JP7610101B2 - DATA REGISTRATION PROCESSING METHOD, DATA REGISTRATION PROCESSING PROGRAM AND DATA REGISTRATION PROCESSING APPARATUS - Google Patents

Description

The present invention relates to a data registration processing method, a data registration processing program, and a data registration processing device.

There are systems that estimate dialogue and support operations based on this dialogue. For example, one possible application would be to estimate the dialogue between a caller who makes a call from an emergency phone on the side of a highway and a controller who receives the call, and display the relevant page of an accident response manual to the controller. In supporting various operations like this, there is a demand to create a database of the dialogue between the two parties, generate sentences such as summaries suitable for carrying out operations based on the dialogue, and use them for searches.

Technologies for generating summaries from such dialogue include, for example, a technology that uses semantic search to extract word connection patterns from conversation data and creates summaries by pattern matching with summary templates. There is also a technology that estimates the type of conversation and extracts parts of the conversation as key points to use as search targets. There is also a technology that takes discourse data as input and outputs flow information consisting of the correspondence between statements of expressions that match a flow expression and flows as discourse semantics. Note that attribute labeling of extracted words based on a dictionary with a 1:1 correspondence between words and labels, for example, attribute labeling of words based on the 5W2H, has become common.

JP 2014-130613 A International Publication No. 2020/036190 JP 2011-134291 A

However, conventional technology has not been able to improve the search accuracy of searching for documents such as necessary summaries from dialogues. Dialogue databases record the text of the dialogues, and also contain words that become noise during searches. For this reason, when searching existing databases, there are cases where the desired information is not found due to differences in the amount of text between the search keywords and the text being searched, making it impossible to improve search accuracy.

Traditionally, semantic search has had issues in two ways: 1. when there is a large difference in the amount of text between the search sentence and the sentence to be searched, and 2. when searching for an overview rather than the full text, such as in a web search.

1. When there is a large difference in the amount of text between the search text and the text to be searched, many users are accustomed to keyword searches and often only search using keywords or short sentences. Generally, the text to be searched, such as in a manual search, is often long. On the other hand, the search text is often short, and unusual words contained in the search text have a large effect on the search results. Even if the text to be searched contains a text that is very similar in terms of semantics and grammar, the other text becomes noise in the distributed representation of the text used in the search, making it difficult to obtain search results with a high degree of similarity. For example, a text that is very similar to the search text (the text to be searched) may be buried in other text and not be found. Similarity can be obtained from the number of matching words in a document, the distance between words in a document, etc. In this regard, it is desirable to realize a search that is independent of the amount of text in the text to be searched.

2. When it comes to summary searches that do not include the full text, such as web searches, general search sites do not look at the full text of the searched sentence, so search results that meet the search request cannot be obtained. Search accuracy differs depending on the content creator (searched site).

Furthermore, conventional methods for estimating the meaning of a sentence have also had problems with the method of analyzing the sentence syntactically and calculating the meaning from the sentence structure. This method itself can convert the meaning of a sentence into a numerical value, making it useful for comparing sentences, and in recent years, methods using deep learning for syntactic analysis (such as Google's BERT and related methods ALBERT and XLNet) have been published, improving the accuracy of the analysis. However, at present, it is not possible to know the meaning of a target sentence from these numerical values or vector values, and a computer cannot understand the meaning.

In one aspect, the present invention aims to estimate the meaning of a text document and improve the search accuracy of a database in which the text document is recorded.

According to one aspect of the present invention, in a data registration process that estimates semantic information contained in an input text document and searches and outputs it, a data registration processing method, a data registration processing program, and a data registration processing device are required that extract from the text document, based on a sentence containing the word, a word that is registered in a knowledge table in which the word, the sentence containing the word, and the attribute of the word are associated with each other, rearrange a group of the extracted multiple words in a predetermined order of the attributes, create a summary consisting of the rearranged group of words, and record the summary in a database in association with the text document.

According to one aspect of the present invention, it is possible to estimate the meaning of a text document and improve the search accuracy of a database in which the text document is recorded.

FIG. 1 is a diagram for explaining an outline of the data registration process according to the present invention. FIG. 2 is a diagram showing the overall configuration of a search system including a data registration processing device according to an embodiment of the present invention. FIG. 3 is a functional block diagram of a search system including a data registration processing device. FIG. 4 is a diagram illustrating an example of a hardware configuration of the data registration processing device. FIG. 5 is an explanatory diagram of the process flow for each function of the data registration processing device. (Part 1) FIG. 6 is an explanatory diagram of the process flow for each function of the data registration processing device. (Part 2) FIG. 7 is an explanatory diagram of the process flow for each function of the data registration processing device. (Part 3) FIG. 8 is a diagram showing the data structure of the knowledge table. FIG. 9 is a diagram showing a specific example of the knowledge table. FIG. 10A is a diagram showing search targets and label classification for knowledge data. FIG. 10B is a diagram showing search targets and label classification for knowledge text. FIG. 11 is an explanatory diagram of data modeling of basic knowledge. FIG. 12 is a table showing specific examples of basic knowledge. FIG. 13 is an explanatory diagram of an example of using basic knowledge search. FIG. 14 is an explanatory diagram of an example of label extraction from a business document and knowledge data label creation. FIG. 15 is an explanatory diagram of an example of automatic labeling and summary creation for a business document. FIG. 16A is an explanatory diagram showing an example of a join order for creating an overview. (Part 1) FIG. 16B is an explanatory diagram showing an example of the join order for creating an overview. (Part 2) FIG. 17 is a flowchart of an example of the overall process of creating knowledge data. FIG. 18A is a flowchart of an example of knowledge data label creation. (Part 1) FIG. 18B is a flowchart of an example of knowledge data label creation. (Part 2) FIG. 18C is a diagram showing an example of a word dictionary. FIG. 19 is a flowchart of an example of creating an overview. FIG. 20A is a flowchart of an example of semantic label creation. (Part 1) FIG. 20B is a flowchart of an example of semantic label creation. (Part 2) FIG. 21 is an explanatory diagram showing an example of how an overview is used by a business application. FIG. 22 is a process flowchart of an example of how an overview is utilized by a business application. FIG. 23 is a diagram showing an example of using a business application based on a dialogue. (Part 1) FIG. 24 is a diagram showing an example of using a business application based on a dialogue (part 2).

Below, the disclosed data registration processing method, data registration processing program, and data registration processing device will be described in detail with reference to the drawings.

Figure 1 is a diagram for explaining an overview of the data registration process according to the present invention. The computer that performs the data registration process according to the embodiment does not understand the meaning of a sentence through structural analysis that uses numerical calculations, but rather imitates human thinking. In the data registration process according to the embodiment, the meaning of a sentence is inferred using natural language text search and text classification.

For example, by defining the structure of the searched sentence and searching a database for the searched sentence that has been labeled with the meaning contained within the sentence, the meaning of the sentence is inferred from the labels of the sentences in the search results and the searched sentence is labeled. Furthermore, a summary is created from the searched sentence. The created summary is registered in the database along with the searched sentence and will be the target of future searches.

The main text structure of this summary is defined as consisting of nine types of attribute text, consisting of the 5W2H, actions, and states, and a summary is automatically generated from long searched text such as dialogue using labels. The 5W2H are: Who, When, Where, What, Why, How, and How much.

The data registration processing device 100 shown in FIG. 1 includes a control unit 101 and a knowledge table 102. The knowledge table 102 holds associations between words, sentences containing the words, and attributes (5W2H + action/state).

The control unit 101 performs the following processes 1 to 3.
1. From the text document (searched sentence) to be analyzed, words registered in the knowledge table 102 are extracted based on sentences containing the words.
2. The extracted words are rearranged in a predetermined order based on the attributes to generate a summary in natural language.
3. The summary consisting of the rearranged group of words is associated with the text document (the sentence to be searched) and recorded in a database, and used for subsequent searches.

Figure 1 shows an example of telephone response/handling work (control work) at a call center that receives calls from emergency phones on an expressway. Dialogue 150 is a text document (searchable text) of a conversation between a caller (speaker B) who reports a vehicle breakdown on the expressway and a controller (speaker A) who receives the call. Here, in a call (telephone call) between the caller and the controller, the caller is calling from a remote location and the call quality is often poor due to noise from other vehicles on the road. The controller also repeats the words spoken by the caller with clear pronunciation. This improves the accuracy of voice recognition by converting only the voice of the controller (speaker A).

