JP4353212B2 - Word string recognition device - Google Patents

Description

  The present invention relates to a word string recognition device capable of recognizing an appropriate word string when, for example, a recognition result candidate group of various signals represented by speech recognition is obtained as a set of discrete word strings. is there.

  2. Description of the Related Art Conventionally, there is known a recognition device that inputs, for example, a voice uttered by a human and outputs a plurality of word string candidates having a high degree of matching compared to dictionary data. It is applied to a control system that recognizes the utterance content and operates the device in accordance with the recognition result. For example, when a user speaks a predetermined word as a voice command, an automobile navigation device that performs an information search operation corresponding to the word has been put into practical use.

  By the way, there are two methods of current speech recognition: continuous word recognition and word spotting. According to the former recognition method, for example, it is possible to recognize each word such as “okazaki, ○○○○○ (store name), ramen”, but in practice, it is impossible to avoid misrecognition. Therefore, in a current navigation device or the like, there are a plurality of candidates as recognition results, and one of them is talked back first to ask the user for confirmation. If the user is different from the content that he / she uttered, the user informs the apparatus that the recognition result is different. Then, the apparatus side presents another recognition result and requests user confirmation again. If a large number of recognition result candidates are prepared and presented one after another, the result can be the recognition result of the content intended by the user, but some of the candidates have no meaning at all. May be included, and it may take a long time to finally reach the appropriate candidate described above.

  In addition, the latter word spotting method is, for example, “Okazaki, ○○○○○ (store name),“ Okazaki, ○○○○○ (store name), This is a speech recognition technology that has been attracting rapid attention in recent years because it can extract the keyword "Ramen, Kuitai". However, in the case of this method, the number of word string candidates generated from the output called a lattice (a word set having time interval information and probability information) is very large, and is narrowed down to a small number of meaningful word strings. It is rare to be done. Moreover, the current number of recognized vocabulary in the method is about 100 words, but it is expected to increase to 1000 or more in the future, and the number of word strings generated from the resulting lattice is enormous. It is thought that it goes up. Therefore, there is a problem similar to (or rather more prominent) the case described in the case of the continuous word string recognition method described above.

  Moreover, with the current speech recognition technology, it has been difficult to return to normal dialogue at the time of misrecognition based on noise or fluctuations in speech environment. For example, if you want to enter a place name, if you enter the mode to select from station names unintentionally, you will not be able to properly recognize other place names. Having a dictionary structure that can be narrowed down to reduce misrecognition is preferable in a sense, and is used in many word recognition methods, but if it falls into an unintended recognition mode as described above, There is a possibility that subsequent desired input cannot be performed. If the user does not know the operation for escaping from the mode, it is difficult to return to a normal dialogue, and the user may be at a loss.

  The present invention has been made in view of these problems, and can prevent inconveniences caused by misrecognition, such as, for example, falling into a recognition mode unintended by the user and being unable to recover, and the user is at a loss. An object of the present invention is to provide a word string recognition device.

In order to achieve the above purpose, we propose the expected outside during the corresponding.

In the word string recognition apparatus according to claim 1, the unexpected time response means predicts whether or not the word string output from the word string output means has an expected content in accordance with the context, that is , predicts in a certain topic. Is determined based on the request keyword indicating what the request in the syntax is, and if it is determined that the request keyword is different from the current topic, Topic change confirmation process that asks to confirm whether the topic has changed, Topic change declaration process that declares that the topic has been changed, Assuming that the previous topic continues, it follows the context One of the context-first correspondence processing is performed. The unexpected time response means executes the topic change declaration process if the word string output from the word string output means after execution of the topic change confirmation process is a content in line with the changed topic.

The word string output means inputs information reflecting the action content of the human being, and outputs a word string candidate having a high degree of matching compared with the recognition dictionary data .

As described above, the context-first correspondence processing is based on the assumption that the previous topic is continuing, and responds according to the context. Accordingly, the unexpected response means may use the topic change confirmation process and the context priority response process properly. As shown in claim 2, even if the topic is changed, the context priority confirmation processing is executed immediately after the change, and the change of the topic is confirmed only when the changed topic continues thereafter. The unexpected time response means is configured to execute the processing.

