KR20080024752A - Dialog management apparatus and method for chatting agent - Google Patents

Abstract

A device and a method for managing dialog of a chatting agent are provided to improve performance of a chatting robot continuously by providing an answer to the chatting robot based on dialog stacks in consideration of dialog context. A dialog analyzer(120) analyzes a speech sound(110) received from a user. A domain determiner(130) determines and verifies domains to determine the domain of the speech sound based on a speech sound analysis result. A dialog expert(140) provides a system speech sound(170) in response to the user speech sound by searching a dialog example database constructed from a dialog corpus. The domain determiner includes a language/semantic quality extractor(121,122) and a keyword extractor(123), and determines the domain by integrating and applying extracted qualities to a quality based probability model. The dialog expert includes a dialog stack(150) storing dialog example session information of the previous user speech sound and dialog/main flow analysis result, and a dialog example selector(160) indexing information such as semantic information of the previous user speech sound and scenario session information added from a scenario based dialog corpus.

Description

Dialog management apparatus and method for chatting agent

1 is a block diagram showing an overall structure of a chat management apparatus for a chat agent reflecting a chat context according to the present invention.

2 is a flowchart for domain determination and verification according to the present invention.

3 is a flowchart illustrating a process of building a dialogue example DB according to the present invention.

4 is a diagram showing an example record of the dialogue example DB according to the present invention.

<Explanation of symbols for the main parts of the drawings>

110 ... user utterance 121 ... language feature extraction unit

122 ... meaning feature extraction section 123 ... keyword feature extraction section

130 Domain Decision 140 Conversationalist

150 ... Call History Stack 160 ... Chat Examples

170.System ignition

The present invention relates to an apparatus and method for managing a conversation between a person and a chatting robot used in an intelligent chat agent. The present invention relates to an apparatus and method for providing a response of a chat robot in consideration of a dialogue context using a conversation stack.

Chat has been widely used in the past for being able to communicate with the other party through a network using a computer or a portable terminal. However, the chat between people has the disadvantage that you can not chat in the absence of the other party, the chat robot was born to overcome this. This necessity is increasing as a means of communication using natural language between humans and computers (robots) in intelligent agents that are required in many fields. However, since most chat robots can answer only when the pattern is exactly matched with the input sentence, a large amount of dialogue examples are required, and it is expensive to build a dialogue example database (DB). In addition, the existing chat robots do not consider the dialogue context, and there is a problem in that one-to-one chat is performed regardless of past information. Therefore, the existing chat robots are inconsistent with the way they communicate with each other. Therefore, it is necessary to improve the chat model of the chat robot.

Therefore, in order to overcome the problems of the prior art described above, the present invention provides a method and method for managing conversations in which a chat robot's answer varies according to past information in consideration of a conversation example DB construction method and a dialogue context for an intelligent chat robot. To present.

In order to achieve the above object, a conversation management method of a chat agent capable of reflecting a conversation context using a discourse stack, the method comprising: (a) analyzing an input user speech and analyzing a domain of the user speech based on the analyzed result; Determining and verifying domains; And (b) searching for a conversation example database (DB) constructed from a conversation corpus and providing a system speech in response to a user speech.

According to another aspect of the present invention, a chat agent management apparatus of a chat agent capable of reflecting a dialogue context using a discourse stack, the apparatus comprising: a dialogue analyzer for analyzing input user speech; A domain determination unit that performs domain determination and verification to determine a domain of user speech based on the result of the dialogue analysis; And a chat expert including a chat expert who searches a chat example database (DB) constructed from a chat corpus and provides system speech in response to user speech.

Next, the present invention will be described with reference to the accompanying drawings.

1 is a general structural diagram of a conversation management system for a chat robot according to the present invention. In FIG. 1, the user utterance 110 receives a user input sentence through a keyboard input or a voice input. When the user utterance 110 comes in, the conversation analyzer 120 determines a domain (topic) and a conversation example through the language feature extractor 121, the semantic feature extractor 122, and the keyword feature extractor 123. Perform dialogue analysis to extract qualities for After the conversation analysis as described above, the domain determination unit 130 performs a domain determination and verification step of determining and verifying the domain (topic) of the user's speech. Thereafter, the domain determiner 130 calls the conversation expert 140 according to the determined domain.