The control unit 101 then searches the dialogue 150 using text sentences obtained by speech recognition processing of words uttered by the controller (speaker A) as the search sentence. During this search, the control unit 101 extracts words registered in the knowledge table 102 and assigns a semantic label La to the words included in the search sentence. The knowledge table 102 holds a large number (1:N) of words that are expected in dialogue during air traffic control operations and have been given a semantic label indicating a typical meaning. For example, when the word is a specific car name XX, N cars and car names are associated and held as the semantic label for one car name XX.

In the example of FIG. 1, the control unit 101 refers to the knowledge table 102 and extracts the semantic label La of the word "fault" from the text document "Oh, it's a malfunction" by the controller (speaker A) in the dialogue 150. The semantic label La is recorded (in the knowledge text table 803 described later) as knowledge text associated with the text sentence.

Next, the control unit 101 generates a summary 180 from the entire dialogue 150. At this time, the control unit 101 refers to the knowledge table 102, rearranges the extracted words in the order of their attributes, and generates a summary 180 equivalent to natural language. The attributes consist of, for example, 5W2H and actions/states, and the words (text sentences corresponding to the semantic label La) are rearranged in the order of the specified attributes.

The knowledge data (knowledge data label) Lb shown in Figure 1 associates the 5W2H + action/status with words. The order of the 5W2H in the illustrated knowledge data label Lb is, from the top, When, Where, Who, What, Why, How, How much, status, and action. Here, the attribute of the word "light truck" in the knowledge data label Lb is "What."

The control unit 101 then mechanically combines the words that are generated when the 5W2H and action/state attributes are rearranged in a predetermined order to create the summary sentence 180. The order in which the attributes are combined is an order that is adapted to the business content, in the above example, the execution of the call center control business (controller A).

The control unit 101 notifies the controller performing control operations of the summary 180, and the controller can immediately present the necessary measures, etc. to the caller in accordance with the summary 180 notified by the control unit 101.

Here, since summary 180 is extracted by inferring the meaning of dialogue 150, the controller does not have to reread the entire lengthy dialogue 150, and the information is concise and useful for processing. This allows the controller to easily use summary 180 to, for example, search for a manual to be presented to the caller, or to input ledger items required for creating a ledger for each caller in the call center control operations, thereby enabling the call center control operations to be carried out efficiently.

As a result, according to the embodiment, the dialogue can be summarized as a summary of a group of words with normalized word order, improving the search accuracy of a database in which summaries are recorded in association with dialogue text documents. When searching for a searched sentence (text document), a summary with a word order defined by attributes is searched for, reducing the variation in the amount of text between the search keywords and the searched sentence, and improving the search accuracy rate.

The data registration process of the embodiment responds to the following requests from users (e.g., the controller) and system developers. Users have a request to be able to use various types of text search. For example, they want to find a sentence that is similar in meaning within a full text (a document with a large amount of text, such as a manual). They want to find a chapter/section that is closest to the summary or main points of the chapter/section. They want to use a search engine in a way that "humans unconsciously draw out knowledge." etc.

The system builders are requesting a data structure that meets each of the above user requirements. To this end, we are responding to requests such as structuring data in a way that is suitable for searching, automating the labeling required for inference, implementing a method for inferring the meaning of a search statement using a search, executing a search using a data structure, and using search results that infer the meaning of a statement on the application side.

Figure 2 is an overall configuration diagram of a search system including a data registration processing device according to an embodiment. Figure 2 shows an example of the control work of the call center described above as a specific example of the system. Controller A accepts a call from informant B who is calling from an emergency phone or the like. The business support PC 201 is operated by controller A and presents information for responding to informant B.

The dialogue between controller A and informant B is acquired by telephone voice acquisition device 202, and converted into dialogue (text document) by voice acquisition PC 203, and recorded in processing server 204. In the illustrated example, voice acquisition PC 203 is configured to transfer the dialogue to voice-to-text server 205, and acquire a voice text document (dialogue) for the dialogue from voice-to-text server 205. Voice-to-text server 205 may be configured using server 205a for voice recognition and server 205b for voice learning.

The processing server 204 corresponds to the data registration processing device 100 shown in FIG. 1. Controller A makes a request to the processing server 204 for a knowledge search required for the business, and the processing server 204 responds to the business support server 206 with knowledge information on various business operations corresponding to the knowledge search request. For example, during a knowledge search, the processing server 204 generates a summary from the dialogue based on the above knowledge table 102, and outputs it to the business support server 206. The processing server 204 also outputs to the business support server 206 a dialogue text obtained by correcting a text document corresponding to the dialogue.

The business support server 206 performs business management processing by executing business applications required for the control work performed by controller A, and outputs a business management screen 210 to controller A's business support PC 201. The above system converts the contents of the call between controller A and informant B into text in real time. Controller A can perform control work while making a call in real time based on the text document of the dialogue and the summary text displayed on the business management screen 210 displayed on the business support PC 201.

Figure 3 is a functional block diagram of a search system including a data registration processing device. It shows the functions corresponding to the system configuration in Figure 2. The following mainly describes the functions of the processing server 204.

The processing server 204 has a voice data control unit 301, a knowledge search engine 302, a knowledge data conversion unit 303, and a knowledge database 304. The control functions of the processing server 204 (the knowledge search engine 302, the knowledge data conversion unit 303, etc.) correspond to the control unit 101 shown in FIG. 1. The knowledge database 304 includes the knowledge table 102 shown in FIG. 1.

The voice data control unit 301 acquires dialogue (text document) equivalent to real-time dialogue from the voice capture PC 203 and outputs it to the knowledge data conversion unit 303.

The knowledge data conversion unit 303 assigns a label to the dialogue (searched sentence) to enable semantic estimation. As described above, the knowledge data conversion unit 303 generates a summary from the dialogue (searched sentence) and records it as knowledge data in the knowledge database 304.

The knowledge search engine 302 performs processing related to searching for dialogue sentences. Based on a search model 320 corresponding to the business application of the business support server 206, the knowledge search engine 302 performs search processing etc. with the dialogue sentences (mainly summary sentences) recorded in the knowledge database 304 as the searched sentences. Then, the knowledge search engine 302 outputs the search results to the business application (business management screen 210) of the business support server 206. The search model 320 is created by a knowledge learning model creation unit 321, which creates and replaces other models (road model, police model, fire model in Figure 1) according to the business content of the client who has introduced the system.

In addition to the above configuration, as shown in FIG. 3, the voice data control unit 301 converts the entire dialogue into text and outputs the data-formatted corrected text (full text) to the source data recording unit 310. The data formatting is performed using a correction dictionary recorded in the knowledge database 304. The source data recording unit 310 records the corrected text (full text) in the knowledge database 304.

The knowledge database 304 has a relational database (RDB) structure. The knowledge data conversion unit 303 organizes the data structure of the knowledge database 304 into a data structure suitable for searching dialogue sentences (searched sentences), thereby enabling searches to be performed that are not dependent on the amount of text in the entire dialogue sentences, thereby improving search accuracy. For example, the knowledge data conversion unit 303 generates data such as the title, characteristic sentence, summary sentence, full text, etc. for each dialogue sentence (searched sentence).

The knowledge table 102 of the knowledge database 304 stores knowledge for generating summaries in advance. This knowledge is a group of information consisting of multiple texts and semantic labels. Using the knowledge table 102, the knowledge search engine 302 is able to search summaries as well as texts. For example, the business application (business management screen 210) of the business support server 206 performs search processes such as short sentence search other than text, title search, characteristic sentence search, and summary search.

The processing server 204 also has the functions of an alive monitoring unit 331 that performs health checks on each function, and a maintenance unit 332 for the knowledge database 304. It also has the functions of a knowledge data management unit 333 that archives knowledge data in the knowledge database 304, etc.

Figure 4 is a diagram showing an example of the hardware configuration of a data registration processing device. The data registration processing device 100 (processing server 204) shown in Figures 2 and 3 can be configured as a general-purpose server or the like made up of the hardware shown in Figure 4.

The data registration processing device 100 includes a CPU (Central Processing Unit) 401, a memory 402, a network interface (IF) 403, a recording medium IF 404, and a recording medium 405. 400 is a bus that connects each part.

