Information processing system, program, and information processing method

Abstract

To solve a problem that a system which resolves an input sentence into words and generates a response sentence has a heavy load, and pre-processing which removes unnecessary words is required, and provide an information processing system in natural language processing, a program, and an information processing method which enable a response in real time, without another pre-processing.SOLUTION: In an information processing system, an information processing method has: a reception step which receives a question sentence from a user; a collation step which takes out a noun phrase from the question sentence to collate the noun phrase with information associated with it; an estimation step which estimates a situation, an object, and an action in the noun phrase, and performs labeling; an extraction step which extracts the front and rear parts of them, about all portions estimated as the action of the associated information; and a creation step which connects the situation, the object, the action, and the part to create a response sentence.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an information processing system, a program, and an information processing method. [Background technology]

[0002] In response to inquiries from users via web pages, automated conversation programs using artificial intelligence, such as chatbots, are generally used. This is because human-assisted inquiry response cannot quickly respond to emergency inquiries from users outside of business hours. Another advantage is that users can handle initial troubleshooting themselves.

[0003] Patent Document 1 describes a system that breaks down an input sentence in a natural language into words, understands the meaning, and generates a response document based on that understanding. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Application No. 2021-534378 Summary of the Invention [Problem to be solved by the invention]

[0005] However, breaking down a message into individual words and understanding their meaning, as disclosed in Patent Document 1, places a heavy load on the entire system, and the system is not designed to immediately output the information needed for troubleshooting. Understanding the other party's intentions requires lengthy communication, and it is not capable of solving problems in real time.

[0006] Furthermore, to enable the system to be used in an automatic conversation program, it is necessary to reduce the time required by selecting and discarding phrases that do not need to be analyzed from the user's questions. Because a separate preprocessing step is required, the system is not efficient.

[0007] In view of the above circumstances, the present invention aims to provide an information processing system, program, and information processing method for natural language processing that enables the collection and response of necessary information in real time without requiring any additional preprocessing. [Means for solving the problem]

[0008] A first aspect of the present invention is 1. A natural language information processing system, comprising: It is configured to perform the following steps: In the reception step, a question is received from the user, In the matching step, a noun phrase is extracted from the question document, and the noun phrase is matched with information associated therewith; In the estimation step, the event, object, and action are estimated and labeled from the noun phrases; In the extraction step, the portions before and after each of all portions estimated as the treatment are extracted from the linked information, In the creation step, the event, the object, the action, and the part are connected together to create a response sentence.

[0009] A second aspect of the present invention is A method for processing information in natural language, comprising: It includes the following steps: In the reception step, a question is received from the user, In the matching step, a noun phrase is extracted from the question document, and the noun phrase is matched with information associated therewith; In the estimation step, the event, object, and action are estimated and labeled from the noun phrases; In the extraction step, the portions before and after each of all portions estimated as the treatment are extracted from the linked information, In the creating step, the event, the object, the action, and the part are connected together to create a response sentence.

[0010] This makes it possible to provide an information processing system, program, and information processing method for natural language that enables real-time responses by performing matching and simple estimation without requiring any additional preprocessing. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is an overall diagram of an information processing system according to an embodiment of the present invention; [Figure 2] FIG. 2 is a block diagram showing the hardware configuration of the server 3. [Figure 3] FIG. 2 is a block diagram showing functions realized by a server 3. [Figure 4] FIG. 2 is a flowchart of the present embodiment. [Figure 5] 10 is an example of labeling information created in this embodiment. [Figure 6] 6 is an example of a response message created from the example of labeling information shown in FIG. 5. DETAILED DESCRIPTION OF THE INVENTION

[0012] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described below with reference to the accompanying drawings. Various features shown in the following embodiments can be combined with each other.

[0013] Incidentally, the program for realizing the software appearing in this embodiment is The program may be provided as a non-transitory computer-readable medium, or may be provided so that it can be downloaded from an external server, or may be provided so that the program is started on an external computer and its functions are realized on a client terminal (so-called cloud computing).

[0014] In this embodiment, the term "unit" refers to, for example, hardware implemented by a circuit in a broad sense. hardware resources and software that can be specifically realized by these hardware resources. In addition, various types of information are handled in this embodiment, and this information can be represented by, for example, physical values ​​of signal values ​​representing voltage and current, high and low signal values ​​as a binary bit set consisting of 0 or 1, or quantum superposition (so-called quantum bits), and communication and calculations can be performed on a circuit in a broad sense.

