KR100377475B1 - Device for Extracting Korean Speech Act using Mood Information - Google Patents

Abstract

The present invention extracts information for dialogue understanding in an intelligent conversation system by automatically extracting speech act information of an input sentence using verbal information of a verb, syntax and semantic code information, and analysis information of a previous spoken sentence in a conversation history memory. Do this. The present invention uses the Korean syntax and semantic codes as inputs to determine whether or not to omit, substitute, or inverted sentence context restoration device to restore the deficiency information, using the Korean syntax and semantic code to extract the predicate logical form propositions Extraction of calligraphy information appearing at the ending ending of verbs by referring to the proposition extractor and the verb ending electronic dictionary, and extracting firepower by using the syntax, meaning, and act code of the previous spoken sentence defined in the calligraphy information and conversation history memory It is a Korean language act extraction apparatus using calligraphy information consisting of a procession extractor which extracts a act of acts using a proposition and an act of hydropower.

Description

Device for Extracting Korean Speech Act using Mood Information}

The present invention relates to a Korean speech act extraction apparatus for automatically extracting speech act information of an input sentence using calligraphy information of a verb appearing as a result of Korean sentence analysis.

Representative examples of natural language processing application systems that are currently being actively developed include dialog voice language translation system, dialog system for database search, and CAI system. For the dialogue voice language translation system, a translation incorporating the intention of the speaker included in the input speech is required. In the dialogue system, the dialogue can be intelligently managed and processed only when the speech act representing the user's intention can be extracted from the input speech.

Many methods for speech act analysis have been studied at home and abroad. Austin, Searle, and others attempted a discordant act of verbal acts based on the type of verb. Allen also analyzed speech acts using vocabulary information that could indicate the intention of the speaker in the input speech, and suggested an indirect speech analysis method. Carberry has set 17 kinds of discourse goals for intention analysis of user input speech in conversation system based on plan recognition, but presents specific method for extracting dialogue goal using interpretation information of input speech. I'm not doing it.

In the Japanese ATR, nine kinds of speech act information were set and statistical data was extracted by manually tagging actual dialogues. The statistical information was used as data for the prediction of the next speech sentence in speech recognition, demonstrating the efficiency in speech recognition error correction. However, a rule-based extraction method using simple surface lexical information was applied for tagging speech acts, and almost depended on the subjective manual work of experts. Therefore, it is impossible to work with a large amount of data, and tagged data also has a disadvantage that it does not have objectivity.

In Korea, speech act analysis based on contextual semantics is being conducted mainly in linguistics, and research on classification of acts of acts and analysis of indirect speech acts is conducted. In addition, an attempt has been made to analyze a user's speech intention using a keyword in a speech language conversation system using a natural language interface, but this requires a subjective and subjective data work that is limited to its application and has a problem in its expandability.

In the dialogue system for hotel reservation, the syntax and semantic expressions obtained from the interpretation of the input speech text were used to analyze the delivery contents of the speaker. The speech power of the previous speech text and the syntax information of the input speech text were used for the speech act analysis. However, there was difficulty in analyzing various kinds of speech acts by using only question-type and plain-type information as syntax information in the description part of input speech.

Therefore, the dialogue application system using the natural language is a method of speech acting on whether the user requires or provides the information in addition to analyzing the user's delivery contents through the syntactic and semantic interpretation of the user's input speech. You should be able to recognize the degree. Therefore, only the traditional syntactic and semantic analysis methods for the speech can not interpret the speech act theory information contained in the input speech sentence, and the speech act analysis apparatus according to the present invention can analyze the speech intention of the speaker.

The present invention uses the calligraphy information, syntax and semantic code information of the verb that appears as a result of Korean sentence analysis, and the dialogue act information of the input sentence by using the analysis information of the previous spoken text in the conversation history memory. The purpose is to perform information extraction.

1 is a hardware block diagram according to the present invention,

2 is a flowchart of a context restoration apparatus in the Korean speech act extraction apparatus to which the present invention is applied;

3 is a classification diagram relating to firepower defined in the Korean dialogue act extraction apparatus.

<Explanation of symbols for main parts of drawing>

1: device for determining context phenomenon 2: device for restoring context

3: Korean sentence proposition extraction device 4: Calligraphy information extraction device

5: hydropower extraction device 6: fire act extraction device

7: verb mother electronic dictionary 8: question mark electronic dictionary

9: speech act extraction table 10: conversation history memory

11: predicate information inspection device 12: essential information inspection device

13: essential information extraction device 14: joiner knowledge base

15: inverted door restoration device 16: substitute word restoration device

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram of an apparatus for extracting Korean dialogue acts using calligraphy information according to the present invention.