The CPU 401 is an arithmetic processing device that functions as a control unit that is responsible for the overall control of the data registration processing device 100. The memory 402 includes a non-volatile memory and a volatile memory. The non-volatile memory is, for example, a ROM (Read Only Memory) that stores the programs of the CPU 401. The volatile memory is, for example, a DRAM (Dynamic Random Access Memory) or an SRAM (Static Random Access Memory) that is used as a work area for the CPU 401.

The network IF 403 is a communication interface for a network NW such as a LAN (Local Area Network), a WAN (Wide Area Network), or the Internet. The data registration processing device 100 is communicatively connected to the network NW via the network IF 403. For example, the data registration processing device 100 is communicatively connected to an external device, such as the voice capture PC 203 and the business support server 206 shown in FIG. 2, via the network NW.

The recording medium IF 404 is an interface for reading and writing information processed by the CPU 401 to and from the recording medium 405. The recording medium 405 is a recording device that supplements the memory 402, and may be a hard disk drive (HDD), a solid state drive (SSD), a universal serial bus (USB) flash drive, or the like.

The CPU 401 executes a program recorded in the memory 402 or the recording medium 405 to realize each function of the data registration processing device 100 (control unit 101). In addition, the memory 402 and the recording medium 405 store and hold information in a database of the data registration processing device 100, for example, the knowledge database 304 shown in FIG. 3, in an updatable manner.

In addition, devices other than the processing server 204 shown in FIG. 2, such as the business support PC 201, the voice capture PC 203, the voice-to-text server 205, and the business support server 206, can also be configured with hardware similar to that shown in FIG. 4.

(Data registration process by function)
5 to 7 are explanatory diagrams of the process flow for each function of the data registration processing device. Fig. 5 shows the process flow of an example of knowledge utilization of dialogue text by a business application of the business support server 206.

The voice during the dialogue is converted to text as dialogue by the voice-to-text server 205 via the voice capture PC 203. The voice data control unit 301 records the text of the dialogue (full text) in the knowledge database 304 (102) via the source data recording unit 310 as dialogue recording data (full text).

The voice data control unit 301 also transmits the text of the dialogue (full text) to the business application (business management screen) 210 of the business support server 206. The business application displays the dialogue (full text) on the business management screen 210.

Then, the business application requests the knowledge search engine 302 of the processing server 204 to perform semantic inference and keyword extraction (basic knowledge search) of the dialogue by operating the business management screen 210 (step S501). In this case, the session management unit 501 of the knowledge search engine 302 manages the request as a session using a Web API (Application Programming Interface) and executes a search process corresponding to the request.

The knowledge search engine 302 performs knowledge classification by executing a search process corresponding to the request, and returns basic business knowledge information to the business application (business management screen 210) (step S502). The business application automatically inputs the search results (business knowledge information) into the corresponding ledger section on the business management screen 210.

After that, when searching for past dialogues on the business management screen 210, the business application makes a request to the knowledge search engine 302 of the processing server 204 (step S503). In this case, the knowledge search engine 302 performs knowledge classification in response to the request for searching past dialogues, and returns business knowledge information of the past dialogues to the business application (business management screen 210) (step S504). The business application then displays similar examples of past dialogues and the business manual on the business management screen 210.

Figure 6 shows the flow of processing for creating knowledge data after the dialogue ends. This section explains the process for creating knowledge data performed by the knowledge search engine 302 and the knowledge data conversion unit 303 of the processing server 204 after the dialogue ends as shown in Figure 5.

When the voice data control unit 301 of the processing server 204 notifies the knowledge data conversion unit 303 of the end of the dialogue, the knowledge data conversion unit 303 reads the dialogue (full text) from the knowledge database 304 and creates knowledge data.

Knowledge database 304 records knowledge data, knowledge text, knowledge labels, knowledge text labels, etc. as knowledge table 102. The data structure of these will be described later. Based on the knowledge table 102, knowledge database 304 creates and outputs a model of search data to knowledge search engine 302.

First, the knowledge data conversion unit 303 converts the dialogue (full text) based on each table in the knowledge table 102 (step S601).

Next, the knowledge data conversion unit 303 infers the type of dialogue content of the dialogue based on a search request from the knowledge search engine 302 (step S602). Next, the knowledge data conversion unit 303 extracts specific keywords from the dialogue based on the basic knowledge in the knowledge table 102 (step S603). Then, the knowledge data conversion unit 303 performs document summarization of the dialogue (step S604). This document summarization includes, for example, the creation of a digest (summary) from the dialogue using general-purpose technology (for example, LexRank made by AI Corporation) and the creation of the above-mentioned summary.

Figure 7 shows an example of knowledge data output for a business application. Based on the business application of the business support server 206, for example, display items and operations on the business management screen 210, a search request for knowledge data for the acquired dialogue is output to the knowledge search engine 302 of the processing server 204.

The knowledge search engine 302 has general affairs knowledge information A to C for each business type through knowledge learning. The business knowledge information has a search model, a classification model, and an index/dictionary. The knowledge search engine 302 then implements search processes A to C corresponding to a search request, and outputs information such as the type of dialogue, past similar examples, and the target business manual to the business application. Note that the search process, business knowledge information, and business knowledge data are shown as A to C as an example, but are not limited to these three. In the example of Figure 7, for example, the learning model application unit 701 applies a learning model suitable for the business content, and selectively applies business knowledge information B and search process B corresponding to the search request.

The maintenance unit 332 also executes a specified maintenance tool to refer to the knowledge database 304, learns knowledge data A to C for each business, and creates business knowledge information A to C.

(About knowledge data structures)
Next, the knowledge data structure of the knowledge table 102 will be described. In the embodiment, the knowledge indicated by the knowledge data structure is a group of information composed of a plurality of texts and semantic labels. This knowledge data structure is structured so that not only text search but also summary search can be performed. Specifically, it is possible to specify the search target as text (for example, summary, title, etc. other than dialogue (full text)).

In the embodiment, the summary is not created artificially, but is generated using a specific method. In order to create this summary, a structure is created in which a label and a category indicating the meaning are assigned to the text and knowledge information containing that text. Specifically, one of nine standard labels for generating an overview of knowledge is assigned (the above "5W2H + action/state"). The structure is created in which a label and a category indicating the typical meaning of the text are assigned to the text.

Figure 8 shows the data structure of a knowledge table, and Figure 9 is a diagram showing a specific example of a knowledge table. In Figure 8, knowledge table 102 shows a total of six tables 801-806 as an ER diagram of a data structure in which multiple tables are associated with each other. In addition, the association status between items in each of tables 801-806 is indicated by 0 (no association), 1 (individual correspondence), or * (multiple correspondence).

The classification category table 801 is interrelated with the knowledge data table 802 and the knowledge text table 803. The classification category table 801 records information (ClassID, ClassType, ParentClassId, Level, Name) on the classification of words contained in the search sentence (dialogue sentence).

The knowledge data table 802 is interrelated with the classification category table 801, the knowledge text table 803, and the knowledge data label table 805. The knowledge data table 802 records the following information: KnowledgeId, Category, Title, Digest (digest text), Summary (summary text), and AllText (full text). Any of 5W2H + action/status is assigned to Category as the category to which the knowledge data applies. In addition, a label with a representative meaning, for example, the name of a car, is assigned to Title.

The knowledge text table 803 has interrelationships with the classification category table 801, the knowledge data table 802, and the knowledge text label table 804. The knowledge text table 803 records information on TextNo (ascending number of a part of text in the entire sentence), TextCategory (classification), TextName, and Text for each KnowledgeId. A simple word belonging to the knowledge data, for example, "XX is the name of a car," is assigned to Text. A semantic label (a label with a meaning equivalent to TextCategory) is assigned to TextName. A type that embodies the text category (5W2H + action/state) to which the knowledge text applies, for example, "Object (What) - Car," is assigned to TextCategory.

The knowledge text label table 804 has a relationship with the knowledge text table 803 and the label table 806. The information on TextNo, LabelID (semantic label ID), and LabelNo for each KnowledgeId is recorded.

The knowledge data label table 805 (corresponding to the knowledge data label 170 in FIG. 1) is interrelated with the knowledge data table 802. The knowledge data label table 805 records the following information: KnowledgeId, LabelType, TypeNo, and LabelItem.

The label table 806 has a relationship with the knowledge text label table 804. The label table 806 records information such as Label ID and Label (semantic label).

