the cautionary term "mecca of" cannot be used in the context and estimates "center" as an alternative.

[0062] Then, the usage caution term detection means 21 determines that for the text "Paris is the mecca of the fashion world", the term "the mecca" is a usage caution term that depends on the context and cannot be used contextually. It uses "center" as an alternative and generates a detection result including the related information described later.

[0063] The reading kana inference means 23 proceeds from step S302, performs a reading kana inference process to generate a detection result (step S304), and proceeds to step S306.

[0064] Specifically, the reading kana inference means 23 detects the Chinese characters included in the text, attaches reading kana, and uses the generated AI information 34 to estimate whether the reading kana is correct in the context of the text. If it is not correct, it estimates the correct reading kana. Then, the reading kana inference means 23 generates the estimation result as a detection result for the user terminal 3 to display on the screen. Details of the reading kana inference means 23 will be described later.

[0065] For example, assume the text is "I was in the middle of eating in the middle". The reading kana inference means 23 extracts Chinese characters from the text and generates text with reading kana attached to the Chinese characters (kana-attached text) "最中(さいちゅう)を食(た)べている最中(さいちゅう)だった".

[0066] Then, the reading kana inference means 23 uses the generated AI information 34 to estimate that the reading kana (さいちゅう) of the first "最中" is incorrect and estimates the correct reading kana (もなか). Also, the reading kana inference means 23 estimates that the reading kana (た) of "食" is correct and estimates that the reading kana (さいちゅう) of the second "最中" is also correct.

[0067] Then, for the text "最中を食べている最中だった", the reading kana inference means 23 generates the correct kana-attached text "最中(もなか)を食(た)べている最中(さいちゅう)だった" as the detection result.

[0068] The input error detection means 24 moves from step S302 to perform input error detection processing to generate a detection result (step S305), and then moves to step S306.

[0069] Specifically, the input error detection means 24 uses the generated AI information 34 to estimate whether there is an input error in the text sentence in context, and if there is an input error, it estimates the correct term. The input error detection means 24 then generates the estimation result as a detection result for the user terminal 3 to display on the screen. Details of the input error detection means 24 will be described later.

[0070] For example, suppose the text is "Please give me some ridge gear." The input error detection means 24 uses the generated AI information 34 to estimate that there is an input error in the term "ridge gear" and that the correct term is "please."

[0071] The input error detection means 24 then generates a detection result indicating that there is an input error in the term "onegear" in the text "yoroshiku onegearshimasu" and that the correct term is "onegai" (please).

[0072] The detection unit 13 moves from steps S303 to S305 and transmits the detection results generated by the word caution detection means 21, the reading inference means 23, and the input error detection means 24 to the user terminal 3 (step S306).

[0073] (Term detection method 21) Next, the usage caution term detection means 21 (usage caution term detection process (step S303)) shown in Figure 2 will be described in detail.

[0074] Figure 4 is a block diagram showing an example configuration of the caution term detection means 21, and Figure 5 is a flowchart showing an example of processing (step S303) of the caution term detection means 21. This caution term detection means 21 includes a verb substitution means 41, a pattern matching means 42, a generation AI model utilization means 43, and a detection result generation means 44.

[0075] The verb replacement means 41 receives the text and the morphologically analyzed text from the morphological analysis unit 12 (step S501). The verb replacement means 41 then extracts morphological terms (verb terms) that indicate the conjugation form of the verb from the morphologically analyzed text, and replaces the morphological terms indicating the conjugation form of the verb with morphological terms indicating the base form of the verb (step S502), thereby generating morphologically analyzed text (text after verb replacement) that includes the base form of the verb. The verb replacement means 41 outputs the input text and the generated text after verb replacement to the pattern matching means 42.

[0076] For example, if the text is "He went crazy and did...", the verb substitution means 41 extracts the morpheme term "crazy" which indicates the conjugated form of the verb, and replaces this term with the morpheme term "to go crazy" which indicates the base form of the verb.

[0077] Each verb has one base form and multiple conjugated forms. This substitution process is performed using pre-configured sets of data (data consisting of the base form and multiple conjugated forms for all verbs), with the base form of the verb and its corresponding conjugated forms forming a set.

[0078] The pattern matching means 42 receives the text and the text after verb substitution from the verb substitution means 41. The pattern matching means 42 then performs pattern matching between the text after verb substitution and each of the multiple usage caution terms (format_key, described later) registered in the usage caution term information 32, and detects one or more matching usage caution terms from the text after verb substitution (step S503).

[0079] The pattern matching means 42 outputs the input text and detected cautionary terms to the generating AI model utilization means 43 and the detection result generation means 44.

[0080] Figure 6 shows an example of the data structure of the cautionary term information 32. This cautionary term information 32 includes complex information for each cautionary term, such as points to note when using it and the category to which the cautionary term belongs. Specifically, the cautionary term information 32 is composed of multiple sets of data, with each set consisting of a cautionary term (phrase), format key (format_key), ruby ​​(ruby), risk type (riskType), description (description), category (category), synonyms (synonyms), context (context), etc.

[0081] In this example, "no mecca" is registered as a term requiring caution in use. Terms requiring caution in use consist of a single morpheme or a term formed by combining multiple morphemes.

[0082] If the part of speech of a term requiring caution is a verb, only the base form of the verb will be registered as a term requiring caution; the conjugated form of the verb will not be registered. The same applies if the term requiring caution includes a verb.

[0083] This reduces the size of the cautionary term information 32, thereby reducing the load on the pattern matching process in step S503 by the cautionary term detection means 21.

[0084] The format key is the morphologically analyzed term obtained by performing morphological analysis on the term requiring caution. The pattern matching in step S503 shown in Figure 5 is performed by the pattern matching means 42 between the text after verb substitution (i.e., the morphologically analyzed text after verb substitution) and this format key.