[0015] In addition, a circuit in the broad sense is a circuit realized by at least an appropriate combination of a circuit, circuitry, processor, memory, etc. That is, it includes an application specific integrated circuit (ASIC), a programmable logic device (e.g., a simple programmable logic device (SPD), a complex programmable logic device (CPLD), and a field programmable gate array (FPGA)), etc.

[0016] 1. Hardware Configuration In Section 1, the hardware configuration of this embodiment will be described.

[0017] 1.1 Information Processing System 1 1 is an overall diagram of an information processing system according to this embodiment. The information processing system 1 includes user terminals 2 (e.g., user terminals 2-1, 2-2, . . . , 2-n) and a server 3, which are connected via a network. These components will be further described. Here, a system exemplified by the information processing system 1 is one that is made up of one or more devices or components.

[0018] 1.2 User terminal 2 The user terminal 2 is not particularly limited as long as it can call up the automatic response program stored in the server 3 via the network, and various devices such as terminal devices typified by edge devices and notebook PCs are possible. The connection to the network can be wired or wireless.

[0019] 1.3 Server 3 The server 3 is connected to the network and is configured to create a response to a question sent from the user terminal 2. The question sent from the user terminal 2 may be in a document format and is not limited to a specific document. Furthermore, the format, such as font, is not limited as long as the matching unit 332, which is one of the functions of the control unit 33 of the server 3, is able to extract a noun phrase from the question and match the noun phrase with the information linked to it.

[0020] 2 is a block diagram showing the hardware configuration of the server 3. The server 3 has a communication unit 31, a storage unit 32, and a control unit 33, and these components are electrically connected via a communication bus 30 inside the server 3. Each component will be further described below.

[0021] The communication unit 31 is preferably a wired communication means such as USB, IEEE1394, Thunderbolt, or wired LAN network communication, but may also include wireless LAN network communication, mobile communication such as 3G / LTE / 5G, Bluetooth (registered trademark) communication, etc. as needed. In other words, the communication unit 31 is a so-called network interface, and is more preferably implemented as a collection of the above-mentioned multiple communication means. The server 3 receives a question from the user terminal 2 via the communication unit 31, and transmits a response created by the control unit 33 to the user terminal 2. Details will be described later.

[0022] The memory unit 32 temporarily stores the question received from the user terminal 2. This can be implemented, for example, as a storage device such as a solid state drive (SSD) that stores various programs stored in the server 3 and executed by the control unit 33, or as a memory such as a random access memory (RAM) that stores temporarily required information (arguments, arrays, etc.) related to program operations. It may also be a combination of these. In particular, the memory unit 32 stores the question received from the user terminal 2 until the "noun phrase matching check" described in step S104 of FIG. 4 is performed and the "response sentence is created and sent" described in steps S109 and S110 of FIG. 4 is performed.

[0023] The control unit 33 processes and controls the overall operations related to the server 3. The control unit 33 is, for example, a central processing unit (CPU) not shown. The control unit 33 recognizes a question input by a user by reading a predetermined program stored in the storage unit 32. In other words, information processing by software stored in the storage unit 32 is specifically realized by the control unit 33, which is an example of hardware, and can be executed as a functional unit included in the control unit 33. These will be described in more detail in the next section. Note that the control unit 33 is not limited to being a single unit, and multiple control units 33 may be provided for each function. A combination of these may also be used.

[0024] The control unit 33 also reads out a predetermined program stored in the storage unit 32, and stores the question received from the user terminal 2 as data to be analyzed without any special preprocessing. Then, it extracts a noun phrase from the question, and compares the noun phrase with the information associated with it to infer and label the event, object, and action from the noun phrase. The labeled information is temporarily stored in a separate area of ​​the storage unit 32. Details will be described later.

[0025] After the estimation and labeling are completed, the control unit 33 extracts all the parts before and after the description labeled as "treatment" from the linked information. Then, the control unit 33 connects the "event," "target," "treatment," and "location" to create a response sentence. Details will be described later.

[0026] The created response sentence is then sent to the user asking the question. Regarding the labeled information stored in a separate area of ​​the storage unit 32, when the "response sentence is cumulatively saved as linked information" process described in step S112 of FIG. 4 is performed, the labeled information is updated and generated as training data linked to the response sentence as a title or header. Note that the training data may be updated and generated using only a portion of the data, and may be selective.

[0027] 2. Functional configuration This section describes the functional configuration of this embodiment. As described above, information processing by software stored in the storage unit 32 is specifically realized by the control unit 33, which is an example of hardware, and can be executed as a functional unit included in the control unit 33.

[0028] 3 is a block diagram showing functions realized by the server 3 (control unit 33). Specifically, the server 3 (control unit 33) includes a reception unit 331, a collation unit 332, an estimation unit 333, an extraction unit 334, a creation unit 335, and a transmission unit 336.