The knowledge table shown in FIG. 9 shows an example in which each table of the knowledge table 102 shown in FIG. 8 is applied to a certain piece of knowledge, for example, the dialogue of the air traffic control operation described above. Note that the description of the classification category table is omitted.

In the knowledge text table 803 in FIG. 9(b), a single piece of knowledge "K001" is set as the KnowledgeId, and a portion of the text in the entire sentence, for example, the ascending numbers "1" to "5" for each spoken sentence, is set as the TextNo. For example, in TextNo "1", "Hello, this is the Expressway Control Center. What's the matter?" is set as the Text (sentence), and since there is no equivalent semantic label in the related knowledge text label table 804, "" (NULL)" is set as the TextName, and "Other conversation" is set as the TextCategory. In addition, in TextNo "3", "Oh, it's a malfunction" is set as the Text (sentence). The TextName (name of the text) is set to "Label003 ("Fault")" in the related knowledge text label table 804 in Figure 9 (d), and the TextCategory is set to "Fault-related."

Here, in the label table 806 in FIG. 9(c), a Label (semantic label) for each LabelID is set, for example, "greeting", "contact", "malfunction", etc., which indicate the meaning of the dialogue. The knowledge text label table 804 in FIG. 9(d) refers to the label table 806, obtains the Label corresponding to the LabelID, and sets it as the TextName of the knowledge text table 803 in FIG. 9(b).

In the knowledge text label table 804 in FIG. 9(d), among the settings in the knowledge text table 803, TextNo, LabelID, and LabelNo are set for each KnowledgeId as information on the text to which a semantic label is assigned. For example, for TextNo "2" of KnowledgeId "K0001", LabelID "Label003 ("Fault")" and LabelNo "1" are set.

The knowledge data label table 805 in FIG. 9(e) is a table in which information for generating a summary sentence is set. In the knowledge data label table 805, the KnowledgeId "K0001" and the 5W2H + action/status consisting of LabelType "When", "Where" ..., "Action", and "Status" are set. In addition, TypeNo and LabelItem are set. For example, in the LabelType "Where" of KnowledgeId "K0001", there are two "Where"s included in the dialogue, and TypeNo "1" and "2" are set, respectively. The Label Item for Type No. "1" is set to "XX Road", and the Label Item for Type No. "2" is set to "Downbound N 29 (equivalent to the emergency telephone identification number)".

The knowledge data table 802 in FIG. 9(a) is interrelated with the classification category table 801, the knowledge text table 803, and the knowledge data label table 805, and records knowledge information obtained by knowledge learning. For example, the Category "Fault response - Level 2" and the Title "2020/08/08 Downstream N29 Fault" are set as the KnowledgeId "K0001" in the knowledge data table 802. A summary sentence based on the above general-purpose technology is set in Digest, a standard summary sentence (above summary sentence) consisting of 5W2H + action/status is set in Summary, and a dialogue sentence (full sentence) is set in AllText.

In the above-described knowledge table 102, each of the tables 801 to 806 has blank fields in the initial state, but information is set in the fields of each of the tables 801 to 806 as knowledge data is created for the dialogue for each search process.

Figure 9(f) is an example of the output of the knowledge table 102, in which a knowledge ID is assigned to each piece of information set in the knowledge data label table 805 in Figure 9(e), and the 5W2H+action/status LabelItems are rearranged in a predetermined order adapted to air traffic control operations. The processing server 204 creates a summary (Summary in Figure 9(a)) based on the output order of the LabelItems in the knowledge table 102 shown in Figure 9(f).

Figure 10A is a diagram showing search targets and label sorting for knowledge data. Figure 10A shows the same contents as the knowledge data table 802 shown in Figure 9(a).

Of the knowledge data in the knowledge data table 802 shown in FIG. 10A, Category is a classification target for text (dialogue), and the setting information "Road control - fault response" indicates the knowledge type label. Title "2020/08/08 Downbound N29 Fault" is the search target (search target 2) when searching for a title. Digest is a "summary" extracted from the dialogue using general-purpose technology, and is the search target for representative sentences (search target 3). Summary is a standard summary sentence (above summary sentence) of the above 5W2H + action/status, and is a summary search (search target 4). This summary sentence can eliminate semantic fluctuations by connecting the 5W2H + action/status in a specified order to make it consistent.

Figure 10B is a diagram showing the classification of search targets and labels for knowledge text. In Figure 10B, the settings of TextCategory and TextName for the knowledge text in the knowledge text table 803 shown in Figure 9(b) indicate semantic labels. In addition, the text is the search target (search target 1) during text search.

In a basic search, a text search of the above search target 1 is performed. This text search can handle sentences that have been divided into short sentences in conventional searches. It can also handle searches of knowledge titles (search target 2), summaries (search target 3), and overviews (search target 4). This makes it possible to handle long texts such as book-like chapters and sections of documents. In addition, by providing various label items (knowledge type labels, semantic labels), it becomes possible to determine the type of knowledge or the meaning of a sentence from the labels.

(Creating basic knowledge using knowledge data structures)
As described above, the basic knowledge is a large number of sentences with representative semantic labels attached to words and short sentences, and the "semantic type" and "semantic label" of a dialogue (text) can be obtained by searching the basic knowledge. Here, 5W2H + action/state is the "semantic type" that indicates a broad classification of the meaning of the sentence, and the text category and text name are the "semantic labels" that indicate the meaning of the sentence.

The number of knowledge elements contained in the dialogue (text) is also found using basic knowledge search. If multiple knowledge elements (knowledge data) are obtained through a search, then there will be that many knowledge elements (meanings). Each knowledge element has a label that indicates its meaning, making it possible to infer complex meanings.

Figure 11 is an explanatory diagram of data modeling of basic knowledge. In the knowledge table 102, basic knowledge is data modeled, and a data structure that allows knowledge to be associated with multiple items (dialogue sentences) is used (see Figure 8, etc.). The processing server 204 (knowledge search engine 302, knowledge data conversion unit 303) generates a large number of short sentences including general parts of speech and short sentences including industry-specific parts of speech as the setting contents (basic knowledge) of the knowledge table 102. For classification labels to classify categories, a list of multiple words that specify 5W2H + actions/states is created, and semantic labels are assigned to items 1 to N.

Then, the processing server 204 (knowledge search engine 302) creates various dictionaries, indexes, and models 1101 (search model 320 in FIG. 3) from the knowledge table 102 of basic knowledge. It also has a data structure in which knowledge is associated with multiple items 1 to N. Each item 1 to N in the knowledge table 102 is associated with specific knowledge items 1 to N in the various dictionaries, indexes, and models 1101 by an associated ID. Also, within the various dictionaries, indexes, and models 1101, associations are made between each piece of knowledge, and between item 2 of a certain piece of knowledge and other pieces of knowledge.

The above data model and data structure make it possible to build a search system that searches and classifies basic knowledge. In other words, by searching for sentences, it becomes possible to obtain classification labels for knowledge that contains similar sentences and semantic labels for similar sentences.

Figure 12 is a diagram showing specific examples of basic knowledge. Specific setting data for the control operations of the above call center is shown for the classification category table 801, knowledge data table 802, and knowledge text table 803 of the knowledge table 102.

The Title "Automobile - Car Type" of the Category "What" in the knowledge data table 802 is associated with multiple TextCategories "What-Automobile - Car Type", TextName, and Text in the knowledge text table 803 (1:N). Similarly, multiple TextCategories "What-Automobile - Car Name", TextName, and Text in the knowledge text table 803 are associated with one Title "Automobile - Car Name" in the Category "What".

The basic knowledge sentences (the above-mentioned knowledge table 102) are simple sentences that contain common words, so they can be easily created in large quantities. By looking at the "semantic type" and "semantic label" of sentences that are highly similar to the search sentence, it is possible to classify the meanings of the search sentence and the words contained in the search sentence. This makes it possible to assign the meanings defined by the computer and program of the data registration processing device 100 and to process each meaning.

In addition, most of the basic knowledge sentences can be reused regardless of the type of business to which they apply, except for certain technical terms. Furthermore, the amount of knowledge can be improved by adding knowledge sentences and semantic types and semantic labels.

(Semantic inference in search and classification processes)
The data entry processing device 100 (knowledge search engine 302) performs searching and classification in the search process (see FIG. 6, etc.). The knowledge search engine 302 can increase the variety of text search and classification (title search, summary search, short sentence/detailed search, etc.), and can stimulate use in various business operations.