[0085] Ruby text is supplementary information such as phonetic readings and annotations for the term requiring caution in use. In this example, the ruby ​​text registered for the term requiring caution, "no mecca," is "marumaru no mecca."

[0086] The risk type indicates the degree to which caution is required when using the cautionary term. For example, "Use with sufficient caution" indicates that the term must be used with sufficient caution, while "Usable" indicates that it can be used. In this example, the risk type "Use with sufficient caution" is registered for the cautionary term "Mecca of".

[0087] The explanation is a sentence that expresses the content or meaning of the cautionary term in question so that users can understand it. In this example, the explanation for the cautionary term "Mecca" is registered as follows: "Avoid using it for religions other than Islam, or using it metaphorically even if it has a positive image... We corrected it to "center" when broadcasting."

[0088] The category indicates the field, category, department, or domain in which the cautionary term is used. For example, "religion," "disease / disability," and "trademark" are registered. In this example, the category "religion" is registered for the cautionary term "Mecca of."

[0089] The synonyms section indicates whether or not a synonym exists for the given term requiring caution, and registers either "No synonyms exist" or "Synonyms exist." If a synonym exists, that synonym is also registered. In this example, the synonym "No synonyms exist" is registered for the term requiring caution "Mecca of".

[0090] The context indicates whether the cautionary term is context-dependent, and is registered as either "true" or "false". "True" indicates that the term is context-dependent (related to the context), while "false" indicates that the term is not context-dependent (not related to the context). In this example, the context "true" is registered for the cautionary term "mecca of". This means that the cautionary term "mecca of" is a context-dependent term.

[0091] Returning to Figures 4 and 5, after step S503, the generating AI model utilization means 43 receives text and cautionary terms from the pattern matching means 42. Then, through processing in steps S504 to S507, if the cautionary terms are context-dependent terms, the generating AI model utilization means 43 uses the generated AI information 34 to estimate whether the cautionary terms can be used in the context, and generates usability information. The generating AI model utilization means 43 outputs the usability information to the detection result generation means 44.

[0092] Figure 7 is a block diagram showing an example configuration of the generation AI model utilization means 43 provided in the usage caution term detection means 21. This generation AI model utilization means 43 includes an instruction sentence generation means 51, a model interface means 52, and a usability information generation means 53.

[0093] The instruction statement generation means 51 receives text and cautionary terms from the pattern matching means 42 and reads the explanation and context corresponding to the cautionary terms from the cautionary term information 32. The instruction statement generation means 51 also receives pre-configured information such as the role, instruction content, and output format that constitute the instruction statement, which will be described later.

[0094] Referring to Figures 5 and 7, the instruction statement generation means 51 determines, based on the read context, whether the usage caution term is a context-dependent term (step S504). If the context is "true", it is determined to be a context-dependent term; if it is "false", it is determined not to be a context-dependent term.

[0095] If the instruction sentence generation means 51 determines in step S504 that the term is context-dependent (step S504:Y), it proceeds to step S505; if it determines that the term is not context-dependent (step S504:N), it proceeds to step S508.

[0096] The instruction statement generation means 51 moves from step S504(Y) and generates an instruction statement consisting of a role, instruction content (including data), output format, etc. (step S505). The data included in the instruction content is text, cautionary terms, and explanations. The instruction statement generation means 51 then outputs the instruction statement to the model interface means 52. This instruction statement indicates that it estimates whether the context-dependent cautionary terms included in the text can be used in the context, and if they cannot be used in the context, it estimates an alternative.

[0097] Figure 8 shows an example of an instruction sentence in the process of detecting words that require caution when used. This instruction sentence indicates the content of the task to be performed by the generating AI model, which is the generated AI information 34, and consists of a role s1-1, instruction content s1-2, output format s1-3, data s1-4, etc.

[0098] Role s1-1 "You are the world's best proofreader.", Instruction content s1-2 "The given text contains words that require judgment on whether they can be used depending on the situation. Based on the description of the words, use lateral thinking to determine whether the words can be used in the context of the text, and if they cannot be used, think of a synonym to replace them.", and Output format s1-3 "The answer is... Now proofread the following text." etc. are pre-set fixed information input by the instruction text generation means 51.

[0099] Furthermore, data s1-4 consists of text s1-41, cautionary terms s1-42, and explanations s1-43 read from cautionary term information 32, all of which are input by the instruction statement generation means 51.

[0100] Text s1-41 corresponds to the "given text" in instruction s1-2, and the cautionary term s1-42 corresponds to the "word for which judgment is required whether it is permissible to use" in instruction s1-2. In addition, explanation s1-43 corresponds to the "description of the term" in instruction s1-2.

[0101] Returning to Figures 5 and 7, after step S505, the model interface means 52 receives an instruction from the instruction generation means 51. Then, the model interface means 52 obtains a response to the instruction using the generated AI information 34 (step S506). In other words, the model interface means 52 outputs the instruction to a server equipped with the generated AI information 34 and obtains a response from the server. The model interface means 52 outputs the response to the availability information generation means 53.

[0102] Figure 9 shows an example of a response to an instruction in the process of detecting words to be used with caution, and corresponds to the example instruction shown in Figure 8. Assume that the pattern matching means 42 shown in Figure 4 detected the word to be used with caution, "mecca of," from the text "Paris is the mecca of the fashion world." Also, assume that the instruction generation means 51 shown in Figure 7 determined that the word to be used with caution, "mecca of," is a context-dependent term.