In the word string recognition device according to claim 3, the unexpected time response means predicts whether or not the word string output from the word string output means has an expected content in accordance with the context. Is determined based on the request keyword indicating what the request in the syntax is, and if it is determined that the request keyword is different from the current topic, Topic conversion confirmation process that asks to confirm whether the topic has been changed, Topic conversion declaration process to declare that the topic has been changed, Assuming that the previous topic is continuing, it follows the context One of the context-first correspondence processing is performed. And even when the topic is changed, the unexpected response method executes the context-first response processing immediately after the conversion, and only after the converted topic continues, confirms the topic change. Execute the process.

By doing so, it is possible to prevent the inconvenience that the user is at a loss due to misrecognition .

By the way, with respect to the word string recognition device, the word string output means inputs information reflecting the action content of the human being , and outputs a word string having a high degree of coincidence compared with the recognition dictionary data. Specifically, for example, the following can be considered.

First, as shown in claim 4 , it may be realized as a speech recognition device that compares a speech input by a person who is a recognition target with dictionary data and outputs a plurality of word string candidates having a high degree of coincidence. For example, it has been put to practical use for voice input of an instruction such as a destination in a car navigation system, and is generally considered as an application target. In other words, speech recognition is based on the acoustic features of the input speech, but the acoustic features of the speech produced by humans vary greatly between individuals, and there are many cases where accurate pronunciation is not made on a daily basis. Is likely to occur.

Further, for example, as shown in claim 5 , it may be realized as a character recognition device that compares a handwritten character string input by a person who is a recognition target with recognition dictionary data and outputs a word string having a high degree of coincidence. . As for the handwritten characters, the individual differences are large as in the case of the above-described acoustic features, and since there are many cases where characters are not written with an accurate typeface on a daily basis, misrecognition is likely to occur.

  Therefore, by preventing such misrecognition, or by taking appropriate measures after the misrecognition has occurred, it is possible to prevent inconveniences such as the user getting lost due to misrecognition. it can.

Various forms of character recognition are conceivable. For example, as well as recognition when a character written with a writing instrument is read with a scanner, as well as on a PDA (personal digital assistant), the screen is viewed with an input pen. Recognition in the case of a character input method such as tracing may be used. Furthermore, as in the case of such voice recognition and character recognition, the contents of the word string are not limited to those directly input to the recognition device, but may be an image recognition device. That is, after recognizing an image that captures a human being as a recognition target as a scene, the dictionary data for converting the scene into a natural language is compared with the recognition scene, and a plurality of word string candidates having a high degree of coincidence are output. An image recognition device. As a specific example, it is conceivable to recognize a sign language pattern from an image in which a human being who is being recognized is sign language, and to output a word string indicating a natural language meaning represented by the sign language pattern. In this case, since the correspondence pattern between the sign language pattern and the word has been established, it is possible to easily output a word string indicating a natural language meaning by the pattern matching. However, also in the sign language pattern in this case, misrecognition still occurs due to individual differences of the sign language person because the words expressed by subtle fingering are different.

  Therefore, in this case as well, even in a sign language pattern recognition device that is likely to include misrecognition, the present invention can be applied to prevent misrecognition or to take appropriate measures after misrecognition has occurred. It is possible to prevent such inconvenience that the user is at a loss due to misrecognition.

Moreover, the ability to implement the process performed by the put that expectations out time corresponding hand stage word string recognition apparatus described above in a computer system, for example, may be provided as a program to be started on the computer system side. In the case of such a program, for example, the program is recorded on a computer-readable recording medium such as a flexible disk, a magneto-optical disk, a CD-ROM, and a hard disk, and is used by being loaded into a computer system and started up as necessary. it can. In addition, the ROM or backup RAM may be recorded as a computer-readable recording medium, and the ROM or backup RAM may be incorporated into a computer system and used.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
First, FIG. 1 is a block diagram conceptually showing the word string recognition device of the embodiment focusing on the function.

