JP6526470B2 - Pre-construction method of vocabulary semantic patterns for text analysis and response system - Google Patents

Description

  The present invention relates to a method of constructing LSP (Lexico Semantic Pattern) knowledge that forms the basis of a system for analyzing and responding to the meaning of natural language texts, and in particular to LSP knowledge for speech recognition systems. It relates to the construction method of

  Technologies that recognize and respond to human speech by machines are applied to various fields of real life. A system is known in which a machine (smartphone) recognizes and responds to human voice or executes various control commands, such as Apple's Siri (registered trademark) and Google Now (Now). There is. Such a system analyzes the input sentence of the user based on text mining technology to grasp the meaning, and generates and outputs a response according to the intention. This is used not only in Siri and Now, but also in various systems of artificial intelligence systems such as robot systems, natural language processing systems such as keyword extraction, text summarization etc.

  In order to analyze the user input text, the question and answer system goes through a natural language analysis process such as morpheme and syntactic analysis. Such research has been conducted for a long time, and has been developed in pattern-based analysis and statistical-based analysis. Among them, pattern-based analysis analyzes sentences by converting words, morphemes and syntax that appear repeatedly in various sentences into patterns of LSP (lexical semantic pattern) forms and giving meaning to the corresponding LSP. Patent Document 1 is a lexical-semantic pattern reconstruction method for recognizing Korean syntax using LSP technology, and this patent has been completed by the inventor of the present invention.

  As a pattern-based analysis method, LSP is a grammatical rule that expresses information such as vocabulary, morpheme, part of speech and syntactic structure, and is well known as a natural language processing method. LSP technology helps to enable one-dimensional syntax analysis for natural languages where syntax analysis is not easily performed. In such a pattern-based analysis method, the system administrator should define the knowledge of LSP in advance in accordance with the input question (question). This is not merely a problem of LSP knowledge construction. When the text analysis process for the input query is finished, the corresponding information is used to go through a response generation process. Therefore, if responses to a question and answers are not constructed in advance efficiently, there is a risk that the provision of response information may fail or an incorrect response may be output.

  In short, in order to provide a high quality service using a text analysis and response system, LSP which is basic knowledge for analyzing questions and answers and response data conforming to the questions should be constructed in advance. Therefore, the inventor of the present invention completed the present invention after long researches on how to efficiently construct the construction of LSP knowledge.

Republic of Korea Patent No. 1409298

  An object of the present invention is to provide a multi-step LSP knowledge construction method to efficiently respond to user queries. It tries to build an environment that provides an effective question and answer system service. Meanwhile, other objects of the present invention, which are not specified, should be considered additionally within the scope of the following detailed description and the effects thereof.

In order to solve such problems, the present invention is a method of pre-building lexical semantic patterns for text analysis and response systems,
A method of pre-building lexical semantic patterns for a text analysis and response system performed by a manager terminal of a question and answer system, comprising:
(A) defining in advance a concept which is a set to which a vocabulary semantic pattern to be matched with an input sentence of the questioner terminal belongs;
(B) collecting sample data which is a target of a lexical meaning pattern and classifying it according to the concept;
(C) the building concept a basic unit is semantic feature semantic feature dictionary defines the (Semantic Description feature) that constitutes the means of one or more entries as a word having the same meaning, one belonging to each semantic feature Structuring into a set of
(D) constructing the lexical semantic patterns for recognizing the sample data by combining predetermined grammatical expressions using the semantic features;
(E) pre-constructing response data responsive to the input sentence of the questioner's terminal for each of the concepts.

  Also, in the method of pre-building lexical semantic patterns for a text analysis and response system according to a preferred embodiment of the present invention, the concept of the step (a) preferably has a hierarchical structure.

  Also, in the method of pre-building lexical semantic patterns for a text analysis and response system according to a preferred embodiment of the present invention, a concept is added if a concept corresponding to sample data is not defined in the step (b). Or further including the step of correcting.

