JP6433765B2 - Spoken dialogue system and spoken dialogue method - Google Patents

Description

  The present invention relates to a voice dialogue system, and more particularly, to a voice dialogue system that dialogues with a person about a topic in a specific domain.

  In recent years, various systems such as personal digital assistants (PDAs) such as smartphones and tablet terminals and car navigation systems have been equipped with a voice recognition function. In a system equipped with a voice recognition function, a user gives an instruction to the system by voice, and the system understands the voice instruction and performs a desired operation. If the system further has a speech synthesis function, the user can interact with the system by voice.

  Conventionally, as a voice dialogue system, a communication gap is solved in real time, and continuous and natural communication is performed (for example, refer to Patent Document 1), a voice related to a replay and confirmation from a speaker side is recognized, There are those that can execute answers and processing based on this (for example, see Patent Document 2). In addition, as an interactive system that generates an appropriate response sentence for the user's voice, it expresses the paraphrase used when including keywords in the response sentence and the types of response sentences Record the response type and the conditions under which the paraphrase word and response type are selected, and determine the paraphrase word and response sentence template for the recognized keyword based on the conditions under which the paraphrase word and response type are selected. In some cases, a response sentence is generated by inserting a paraphrase into the response sentence template (see, for example, Patent Document 3).

JP 2012-181697 A JP 2010-197858 A JP 2008-39928 A

  In recent years, audio recognition functions are being installed in AV systems such as television devices. However, the current AV system is only a command-based speech recognition system that understands the utterances of commands such as “Volume up” and “Turn off” and operates as expected. For this reason, even if the user wants to know the content while viewing the content, the necessary information cannot be obtained by voice dialogue with the system. Further, the information that the user wants to know is not limited to simple information that can be answered in 5W1H, but may be complex information such as opinions and impressions regarding the content.

  In view of the above problems, an object of the present invention is to provide a voice dialogue system capable of voice conversation with a person on a topic in a specific domain.

  A speech dialogue system according to one aspect of the present invention is a speech dialogue system that performs speech dialogue with a person on a topic in a specific domain, and a speech recognition unit that recognizes a speech of a speaker and generates a spoken sentence. Analyzing the utterance and understanding what information the speaker wants to know in the domain; and depending on the content of the information that the speaker wants to know, the domain A dialogue management unit that retrieves the target information by searching either the first information source that holds various information in the domain or the second information source that holds the impression of others in the domain; Is a system including a response sentence generation unit that generates a response sentence for the utterance of the speaker and a voice synthesis unit that synthesizes the response sentence and generates a speech signal.

  In addition, a voice dialogue method according to another aspect of the present invention is a method of voice dialogue with a person on a topic in a specific domain, in which a voice recognition unit recognizes a utterance of a speaker and utters a spoken sentence. The intention understanding unit analyzes the utterance sentence to understand what information the speaker wants to know in the domain, and the dialog management unit information that the speaker wants to know. Depending on the content of the information, the first information source that holds various information in the domain and the second information source that holds the impression of others in the domain are searched to obtain target information, In this method, a response sentence generation unit generates a response sentence for the utterance of the speaker using the target information, and a voice synthesis unit generates a voice signal by voice synthesis of the response sentence.

  According to these systems or methods, when a user (speaker) speaks something to the spoken dialogue system, the user understands what information the speaker wants to know in a particular domain and wants to know Depending on the content of the information, the first information source that holds various information in the domain and the second information source that holds the impression of others in the domain are searched to obtain the target information. A response sentence using information is returned to the user as voice. Thereby, the user can acquire not only simple information that can be answered in 5W1H but also complicated information such as opinions through dialogue with the system.

  The first information source may be a website on the Internet.

  According to this, since information on websites on the Internet is frequently added / updated, by using a website on the Internet as the first information source, the voice interactive system can always obtain the latest information. it can.

  The voice dialogue system may include an impression database as the second information source, and the dialogue management unit obtains a user review from a website on the Internet regularly or irregularly, and A sentence including an evaluation expression may be extracted from the review and registered in the impression database.

  According to this, not a user review published on a website on the Internet but a sentence including an evaluation expression useful as an impression is extracted and registered in the impression database. In addition, a user review newly posted on a website on the Internet can be taken into the impression database regularly or irregularly to update the impression database.

