JP2015148758A - Voice interactive system and voice interactive method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make a more appropriate reply for the intension of a user when continuing a natural conversation with the user.SOLUTION: A voice interactive system 100 which repeats interaction a plurality of times by combining a speech from a user 200 with a reply to the speech includes: a reply sentence database 170 which stores a plurality of reply sentences indicating reply candidates; an interaction recording section 160 which records the content of a predetermined number of past interactions with the user; a keyword extraction section 110 which extracts a keyword in the content of the interaction from the interaction recording section 160; a word extraction section 130 which extracts a word included in a speech which is input by the user; a selection section 140 which selects a reply sentence corresponding to the extracted word and the extracted keyword from among the reply sentences stored in the reply sentence database; and an output section 150 which outputs the selected reply sentence to the user by voice, as a reply to the speech.

Description

  The present invention relates to a voice dialogue system and a voice dialogue method, and more particularly, to a voice dialogue system and a voice dialogue method for performing continuous dialogue with a user.

  As a voice dialogue system in which a user and a robot talk by voice, for example, there is a technique according to Patent Document 1. The technique according to Patent Document 1 extracts keywords from user's utterance contents, searches a conversation database or the like with the extracted keywords, and outputs the hit contents as a response by voice.

JP 2006-171719 A

  Here, when exchanging a plurality of conversations on a topic with each other, it is common that the next utterance is performed based on words or the like that have appeared in the previous conversation. Therefore, even an important word in a conversation topic may be omitted in subsequent utterances.

  In Patent Document 1, only the expression included in the utterance subject to dialogue is analyzed, and the word obtained by the analysis is searched as a keyword. Therefore, content different from the user's intention is hit as a response. there is a possibility.

  For example, after the user says, “What is Jupiter made of?” And the robot responds, “Jupiter is made of gas and is the largest planet in the solar system.” , The robot may respond such as “Saturn's diameter is about XX kilometers”. However, the "diameter" in the user's second utterance "Tell me the diameter" was originally intended to be "the diameter of Jupiter", but it is called "Jupiter" because it is a continuation of the previous dialogue. The words are omitted. Therefore, “Jupiter” is not included in the search keyword, indicating that a response regarding a certain diameter has been hit from the database. As described above, in the technique according to Patent Document 1, when an important keyword in a specific topic is omitted in the input from the user, the omitted keyword is not considered, and thus the user's intention is accurately determined. There is a problem that cannot be grasped.

  The present invention has been made to solve such a problem, and in a case where a natural dialogue with the user is continuously performed, a voice dialogue for realizing a more accurate response to the intention of the user. It is an object to provide a system and a voice interaction method.

A spoken dialogue system according to a first aspect of the present invention includes:
A spoken dialogue system that continuously conducts a dialogue that is a combination of an utterance from a user and a response to the utterance a plurality of times,
A response sentence database in which a plurality of response sentences indicating candidate responses are stored in advance;
A dialog recording unit for recording the past dialog content with the user a predetermined number of times;
A keyword extraction unit for extracting keywords in the dialogue content from the dialogue recording unit;
A word extraction unit that extracts a word included in the utterance when a new utterance is input from the user;
A selection unit that selects a response sentence corresponding to the extracted word and the extracted keyword among a plurality of response sentences stored in the response sentence database;
An output unit that outputs the selected response sentence to the user by voice as a response to the utterance;
Is provided.

The voice interaction method according to the second aspect of the present invention includes:
A voice dialogue method using a voice dialogue system that continuously performs a dialogue that is a combination of an utterance from a user and a response to the utterance a plurality of times,
The spoken dialogue system includes:
A response sentence database in which a plurality of response sentences indicating candidate responses are stored in advance;
A dialogue recording unit that records a past number of dialogues with the user a predetermined number of times,
Extracting keywords in the dialogue content from the dialogue recording unit,
When a new utterance is input from the user, a word included in the utterance is extracted,
Selecting a response sentence corresponding to the extracted word and the extracted keyword from a plurality of response sentences stored in the response sentence database;
The selected response sentence is output to the user by voice as a response to the utterance.