The conversation expert 140 is divided by domain, and each conversation expert 140 has a unique discourse record stack 150. The discourse record stack 150 has information such as an analysis result of previous user utterances and an answer of a chat robot. The conversation example selection unit 160 searches for a conversation example most similar to the current conversation situation from the conversation example database DB. Here, the conversation situation collectively means the current conversation state information for searching the conversation example. That is, domain, speech act and semantic analysis result, discourse record, scenario session information and the like. Among the selected conversation example candidates, the user utterance similarity is measured to finally select the conversation example having the highest similarity, and the answer corresponding to the example is the system utterance 170.

In the present invention, by extracting various qualities from the user speech, the domain (topic) of the speech is determined and the conversation example is selected. Determining the domain is to index the dialogue example scenario in the chat by domain, to select an appropriate answer among the answer candidates searched by domain, and to exclude the chat example irrelevant to the domain from the search candidate. Also, extracting semantic feature is to process conversation example DBs of existing chat robot based on vocabulary and process with different example even though it is semantically same user utterance. This has the disadvantage of requiring a large amount of dialogue examples to build the dialogue example DB. Therefore, the present invention builds a dialogue example DB based on semantics so that the same example sentence can be processed when the user speech is semantically identical. This reduces the cost and increases the efficiency by constructing a semantic based dialogue example DB. In addition, when constructing a dialogue example DB to consider the dialogue context, a scenario-based chat dialogue corpus (Corpus) is used. When corpus records session information for each scenario and indexes conversation examples, indexing is performed by adding semantic information of previous user utterances (results and driving analysis results) and session information to current user utterances. By doing so, the conversation example can be searched using the information of the discourse recording stack 150.

Referring to FIG. 2, the steps performed by the domain determiner 130 shown in FIG. 1 are described in detail. The user's speech is extracted through morphological analysis and part-of-speech tagging (step 210) to extract linguistic features (step 220). Linguistic qualities here are available in a variety of forms for the description. In general, linguistic qualities are features that help determine domain and language understanding, and extract morpheme prototypes, morpheme N-grams, and parts of speech tags. The speech act and meaning of the user's speech are analyzed based on the analyzed linguistic qualities (step 230), and the semantic qualities of the current user's speech acts and intentions and components are extracted (step 240). The keyword included in the user utterance and the weight of each domain of the keyword are extracted through the keyword model (step 250). Here, keyword qualities are domain classification results using only TF * IDF values, and N keywords in the order of the highest weight among user utterances and N domains having the highest weight by TF * IDF values of the keywords are extracted. Here, TF * IDF is frequently used in information retrieval, and abbreviation of Term Frequency and Inverse Document Frequency, respectively, to extract keywords using this. TF is literally the frequency of words coming out, and IDF is the inverse of how many documents the words come from. That is, surveys or mothers can't be keywords (see the related web page on this one at http://abolapia.egloos.com/207027).

In determining the domain, if the current user's speech is the same as the previous domain, there is a persistence of the conversation. Therefore, it can be assumed that the previous conversation experts continue to process user utterances, considering that they are an extension of the previous conversation without domain switching. In order to verify whether the current utterance is the continuation of the previous utterance, the conditional probabilities of immediately preceding and the current speech act and intention are calculated and compared with the set reference value (step 260). If the probability value is higher than the reference value, the domain decision (step 270) is omitted, and the dialogue expert 140 immediately before the reference value is lower. If the value is lower than the reference value, the linguistic and extracted semantic features (step 220) and semantic features (step 240) are extracted. In operation 270, the keyword feature (250) is combined to determine a domain having the highest probability using a probability model.

3 is a flowchart illustrating a process of constructing a dialogue example DB described above. The user speech is extracted from the scenario-based chat conversation corpus 310 (step 320), and the corresponding scenario and speech session information are extracted (step 330). The previous semantic information is extracted (step 340). The previous semantic information is a combination of speech acts and driving analysis results of previous user speeches. In operation 350, current speech act information, driving information, and domain information corresponding to a scenario are extracted. Finally, the system ignition extraction (step 360). Finally, the extracted information is indexed to create a scenario-based chat dialogue database 370.

4 are record examples of a scenario based chat conversation example DB indexed from a conversation corpus. The dialogue example searches the dialogue example with the scenario session information, previous semantic information, current speech information, current driving information, domain information, and the like and selects the dialogue example by measuring the edit distance with the user's speech. Done. If necessary, you can relax the search conditions and find a conversation example with partial matches if you cannot find a full match. For example, if a conversation example is not found in a given scenario, the conversation example DB is searched except for the scenario session information and previous semantic information.