  ADVANTAGE OF THE INVENTION According to this invention, the audio | voice dialog system which can carry out a dialog with a person about the topic in a specific domain is implement | achieved. Thereby, useful information that the user desires can be provided to the user through a simple interaction called dialogue.

1 is a block diagram of a voice interaction system according to an embodiment of the present invention. Flow chart showing operation of spoken dialogue system External view of a television apparatus which is an embodiment of a voice interaction system

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. FIG. 1 is a block diagram of a voice interaction system according to an embodiment of the present invention. The speech dialogue system 10 according to the present embodiment includes a speech recognition unit 11, an intention understanding unit 12, a dialogue management unit 13, a response sentence generation unit 14, and a speech synthesis unit 15. About a topic in a specific domain It interacts with the speaker by voice. The domain refers to the content, field, genre, etc. of dialogue handled by the voice dialogue system 10, for example, specific topic areas such as movies, music, cooking, and weather. For example, if the domain is a movie, the speaker can interact with the voice interaction system 10 about the movie.

  In the following description, the case where the domain is a movie will be described, but the domain targeted by the present invention is not limited to a movie. Further, the number of domains that the voice dialogue system 10 can handle is not limited to one, and the voice dialogue system 10 can deal with a plurality of domains.

  The voice recognition unit 11 recognizes the utterance of the speaker and generates an utterance sentence. The utterance of the speaker is input to the voice recognition unit 11 as a voice signal from a microphone (not shown). A known technique can be used as the voice recognition function of the voice recognition unit 11. The generated utterances vary, such as “Who is the director?”, “When was the movie released?”, “Who is appearing?”, “Tell us what you think of this movie”.

  The intention understanding unit 12 analyzes the utterance sentence generated by the voice recognition unit 11 and understands what information the speaker wants to know in the domain, that is, the intention of the speaker. Here, there are many cases where the purpose of the utterance sentence is ambiguous, such as “who is the director?”. In addition, for example, in addition to “who is performing?”, Various expressions such as “who is performing” and “who is appearing” are used as utterances for asking performers. The intent understanding unit 12 refers to the intent understanding model 121 and the problem solving knowledge 122 to correctly understand the intention of the utterer included in the utterance sentence whose purpose is ambiguous or the utterance sentence of various expressions. it can.

  The intent understanding model 121 is an utterance sentence example collection in which various expressions are collected for each intention. The intention understanding model 121 is a database of what kind of intention a word or expression pattern of an utterance has. By referring to the intention understanding model 121, the intention understanding unit 12 can correctly understand the intention of the speaker without being influenced by superficial language even if various expressions are used in the utterance sentence.

  The problem solving knowledge 122 stores correspondences between various utterance expressions (problems) and their solutions. Depending on the content of the information the speaker wants to know, the problem can be classified into two types. The first is a simple question of inquiring about the director's name, release date, and performers. The other is a complicated problem inquiring about the impression of the movie. The solution to a simple problem is “search the website”. On the other hand, the solution to the complicated problem is “search the impression database”. Thus, the intention understanding unit 12 can identify the problem and its solution by referring to the problem solving knowledge 122.

  The dialogue management unit 13 receives the intention understanding result from the intention understanding unit 12, searches one of the two information sources according to the content of the information that the speaker wants to know, and obtains the target information. get. One source of information is a website on the Internet. For example, DBpedia (http://en.dbpedia.org/) can be used as the website. DBpedia is a data set in which the information of Wikipedia (http://en.wikipedia.org/) is structured, and desired information can be searched by SPARQL of an RDF (Resource Description Framework) query language. Another information source is the impression database 131. The impression database 131 stores various impressions of others in the domain. In the case of this embodiment, the impression database 131 stores impressions of other people regarding various movies.

  The dialogue manager 13 determines the system behavior according to the dialogue scenario 132. The dialogue scenario 132 describes the flow of dialogue such as what kind of interaction should be performed when various intentions are input. For example, when the utterance intention is “I want to know the name of the director” and the solution is “Search the website”, the dialogue management unit 13 searches the website on the Internet as a system action and searches for the director name. To get. On the other hand, when the utterance intention is “I want to hear the impression of the movie” and the solution is “Search the impression database”, the dialogue management unit 13 searches the impression database 131 as a system action and provides it to the speaker. Get the right impression to do.