[0103] The instruction statement generation means 51 generates an instruction statement consisting of the role s1-1, instruction content s1-2, output format s1-3, data s1-4, etc., as shown in Figure 8. The data s1-4 is a text s1-41 "Paris is the mecca of the fashion world", a word requiring judgment s1-42 "the mecca of", and a description s1-43 "Non-Islamic...broadcast."

[0104] When such an instruction is output to a server equipped with generated AI information 34, the answer obtained is {"available":0, "alternative":"center"}. In other words, for an instruction such as estimating whether the context-dependent term "mecca of fashion" contained in the text "Paris is the mecca of fashion" is usable in the context, the answer indicates that the context-dependent term "mecca of fashion" is "not usable in the context," and that its alternative is "center."

[0105] For example, if the text is "Roads for pilgrims to access Mecca were built," and the instruction is to estimate whether the contextually relevant term "Mecca" is usable in this context, then the answer obtained is {"available":1, "alternative":""}. This answer indicates that the contextually relevant term "Mecca" is usable in this context.

[0106] Returning to Figures 5 and 7, after step S506, the availability information generation means 53 receives a response from the model interface means 52. Based on the response, the availability information generation means 53 generates availability information (step S507). The availability information generation means 53 outputs the availability information to the detection result generation means 44 and proceeds to step S508.

[0107] In the example described above, if {"available":0, "alternative":"central"} is entered as the answer, the availability information generation means 53 generates availability information indicating that the context-dependent term "mecca" is "unavailable in this context" and that its alternative is "central".

[0108] Furthermore, if the availability information generation means 53 inputs {"available":1, "alternative":""} as the answer, it generates availability information indicating that the context-dependent term "Mecca" is a term that is "usable in the context".

[0109] Returning to Figures 4 and 5, the detection result generation means 44 receives availability information from the generated AI model utilization means 43, and also receives text and cautionary terms from the pattern matching means 42.

[0110] The detection result generation means 44, moving from step S504(N) or step S507, generates positional information indicating the location of the term in the text for the term requiring caution, and reads related information from the term requiring caution information 32 (step S508). The related information is information such as furigana, risk type, explanation, category, synonyms, etc., corresponding to the term requiring caution in the term requiring caution information 32 shown in Figure 6.

[0111] The detection result generation means 44 generates a detection result for each caution term, including text, location information, availability information (if the caution term is a context-dependent term), and related information, and outputs this to the communication unit 11 (step S509).

[0112] Then, the communication unit 11 transmits the detection result generated by the detection result generation means 44 of the caution term detection means 21 to the user terminal 3, and the caption production support device 2 displays text on the user terminal 3 in a form that clearly indicates the position of the term corresponding to the caution term, using the position information included in the detection result.

[0113] Furthermore, the caption creation support device 2 displays availability information on the user terminal 3 according to the position where the term corresponding to the context-dependent term requiring caution in the text is displayed. In addition, the caption creation support device 2 displays related information on the user terminal 3 according to the position where the term corresponding to the term requiring caution in the text is displayed.

[0114] Furthermore, when the communication unit 11 transmits the detection results generated by the kanji detection means 22 to the user terminal 3, the teleprompter production support device 2 displays text on the user terminal 3 in a form that clearly indicates that the kanji that matches the grade-specific detection conditions included in the detection results match those grade-specific detection conditions.

[0115] FIG. 10 is a diagram showing an example of a display screen of the user terminal 3 (usage caution term detection process), and shows an example in which the detection result generated by the usage caution term detection process of the subtitle production support device 2 is displayed when assisting in the business of producing subtitles for an educational program.

[0116] This example of the display screen shows that the text input by the user operation is "Paris is the mecca of the fashion world" (see the text input area h2). Also, the detected usage caution term is "the mecca", and it shows that the auxiliary verb "is" following this is also a usage caution term (see i1 in the result display area h3).

[0117] Also, this example of the display screen shows that when kanji learned by third graders and above are used as the kanji with ruby, and kanji learned by junior high school students and above are made into hiragana, the kanji that matches this is "world" (see j1 in the result display area h3).

[0118] Also, on the left side of the example of the display screen, the detection conditions by grade are displayed (see the condition input area h1). The detection conditions by grade are conditions that reflect, for example, following the production rules of an educational program when producing subtitles for an educational program.

[0119] In this example of the display screen, the detection conditions by grade are that kanji learned by third graders and above are used as the kanji with ruby, and kanji learned by junior high school students and above are made into hiragana. Also, an item for specifying whether to align all alphanumeric characters and symbols in the text to full-width or half-width is displayed. Furthermore, items for specifying detecting risk words (usage caution terms), items for specifying detecting input errors by AI, and items for specifying saving the detection result in text format are displayed. In this example of the display screen, detecting risk words, etc. are specified.

[0120] First, the user inputs grade-level detection conditions by using the keyboard and mouse on user terminal 3. Specifically, when detecting kanji by grade level, the user turns on the "Detect kanji by grade level" switch in the condition input area h1 and specifies the grade levels for "Kanji to add furigana to" and "Kanji to convert to hiragana."

[0121] Furthermore, the user can turn on the switch to detect risk words, which are terms requiring caution when using the language. The user then enters the text to be detected in the text input area h2.

[0122] If the user terminal 3 is equipped with the kanji detection means 22 and grade-specific kanji information 33 shown in Figure 2, the kanji detection means 22 extracts kanji from the text according to the text entered in accordance with the user's operation. Then, for each extracted kanji, the kanji detection means 22 identifies the grade in which the student is receiving instruction using the grade-specific kanji information 33, determines whether or not it matches the grade-specific detection conditions, and generates a detection result that includes kanji that match the grade-specific detection conditions.

[0123] The user terminal 3 (display unit) displays the text on the screen in the result display area h3, and also displays the kanji characters indicated by the detection results contained in the text in the form corresponding to "kanji with furigana" and "kanji to be converted to hiragana," respectively.