In addition, for meaning inference, the knowledge search engine 302 is used in place of the human brain and is utilized in various business situations. By using the above knowledge data structure (knowledge table 102) and the knowledge search engine 302, meaning inference of business texts, for example, creating summaries and labeling dialogue texts, can be performed, thereby promoting the self-improvement of business knowledge.

Figure 13 is an explanatory diagram of an example of the use of basic knowledge search. For example, a business application (business management screen 210) causes the knowledge search engine 302 to perform a basic knowledge search for the text of speaker B in a business document (dialogue text), "I had a collision on the Chugoku Expressway" (step S1301). At this time, the knowledge search engine 302 refers to the basic knowledge (knowledge table 102) and obtains two search results.

Assume that the first result obtained through basic knowledge search is KnowledgeId "K0003", LabelType "Where", TextName "Expressway", and text "The Chugoku Expressway is an expressway." The second result obtained is KnowledgeId "K0012", LabelType "Action", TextName "Accident", and text "A collision accident is a type of accident." "Where" and "Expressway", as well as "Action" and "Accident" are semantic types (semantic labels). Furthermore, "The Chugoku Expressway is an expressway." and "A collision accident is a type of accident." are similar sentences to be searched.

Then, the knowledge search engine 302 responds with the search results for the business application (business management screen 210) with the semantic label "Where" being "Expressway" and the semantic label "Action" being "Accident" (step S1302).

This allows the business application (business management screen 210) to recognize that "Where" in the dialogue is "expressway" and "Action" is "accident." The business application (business management screen 210) can then assign semantic types and semantic labels to business documents (dialogue), automating the labeling of document classification.

In this way, the business application (business management screen 210) can obtain the semantic labels of the search text by searching the data registration processing device 100 (knowledge search engine 302) for text used in business from basic knowledge. By assigning semantic labels to the search text (business text, dialogue), it becomes possible to infer the meaning of the search text. In addition, multiple search results can be recognized as containing multiple different meanings of different types (5W2H + action/state). Then, the business application (business management screen 210) can use the obtained semantic labels to automatically classify business text using the labels.

(Automatic labeling of business text information and creation of summaries)
The data registration processing device 100 automatically labels business documents and extracts words with standard labels through a basic knowledge search. Here, the number and type of meanings contained in each sentence (short sentence) in the business document and the corresponding words are extracted. Then, standard labels of 5W2H + action/state are created to create an overview of multiple sentences (long sentences). This automatic labeling allows categorization of new business documents during business document search, for example, extraction of groups of similar business documents and categories.

The data registration processing device 100 also realizes a search for a long summary when creating a summary of a business document. That is, the data registration processing device 100 creates a summary from a business document (dialogue) using general-purpose technology, and also creates a summary using standard labels of "5W2H + action/status."

When labeling the business text and creating the summary, the data registration processing device 100 uses the search results of basic knowledge (knowledge table 102) to assign semantic labels to each sentence in the business text. Note that basic knowledge text is created as basic knowledge, and knowledge data is generated (knowledge table 102 is created) in advance.

As a result, the data registration processing device 100 can create digests and summaries of business documents by using the search results of basic knowledge. As a result, as described above, the use of summaries makes it easier to categorize new business documents. In addition, since the data registration processing device 100 mechanically and automatically assigns semantic labels to business documents, these can be used for subsequent document searches and classification. In addition, it becomes possible to determine closely related actions (such as selecting the type of business ledger or presenting a business manual) based on the results of the category classification.

Figure 14 is an explanatory diagram of an example of label extraction from a business document and knowledge data label creation. An example is a dialogue between a controller (speaker A) and a caller (speaker B) in the business application (business management screen 210) described above.

The knowledge data conversion unit 303 of the data registration processing device 100 outputs the dialogue 150 to the knowledge search engine 302 by searching each sentence of the dialogue 150 (step S1401). The knowledge search engine 302 refers to the basic knowledge sentences (knowledge table 102) (step S1402) and outputs the search results to the knowledge data conversion unit 303 (step S1403). For example, the search results for a text document (area D in the figure) containing Speaker A's "Oh, it's N29 downward, right?" include information on the meaning of the words "downward" and "N29 (emergency call)."

The knowledge data conversion unit 303 extracts meanings and words from the obtained search results (step S1404) to create a knowledge data label (the above-mentioned knowledge data label table) 805 (step S1405). It also creates a semantic label La (step S1406). The created semantic label La is recorded in the knowledge text table 803.

As the knowledge data conversion unit 303 continues to search for dialogue, words based on 5W2H + action/state are accumulated in the knowledge data label table 805, and can be used to create a summary.

The knowledge data conversion unit 303 also associates the created semantic label La (803) with the dialogue 150. Of the text of each (speaker A) in the dialogue 150, semantic labels are assigned to text that is necessary for business. For example, the text of speaker A in the dialogue 150, "Ah, it's N29 downward, right?" is assigned the semantic labels La "downward" and "N29". As shown in FIG. 14, the words associated with multiple semantic labels La correspond to areas K in multiple text documents of speaker A in the dialogue 150.

Figure 15 is an explanatory diagram of an example of automatic labeling of business documents and creation of a summary sentence. The data registration processing device 100 (knowledge search engine 302, knowledge data conversion unit 303) extracts a summary sentence 1501 from the dialogue sentence 150 (full sentence) using general-purpose technology. The data registration processing device 100 records the extracted summary sentence 1501 in the Digest of the knowledge data table 802 (step S1501). The data registration processing device 100 also sets the same category (in the illustrated example, "Road control - breakdown response") that was at the top of the search results obtained in the past dialogue sentence search to the Category of the knowledge data table 802. The Title can be set by extracting specific items of the knowledge data label (date, time, place, event, etc.), or it can be set manually.

The data registration processing device 100 also mechanically combines the labels in the created knowledge data label table 805 to create a summary 180. The data registration processing device 100 records the created summary 180 in the Summary of the knowledge data table 802 (step S1502).

The data registration processing device 100 also assigns a label to the text of the dialogue 150 and records it in the TextCategory and TextName fields of the knowledge text table 803 (step S1503).

Figures 16A and 16B are explanatory diagrams showing examples of the order of combining summary sentences. The data registration processing device 100 can change the order of combining summary sentences depending on the priority of the business type, etc. For example, Figure 16A (a) shows an example of creating a summary sentence corresponding to the above-mentioned traffic control business, Figure 16B (b) shows an example of creating a summary sentence corresponding to an emergency command business, and Figure 16B (c) shows an example of creating a summary sentence corresponding to a fire command business.

In the summary sentence for traffic control operations shown in FIG. 16A(a), i.e., in the knowledge data label table 805, the 5W2H + action/state is combined in the following order: When, Where, Who, What, Why, How, How much, state, action. In the summary sentence 1601 (180) combined in this order, Where has a higher priority after When, so the locations required for traffic control operations can be quickly identified.

In the knowledge data label table 805 for emergency command tasks shown in FIG. 16B(b), the 5W2H + action/status are combined in the following order: When, Why, Who, status, How much, Where, How, What, action. In the summary sentence 1602 (180) combined in this order, Why, Who, and status have high priority after When, so the status of people required for emergency command tasks can be known quickly.

In the knowledge data label table 805 for firefighting command duties shown in FIG. 16B(c), the 5W2H + action/status are combined in the following order: When, Why, Where, What, status, Who, How much, How, action. In the summary statement 1603 (180) combined in this order, Why, Where, and What have the highest priority after When, so the status of the firefighting target, which is necessary for firefighting command duties, can be known quickly.

(Data processing by data registration processing device)
17 is a flow chart showing an example of the overall process of creating knowledge data, which shows an example of the process of creating knowledge data (corresponding to FIGS. 14 and 15) performed by the CPU 401 of the data registration processing device 100.

First, the CPU 401 receives a notification of the end of a series of dialogues from the voice capture PC 203 (step S1701). Next, the CPU 401 creates an ID (knowledge ID) for the series of dialogues (step S1702). Next, the CPU 401 creates a record in the knowledge data table 802 (step S1703), and creates a record in the knowledge text table 803 (step S1704).

Next, the CPU 401 creates a knowledge data label table 805 (step S1705). Next, the CPU 401 creates a summary from the dialogue (step S1706) and records the summary in the knowledge data table 802 (step S1707).