As described above, according to the present invention, in constructing a chat agent conversation management apparatus, a semantic based conversation example DB may be constructed to create a small amount of conversation examples, and a scenario based conversation example may be used to reflect a conversation context. The chat robot conversation management device can be implemented. This methodology can be performed by using the probabilistic model and indexed DB that are automatically learned from the dialogue corpus, and can continuously improve the performance of the chat robot.

The chat robot may be usefully used as an interactive short message service of a mobile phone, an intelligent robot, a home robot, or an intelligent agent such as a chat robot by analyzing a user's speech in real time and responding appropriately. Those skilled in the art will understand that various modifications and equivalent other embodiments are possible from this. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

Claims (10)

 In the chat management method of the chat agent that can reflect the context of the conversation using a discourse stack, (a) domain determining and verifying an input user utterance and determining a domain of the user utterance based on the analyzed result; And (b) searching for a conversation example database (DB) constructed from a conversation corpus and providing system speech in response to user speech. The method of claim 1, wherein step (a) is a probability-based domain determination step combining various qualities. (a1) using language features using natural language processing and semantic features through language understanding to determine a domain of user speech; (a2) extracting keyword features using a keyword model; And (a3) integrating the extracted features and applying them to a feature-based probability model to determine a domain. 3. The method of claim 2, wherein step (a) is a step of verifying domain persistence based on semantic qualities. The conditional probability value of the semantic qualities of previous user utterances and the semantic qualities of current utterances is expressed by using user's speech acts or intention information. Verifying that the conversation persists in the same domain using the method. The method of claim 1, further comprising the step of constructing a dialogue example database (DB) applied to the step (b), wherein the establishing the dialogue example database (DB) comprises the meaning of user utterance in constructing a dialogue-based example conversation DB. As indexing based on information, (c1) indexing the conversation example DB using the previous meaning information as an index key to use the dialogue act, driving and domain information as the semantic information, and to reflect the dialogue context; And (c2) Automatically extract user utterance, extract scenario and utterance session, extract previous semantic information, current dialogue and driving, domain extraction and system utterance through a series of processes for constructing the scenario sample DB of the collected scenario-based chat conversation corpus A conversation management method for a chat agent, comprising extracting. The method of claim 1, wherein step (b) reflects a conversational context using a speech stack, (d1) using a conversation stack recording recorded example session information, dialogue act analysis result, and driving analysis result of previous user speech to represent previous information of the user speech; And (d2) Conversation management method for a chat agent including adding and indexing information such as semantic information and scenario session information of a previous user's utterance from a scenario-based conversation corpus in order to consider a conversation context in constructing a conversation example DB. .  In the chat management apparatus of the chat agent that can reflect the chat context by using a discourse stack, A conversation analyzer analyzing the input user speech; A domain determination unit that performs domain determination and verification to determine a domain of user speech based on the result of the dialogue analysis; And A conversation management device for a chat agent including a chat expert who searches a chat sample database (DB) constructed from a chat corpus and provides system speech in response to user speech. The apparatus of claim 6, wherein the domain determiner comprises: a language feature extracting unit and a semantic feature extracting unit for extracting a language feature using natural language processing and a semantic feature through language understanding to determine a domain of a user speech; And And a keyword extracting unit for extracting keyword features using a keyword model, and determining the domain by integrating the extracted features and applying the feature to a probability-based probability model. 8. The method of claim 7, wherein the domain determiner is a domain continuity verification step based on semantic qualities, and uses conditional probability values of the semantic qualities of previous user speeches and the semantic qualities of the current speech using user speech acts or intention information. And verifying that the conversation continues in the same domain. 8. The conversation example database (DB) according to claim 6, wherein the conversation example database (DB) uses dialogue acts, driving and domain information as the semantic information, and indexes the dialogue example DB using the previous semantic information as an index key to reflect the dialogue context; Establish the collected scenario-based chat conversation corpus by automatically extracting user speech, extracting scenarios and speech sessions, extracting previous semantic information, extracting current speech and driving, domain extraction, and system speech with a series of processes for constructing a conversation example DB. And a conversation managing device for a chat agent. The method of claim 6, wherein the conversation expert uses a conversation stack to reflect the conversation context. Discourse stack recording recorded example session information, dialogue act analysis result, and driving analysis result of previous user speech to represent previous information of user speech; And In order to consider the dialogue context in constructing the dialogue example DB, a chat example selection unit for adding and indexing information such as the semantic information of the previous user's utterance and scenario session information from the scenario-based dialogue corpus is included. Dialogue management device.

Images