[0029] The receiving unit 331 is configured to receive a question from a user. For example, the receiving unit 331 receives a question posted on the web from the user terminal 2 or an e-mail sent directly.

[0030] The matching unit 332 is configured to extract a noun phrase from a question sent from the user terminal 2 and match the noun phrase with information linked to it. The linked information is, for example, a collection of responses individually created in response to questions previously submitted by users, i.e., a collection of so-called Q&As. The matching checks whether the noun phrase is included in the linked information.

[0031] If the above noun phrases are found in the linked information after the above matching, the estimation unit 333 estimates them as parts corresponding to "event," "object," and "treatment." Then, as will be described later, while separately considering the existence of synonymous phrases of the above noun phrases, their overlap, frequency of appearance, etc., the estimated parts are labeled as "event," "object," and "treatment" as in the example shown in Figure 5.

[0032] On the other hand, if the result of the comparison shows that the noun phrase is not included in the linked information, it is possible to compare again with a phrase that is synonymous with the noun phrase set in advance.

[0033] Then, by checking the range of synonyms that have been set and by repeatedly checking, the parts that correspond to "situation," "object," and "treatment" are inferred, and these inferred parts are labeled as "situation," "object," and "treatment," as in the example shown in Figure 5.

[0034] The linked information may be based on learned data such as a collection of so-called Q&As that have been learned in advance as training data. Furthermore, the trained data may be generated or updated by further adding, as training data, the prepared response sentences described below or the above-mentioned response sentences individually prepared in response to questions previously submitted by users.

[0035] The timing for generating or updating the learned data is when the "response sentences are cumulatively saved as linked information" described in step S112 of Fig. 4 is performed together with the labeled information stored in a separate area of ​​the storage unit 32. Similarly, the generation or update of the teacher data may be performed using only a portion of the response sentence data, and the labeled information may be selective.

[0036] Furthermore, it is more preferable that the estimation and labeling in the estimation unit 333 be performed by machine learning based on the above-mentioned training data. There are no particular limitations on the machine learning algorithm, and k-nearest neighbor method, logistic regression, support vector machine, neural network, topic model, Gaussian mixture model, etc. may be appropriately adopted.

[0037] The extraction unit 334 extracts all parts before and after the part labeled as "treatment" from the linked information. In order to reduce the load on the creation unit 335 (described later) and to improve the readability of the response sentence, it is possible to not extract some of the duplicated parts labeled as "treatment." The program stored in the storage unit 32 can also be set in advance to eliminate duplicates.

[0038] The creation unit 335 creates a response sentence by adding the extraction results to the labeled parts "event," "target," and "action" as shown in Figure 5. Because "event," "target," and "action" are set as words or phrases, it is preferable to have a separate automatic creation program to combine them with the extracted parts to create a highly readable document. Considering the time required for document creation, it is more preferable that the administrator of the server 3 be able to select whether or not to launch the automatic creation program. When machine learning is used in the estimation unit 333, it is even more preferable to use machine learning such as a neural network for document creation, as this increases the compatibility of the entire system.

[0039] The sending unit 336 is configured to send the response text created above to the user. When the response text is sent in document format as an attachment to an e-mail or the like, it is desirable that a program for automatically creating a title for the e-mail or the like be stored in the storage unit 32. This improves work efficiency. Of course, the so-called automatic title creation program and the above-mentioned automatic document creation program may be operated together.

[0040] 3. Information Processing Method This section describes the information processing method of the information processing system 1 in natural language described above. This information processing method includes the following steps: In the reception step, a question sentence is received from a user. In the matching step, a noun phrase is extracted from the question sentence, and the noun phrase is matched with the information linked to it. In the estimation step, an event, object, and action are estimated from the noun phrase and labeled. In the extraction step, the parts before and after all parts of the linked information that are estimated as the action are extracted. In the creation step, a response sentence is created by connecting the event, object, action, and part.

[0041] 4 is a flowchart showing the flow of information processing in natural language executed by the information processing system 1. Below, each flow of this flowchart will be explained, particularly taking as an example a case where the question is about an internal combustion engine.

[0042] When a user asks a manufacturer a question about a problem with an internal combustion engine, the user creates a question on the web as shown in Fig. 5 (step S101) and sends it via a network to a server 3 owned by the manufacturer. The manufacturer accepts the question from the user sent to server 3 (step S102). As long as the question is sent via a network, it is not limited to being accepted by an automated conversation system such as a chatbot, and it may also be accepted by email or the like.