Next, the CPU 401 creates a semantic label La (step S1708), records the semantic label La in the knowledge text table 803 (step S1709), and ends the above processing.

Figures 18A and 18B are flowcharts of an example of knowledge data label creation. They show a detailed example of the process of creating the knowledge data label table 805 (step S1705 in Figure 17) performed by the CPU 401 of the data registration processing device 100.

As shown in FIG. 18A, first, the CPU 401 acquires a dialogue (step S1801). Next, the CPU 401 executes a basic knowledge search using the dialogue as a search sentence (step S1802). This process corresponds to step S1401 of the short sentence search shown in FIG. 14.

Next, the CPU 401 determines whether there are any search results for the basic knowledge search of step S1802 (step S1803). If there are any search results (step S1803: Yes), the CPU 401 proceeds to the process of step S1804, and if there are no search results (step S1803: No), the CPU 401 ends the above process.

In step S1804, the CPU 401 executes a loop process the number of times equal to the number of search results (step S1804). In this loop process, the CPU 401 determines whether the score is equal to or greater than a constant value (step S1805). The score corresponds to the weight of the semantic label assignment to the word. If the score is equal to or greater than the constant value (step S1805: Yes), the CPU 401 proceeds to the process of step S1806, and if the score is less than the constant value (step S1805: No), the CPU 401 proceeds to the process of step S1807.

In step S1806, the CPU 401 creates a knowledge data label table 805 (step S1806). This process corresponds to step S1405 in FIG. 14. After this, the CPU 401 determines whether the loop processes have been executed for the number of search results (step S1807), and if there are any unexecuted processes, the process returns to step S1804, and if all have been executed, the process ends.

Figure 18B shows the details of the process of creating the knowledge data label table 805 in Figure 18A (step S1806). First, the CPU 401 identifies one of the nine categories (attributes) of the search statement (5W2H + action/state) (step S1811).

Next, the CPU 401 divides the dialogue into words by segmentation, normalization, etc. (step S1812). The CPU 401 then executes a loop process the number of times equal to the number of divided words (step S1813). In this loop process, the CPU 401 searches for words in the "word dictionary corresponding to the category" (step S1814).

Figure 18C is a diagram showing an example of a word dictionary. As shown in Figure 18C (a), the word dictionary is a dictionary in which known words are preregistered, and is a hierarchical structure of dictionaries corresponding to categories (5W2H + action/state) and dictionaries corresponding to text names (e.g., date, time, duration). For example, the branch of the category "When" has a hierarchy in which the text names "date," "time," and "duration" are connected from the top.

As shown in FIG. 18C(b), the "word dictionary corresponding to categories" 1820 contains registered words corresponding to a certain category "vehicle type," such as "large cargo vehicle, bus, passenger car, motorcycle, truck, 1-ton vehicle, 2-ton vehicle, rental car, 1-box vehicle, etc."

Returning to FIG. 18B, after processing step S1814, CPU 401 determines whether or not there is a matching word (step S1815). If the word is in the "word dictionary corresponding to the category" (step S1815: Yes), CPU 401 proceeds to processing step S1816, and if the word is not in the "word dictionary corresponding to the category" (step S1815: No), CPU 401 proceeds to processing step S1817.

In step S1816, the CPU 401 records the words for the identified category in the knowledge data label table 805 (step S1816). After this, the CPU 401 determines whether the loop processing has been executed for the number of divided words (step S1817), and if there are any unexecuted loops, the CPU 401 returns to the processing of step S1813, and if all loops have been executed, the CPU 401 ends the above processing.

Figure 19 is a flowchart of an example of summary creation. It shows a detailed example of the process of creating a summary 180 (step S1706 in Figure 17) performed by the CPU 401 of the data registration processing device 100. The CPU 401 obtains knowledge data labels from the knowledge data label table 805 created above (step S1901), and obtains a template corresponding to the summary combination definition (step S1902). This template defines the order of words recorded in the knowledge data label table 805, i.e., the output order of the summary, and is a template for creating a summary with the word order for each business content described in Figures 16A and 16B.

Next, the CPU 401 performs loop processing of the nine attributes of 5W2H+action/status (step S1903). For example, it is determined whether or not a knowledge label (knowledge data label, for example, "Where") of one of the nine attributes selected is in the knowledge data label table 805 (step S1904). If the result of the determination is that there is a knowledge label (step S1904: Yes), the CPU 401 proceeds to processing of step S1905, and if there is no knowledge label (step S1904: No), it proceeds to processing to determine whether the attribute loop processing has ended (step S1908).

In step S1905, a loop process for the number of labels is performed (step S1905). In this loop process for the number of labels, the CPU 401 reflects the knowledge label in the general combination definition (template) (step S1906). That is, the knowledge label is recorded in the template. After this, the CPU 401 judges whether the loop process for the number of labels has been performed (step S1907), and if there are any unexecuted attributes, the process returns to the processing of step S1905, and if all have been executed, the process proceeds to the processing of step S1908. In step S1908, it is judged whether the loop process for nine attributes has been performed (step S1908), and if there are any unexecuted attributes, the process returns to the processing of step S1903, and if all have been executed, the above processing ends.

Figures 20A and 20B are flowcharts of an example of semantic label creation. They show a detailed processing example of the creation (step S1708 in Figure 17) of semantic label La (see Figure 14) performed by the CPU 401 of the data registration processing device 100. First, as shown in Figure 20A, the CPU 401 acquires a dialogue (step S2001). Next, the CPU 401 executes a basic knowledge search using the dialogue as a search sentence (step S2002). This process corresponds to step S1401 of the short sentence search shown in Figure 14.

Next, the CPU 401 determines whether there are any search results for the basic knowledge search of step S2001 (step S2003). If there are any search results (step S2003: Yes), the CPU 401 proceeds to the process of step S2004, and if there are no search results (step S2003: No), the CPU 401 ends the above process.

In step S2004, the CPU 401 executes a loop process the number of times equal to the number of search results (step S2004). In this loop process, the CPU 401 determines whether the score is equal to or greater than a constant value (step S2005). If the score is equal to or greater than the constant value (step S2005: Yes), the CPU 401 proceeds to the process of step S2006, and if the score is less than the constant value (step S2005: No), the CPU 401 proceeds to the process of step S2007.

In step S2006, the CPU 401 creates a semantic label La (step S2006). This process corresponds to step S1406 in FIG. 14. After this, the CPU 401 determines whether the loop process has been performed for the number of searches (step S2007), and if there are any unexecuted loops, the process returns to step S2004, and if all loops have been executed, the process ends.

Figure 20B shows the details of the process of creating the semantic label table La (step S2006) in Figure 20A. First, the CPU 401 obtains the text name of the search result (step S2011).

Next, the CPU 401 divides the dialogue into words by segmentation, normalization, etc. (step S2012). The CPU 401 then executes a loop process the number of times equal to the number of divided words (step S2013). In this loop process, the CPU 401 searches for words in the "word dictionary corresponding to the text name" (step S2014). At this time, the word dictionary referenced has the same contents as the "word dictionary corresponding to the category name" 1820 shown in FIG. 18C (b).

After processing step S2014, the CPU 401 determines whether or not there is a matching word (step S2015). If the word is in the "word dictionary corresponding to the text name" (step S2015: Yes), the CPU 401 proceeds to processing step S2016. On the other hand, if the word is not in the "word dictionary corresponding to the text name" (step S2015: No), the CPU 401 proceeds to processing step S2017.

In step S2016, the CPU 401 records the word corresponding to the semantic label as semantic label La in the knowledge text table 803 (step S2016). After this, the CPU 401 determines whether the loop processing has been performed for the number of divided words (step S2017), and if there are any unexecuted loops, the process returns to step S2013, and if all loops have been executed, the process ends.

Figure 21 is an explanatory diagram showing an example of how a summary text is used by a business application. This explains an example of how the summary text is used when the business support server 206 executes the business application for the above-mentioned control business.

The business application of the business support server 206 adds the dialogue sentence 150 sentence by sentence and performs an overview search on the knowledge search engine 302 of the data registration processing device 100 (step S2101). The knowledge search engine 302 references the dialogue history sentence 2100 (step S2102) and outputs the search results to the business support server 206 (step S2103). For example, the overview search results include the text document "This is the Expressway Control Center. What's the matter? Oh, it's broken," the words "expressway," "control," "center," "how," "breakdown," and the Category "road control-breakdown."