[0124] Furthermore, when user terminal 3 receives text input in the text input area h2, for example, when the Enter key is pressed, it sends the text to the caption creation support device 2. In this case, user terminal 3 sends the newly entered text to the caption creation support device 2 each time the Enter key is pressed.

[0125] As a result, the text caption production support device 2 performs processing to detect words requiring caution, and the detection results are transmitted from the text caption production support device 2 to the user terminal 3.

[0126] The user terminal 3 (display unit) displays text on the screen in the result display area h3 based on the detection results transmitted from the teleprompter production support device 2, and also highlights the context-dependent term "no mecca da" (no mecca da), which includes the auxiliary verb "da" (see i1).

[0127] Here, the user terminal 3 (display unit) moves the cursor to the position of the term "Mecca of". The user terminal 3 (display unit) displays information on whether the context-dependent term "Mecca of" is usable, near the position of the term "Mecca of", and without overlapping the text (see k1-1 in result display area k1), and also displays related information (see k1-2 in result display area k1).

[0128] In this way, users can recognize context-dependent cautionary terms contained in the text they input, and can recognize whether or not those context-dependent cautionary terms are usable in the context (in this example, they are not usable ("may be a risky use")), and if they are not usable in the context, they can recognize alternatives (in this example, "center").

[0129] Furthermore, users can recognize the degree of caution required for the use of cautionary terms, their meaning, the presence or absence of synonyms, and their categories. In addition, users can recognize kanji characters for each grade level, in accordance with the curriculum guidelines, as either kanji to be given furigana (phonetic readings) or kanji to be written in hiragana.

[0130] This allows users, for example, to use the subtitle production support system 1 when creating subtitles for broadcast programs to check the content displayed on the subtitles during the production stage, rather than during the preview stage. Therefore, rework in the subtitle production process during the preview stage can be reduced.

[0131] As described above, according to the teleprompter production support device 2 of the embodiment of the present invention, the verb replacement means 41 of the caution term detection means 21 replaces morphological terms indicating the conjugated form of a verb with morphological terms indicating the base form of a verb in the text after morphological analysis, thereby generating the text after verb replacement. The pattern matching means 42 performs pattern matching between the text after verb replacement and each of the multiple caution terms registered in the caution term information 32, thereby detecting the caution terms.

[0132] The generating AI model utilization means 43, when the term requiring caution is a context-dependent term, uses the generated AI information 34 to estimate whether the term requiring caution can be used in the context of the text, estimates an alternative if it cannot be used, and generates usability information.

[0133] The detection result generation means 44 reads relevant information from the caution term information 32 for the caution term and generates a detection result that includes text, availability information, related information, etc. This detection result is transmitted to the user terminal 3 and displayed on the screen.

[0134] In this way, it becomes possible to detect cautionary terms in a context-aware manner, which cannot be achieved by simply comparing them with dictionaries such as the cautionary term information 32. In other words, it is possible to detect whether or not cautionary terms included in text entered by the user can be used in the context of the text. As a result, the user can determine whether or not context-dependent cautionary terms can be used in the context, and if not, recognize alternatives.

[0135] Furthermore, when users create on-screen text for broadcast programs, the system can detect context-dependent terms requiring caution during the production stage, rather than during the preview stage, and determine whether they are usable in the given context. This reduces rework in the on-screen text creation process during the preview stage.

[0136] (Method 23 for inferring pronunciation) Next, we will explain in detail the reading inference means 23 (and the reading inference processing (step S304)) shown in Figure 2.

[0137] Figure 11 is a block diagram showing an example configuration of the reading inference means 23, and Figure 12 is a flowchart showing an example of processing by the reading inference means 23 (step S304). This reading inference means 23 includes a reading addition means 61, a generation AI model utilization means 62, and a detection result generation means 63.

[0138] The reading kana addition means 61 receives text from the morphological analysis unit 12 (step S1201). The reading kana addition means 61 then detects kanji from the text and adds reading kana to the detected kanji to generate text with kana (step S1202). The reading kana addition means 61 outputs the text with kana to the generation AI model utilization means 62.

[0139] Furthermore, the reading addition means 61 may add readings to all detected kanji, or it may add readings only to kanji that are difficult to identify as a single reading, for example, because there are multiple candidate readings.

[0140] The generation AI model utilization means 62 receives the kana-added text from the reading kana addition means 61, and through processing in steps S1203 to S1205, uses the generation AI information 34 to estimate whether the reading kana included in the kana-added text is correct, and generates the corrected kana-added text. The generation AI model utilization means 62 then outputs the corrected kana-added text to the detection result generation means 63.

[0141] Figure 13 is a block diagram showing an example configuration of the generation AI model utilization means 62 provided in the reading inference means 23. This generation AI model utilization means 62 includes an instruction sentence generation means 71, a model interface means 72, and a proofreading text extraction unit 73.

[0142] The instruction text generation means 71 receives the kana-annotated text from the phonetic transcription means 61, and also receives the roles, instruction content, output format, etc., that constitute the instruction text, which will be described later, as pre-configured.

[0143] The instruction statement generation means 71 generates an instruction statement consisting of the input role, instruction content (including data), output format, etc. (step S1203). The data included in the instruction content is kana-annotated text. The instruction statement generation means 71 then outputs the instruction statement to the model interface means 72. This instruction statement indicates an instruction to estimate whether the kana-annotated text is correct or not, and if it is incorrect, to estimate the correct kana.

[0144] Figure 14 shows an example of an instruction sentence in the reading comprehension process. This instruction sentence indicates the content of the task to be performed by the generating AI model, which is the generating AI information 34, and consists of the role s2-1, instruction content s2-2, output format s2-3, data s2-4 (not shown in Figure 14 and omitted), etc.