  The dialogue management unit 13 can collect user reviews from websites on the Internet and register them in the impression database 131. For example, Yahoo! You can use movies (http://movies.yahoo.co.jp/). The website has reviews from many people about various movies.

  For example, the dialogue management unit 13 can collect user reviews from websites and store them in the impression database 131 as follows. First, the dialogue management unit 13 uses Yahoo! Get user reviews for each movie from the movie site. There are multiple user reviews for each movie. Then, the dialogue management unit 13 performs morphological analysis of each sentence of the user review using the morphological analysis engine, further removes unnecessary characters, and extracts only the opinion sentence including the evaluation expression using the opinion extraction tool. An opinion (evaluation expression) extraction tool (http://alaginrc.nict.go.jp/opinion/) provided by the National Institute of Information and Communications Technology can be used as the opinion extraction tool. Data of evaluation expressions is extracted for each user review using the tool. The evaluation expression is an expression representing a positive or negative opinion, reputation, or recommendation such as “XX is good” or “XX is boring”. The extracted data set is the impression. That is, the impression in this specification is a set of sentences including evaluation expressions.

  Furthermore, the dialogue management unit 13 extracts feature amounts from the extracted impressions. Specifically, an adjective TF (Term Frequency) and a noun TF are calculated as the usage tendency of adjectives and nouns for each impression. Furthermore, the dialogue management unit 13 calculates TF-IDF (Term Frequency-Inverse Document Frequency) for nouns. Then, the dialogue management unit 13 registers these adjectives TF, nouns TF, and nouns TF-IDF together with the extracted impressions in the impression database 131. The adjectives TF, nouns TF, and nouns TF-IDF are used as indices when searching for suitable impressions from the impression database 131.

  Since user reviews on the website are frequently added and new movies are released one after another, it is desirable that the dialog management unit 13 updates the impression database 131 regularly or irregularly. For example, the impression database 131 may be updated every day at a predetermined time, or may be updated every time an instruction is received from the user.

  The response sentence generation unit 14 generates a response sentence for transmitting the system action result of the dialogue management unit 13 to the speaker. Specifically, the response sentence generation unit 14 generates a response sentence for the utterance of the speaker by using the target information acquired by the dialogue management unit 13 from various information sources. For example, when the utterance of the speaker is “When was the theater released?” And the dialogue management unit 13 acquires “January 20, 2014” as the target information, the response sentence generation unit 14 A response sentence such as “This movie was released to the theater on January 20, 2014” is generated.

  The voice synthesizer 15 synthesizes the response sentence generated by the response sentence generator 14 and generates a voice signal. The audio signal is output to a speaker (not shown) and transmitted to the speaker as sound. A known technique can be used as the speech synthesis function of the speech synthesis unit 15.

  Next, the processing flow of the voice interaction system 10 will be described. FIG. 2 is a flowchart showing the operation of the voice interaction system 10.

  First, the speech recognition unit 11 recognizes the speech of the speaker and generates an utterance sentence (S1). Then, the intention understanding unit 12 analyzes the generated utterance sentence to understand the intention of the speaker (S2).

  If the speaker's intention is a 5W1H question and the information that the speaker wants to know is simple information (YES in S3), the dialog management unit 13 can display a website on the Internet (for example, Yahoo! Movie). Search for the desired information. On the other hand, when the information that the speaker wants to know is complex information such as an impression (NO in S3), the dialogue management unit 13 searches the impression database 131 and acquires an appropriate impression (S5).

  When the dialogue management unit 13 acquires the target information, the response sentence generation unit 14 generates a response sentence for the utterance of the speaker using the target information acquired by the dialog management unit 13 (S6). Then, the speech synthesizer 15 performs speech synthesis on the generated response sentence to generate a speech signal, and outputs speech from the speaker (S7).

  As described above, according to the present embodiment, a conversation with a person can be performed on a topic in a specific domain. Accordingly, useful information desired by the user can be provided to the user through a simple interaction called dialogue, and the convenience of the user can be improved.

  The intent understanding model 121, the problem solving knowledge 122, the impression database 131, and the dialogue scenario 132 may be arranged in a storage device (not shown) in the voice dialogue system 10, or an external storage device (not shown). May be arranged. Further, these may be arranged in the same storage device, or may be distributed in a plurality of different storage devices. Further, the voice recognition unit 11, the intention understanding unit 12, the dialogue management unit 13, the response sentence generation unit 13, and the voice synthesis unit 15 may be mounted on the same server device, or the functions are distributed to a plurality of server devices. It may be. Distributed knowledge and functions can communicate with each other through a network.