The business application 206 performs category classification/search from the obtained search results (step S2104). At this time, the business application 206 can select the type of record book corresponding to the search result category "road control-breakdown" (step S2105), and perform the process of displaying the business manual-breakdown response (step S2106).

Figure 22 is a process flowchart of an example of using a summary by a business application. Figure 22 shows an example of summary creation executed by various business applications (CPU 401 installed in the business support server 206) shown in Figure 21 etc.

First, the CPU 401 acquires a dialogue (step S2201). Next, the CPU 401 executes a category search of the past dialogue history using the dialogue as a search query (step S2202). This process corresponds to step S2101 in FIG. 21.

Next, the CPU 401 determines whether or not there are any search results from the search executed in step S2201 (step S2203). If there are any search results (step S2203: Yes), the CPU 401 proceeds to the process of step S2204, and if there are no search results (step S2203: No), the CPU 401 ends the above process.

In step S2204, the CPU 401 executes a loop process the number of times equal to the number of categories (step S2204). In this loop process, the CPU 401 determines whether the score is equal to or greater than a constant value (step S2205). If the score is equal to or greater than the constant value (step S2205: Yes), the CPU 401 proceeds to step S2206, and if the score is less than the constant value (step S2205: No), the CPU 401 proceeds to step S2207.

In step S2206, the CPU 401 executes a category determination action (step S2206). This process corresponds to steps S2105 and S2106 in FIG. 21. The category determination action can perform processes such as: 1. Present the relevant section of the business manual to the user; 2. Determine the type of record ledger and start automatically inputting the search results into the selected record ledger; 3. Search for materials related to the business and present them to the user.

After this, the CPU 401 determines whether the loop processing has been performed for the number of categories (step S2207), and if there are any unexecuted processing, the process returns to step S2204, and if all processing has been performed, the process ends.

Figures 23 and 24 are diagrams showing examples of using business applications based on dialogue. Figure 23 shows an example in which the business support server 206 sets the display contents and display items required for control operations on the business management screen 210 based on the search results of the data registration processing device 100.

The business application of the business support server 206 performs a dialogue history search (step S2301), a basic knowledge search (step S2302), and a business manual search (step S2303) on the data registration processing device 100 (knowledge search engine 302).

For the dialogue history search (step S2301), the data registration processing device 100 (knowledge search engine 302) refers to the dialogue history data model 2301 and returns the search results to the business application. For the basic knowledge search (step S2302), the data registration processing device 100 (knowledge search engine 302) refers to the basic knowledge data model 2302 and returns the search results to the business application. For the business manual search (step S2303), the data registration processing device 100 (knowledge search engine 302) refers to the business manual data model 2203 and returns the search results to the business application. Here, the data registration processing device 100 (knowledge search engine 302) updates the dialogue history data model 2301 based on the new dialogue history from the business application.

The business application displays various display screens corresponding to the search results on the business management screen 210. For example, when the dialogue log/keyword input screen 2311 is displayed on the business management screen 210, the dialogue 150 and the keywords (indicated by thick lines in the figure) within the dialogue 150 are displayed as the search results of the data registration processing device 100. The data registration processing device 100 (e.g., the knowledge search engine 302) searches whether the dialogue 150 contains important keywords that have been set in advance, and the business application highlights keyword K from the search results on the business management screen 210.

The business application also automatically selects a record book (action item 2312) that corresponds to the dialogue content of the dialogue 150 (see step S2105 in FIG. 21) and displays it on the business management screen 210. The action item 2312 consists of, for example, "type," "date and time," "route," and "direction" (corresponding to each label) as confirmation and communication items for air traffic control work, and confirmed information (text corresponding to each label).

For example, the business application can display action items 2312 on the business management screen 210 based on the results of a basic knowledge search (step S2302) of the dialogue 150 in the data registration processing device 100 and a dictionary search for each item. The business application can then automatically input the search results into each input item of the action item 2312, such as confirmation and communication items for control operations, and confirmed information.

The business application can automatically select a business manual 2313 corresponding to the dialogue content from the search results of the business manual search (step S2303) of the dialogue 150 in the data registration processing device 100. Then, the selected business manual 2313 can be displayed on the business management screen 210.

The business application can also display a camera image 2314 containing image capture information on the business management screen 210 from the basic knowledge search (step S2302) of the dialogue 150 in the data registration processing device 100 and the search results of a camera-compatible dictionary. The camera image 2314 is an image captured by a camera corresponding to the "location" included in the dialogue 150.

As a result, the business application can display various display contents appropriate to the meaning of the dialogue 150 on the business management screen 210 in cooperation with the data registration processing device 100. Controller A, who performs management and control operations, can smoothly and efficiently carry out control operations for the other party (notifier B) while looking at the display contents displayed on the business management screen 210. For example, by having the business application automatically input various search results from the data registration processing device 100, the controller can eliminate input operations on the business management screen 210 and concentrate on the call with notifier B.

FIG. 24 also shows an example in which the business support server 206 creates a report on a control task on the task management screen 210 based on the search results of the data registration processing device 100. For example, assume that controller A operates the business application to display report 2401 on the task management screen 210. In this case, the business application can automatically input each input field of the report 2401 based on the basic knowledge search (step S2302) of the dialogue sentence 150 in the data registration processing device 100 and the search results of the dictionary for each item. For example, the business application automatically inputs the summary sentence 180 into the accident response report sentence 2411 of the report sentence 2401, and automatically inputs the text of the summary sentence 180 by label into each item of the report content 2412.

This allows Controller A to check whether he has missed any necessary information from Caller B while looking at the summary 180 entered into report 2401.

In the embodiment described above, the dialogue 150 has been converted into text based on a phone call, but this is not limited to the above. Even if the dialogue 150 is a chat message, a summary or the like can be generated by a similar knowledge search.

According to the embodiment described above, the data registration processing device estimates the meaning of the input text document and performs a search. This data registration processing program extracts words registered in a knowledge table that associates words, sentences containing the words, and attributes of the words from the text document based on sentences containing the words, rearranges a group of the extracted words in a predetermined order of attributes, creates a summary consisting of the rearranged word group, and records the summary in a database in association with the text document. This normalizes the entire text document and creates a summary summarized by the word group, thereby improving the search accuracy of the database in which the text document is recorded. By extracting words registered in the knowledge table based on sentences containing the words, a summary based on knowledge of 1:N correspondence other than 1:1 correspondence can be created, and the summary can clearly indicate the meaning of the text document.

The data registration processing device also includes a process for converting the voice of the dialogue into text through speech recognition and generating a text document. This makes it easy to indicate the meaning of the voice exchange in the dialogue.

The system may also include a process for acquiring the audio of the dialogue from the contents of a telephone conversation. This allows the audio that forms the basis of a text document to be generated from the contents of a telephone conversation. It also makes it possible to create a summary that corresponds to the contents of the conversation in real time.

The dialogue voice may also be obtained from the voice of the operator in the telephone conversation. In this way, by generating a text document based on the voice of the system operator rather than the other party, the accuracy of voice recognition can be improved without being affected by the quality of the call or the speaking habits of different parties.

In addition, by recording in advance in the knowledge table information on the semantic type and label of words that are highly similar to the text document, when extracting words, the meanings of words contained in the text document can be obtained based on the semantic type and label information. This means that each extracted word has a meaning, and a summary sentence made up of a group of words can provide a correct summary.

Attributes include the 5W2H (When, Where, Who, What, Why, How, How much), states, and actions, and words can be extracted by searching for documents that contain registered words for each attribute that have been pre-registered in the knowledge table. This allows the word order of the summary to be normalized, making it possible to indicate the meaning of the entire text document from just the summary.

The words can be rearranged in a predetermined attribute order that corresponds to the type of business for which the summary is used. For example, this allows for the attributes of the words that should be positioned at the beginning of the summary to be different for different types of business, such as traffic control, emergency command, and firefighting command, making it possible to efficiently carry out each different business.

In addition, when analyzing new text documents, the summary stored in the database will be included. This will improve the accuracy of searches using the database. The summary will eliminate semantic variation by keeping the meaning and order of words consistent. This will make it possible to perform summary searches of books, such as chapters and sections of long documents, and long dialogues.