[0145] Role s2-1 "You are the world's best proofreader.", instruction content s2-2 "Check the readings of the kanji written in parentheses in the given text (input), use lateral thinking to detect whether the readings are correct according to the context, and correct them if there are any errors in the readings. If you infer that there are no input errors, return the readings in parentheses as they are.", and output format s2-3 "The answer is... Now, proofread the following text." etc. are pre-set fixed information input by the instruction text generation means 71.

[0146] Furthermore, the data s2-4, which is not shown in the diagram, is the kana-annotated text s2-41 input by the instruction sentence generation means 71. The kana-annotated text s2-41 corresponds to the "given sentence (input)" in the instruction content s2-2.

[0147] Returning to Figures 12 and 13, the model interface means 72 receives an instruction from the instruction generation means 71 after step S1203. Then, the model interface means 72 obtains a response to the instruction using the generated AI information 34 (step S1204). In other words, the model interface means 72 outputs the instruction to a server equipped with the generated AI information 34 and obtains a response from the server. The model interface means 72 outputs the response to the proofreading text extraction unit 73.

[0148] Figure 15 is a diagram showing an example of a response to an instruction in the reading analogy process, and corresponds to the example instruction shown in Figure 14. Assume that the reading addition means 61 shown in Figure 11 generated the kana-accompanied text "saichuu o tabeteiru maichuu datta" from the text "saichuu o tabeteiru maichuu datta".

[0149] The instruction statement generation means 71 generates an instruction statement consisting of the role s2-1, instruction content s2-2, output format s2-3, data s2-4, etc., as shown in Figure 14. The data s2-4 is a sentence (input) kana-annotated text s2-41 "I was in the middle of eating saichu."

[0150] When such an instruction is output to a server equipped with generated AI information 34, the answer obtained is {"input":{"text":"I was in the middle of eating saichu"},"output":{"ruby":"I was in the middle of eating monaka"}}. In other words, for an instruction such as estimating whether the readings "saichu", "ta", and "saichu" in the kana-annotated text "I was in the middle of eating saichu" are correct, the answer indicates that the reading of the first "saichu" is not "saichu" but correctly "monaka", and that the kana-annotated text containing the correct reading after correction is "I was in the middle of eating monaka".

[0151] For example, if the instruction is to estimate whether the readings "odo" and "saichuu" in the kana-annotated text "odotte iru saichuu datta" are correct, and the answer obtained is {"input":{"text":"odotte iru saichuu datta"}, "output":{"ruby":"odotte iru saichuu datta"}}, this answer indicates that the readings in the kana-annotated text being estimated are correct, and that the kana-annotated text containing the corrected readings after correction is the same as the kana-annotated text being estimated.

[0152] Returning to Figures 12 and 13, the proofreading text extraction unit 73 receives the answer from the model interface means 72 after step S1204. The proofreading text extraction unit 73 then extracts the proofread text with kana (text with kana including the correct reading) from the answer (step S1205). The proofreading text extraction unit 73 outputs the proofread text with kana to the detection result generation means 63.

[0153] In the example above, if the corrected text extraction unit 73 receives {"input":{"text":"I was in the middle of eating a monaka"}, "output":{"ruby":"I was in the middle of eating a monaka"}} as input, it will extract the corrected text with kana characters, "I was in the middle of eating a monaka", from the input.

[0154] Furthermore, if the corrected text extraction unit 73 receives {"input":{"text":"dancing in the middle of dancing"}, "output":{"ruby":"dancing in the middle of dancing"}} as input, it will extract from that input the kana-attached text "dancing in the middle of dancing" which contains the correct reading, even though it is not a corrected kana-attached text, as the corrected kana-attached text.

[0155] Returning to FIGS. 11 and 12, the detection result generation means 63 inputs the corrected kana text from the generation AI model utilization means 62, generates this as a detection result, and outputs the detection result to the communication unit 11 (step S1206).

[0156] Then, by the communication unit 11 transmitting the detection result generated by the detection result generation means 63 of the reading kana inference means 23 to the user terminal 3, the telop production support device 2 causes the user terminal 3 to display on the screen a kana text including the correct reading kana, which is the corrected kana text.

[0157] Also, as described above, by the communication unit 11 transmitting the detection result generated by the kanji detection means 22 to the user terminal 3, the telop production support device 2 causes the user terminal 3 to display on the screen, in a form in which it is明示 that the kanji符合 the academic year - specific detection conditions included in the detection result, the text for the kanji that符合 the academic year - specific detection conditions.

[0158] FIG. 16 is a diagram showing an example of a display screen of the user terminal 3 (reading kana inference process), and shows an example in which the detection result generated by the reading kana inference process of the telop production support device 2 is displayed when assisting in the task of producing a telop for an educational program.

[0159] This display screen example shows that the text input by the user operation is "最中を食べている最中だった" (refer to the text input area h2). Also, it shows that the kanji that符合 the academic year - specific detection conditions shown on the left side of the display screen example (refer to the upper part of the condition input area h1) are "最中", "食", and "最中" (refer to j2, j3, and j4 in the result display area h3).

[0160] As described in FIG. 10, the user inputs the academic year - specific detection conditions by operating the keys and mouse using the user terminal 3, and designates to turn on a switch or the like for detecting usage caution terms, which are risk words.

[0161] Furthermore, when user terminal 3 receives text input in the text input area h2 (in this example, "I was in the middle of eating monaka"), and for example, when the Enter key is pressed, the text is sent to the caption creation support device 2.

[0162] As a result, the text caption production support device 2 performs reading inference processing, etc., and the detection results are transmitted from the text caption production support device 2 to the user terminal 3.