  As described above, in each aspect of the present invention, not only words included in the utterance contents at the time of the user's utterance but also words (keywords) included in the latest conversation contents, response sentences corresponding to these words are included. select. Therefore, even if the user has made an utterance omitting an important keyword based on the content of the previous conversation, the user's intention can be grasped, and a response sentence according to the content of the previous conversation is output. can do.

  According to the present invention, it is possible to provide a voice dialogue system and a voice dialogue method for realizing a more accurate response to a user's intention when a natural dialogue with the user is continuously performed.

It is a figure which shows the structure of the speech dialogue system concerning Embodiment 1 of this invention. It is a flowchart which shows the flow of the response process concerning Embodiment 1 of this invention. It is a figure which shows the structure of the voice dialogue system concerning Embodiment 2 of this invention. It is a flowchart which shows the flow of the response process concerning Embodiment 2 of this invention.

  Hereinafter, specific embodiments to which the present invention including the above-described aspects are applied will be described in detail with reference to the drawings. In the drawings, the same elements are denoted by the same reference numerals, and redundant description will be omitted as necessary for the sake of clarity.

<Embodiment 1 of the Invention>
FIG. 1 is a diagram showing a configuration of a voice interaction system 100 according to the first exemplary embodiment of the present invention. The spoken dialogue system 100 is an information system that continuously performs dialogue with the user 200 a plurality of times. Here, the dialogue refers to a combination of an utterance from the user and a response to the utterance. The voice interactive system 100 may be, for example, an interactive robot. The spoken dialogue system 100 includes a keyword extraction unit 110, an utterance reception unit 120, a word extraction unit 130, a selection unit 140, an output unit 150, a dialogue recording unit 160, and a response sentence database (DB) 170. .

  The dialogue recording unit 160 records past dialogue contents with the user 200 a predetermined number of times. Here, the dialogue content includes text obtained by converting speech from the user 200 to the voice dialogue system 100 into text data, and text data that is a response from the voice dialogue system 100 to the user 200. That is, the dialogue recording unit 160 records and holds history data for the past several times of dialogues that are continuously performed with the user 200 regarding a specific topic.

  The response sentence DB 170 is a database in which a plurality of response sentences indicating candidates for responses to the utterances from the user 200 are stored in advance. The dialogue recording unit 160 and the response sentence DB 170 are realized by an internal storage device (not shown) such as a robot related to the voice dialogue system 100. However, the dialogue recording unit 160 and the response sentence DB 170 may be realized in an external storage device such as the robot.

  The keyword extraction unit 110 extracts keywords in the past dialogue contents from the dialogue recording unit 160. Here, it is assumed that the keyword extraction unit 110 uses, as keywords, typical words related to topics in a series of conversations among words included in past conversation contents recorded in the conversation recording unit 160. That is, it is desirable that the keyword extraction unit 110 uses not all the words in the dialogue recording unit 160 but some words, and the characteristic words in the continuous dialogue as keywords. For example, of the past dialog contents, words that appear frequently in the latest several times may be used as keywords in the next utterance. Or it is good also considering a word with high importance in a predetermined standard among the contents of past dialogue as a keyword.

The utterance reception unit 120 receives an utterance input from the user 200, converts the utterance into text data, and stores the text data in the dialogue recording unit 160. When a new utterance is input from the user 200, the word extraction unit 130 extracts a word included in the utterance. Here, it is assumed that the word extraction unit 130 extracts all or some of the words included in the input utterance. When extracting some words, the word extraction unit 130 may extract words that are highly important based on a predetermined criterion. The selection unit 140 selects a response sentence corresponding to the extracted word and the extracted keyword from the plurality of response sentences stored in the response sentence DB 170. Here, as a response sentence corresponding to the extracted word and the extracted keyword, a response sentence including the word and the keyword itself, or a word or keyword that is not included, is related to the word or keyword. It is good also as a response sentence containing another word. For example, the selection unit 140 may specify another word related to the word or the keyword and select a response sentence including the specified word.
The output unit 150 converts the selected response sentence into a voice as a response to the utterance and outputs it to the user 200.