[0043] The server 3 performs a morphological analysis on the received question from the user (step S103). After that, the server 3 is configured to compare the morphologically analyzed results, particularly noun phrases, with the information linked to them. The linked information may be, for example, a collection of responses individually created in response to questions previously submitted by users, such as a collection of so-called Q&As. The comparison involves checking whether the noun phrase is included in the linked information (step S104).

[0044] If the question is about a cylinder liner in an internal combustion engine, first, linking information containing the word "cylinder liner" is extracted. Then, the presence or absence of the word "wear" contained in the noun phrase extracted as a result of the morphological analysis is checked against the extracted linking information. If the word "wear" is found in the linking information as a result of the check, "cylinder liner" is set as the "object" and "wear" as the "event" (step S105).

[0045] In this way, the part of the linked information containing "cylinder" and "wear" that is estimated as the "treatment" part is selected from files or documents. Specifically, verb phrases and adjective phrases are estimated and selected. In this example, "cylinder liner" is set as the "object" and "wear" as the "event," but in this case, the words "appreciate," "confirm," "judge," "suspect," "recommend," and "set" are inferred and selected as the "action" (step S107).Then, the relevant parts are labeled as "event," "object," and "action."

[0046] Furthermore, when the linked information is a collection of so-called Q&As, the word "thank you," which is estimated as a "treatment" as described above, appears in almost all "situations" and "objects." On the other hand, words such as "judgment" that are estimated as "actions" as described above tend to be expressed only in relation to things that can be seen with the naked eye in an internal combustion engine, and it is preferable to make a difference in the order and number of times that they are written as "actions" when creating the response sentences described below.

[0047] In step S105, if the word "wear" contained in the extracted noun phrase is not included in the linking information as a result of the morphological analysis as described above, it is possible to include "wear and tear" as a synonym of "wear" and perform matching again (step S106). The setting range of synonyms and the number of loops in step S106 when the synonym is not included in the linking information may be freely set by the administrator of server 3.

[0048] Once the "event," "target," and "action" have been labeled as shown in Figure 5, all of the information before and after the "action" part is extracted from the linked information (step S108). These are then combined to create the example response shown in Figure 6 (step S109), which is then sent to the user as a response (step S110). This makes it possible to create a response relatively quickly using simple estimations such as those described above, without requiring any additional preprocessing. Figure 6 shows an example response created based on the example labeling information shown in Figure 5, where the "target" is "cylinder liner" and the "event" is "wear."

[0049] When an internal combustion engine problem like the one described above occurs on board a ship without a communication environment, a simple chatbot is needed to enable crew members to quickly troubleshoot the problem, making this information processing method particularly useful.

[0050] As mentioned above, the information linked to the labeling information is like a trouble case about a product that the manufacturer has and that is related to questions from users. There are no restrictions on the format or data format as long as the above matching is possible.

[0051] 4, past responses from other users may be added to the linking information. This makes it possible to improve the accuracy of the noun phrase matching check in step S104 and the re-matching in step S106. Furthermore, since "event," "object," and "treatment" are set by words or phrases when creating a response, it is preferable to automatically cut out overlapping parts in order to combine the above labeling information parts to create a highly readable document.

[0052] The user accepts the response sent in this way (step S110), and if no new questions arise about the response content, the process ends. On the other hand, if a new question arises, the process starts again from step S101 and is repeated in the same way.

[0053] When the linked information is generated or updated as training data in machine learning, it is stored in the storage unit 32 (step S112) at the same time as step S110. At this time, the question sentence from the user that was temporarily stored in the storage unit 32 has been deleted, and it goes without saying that the linked information is generated or updated as training data using the items labeled by the above estimation as so-called headers.

[0054] This completes the process (step S113), and the process continues in a loop from step S102, where a question from the user is accepted. This is particularly useful as an information processing method that can quickly answer questions from linked information that aggregates so-called Q&As, in situations where high-performance computer resources cannot be used.

[0055] 4.Other The information processing system 1 according to this embodiment may adopt the following aspects: In the information processing system 1, the estimation and labeling in the estimation step are based on learned data in which the linked information is learned in advance as training data. This not only increases the probability of matching in the verification check (step S104), but also improves the accuracy of the estimation. Furthermore, this is more preferable because it eliminates the variations that occur when a response sentence is created by a person.

[0056] The teacher data is generated or updated in the information processing system 1 using the above-mentioned machine learning. By updating the teacher data, response sentences are accumulated as linked information as in step S110, which is more preferable in that it not only increases the matching probability of the above-mentioned comparison check (step S104), but also improves the accuracy of the above-mentioned estimation.