As shown in FIG. 1, the presence or absence of a predicate component, a substitute word, and a position of the predicate are determined by checking the syntax and semantic code of the input sentence analyzed by the Korean sentence analyzing apparatus to determine whether the sentence is omitted, the substitute sentence, or the inverted sentence. The restoration of the deficient information for the input speech is determined through the contextual phenomenon determining apparatus 1. When it is determined that the context developing process is performed by the context developing apparatus, the predicate information extracting apparatus 12 uses the conversation history memory (1) when the predicate component is omitted through the predicate information checking apparatus 11 for processing the omitted statement. 10) The predicate component is restored by referring to the predicate component of the previous utterance statement stored in the image. After the predicate component restoring process is completed, the essential case component of the predicate component is restored in the essential case information extracting device 13 with reference to the joiner knowledge base 14. If the predicate component is not located at the end, the inverted sentence restoring device 15, which acts to position the corresponding narrative structure as the end, and the surrogate component are referred to the surrogate component by referring to the information on the knowledge base and the conversation history memory. After passing through the terminology restoring apparatus 16 to replace the corresponding information, the reference information preparation process for extracting speech act information is finished.

When the information lacked by the above process is restored, the propositional rule regarding the input speech sentence in the predicate logical form is extracted by the Korean sentence proposition extracting apparatus 3 using the restored syntax and semantic code. At this time, the proposition makes the component vocabulary and semantic code information of each item of the syntax code as an item of the predicate logical expression. The calligraphy information extractor 4 extracts calligraphy information from the calligraphy information extractor 4 using the ending-to-call relationship table pre-built in the verb ending electronic dictionary. Conversation history memory, which can refer to dialogue acts, and question-marking electronic dictionaries that can identify questionable acts, extracts hydrothermal power, which is a component of conversation acts. Finally, the acting information extracted from the Korean sentence proposition extracting apparatus 3 by the act of act extraction apparatus 6 and the form of hydropower analyzed by the apparatus of hydropower extraction apparatus are finally used by using the table of relations of acting power vs. act of dialogue. Create

In the present invention, the actual dialogue is analyzed, and as shown in FIG. 3, the hydroelectric power is classified into four types of expression, request, provision, and the like based on the superficial lexical information. The expression is defined as the idiomatic expression that appears in the dialogue, and is classified into the dialogue initiation, conversation termination, and everyday conversation. And the request indicates the intention of the speaker to ask the listener to provide information or action. It consists of three information requests and one action request.

Among the information requests, the authenticity information request and the interrogation information request are divided into the presence or absence of the interrogator, and the confirmation information request is defined as the intention to call for confirmation by referring back to the phrase shown in the previous speech. The presentation indicates the speaker's intention to utter the speaker's information or action or promises to provide the action. It is divided into five types of provision. Others were classified as fragments of speech because of signs of surprise or halting speech. This hydroelectric classification can be further subdivided according to the application of each application system.

In the present invention, each spoken sentence of the conversation corpus was analyzed manually to investigate the correlation between the firepower and the Korean ending endings defined in FIG. As a result, it is true that hydropower is influenced by the ending ending according to calligraphy, but it shows characteristics of each hydropower depending on the type of vocabulary appearing in the utterance and the relationship with previous hydropower.

It shows the kind of ending ending that appears in the speech of authenticity information request, information provision, action request and promise provision, which are the final ending powers, which are found in the analysis of actual dialogues. It is also frequently found that the ending endings appear in the dialogue with the ending ending "Yo". The type of calligraphy that appears in the request for authenticity information is the questionable ending ending and the questionable narrative survey, and the informational presentation shows the plain ending and the admiration ending ending and the narrative survey. In the present invention, the narrative investigation "Ida" is not a ending ending, but it is used as a terminology, so that the ending ending and its function are the same, and the type of calligraphy appears. In the demand for action, the command type, the invitation type, and the permission type appear, and the promise offer is the promised mother. The interrogative information request is the same kind of mother that appears in the authenticity information request, but differs in that the interrogative sentences contain interrogative sentences. Therefore, the questionnaire is classified into question statement, question sentence sentence, question adverb, and question word, so that the electronic dictionary (8) of question mark can be distinguished.