  In addition, as specification of the other word relevant to a word, you may use the technique of a predicate term structure analysis, for example. In that case, the selection unit 140 selects a sentence having a similar analyzed structure. Here, as the predicate term structure analysis, for example, “Koiichiro Yoshino et al.,“ Spoken dialogue system that recommends information based on similarity of predicate terms ”, Information Processing Society of Japan, Vol. 2011-SLP-87, No. 11 ″ can be applied. In that case, the word extraction unit 130 performs morphological analysis on the target text data, and extracts a plurality of words included in the text data. First, a response sentence including all of the extracted search words that are a plurality of words and keywords is searched from the response sentence DB 170. If the search cannot be performed, the selection unit 140 selects a part of the search word, The response sentence including the selected search word is searched from the response sentence DB 170. At this time, for example, the selection unit 140 calculates the degree of association between elements by a predetermined method for each search word, and the degree of association is higher. Search word This may be selected as a part of the search term, thereby improving the accuracy of the selected response sentence, although the same applies if the word extraction unit 130 is replaced with the keyword extraction unit 110 in the above. Techniques other than predicate term structure analysis may be applied.

  FIG. 2 is a flowchart showing a flow of response processing according to the first exemplary embodiment of the present invention. As a premise, it is assumed that records of a plurality of dialogues (utterances and responses) regarding a specific topic are already held in the dialogue recording unit 160 between the user 200 and the voice dialogue system 100.

  First, the keyword extraction unit 110 refers to the dialogue recording unit 160 and extracts keywords from past dialogue records (S11). In addition, the utterance receiving unit 120 receives a new utterance and records text data converted from the received utterance in the dialogue recording unit 160 (S12). Subsequently, the word extraction unit 130 extracts words from the received utterance text data by morphological analysis or the like (S13). In the first embodiment, the order of step S11 and steps S12 and S13 does not matter.

  Next, the selection unit 140 selects a response sentence corresponding to the keyword extracted in step S11 and the word extracted in step S13 from the response sentence DB 170 (S14). Then, the selection unit 140 records the selected response sentence in the dialogue recording unit 160 (S15). Further, the output unit 150 converts the selected response sentence into speech and outputs it to the user 200 (S16).

  As described above, in the first embodiment of the present invention, not only the words included in the utterance content received from the user 200 in step S12 but also the words included in the conversation content recorded in the dialog recording unit 160 before that are included. Then, the response sentence corresponding to these is selected. Therefore, even if the user has made an utterance omitting important keywords based on the content of the previous conversation, the user's intention can be accurately grasped, and a response sentence in accordance with the content of the previous conversation. Can be output.

<Embodiment 2 of the Invention>
The second embodiment of the present invention is an improvement over the first embodiment described above. In the first embodiment, every time an utterance from the user 200 is received, the response sentence DB 170 is searched to obtain a response sentence. That is, each time a word is input, it is necessary to collate the input word with the response sentence in the response sentence DB 170. For this reason, a certain amount of processing time is required from receiving an utterance to responding. In order to continue a more natural conversation with the user 200, it is desirable to shorten the response time.

  Here, a large number of response sentences are stored in the response sentence DB 170 of FIG. 1 in order to correspond to utterances on various topics with the user 200. However, when a conversation is continuously performed on a specific topic, the response sentence that is actually selected is limited to that related to the specific topic. Therefore, in the second embodiment of the present invention, response sentence candidates are selected in advance from the response sentence DB 170 based on the history of conversations that have already been performed before accepting a new utterance from the user 200. It is. For example, a response sentence candidate of an utterance that will be accepted next is searched in parallel with the immediately preceding response process and stored as a cache. Thereby, it is not necessary to collate with the response sentence DB 170 in which a large amount of data is stored after receiving a new utterance, and the response processing time can be shortened.