[0057] In the information processing system 1, the linked information also includes the results obtained by text mining the response sentence created in step S109 and the response sentence created separately in the past. The linking information is associated with noun phrases extracted from the user's question, but it is preferable if it is possible to examine from the results of text mining whether the so-called title or header part is the noun phrase in question, as this will improve the accuracy of the above estimation.

[0058] In the information processing system 1, the question text and the linked information are composed of an email sent from the user and a response text to an email that was separately created in the past. Unlike typical character limit data, this method is preferable because it allows for more information to be collected and improves the accuracy of the estimation. Questions sent directly to the manufacturer via email are also preferable because they can be easily accumulated as training data for machine learning without requiring special attention.

[0059] Furthermore, since emails are written in colloquial language, they are exchanged repeatedly and have a variety of writing styles, but it is relatively easy to extract noun phrases using the information processing method described above. It is also suitable for the text mining method described above, which makes it possible to create a wider range of linking information, making it even more preferable.

[0060] Furthermore, when so-called Q&A data is used as the linking information, it is even more preferable because it makes it possible to automatically generate the linking information from trouble cases that have occurred in the past through email or other exchanges.

[0061] Finally, while various embodiments of the present invention have been described, these are presented by way of example only and are not intended to limit the scope of the invention. The novel embodiments may be embodied in various other forms, and various omissions, substitutions, and modifications may be made without departing from the spirit of the invention. Such embodiments and modifications are intended to be included within the scope and spirit of the invention, as well as within the scope of the inventions and their equivalents as defined in the accompanying claims. [Explanation of symbols]

[0062] 1: Information processing system 2: User terminal 3: Server 30: Communication bus 31: Communications Department 32: Storage section 33: Control section 331: Reception 332: Matching section 333: Estimation section 334:Extraction part 335: Creation Department 336: Transmission unit

Claims

1. 1. A natural language information processing system, comprising: It is configured to perform the following steps: In the reception step, a question is received from the user, In the matching step, a first noun phrase and a second noun phrase are extracted from the question sentence by morphological analysis, and the second noun phrase is matched with information linked to the first noun phrase; In the estimation step, a part corresponding to an event, an object, or a procedure is estimated from the linked information including the second noun phrase identified by the matching, and the estimated part is labeled; In the extraction step, for all parts estimated as the treatment from the linked information, parts before and after the treatment are extracted; In the creation step, the event, the object, the action, and the part are connected to create a response sentence; The estimation and labeling in the estimation step are based on learned data in which the linked information is learned in advance as training data.

2. 2. The information processing system in natural language according to claim 1, the estimating step further comprises a learning step; In the learning step, the created response sentences and response sentences created separately in the past are further added as training data to generate or update the learned data.

3. In the information processing system for natural language according to any one of claims 1 to 2, The linked information also includes the results obtained by text mining the created response sentence and the response sentence created separately in the past.

4. In the information processing system for natural language according to any one of claims 1 to 2, The question and the linked information consist of an email sent by the user and an email response created separately in the past.

5. 4. The information processing system in natural language according to claim 3, The question and the linked information consist of an email sent by the user and an email response created separately in the past.

6. A method for processing information in natural language, comprising: It includes the following steps: In the reception step, a question is received from the user, In the matching step, a first noun phrase and a second noun phrase are extracted from the question sentence by morphological analysis, and the second noun phrase is matched with information linked to the first noun phrase; In the estimation step, a part corresponding to an event, an object, or a procedure is estimated from the linked information including the second noun phrase identified by the matching, and the estimated part is labeled; In the extraction step, for all parts estimated as the treatment from the linked information, parts before and after the treatment are extracted; In the creation step, the event, the object, the action, and the part are connected to create a response sentence; The method, wherein the estimation and the labeling in the estimation step are based on trained data in which the linked information is trained in advance as training data.

7. 7. The information processing method according to claim 6, the estimating step further comprises a learning step; In the learning step, the created response sentences and response sentences created separately in the past are further added as training data to generate or update the learned data.

8. In the information processing method for natural language according to any one of claims 6 to 7, The method, wherein the linked information also includes results obtained by text mining the created response sentence and the response sentence created separately in the past.

9. In the information processing method for natural language according to any one of claims 6 to 7, The method, wherein the question and the linked information are composed of an email sent by the user and a response to an email that was separately created in the past.

10. 9. The method for processing information in natural language according to claim 8, The method, wherein the question and the linked information are composed of an email sent by the user and a response to an email that was separately created in the past.

11. A program, A method for making a computer execute each step of the information processing system in natural language according to any one of claims 1 and 2.

Images