<Example>
Next, an embodiment of the voice interaction system 10 will be described. FIG. 3 shows the external appearance of a television apparatus that is an embodiment of the voice interaction system 10. The voice interaction system 10 according to the present embodiment can be mounted on, for example, a television device 100 including a display panel 101, a microphone 102, and a speaker 103.

  The television apparatus 100 includes a television tuner (not shown) and can receive a terrestrial digital broadcast 104 and a BS (Broadcasting Satellite) / CS (Communication Satellite) digital broadcast 105. The terrestrial digital broadcast 104 and the BS / CS digital broadcast 105 are input to the television apparatus 100 through an antenna terminal (not shown).

  In addition, the television device 100 can be connected to the Internet 106 by wire or wirelessly. The television apparatus 100 can receive content such as VOD (Video On Demand) from a content server 107 on the Internet 106. The television apparatus 100 can be connected to a recorder 108 such as a DVD (Digital Versatile Disc) / BD (Blu-ray Disc). The television device 100 can reproduce content recorded on a hard disk device (not shown) mounted on the recorder 108 or an optical disc such as a DVD or a BD. The television apparatus 100 may have a recorder function.

  The content received by the television device 100 is displayed on the display panel 101, and the user 200 (speaker) can ask the television device 100 various questions regarding the content while viewing the content. The television apparatus 100 collects the voice uttered by the user 200 with the microphone 102, understands the content of the utterance of the user 200, generates an appropriate answer to the question of the user 200, and utters the answer from the speaker 103. To output. In this manner, the user 200 can interact with the television apparatus 100 by voice.

  Although the embodiment of the present invention has been described above, the present invention is not limited to the configuration of the above embodiment, and various modifications can be made. For example, the voice interaction system 10 according to the present embodiment can be mounted on, for example, a smartphone or a PC having a content reproduction function in addition to the television device 100.

  Moreover, the structure shown by the said embodiment is only one Embodiment of this invention, and is not the meaning which limits this invention to the said structure.

DESCRIPTION OF SYMBOLS 10 Speech dialogue system 12 Speech recognition part 13 Intent understanding part 14 Dialog management part 15 Response sentence generation part 16 Speech synthesis part 131 Impression database 200 User (speaker)

Claims (6)

A spoken dialogue system that speaks with people about topics in a specific domain,
A speech recognition unit that recognizes a speaker's speech and generates an utterance sentence;
An intention understanding unit that analyzes the spoken sentence and understands what information the speaker wants to know in the domain;
Depending on the content of the information that the speaker wants to know, either the first information source that holds various information in the domain or the second information source that holds the impression of others in the domain is searched. A dialogue management unit that obtains desired information;
A response sentence generation unit that generates a response sentence for the utterance of the speaker using the target information;
A speech dialogue system comprising: a speech synthesizer for speech synthesis of the response sentence to generate a speech signal. The spoken dialogue system according to claim 1, wherein the first information source is a website on the Internet. An impression database as the second information source;
The dialog management unit acquires a user review from a website on the Internet regularly or irregularly, extracts a sentence including an evaluation expression from the user review, and registers the sentence in the impression database. The spoken dialogue system according to claim 1 or 2. A way to talk with people about topics in a specific domain,
The speech recognition unit generates speech by recognizing the speech of the speaker,
An intent understanding unit analyzes the utterance sentence to understand what information the utterer wants to know in the domain,
The dialogue management unit has a first information source that holds various information in the domain and a second information source that holds impressions of others in the domain according to the content of information that the speaker wants to know. Search for one to get the information you want,
A response sentence generation unit generates a response sentence for the utterance of the speaker using the target information;
A voice dialogue method, wherein a voice synthesizer synthesizes the response sentence with voice to generate a voice signal. 5. The voice interaction method according to claim 4, wherein the first information source is a website on the Internet. The second information source is an impression database;
The dialog management unit acquires a user review from a website on the Internet regularly or irregularly, extracts a sentence including an evaluation expression from the user review, and registers the sentence in the impression database. The voice interaction method according to claim 4 or 5.

Images