  FIG. 3 is a diagram showing a configuration of the voice interaction system 100a according to the second embodiment of the present invention. FIG. 3 is an improvement of the above-described voice interaction system 100. The same components as those in the voice interaction system 100 are denoted by the same reference numerals, and detailed description thereof is omitted.

  The voice interaction system 100a differs from the voice interaction system 100 in that the keyword extraction unit 110 is replaced with the keyword extraction unit 110a and the selection unit 140 is replaced with the selection unit 140a, and the word importance DB 180 and the cache 190 are added.

  The word importance DB 180 is a database in which importance based on a predetermined standard is defined for each of a plurality of words. The predetermined standard is, for example, a standard according to each of a plurality of topics and themes. Further, the importance is calculated based on the appearance frequency of words for each word included in a document set in a certain topic. For example, the importance can be calculated using a known technique such as tf-idf (tf: Term Frequency, idf: Inverse Document Frequency). Alternatively, the importance of each word may be set in advance by a human being for each topic.

  The cache 190 is a partial storage unit that can store a part of a plurality of response sentences. That is, the amount of data stored in the cache 190 is smaller than the amount of data stored in the response sentence DB 170. The cache 190 may be realized by a storage device that is faster than the storage device that implements the response sentence DB 170, such as a primary storage device.

  In addition to the function of the keyword extraction unit 110, the keyword extraction unit 110a refers to the word importance DB 180 when a plurality of words are extracted as keywords from the dialogue recording unit 160, and among the extracted plurality of words, the keyword extraction unit 110a A word whose degree is a predetermined value or more is set as a keyword. The predetermined value can be set arbitrarily. The keyword extraction unit 110a is executed before a new utterance from the user is input. For example, it is executed in parallel with the processing of the selection unit 140a and the output unit 150 by the immediately preceding utterance. Alternatively, it may be executed when the user 200 has a certain time or more before the next utterance is made.

  The selection unit 140a includes a related sentence selection unit 141 and a response sentence selection unit 142. The related sentence selection unit 141 selects a plurality of response sentences (related sentences) including the keywords extracted from the response sentence DB 170 before a new utterance from the user is input. Then, the related sentence selection unit 141 stores the selected plurality of response sentences in the cache 190. After the utterance from the user 200 is newly input, the response sentence selection unit 142 selects a response sentence corresponding to the word included in the utterance from the cache 190. Note that the method of selecting a response sentence corresponding to a word in the response sentence selection unit 142 is to specify another word related to the word, and to select a response sentence including the specified other word, as in the first embodiment. You may make it select.

  In the present embodiment, the keyword extraction unit 110a and the word importance DB 180 are not essential. Even in this case, the processing time can be shortened. And the precision which selects a response sentence can be improved by using the keyword extraction part 110a and the word importance DB180.

  FIG. 4 is a flowchart showing a flow of response processing according to the second embodiment of the present invention. In the following, description of processing equivalent to that in FIG. 2 is omitted. The keyword extraction unit 110a extracts keywords from past dialogue records and word importance (S11a). That is, the keyword extraction unit 110a extracts a plurality of words from the dialogue recording unit 160 during or after the response process for the immediately preceding utterance and before the next utterance is input. Then, the keyword extraction unit 110a sets keywords whose importance is equal to or higher than a predetermined value for each extracted word.

  Next, the related sentence selection unit 141 selects a plurality of response sentences including keywords from the response sentence DB 170 as related sentences (S11b). The related sentence selection unit 141 selects a part of the response sentence in the response sentence DB 170. Then, the related sentence selection unit 141 stores the selected related sentence in the cache 190 (S11c).

  Thereafter, steps S12 and S13 are executed as in FIG. Then, the response sentence selection unit 142 selects a response sentence corresponding to the word extracted from the cache 190 in step S13 (S14a). Thereafter, steps S15 and S16 are executed as in FIG.