[0163] Furthermore, when "Save results as text" is selected based on the user's keyboard and mouse operations, the corrected text with kana (including the correct reading) "I was eating a monaka" is displayed (see result display area k2), and can be saved by the user.

[0164] In this way, users can determine that the text they input contains kana characters with the correct readings.

[0165] This allows, for example, when a user creates on-screen text for a broadcast program, to use the on-screen text creation support system 1 to check the pronunciation of kanji characters related to the content to be displayed in the on-screen text during the production stage, rather than during the preview stage. Therefore, it is possible to reduce rework in the on-screen text creation process during the preview stage.

[0166] As described above, according to the teleprompter production support device 2 of the embodiment of the present invention, the reading kana addition means 61 of the reading kana guessing means 23 generates text with kana by adding reading kana to the kanji detected from the text.

[0167] The generation AI model utilization means 62 uses the generation AI information 34 to estimate whether the readings in the kana-annotated text are correct or not. If they are incorrect, it estimates the correct readings and generates the corrected kana-annotated text. The detection result generation means 63 generates the corrected kana-annotated text as the detection result. This detection result is transmitted to the user terminal 3 and displayed on the screen.

[0168] In this way, it becomes possible to detect terms requiring caution in context, which cannot be achieved by simply comparing them with dictionaries such as the "Terms to Use with Caution Information 32". In other words, it is possible to detect the correct reading of kanji characters included in text entered by the user. This allows the user to determine the correct reading of kanji characters included in the text.

[0169] Furthermore, when users create on-screen text for broadcast programs, they can determine the correct reading of kanji characters during the production stage rather than the preview stage, thus reducing rework in the on-screen text creation process during the preview stage.

[0170] (Input error detection means 24) Next, we will explain in detail the input error detection means 24 (and the input error detection process (step S305)) shown in Figure 2.

[0171] Figure 17 is a block diagram showing an example configuration of the input error detection means 24, and Figure 18 is a flowchart showing an example of processing by the input error detection means 24 (step S305). This input error detection means 24 includes a generated AI model utilization means 81 and a detection result generation means 82.

[0172] The generation AI model utilization means 81 receives text and the text after morphological analysis from the morphological analysis unit 12 (step S1801). Then, through processing in steps S1802 to S1804, the generation AI model utilization means 81 uses the generation AI information 34 to estimate whether or not there are input errors in the text and generates proofreading information. The generation AI model utilization means 81 outputs the proofreading information to the detection result generation means 82.

[0173] Figure 19 is a block diagram showing an example configuration of the generation AI model utilization means 81 provided in the input error detection means 24. This generation AI model utilization means 81 includes an instruction statement generation means 91, a model interface means 92, and a calibration information generation means 93.

[0174] The instruction text generation means 91 receives text and the text after morphological analysis from the morphological analysis unit 12, as well as pre-configured instructions such as the role, instruction content, and output format that constitute the instruction text, which will be described later.

[0175] The instruction statement generation means 91 generates an instruction statement consisting of the input role, instruction content (including data), output format, etc. (step S1802). The data included in the instruction content is the text and the text after morphological analysis. The instruction statement generation means 91 then outputs the instruction statement to the model interface means 92. This instruction statement estimates whether or not there are input errors in the text sentence, and if there are input errors, it estimates the correct term and indicates the location of the error.

[0176] Figure 20 shows an example of an instruction statement in the input error detection process. This instruction statement indicates the content of the task to be performed by the generating AI model, which is the generated AI information 34, and consists of a role s3-1, instruction content s3-2, output format s3-3, data s3-4 (not shown in Figure 20 and omitted), etc.

[0177] The roles s3-1 "You are the world's best proofreader," the instructions s3-2 "Use lateral thinking to detect input errors such as misspellings, omissions, or incorrect kanji conversions in the given text, and proofread it if input errors are found. If you infer that there are no input errors, return the input text as is," and the output format s3-3 "The answer is... Now, proofread the following text," etc., are pre-set fixed information input by the instruction text generation means 91.

[0178] Furthermore, the data s3-4, which is not shown, is the text s3-41 input by the instruction sentence generation means 91 and the text s3-42 after morphological analysis. Text s3-41 and the text s3-42 after morphological analysis correspond to the "given sentence" in the instruction content s3-2.

[0179] Returning to Figures 18 and 19, the model interface means 92 receives an instruction from the instruction generation means 91 after step S1802. Then, the model interface means 92 obtains a response to the instruction using the generated AI information 34 (step S1803). In other words, the model interface means 92 outputs the instruction to a server equipped with the generated AI information 34 and obtains a response from the server. The model interface means 92 outputs the response to the proofreading information generation means 93.

[0180] Figure 21 is a diagram showing an example of a response to an instruction in the input error detection process, and corresponds to the example instruction shown in Figure 20. Assume that the text "よろしくおねギアします" and the corresponding morphologically analyzed text are input to the morphological analysis unit 12 shown in Figure 2.

[0181] The instruction text generation means 91 generates an instruction text consisting of the role s3-1, instruction content s3-2, output format s3-3, data s3-4, etc., as shown in Figure 20. The data s3-4 is the text s3-41 "Thank you for your cooperation" and the corresponding morphologically analyzed text s3-42.

[0182] When such an instruction is output to a server equipped with generated AI information 34, the answer is a list [{"item":{"incorrect":"onegear","predict":"onegai"},"position":{"start":8,"end":15}}]. In other words, for an instruction such as estimating whether there is an input error in the text "yoroshiku onegearshimasu", the answer indicates that there is an input error in the term "onegear", the correct term is "onegai", the starting position of that term in the text is the 8th position, and the ending position is the 15th position.