Text documents may also be classified based on semantic type and label. This makes it possible to easily and accurately search and classify text documents.

It is also possible to create a summary of the entire text document, record the summary in the knowledge table, and use it for text search. Text search can be used to search for sentences in which the entire text has been divided into short sentences.

The summary can also be output to the device of the operator who is conducting the conversation. This allows the operator to obtain a summary that corresponds to the content of the conversation, and even if the conversation takes a long time (the amount of text document increases), the operator can obtain a summary that appropriately summarizes the entire text document, allowing them to carry out their work efficiently.

The system may also be configured to connect a voice-to-text device that converts the voice of a conversation into text using voice recognition, and a business device where an operator performs control tasks for the other party in the conversation, to the data registration device. This makes it possible to output a summary corresponding to the content of the conversation to the business device.

For these reasons, according to the embodiment, for example, an overview obtained by summarizing an entire dialogue as a group of words with normalized word order is an overview in which important words are extracted and noise can be removed. This makes it possible to improve the accuracy of searches in databases in which text documents such as dialogues are recorded. It also makes it possible to improve the efficiency of searches for long texts such as dialogues, enabling effective use of text.

The method for data registration processing described in the embodiment of the present invention can be realized by having a processor of a server or the like execute a prepared program. This method is recorded on a computer-readable recording medium such as a hard disk, flexible disk, CD-ROM (Compact Disc-Read Only Memory), DVD (Digital Versatile Disk), or flash memory, and is executed by being read from the recording medium by a computer. This method may also be distributed via a network such as the Internet.

The following additional notes are provided with respect to the above-described embodiment.

(Appendix 1) A data entry processing method for estimating semantic information contained in an input text document and retrieving and outputting the information,
extracting, from the text document, a word registered in a knowledge table in which the word, the sentence containing the word, and an attribute of the word are associated with each other, based on a sentence containing the word;
rearrange the extracted words in a predetermined order according to the attributes;
creating an overview consisting of the rearranged group of words, and recording the overview in a database in association with the text document;
A data registration processing method, the processing being executed by a computer.

(Additional Note 2) The dialogue speech is converted into text by speech recognition, and the text document is generated.
2. The data registration processing method according to claim 1, further comprising the steps of:

(Additional Note 3) The voice of the dialogue is obtained from the contents of a telephone conversation.
3. The data registration processing method according to claim 1 or 2, further comprising the steps of:

(Additional Note 4) The voice of the dialogue is obtained from the voice of the operator in the telephone conversation.
4. The data registration processing method according to claim 3, further comprising the steps of:

(Additional Note 5) The knowledge table stores information on the meanings and labels of words that have a high similarity to the text document,
In the word extraction, the meanings of words included in the text document are obtained based on the meaning type and label information.
5. A data registration processing method according to any one of claims 1 to 4.

(Additional Note 6) The attributes include 5W2H (When, Where, Who, What, Why, How, How much), states, and actions.
The extraction of the words is performed by searching for documents including the registered words for each of the attributes that are registered in advance in the knowledge table.
6. A data registration processing method according to any one of claims 1 to 5.

(Additional Note 7) The words are rearranged in a predetermined order of the attributes corresponding to the type of business in which the summary is used.
7. A data registration processing method according to any one of claims 1 to 6.

(Appendix 8) When analyzing the new text document, the summary stored in the database is included.
7. A data registration processing method according to any one of claims 1 to 6.

(Supplementary Note 9) Classifying the text document based on the semantic type and label.
9. The data registration processing method according to any one of claims 5 to 8, further comprising the steps of:

(Appendix 10) A data registration processing method according to any one of appendices 1 to 9, characterized in that a summary of the entire text document is created, the summary is recorded in the knowledge table, and used for text search.

(Additional Note 11) The summary is output to an apparatus of an operator who performs the dialogue.
11. The data registration processing method according to any one of claims 1 to 10, further comprising the steps of:

(Appendix 12) A data registration processing program for estimating semantic information contained in an input text document and retrieving and outputting the information,
extracting, from the text document, a word registered in a knowledge table in which the word, the sentence containing the word, and an attribute of the word are associated with each other, based on a sentence containing the word;
rearrange the extracted words in a predetermined order according to the attributes;
creating an overview consisting of the rearranged group of words, and recording the overview in a database in association with the text document;
A data registration processing program that causes a computer to execute a process.

(Appendix 13) A data registration processing device that estimates semantic information contained in an input text document and searches and outputs the information,
a knowledge table in which words, sentences containing the words, and attributes of the words are associated with each other in advance;
a control unit that extracts the words registered in the knowledge table from the text document based on a sentence including the words, rearranges a group of the extracted words in a predetermined order of the attributes, creates a summary consisting of the rearranged group of words, and records the summary in a database in association with the text document;
A data registration processing device comprising:

(Additional Note 14) The control unit
a speech-to-text conversion device that converts the text document into text by speech recognition of a dialogue;
A service device is connected to the other party of the conversation, and an operator controls the other party.
outputting the summary corresponding to the content of the dialogue to the business device;
14. The data registration processing device according to claim 13.

100 Data registration processing device 101 Control unit 102 Knowledge table 150 Dialogue text 160 Text label 170 Knowledge data label
180 Summary 204 Processing server 206 Business support server (business application)
210 Business management screen 302 Knowledge search engine 303 Knowledge data conversion unit 304 Knowledge database 401 CPU (control unit)
402 Memory 403 Network interface (IF)
404 Recording medium IF
801 Classification category table 802 Knowledge data table 803 Knowledge text table 804 Knowledge text label table 805 Knowledge data label table 806 Label table A Controller (speaker)
B. Reporter (speaker)

Claims (13)

A data entry processing method for estimating semantic information contained in an input text document and retrieving and outputting the information,
extracting, from the text document, a word registered in a knowledge table in which the word, the sentence containing the word, and an attribute of the word are associated with each other, based on a sentence containing the word;
rearrange the extracted words in a predetermined order according to the attributes;
creating an overview consisting of the rearranged group of words, and recording the overview in a database in association with the text document;
A data registration processing method, the processing being executed by a computer. Converting the dialogue into text through speech recognition and generating the text document;
2. The data registration processing method according to claim 1, further comprising the step of: Acquiring the audio of the dialogue from the contents of a telephone conversation;
3. The data registration processing method according to claim 2, further comprising a process of: The voice of the dialogue is acquired from the voice of the operator in the telephone conversation.
4. The data registration processing method according to claim 3, further comprising the step of: The knowledge table stores information on the meanings and labels of words that have a high similarity to the text document,
In the word extraction, the meanings of words included in the text document are obtained based on the meaning type and label information.
5. The data registration processing method according to claim 1, The attributes include 5W2H (When, Where, Who, What, Why, How, How much), state, and action;
The extraction of the words is performed by searching for documents including the registered words for each of the attributes that are registered in advance in the knowledge table.
6. The data registration processing method according to claim 1, the words are rearranged in a predetermined order of the attributes corresponding to a business type in which the summary is used;
7. The data registration processing method according to claim 1, When analyzing new text documents, the summaries stored in the database are included.
7. The data registration processing method according to claim 1, document classifying the text documents based on the semantic types and labels;
9. The data registration processing method according to claim 5, further comprising the step of: The data registration processing method according to any one of claims 1 to 9, characterized in that a summary of the entire text document is created, the summary is recorded in the knowledge table, and used for text search. outputting the summary to an apparatus of an operator who is to carry out the dialogue;
5. The data registration processing method according to claim 2, further comprising the step of: A data entry processing program that estimates semantic information contained in an input text document and searches and outputs the information,
extracting, from the text document, a word registered in a knowledge table in which the word, a sentence containing the word, and an attribute of the word are associated with each other, based on a sentence containing the word;
rearrange the extracted words in a predetermined order according to the attributes;
creating an overview consisting of the rearranged group of words, and recording the overview in a database in association with the text document;
A data registration processing program that causes a computer to execute a process. A data entry processing device that estimates semantic information contained in an input text document and searches and outputs the information,
a knowledge table in which words, sentences containing the words, and attributes of the words are associated with each other in advance;
a control unit that extracts the words registered in the knowledge table from the text document based on a sentence including the words, rearranges a group of the extracted words in a predetermined order of the attributes, creates a summary consisting of the rearranged group of words, and records the summary in a database in association with the text document;
A data registration processing device comprising:

Images