  As described above, in this embodiment, response sentence candidates are listed in advance based on past conversation records, and then, when uttered, responses are selected from the listed (restricted) candidates. Since the sentence is selected, the processing time can be shortened. Therefore, even if an utterance in which an important keyword is omitted based on the contents of the conversation so far, the user's intention can be accurately grasped in a short time.

<Other embodiments of the invention>
Note that the present invention is not limited to the above-described embodiment, and can be modified as appropriate without departing from the spirit of the present invention. For example, in the above-described embodiment, the present invention has been described as a hardware configuration, but the present invention is not limited to this. The present invention can also realize arbitrary processing by causing a CPU (Central Processing Unit) to execute a computer program.

  In the above example, the program can be stored and supplied to a computer using various types of non-transitory computer readable media. Non-transitory computer readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (for example, flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (for example, magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / W, DVD (Digital Versatile Disc), BD (Blu-ray (registered trademark) Disc), semiconductor memory (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM ( Random Access Memory)). The program may also be supplied to the computer by various types of transitory computer readable media. Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves. The temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

DESCRIPTION OF SYMBOLS 100 Speech dialogue system 100a Voice dialogue system 110 Keyword extraction part 110a Keyword extraction part 120 Speech reception part 130 Word extraction part 140 Selection part 140a Selection part 141 Related sentence selection part 142 Response sentence selection part 150 Output part 160 Dialog recording part 170 Response sentence DB
180 word importance DB
190 cache 200 users

Claims (5)

A spoken dialogue system that continuously conducts a dialogue that is a combination of an utterance from a user and a response to the utterance a plurality of times,
A response sentence database in which a plurality of response sentences indicating candidate responses are stored in advance;
A dialog recording unit for recording the past dialog content with the user a predetermined number of times;
A keyword extraction unit for extracting keywords in the dialogue content from the dialogue recording unit;
A word extraction unit that extracts a word included in the utterance when a new utterance is input from the user;
A selection unit that selects a response sentence corresponding to the extracted word and the extracted keyword among a plurality of response sentences stored in the response sentence database;
An output unit that outputs the selected response sentence to the user by voice as a response to the utterance;
A voice dialogue system comprising: A partial storage unit capable of storing a part of the plurality of response sentences;
The selection unit includes:
Before a new utterance from the user is input, select a plurality of response sentences including the extracted keyword from the response sentence database;
Storing the selected plurality of response sentences in the partial storage unit;
The speech dialogue system according to claim 1, wherein a response sentence including a word included in the utterance is selected from the partial storage unit after a new utterance from the user is input. A word importance database in which importance based on a predetermined criterion is defined for each of a plurality of words;
The keyword extracting unit refers to the word importance database when a plurality of words are extracted as the keyword from the dialogue recording unit, and the importance is equal to or higher than a predetermined value among the extracted words. The spoken dialogue system according to claim 1, wherein a word is the keyword. The selection unit specifies the extracted word or another word related to the extracted keyword, does not include any of the extracted word or the extracted keyword, and The voice dialogue system according to any one of claims 1 to 3, wherein a response sentence including a word is selected as the corresponding response sentence among the plurality of response sentences. A voice dialogue method using a voice dialogue system that continuously performs a dialogue that is a combination of an utterance from a user and a response to the utterance a plurality of times,
The spoken dialogue system includes:
A response sentence database in which a plurality of response sentences indicating candidate responses are stored in advance;
A dialogue recording unit that records a past number of dialogues with the user a predetermined number of times,
Extracting keywords in the dialogue content from the dialogue recording unit,
When a new utterance is input from the user, a word included in the utterance is extracted,
Selecting a response sentence corresponding to the extracted word and the extracted keyword from a plurality of response sentences stored in the response sentence database;
A voice dialogue method for outputting the selected response sentence to the user by voice as a response to the utterance.

Images