[0183] Furthermore, if the instruction is to estimate whether or not there is a typographical error in the text "Thank you in advance," the answer will be an empty list []. This answer indicates that there is no typographical error in the text "Thank you in advance."

[0184] Returning to Figures 18 and 19, after step S1803, the calibration information generation means 93 receives a response from the model interface means 92 and generates calibration information from the response (step S1804). The calibration information generation means 93 outputs the calibration information to the detection result generation means 82.

[0185] In the example described above, if the input of the list [{"item":{"incorrect":"ridge gear","predict":"please"},"position":{"start":8,"end":15}}] is entered as the answer, the proofreading information generation means 93 generates proofreading information indicating that there is an input error in the term "ridge gear", the correct term is "please", and that the starting position of that term in the text is the 8th position and the ending position is the 15th position.

[0186] Furthermore, if no input errors are detected, an empty list [] is returned as the response to the generated AI information 34, so the proofreading information generation means 93 inputs an empty list [] as the response. In other words, the proofreading information generation means 93 inputs the response list, detects the number of errors from the number of elements in the list, and if the number of elements is zero (in the case of an empty list []), the detection result generation means 82 shown in Figure 17 generates a detection result indicating that there are no input errors in the text and sends it to the user terminal 3.

[0187] Returning to Figures 17 and 18, the detection result generation means 82 receives correction information from the generated AI model utilization means 81, as well as text and the text after morphological analysis from the morphological analysis unit 12. The detection result generation means 82 then generates a detection result including the text and correction information, and outputs the detection result to the communication unit 11 (step S1805). If no input errors are detected in the generated AI information 34, the detection result generation means 82 outputs the text and a detection result indicating that there are no input errors in the text to the communication unit 11.

[0188] Then, the communication unit 11 transmits the detection result generated by the detection result generation means 82 of the input error detection means 24 to the user terminal 3, so that the caption production support device 2 displays the correct term on the screen if there is an input error in the text.

[0189] Furthermore, as described above, the communication unit 11 transmits the detection results generated by the kanji detection means 22 to the user terminal 3, causing the teleprompter production support device 2 to display text on the user terminal 3 in a form that clearly indicates that the kanji that matches the grade-specific detection conditions included in the detection results match those grade-specific detection conditions.

[0190] Figure 22 shows an example of the display screen of user terminal 3 (input error detection processing), illustrating an example where the detection results generated by the input error detection processing of the caption production support device 2 are displayed when assisting with the task of producing captions for educational programs.

[0191] This example screen shows that the text entered by the user is "yoroshiki onegear shimasu" (see text input area h2). It also shows that the kanji characters that match the grade-specific detection conditions shown on the left side of the example screen (see the top of condition input area h1) are "o" and "ne" (see j5, l1 in result display area h3).

[0192] As explained in Figures 10 and 16, the user inputs grade-level detection conditions and specifies to turn on switches, etc., for detecting risk words, which are terms requiring caution when using.

[0193] Furthermore, when user terminal 3 receives text input in the text input area h2 (in this example, "よろしくおねギアします"), and for example, when the Enter key is pressed, the text is sent to the caption creation support device 2.

[0194] As a result, the text overlay production support device 2 performs input error detection processing, and the detection results are transmitted from the text overlay production support device 2 to the user terminal 3.

[0195] The user terminal 3 (display unit) highlights the term "Onegiashimasu," which includes the auxiliary verb and auxiliary verb "shimasu," in response to the text "Yoroshiku Onegiashimasu" (see m1).

[0196] Here, the user terminal 3 (display unit) moves the cursor to the position of the term "ridge gear" according to the user's operation. Then, the user terminal 3 (display unit) generates the term "ridge gear" with input errors from the detection results as the detected term, and generates the correct term "please" as the suggested correct term. Then, the user terminal 3 (display unit) displays the detected term "ridge gear" and the suggested correct term "please" on the screen near the position of the term "ridge gear" in the result display area h3, without overlapping the text (see result display area k3).

[0197] In this way, users can determine whether or not there are input errors in the text they have entered, and if there are errors, they can see suggested corrections.

[0198] This allows, for example, when a user creates on-screen text for a broadcast program, to use the on-screen text creation support system 1 to check for input errors in the content to be displayed on the on-screen text during the production stage, rather than during the preview stage. Therefore, rework in the on-screen text creation process during the preview stage can be reduced.

[0199] As described above, according to the teleprompter production support device 2 of the embodiment of the present invention, the generation AI model utilization means 81 of the input error detection means 24 uses the generation AI information 34 to estimate whether or not there is an input error in the text, and if there is an input error, it estimates the correct text and generates proofreading information. The detection result generation means 82 generates a detection result including the text and proofreading information. This detection result is transmitted to the user terminal 3 and displayed on the screen.

[0200] In this way, it becomes possible to detect terms requiring caution in context, which cannot be achieved by simply comparing them with dictionaries such as the "Terms to Use with Caution Information 32". In other words, it is possible to detect whether or not there are input errors in the text entered by the user, and if there are input errors, it is possible to detect suggested corrections. As a result, the user can determine whether or not there are input errors in the text and recognize suggested corrections if there are errors.

[0201] Furthermore, since users can identify input errors during the production phase rather than the preview phase when creating on-screen text for broadcast programs, rework during the preview phase of on-screen text production can be reduced.

[0202] Although the present invention has been described above with reference to embodiments, the present invention is not limited to the above embodiments and can be modified in various ways without departing from the technical concept.

[0203] For example, the teleprompter production support system 1 shown in Figure 1 is used to support the task of producing teleprompters, but the present invention is not only applicable to the teleprompter production support system 1, but also to systems that support the task of producing scripts for broadcast programs, etc. In short, the present invention is applicable to various systems that estimate whether there are terms in the text that cannot be used in context, whether the phonetic readings added to the text are correct, and whether there are input errors in the text, and present the estimated detection results to the user.