  Also, in the method of pre-building lexical semantic patterns for a text analysis and response system according to a preferred embodiment of the present invention, the lexical semantic patterns of the step (d) are grammatical expressions for recognizing the sample data and the steps Use the semantic features of (c).

  Also, in the method of pre-building lexical semantic patterns for a text analysis and response system according to a preferred embodiment of the present invention, the response data of the step (e) should be changed according to the input sentence of the questioner terminal Is preferably designated as a variable.

  According to the present invention, there is an advantage that LSP knowledge of the question and answer system can be constructed effectively. Of course, knowledge management and maintenance can be performed efficiently. On the other hand, even if effects are not explicitly mentioned here, the effects described in the following specification expected by the technical features of the present invention and the provisional effects thereof are described in the specification of the present invention. It is added that it is treated the same as yours.

It is a block diagram showing an example of composition of a system in a certain scenario using a question and answer system based on LSP (lexical semantic pattern). Figure 2 is a flow chart schematically illustrating the overall process of the invention according to a preferred embodiment. It is a figure which shows the example of a screen structure of an administrator terminal in constructing | assembling a caption by this invention. It is a figure which shows an example of the semantic feature dictionary table 200 which defined the semantic feature by the method of this invention. It is a figure which shows the structural example of the entry table 201 with respect to the 500th semantic feature "meeting" of FIG. It is a figure which shows the structural example of the LSP construction | assembly table 300 produced | generated by the method of this invention. Incidentally, it is clarified that the attached drawings are illustrated as a reference to assist the understanding of the technical idea of the present invention, and the scope of the present invention is not limited thereby.

  Hereinafter, specific contents for carrying out the present invention will be described with reference to the attached drawings. Then, in the description of the present invention, when it is determined that the related known functions are obvious to those skilled in the art and there is a fear that the gist of the present invention may be unclear. I omit it.

  FIG. 1 illustrates a scenario of usage scenarios of a question-and-answer system based on LSP (lexical semantic pattern). Particularly, the illustrated example is mainly the case where the input sentence of the questioner is formed by speech, but the LSP knowledge construction method of the present invention is used even when the input sentence of the questioner is input not by speech but by text. Can.

  When the questioner accesses the user device 10 and inputs an input sentence, it is converted into text through the speech recognizer 11 built in the user device 10. The question-and-answer system 12 of the user device 10 analyzes the input sentence using the pre-established LSP knowledge information, and presents the question-and-answer person with the response data matching the question via the response output unit 13.

  Data and program code necessary for driving the question and answer system 12 are built in the storage device. In a preferred embodiment of the present invention, it is built on the memory of the user device 10 as shown in FIG. Also, in another preferred embodiment of the present invention, the storage device in which the question and answer system 12 is built is located outside the user device 10, in which case the user device 10 is connected to an external storage device via network communication. The response data is output to the questioner using the question and answer system 12 which is resident permanently.

  The method of the present invention relates to a method of gradually establishing LSP knowledge constituting the question and answer system 12 in advance, which is implemented by the manager terminal 20. The LSP knowledge described below includes databases used to analyze the questioner's input sentences and to extract response data and software modules for using the data stored in those databases. On the other hand, when actually using LSP knowledge constructed in advance according to the present invention, the functions and operational relationships of hardware and software in various modifications of FIG. 1 or FIG. It includes transformation and various improvements.

  FIG. 2 illustrates the overall process of the LSP knowledge construction method according to a preferred embodiment of the present invention. This is also the process of LSP dictionary construction method for text analysis and response system. Each of these steps is performed by an administrator terminal which is a computing system in which hardware / software modules are built.

  First, construct a concept (S10). A concept acts as a set to which lexical semantic patterns belong, and is a unit that determines which response to take for an input sentence. That is, the response sentence output to the user is a response sentence registered in a concept to which an arbitrary LSP matched with the input sentence of the questioner terminal belongs.