Firepower Appearance Characteristics Firepower Appearance Characteristics Initiation of conversation Lexical Provide information Ending End of conversation Lexical Answer Ending endings + context Conversation Lexical A suggestion Ending endings + vocabulary Authenticity Information Request Ending Promise Ball Ending Interrogation Information Request Ending endings + vocabulary Provide confirmation Context + Vocabulary Request Information Context + Vocabulary Other Lexical Act request Ending

In the hydroelectric power extracting apparatus 5, when the ending ending is present, the ending ending writing method of the input speech sentence is analyzed and the processing is performed according to the characteristics of each writing method. At this time, the calligraphy is divided into seven types, and the hydropower analysis according to the calligraphy is based on the relationship between the hydropower in Table 1 and the ending ending calligraphy.

The method of extracting firepower using calligraphy information first checks whether an input speech is an idiomatic conversational expression. The type of calligraphy appearing in the ending ending is determined by referring to the verb ending dictionary and branches to the processing unit corresponding to each calligraphy. If the type of calligraphy is written or admired, the determination of auxiliary verbs determines the auxiliary link "high" and the auxiliary adjective "want" to express hope. In this case, when the previous fire power information on the conversation history memory 10 is an information request and the first spoken text of the user, the current fire power is determined as providing a response to the previous information request. The use of such calligraphy information enables more accurate hydropower extraction by effectively limiting the number of hydropower extraction candidates.

In general, a dialogue act is composed of a conversation power indicating the speaker's utterance intention and a proposition indicating the contents of the utterance. In the dialogue act extraction unit 6 of the present invention, a dialogue act code is finally generated using the dialogue act extraction table 9 of Table 2. do.

Thermal power classification Hydropower subclass Mad row expression Dialogue expression ConvinceByInform (S, H, P) demand Interrogation Information Request MotivatebyRequest (S, H, ConvincebyInformwh (H, S,? XP)) Authenticity Information Request MotivatebyRequest (S, H, ConvincebyInformIf (H, S, P) Request Information MotivatebyRequest (S, H, ConvincebyInformIf (H, S, P) Act request MotivatebyRequest (S, H, Act (H, P)) offer Provide information ConvinceByInform (S, H, P) Provide response ConvinceByInform (S, H, P) Suggestion ConvinceByInform (S, H, Act (S, P)) Promise ConvinceByInform (S, H, P) Provide confirmation ConvinceByInformIf (S, H, P) Other Other ConvinceByInform (S, H,? XP)

In Table 2 above, 'S' means Speaker, 'H' means Hearer, and 'P' means Proposition.

The present invention as described above can be used as a Korean sentence comprehension device and its method, which is a core technology such as a dialogue system, a CAI system, a dialogue voice language translation system, etc. It can be used as an automatic dialogue extractor for building statistical information.

Claims (5)

Context retrieval means for determining whether to abbreviate sentence, substitute sentence, inverted sentence by inputting Korean syntax and semantic code and to restore the deficiency information; Korean sentence proposition extracting means for extracting a proposition in predicate logical form using Korean syntax and semantic code; Means for extracting calligraphy information appearing at the ending ending of a verb by referring to an electronic dictionary of verbs and extracting firepower using phrases, meanings, and act act codes of previous spoken sentences defined on the calligraphy information and conversation history memory; Korean language act extraction apparatus using calligraphy information, characterized in that consisting of a speech act extraction means for extracting the act of acts using propositions and the power of hydropower using the act of act extraction table composed of the relationship between act of hydropower and act of act. The method of claim 1, The context restoration means, Predicate information extraction device (1), predicate information extraction device (11), predicate information extracting device (12) using conversation history memory (10), and joiner knowledge base (14) for determining whether or not an abbreviation statement, substitute statement, or inverted statement is present. Korean speech act extraction apparatus using calligraphy information, characterized in that consisting of the essential information extracting device (13), tochimun restoring device (15), the substitute term restoring device (16) for restoring the essential case components. The method of claim 1, The Korean sentence proposition extracting means, Korean speech act extraction apparatus using calligraphy information comprising a Korean sentence proposition extracting apparatus (3) for extracting a proposition in the form of predicate logic using a Korean syntax and semantic code. The method of claim 1, Means for extracting the thermal power, Calligraphy information extraction device (4) for extracting calligraphy information appearing at the ending ending of the verb with reference to the verb ending electronic dictionary (7), calligraphy information, the question mark electronic dictionary (8) and the conversation history memory (10) Korean speech act extraction apparatus using calligraphy information, characterized in that consisting of a hydropower extraction unit (5) for extracting the hydropower using the syntax, meaning, speech act code of the previous speech. The method of claim 1, The speech act extraction means, Korean language act extraction apparatus using calligraphy information, characterized by consisting of a speech act extraction unit (6) for extracting the act of acts using propositions and the power of hydropower using a dialogue act extraction table (9) composed of the relationship between the hydropower and the act of acts.