[0204] Furthermore, a standard computer can be used as the hardware configuration for the teleprompter production support device 2 and user terminal 3 according to the embodiment of the present invention. The teleprompter production support device 2 and user terminal 3 are composed of a computer equipped with a CPU, a volatile storage medium such as RAM, a non-volatile storage medium such as ROM, and an interface.

[0205] The functions of the teleprompter production support device 2, including the communication unit 11, morphological analysis unit 12, detection unit 13 (means for detecting words requiring caution 21, means for detecting kanji characters 22, means for inferring readings 23 and means for detecting input errors 24), and storage unit 14, are each realized by having the CPU execute a program that describes these functions. The same applies to the functions of the display unit and other components of the user terminal 3.

[0206] These programs are stored on a storage medium and are read and executed by the CPU. These programs can also be stored and distributed on storage media such as magnetic disks (HDDs, etc.), optical disks (CD-ROMs, DVDs, etc.), and semiconductor memory (SSDs, etc.), and can be transmitted and received over a network. [Explanation of Symbols]

[0207] 1. Subtitle Production Support System 2. Subtitle Production Support Device 3. User terminals 4 Network 11 Communications Department 12 Morphological analysis section 13 Detection unit 14 Storage section 21. Detecting Terms of Use 22 Kanji detection means 23. Means of inferring pronunciation by analogy 24 Input error detection means 31. Dictionary for morphological analysis 32 Cautionary terms information 33 Kanji information by grade level 34. Generated AI Information (Large-Scale Language Models) 41 Verb substitution means 42 Pattern matching means 43,62,81 Generative AI Model Utilization Methods 44,63,82 Detection result generation means 51,71,91 Instruction sentence generation means 52,72,92 Model Interface Means 53 Availability information generation means 61 Means for adding phonetic readings 73 Proofreading Text Extraction Unit 93 Calibration information generation means Roles of s1-1, s2-1, s3-1 s1-2,s2-2,s3-2 Instruction content s1-3,s2-3,s3-3 output format s1-4, s2-4, s3-4 data s1-41 Text s1-42 Cautionary terms used s1-43 Description s2-41 Text with kana characters s3-41 Text s3-42 Text after morphological analysis h1 Condition input area h2 Text input area h3,k1,k2,k3 Result display area

Claims

1. In a term detection device that receives text entered by a user from a user terminal, detects predetermined terms from the text, and transmits the detection results to the user terminal, A memory unit stores multiple terms that require caution when used as "terms to be used with caution," and this memory unit stores multiple terms as "terms to be used with caution" information. A morphological analysis unit performs morphological analysis on the aforementioned text and generates the text after morphological analysis. A detection unit that, based on the morphologically analyzed text generated by the morphological analysis unit and the usage caution term information stored in the storage unit, detects usage caution terms contained in the text as detected usage caution terms, estimates whether the detected usage caution terms can be used in the context of the text using a large-scale language model, and generates information as the detection result whether the detected usage caution terms can be used in the context of the text. A term detection device characterized by being equipped with the following features.

2. In the term detection device according to claim 1, The detection unit has a means for detecting terms to be used with caution. The aforementioned means for detecting terms to be used with caution is: Instruction text generation means generates an instruction text that uses the text, the detected caution term, and the explanatory information about the detected caution term contained in the caution term information stored in the memory unit to estimate whether the detected caution term can be used in the context of the text, and, along with the information on whether it can be used, generates an instruction text that includes an output format to obtain an alternative to the detected caution term as a response if it is estimated that it cannot be used. A model interface means that outputs the instruction generated by the instruction generation means to the large-scale language model and obtains the answer from the large-scale language model, A term detection device comprising: a detection result generation means that generates location information of the detected caution term contained in the text; and a detection result generation means that generates information on whether the detected caution term contained in the response obtained by the model interface means is available, an alternative if it is not available, and the location information.

3. In the term detection device according to claim 2, The detection unit is Furthermore, the system includes a means for generating a text with kana by adding phonetic readings to the text, and for estimating whether the phonetic readings included in the text with kana are correct in the context of the text using the large-scale language model. The aforementioned means for inferring the reading of the characters is, Instruction generation means for generating an instruction statement that includes instructions for estimating whether the reading of the aforementioned kana-annotated text is correct in the context of the text, and an output format for obtaining a response that returns the kana-annotated text as is if it is estimated to be correct, and returns the kana-annotated text corrected to the correct reading if it is estimated to be incorrect. A model interface means that outputs the instruction generated by the instruction generation means to the large-scale language model and obtains the answer from the large-scale language model, A term detection device comprising: a detection result generation means for extracting the kana-scripted text or the corrected kana-scripted text from the response obtained by the model interface means, and generating the detection result including the kana-scripted text or the corrected kana-scripted text.

4. In the term detection device according to claim 2, The detection unit is Furthermore, it includes an input error detection means that uses the large-scale language model to estimate whether or not there are user input errors in the text, The aforementioned input error detection means is Instruction text generation means generates an instruction text that uses the aforementioned text and the morphologically analyzed text generated by the morphological analysis unit to generate an instruction text that includes an output format for obtaining a response that, if an input error is estimated to exist, returns the term with the input error, the correct term after correction of that term, and the positional information of the term with the input error in the text, and if no input error is estimated to exist, returns the text as is. A model interface means that outputs the instruction generated by the instruction generation means to the large-scale language model and obtains the answer from the large-scale language model, A term detection device comprising: a detection result generation means for generating the detection result which includes the corrected term and location information included in the response obtained by the model interface means, or the text included in the response.

5. A program for causing a computer to function as a term detection device according to any one of claims 1 to 4.

Images