  Also, preferably, the concept defined and constructed in S10 has a hierarchical structure. FIG. 3 shows the concept generation screen 100 and shows that a plurality of concepts form a hierarchical structure. For example, while defining the common concept, period, time, place, attendee and filtering concepts are defined as subcategories thereof, and period, time, place and attendee are further defined and registered in the filtering concept. In this way, in constructing a concept, the meaning of sentences is classified and defined while forming a hierarchy from a large category meaning expression to a detailed meaning. Each concept may or may not have an LSP. However, in the present invention, the LSP generated below necessarily belongs to the concept.

  By pre-structuring so that many LSPs belong to a concept in this way, LSPs become a set of concepts. Therefore, it is possible to manage more effectively by bundling LSPs capable of analyzing text of similar content into one concept.

  In order to construct an LSP belonging to each concept, it is necessary to secure sample data which is a target text. Collect sample data and classify it according to the concept S20. The more sample data is collected, the more sophisticated the concept and LSP can be constructed. This has a direct impact on the performance of the question and answer system. The collected sample data is classified according to the constructed concept, but if it is difficult to classify the collected sample data into a specific concept, ie, there is no concept corresponding to the collected sample data Add or modify concepts in

For illustrative purposes, sample data such as the following sentence is illustrated:
(A) "Teach me a cafeteria that has good lunches for team meetings"
(B) "I'm going to do a task work shock, but what is the best?"
(C) "I have to make a schedule for the task meeting ..."

  In order for the question and answer system to be effectively implemented, it is necessary to structure vocabulary having different but the same meaning. Therefore, a semantic feature dictionary, which is a basic unit constituting the meaning of the concept, is defined to construct a semantic feature dictionary S30.

  A semantic feature is one of the basic units constituting an LSP, and a semantic feature dictionary is a collection of one or more entries having the same meaning into one set.

  In the case of the sentence (A), the sentences of the sample data are composed of semantic features such as "request", "restaurant", and "purpose". Each semantic feature includes, for example, entries such as "request (teaching)", "restaurant (dining)", and "purpose (meal, lunch)". In the case of sentence (B), semantic features such as "meeting", "purpose" and "where", and sentence (C) is made up of semantic features such as "purpose" and "desire". The concept encompassing the above sentence is "reservation". After all, from some sample sentences, this concept consists of semantic features such as "request", "restaurant", "meeting", "purpose", "where" and "hope".

  In Fig. 3, the meal concept under "reservation" is the concept of "reservation of restaurant", and this concept consists of semantic features such as "restaurant", "request", "hope" and "where" The meeting concept under “reservation” may additionally include the “meeting” semantic feature instead of the “restaurant” semantic feature.

This will be further described using FIG. FIG. 4 shows an example of a semantic feature dictionary table 200 in which the semantic features are defined. The "meeting" which is the 500th semantic feature 201 in the semantic feature dictionary table 200 will be described as an example.

  "Conference", "Issue + Meeting", "Meeting", "Meeting", "Issue + Meeting", "Discussion" and "Discussion" have the same meaning. Therefore, these words can be bundled into an entry of semantic feature 201 "meeting_n" and classified into one entry as in entry table 210 of FIG. 5, and this entry is a set of subclasses of semantic feature 201 meeting_n. Structured to be

  Such semantic features have a dictionary-like role, and the semantic features become a set of entries, since lexical entries having the same meaning are added to the semantic features thus defined. The domain includes the semantic features of keywords and the semantic features of narrative expressions in the domain.

  In the lexical semantic pattern, the symbol "@" is used to express the semantic feature "@meeting_n". When construction of the semantic feature dictionary is completed, a vocabulary semantic pattern (LSP) for sample data collected and classified previously using the dictionary is constructed S40.

  In constructing LSPs, expressions such as clauses, morphemes, syllables, dictionaries, variables and various operators based on various grammatical expressions as well as semantic features can be used. As described above, in the present invention, LSP is made to belong to any concept.

  By executing S30 first, an LSP representing one representative sentence pattern can recognize sentences only in combination of semantic feature entries constituting the LSP.

  FIG. 6 shows an example of the LSP construction table 300 of the present invention. The LSP construction table 300 is a part of representative sentence LSPs related to the example sentences (A), (B) and (C) of the sample data of S20. The basic structure of LSP includes vocabulary, part of speech, and morpheme, and the meanings of the symbols (operators and part of speech) used in FIG.

  The input sentence of the questioner is analyzed by the question and answer system, and when the analysis result matching LSP syntax is retrieved, the question and answer system outputs the corresponding response data to the questioner. For that purpose, the response data are constructed in advance S50.

Preferably, response sentences are constructed in advance for each concept. The fact that the input sentence of the questioner is matched to the LSP means that the concept that the sentence corresponds to is specified. This is because the LSP syntax constructed in S40 by pre-execution in S20 is classified according to each concept. Therefore, it is preferable that the response data constructed in advance in S50 be classified and registered for each concept. Under the present circumstances, it is preferable to designate the part which should change according to an input sentence as a variable. The response data corresponding to the example sentence of the sample data presented in S20 is registered in advance as follows.
(A) "There is @restaurant in nearby @distance"
(B) "There is no better place than @distance for a workshop location."
(C) "The schedule of @project is @when"

  Each block described so far represents a module, segment or portion of code that includes one or more executable instructions for performing a particular logical function. Also, it should be noted that in some alternative implementations, the functions mentioned in the block may occur out of order. For example, two blocks shown in succession may in fact be performed substantially simultaneously, or may be performed in the reverse order depending on the function they sometimes hit. For example, the collection of sample data in S20 may be performed before performing S10, or may be performed after S30. Also, although it is natural that S20 comes before S40, it is also conceivable to add sample data as needed.

  Incidentally, the pre-construction method of lexical semantic patterns for a text analysis and response system according to various preferred embodiments of the present invention is a computer readable medium embodied in program instructions executed through various computer means. Is recorded in The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the computer software art. Examples of computer readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magnetic disks such as floppy disks. Included are optical media and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like. Examples of program instructions include machine language code such as produced by a compiler, as well as high-level language code executed by a computer using an interpreter or the like. A hardware device may be configured to act as one or more software modules to perform the operations of the present invention, and vice versa.

  The scope of protection of the present invention is not limited to the description and expression of the embodiments explicitly described above. In addition, it is added again that the scope of protection of the present invention is not limited by obvious changes and substitutions in the technical field to which the present invention belongs.

Claims (4)

A method of pre-building lexical semantic patterns for a text analysis and response system performed by a manager terminal of a question and answer system, comprising:
(A) defining in advance a concept which is a set to which a vocabulary semantic pattern to be matched with an input sentence of the questioner terminal belongs;
(B) collecting sample data which is a target of a lexical meaning pattern and classifying it according to the concept;
(C) the building concept a basic unit is semantic feature semantic feature dictionary defines the (Semantic Description feature) that constitutes the means of one or more entries as a word having the same meaning, one belonging to each semantic feature Structuring into a set of
(D) constructing the lexical semantic patterns for recognizing the sample data by combining predetermined grammatical expressions using the semantic features;
(E) pre-building response data corresponding to the input sentence of the questioner terminal according to the concept, and pre-constructing a lexical semantic pattern for a text analysis and response system. The method according to claim 1, wherein the concept of step (a) has a hierarchical structure. The pre-construction of lexical semantic patterns for the text analysis and response system according to claim 1, further comprising the step of adding or modifying the concept in the step (b) if the concept corresponding to the sample data is not defined. Method. The method according to claim 1, wherein the response data in the step (e) designates a part to be changed according to the input sentence of the questioner terminal as a variable.

Images