  A signal represented by speech input via the user interface is subjected to a predetermined recognition process in a speech recognition unit or other signal system recognition unit, and a word string candidate is output. The voice recognition unit has functions of voice recognition and word string generation, recognizes a voice input corresponding to a certain utterance using a recognition dictionary (recognition vocabulary dictionary), and collects words registered in the recognition dictionary Get word string candidates as This is the voice recognition function here, but at this point, a plurality of candidate words may be shown at the same time. Thereafter, by using the word string generation function, a plurality of word strings that do not overlap in terms of time interval are generated based on the words obtained by the voice recognition function, and are output to the dialogue management means. At this time, the likelihood is also output along with the word string. As described above, a speech recognition result for one speech utterance usually has a plurality of candidates with likelihood information, and is referred to as “well-up” in speech recognition.

  Then, the dialogue management means recognizes the word string candidate accompanied by the likelihood information by applying the context and semantic constraint information, and dynamically updates the recognition dictionary using the dictionary construction means. Furthermore, the dialogue management unit performs display control for the display system, and controls the audio output system and the device control system via the screen control unit.

  The above is a block diagram conceptually showing the word string recognition device focusing on the function. Next, it is a block diagram showing a configuration when the word string recognition device is applied to an in-vehicle control system. This will be described with reference to FIG. This control system is a system that is mounted on an automobile (vehicle) and controls various devices mounted on the vehicle while talking with a vehicle occupant (mainly a driver) as a user. .

  As shown in FIG. 2, the control system of the present embodiment includes a control device 1, an input device 3 for a user to input various commands and data by an external operation, and a microphone 5 for inputting sound. A speaker 7 for outputting sound, a display (display device) 8 for displaying an image, a known navigation device 9 for detecting the current position (current location) of the vehicle, route guidance, and the like. An air conditioner 13 for controlling air conditioning, an audio device 15 including a cassette tape recorder, a CD (compact disc) player, an MD (mini disc) player, a radio, a television, and the like, and a well-known VICS (Vehicle Information and Communication System) Wireless data communication between broadcast terminals and Internet broadcast terminals, which are the connection points for the Internet A communication device 17 to perform, a vehicle operation state such as a vehicle speed and acceleration / deceleration state, various sensors 19 for detecting the temperature inside and outside the vehicle, and the presence / absence of raindrops, a vehicle door lock, a window glass (power window), an engine , And other control devices (not shown) for controlling the brake device and the like.

  The navigation device 9 is a well-known GPS device for detecting the current position of the vehicle, a CD-ROM storing route guidance data such as map data, place name data, and facility name data, and data from the CD-ROM. A CD-ROM drive for reading out and an operation key for a user to input a command. The navigation device 9 is optimal for reaching the current position of the vehicle and the destination when, for example, a command for instructing the route to the destination and the destination is input via the operation key from the user. A road map including various routes is displayed on the display 8 to provide route guidance. The display 8 not only displays a road map for route guidance by the navigation device 9, but also displays various images such as an information search menu, and the audio device 15 is set to a television mode. Then, the received image of the television received by the television tuner provided in the audio device 15 is displayed.

  The control device 1 includes a system control unit 21 mainly composed of a microcomputer including a CPU, a ROM, a RAM, and the like, and an interface (I / O) for inputting commands and data from the input device 3 to the system control unit 21. F) 23, a voice input unit 25 for converting a voice signal input from the microphone 5 into digital data and inputting the digital data to the system control unit 21, and a text data output from the system control unit 21 converted into an analog voice signal Then, the voice synthesizing unit 27 that outputs to the speaker 7 and causes the speaker 7 to ring, the screen control unit 28 that controls the display screen on the display 8, the navigation device 9, the air conditioner device 13, the audio device 15, and the communication device. 17. Various sensors 19 and other control devices and the system control unit 21 are connected so that data communication is possible. That device control interface and a (device control I / F) 29.

  In addition, the control device 1 includes an Internet address database 31 that stores an Internet address (Internet address) and a search control unit 33 in order to search and acquire desired information from the Internet by the communication device 17. Yes. Then, when the system control unit 21 outputs a search keyword representing the search content (content) to the search control unit 33, the search control unit 33 operates the communication device 17 via the device control I / F 29, and the Internet broadcast terminal. The information corresponding to the search keyword is searched for, and the search result is input to the system control unit 21. The Internet address database 31 stores Internet addresses used in the past by the search control unit 33 according to commands from the system control unit 21, and the search control unit 33 searches the system control unit 21 in the past. When the same search keyword as the keyword is received, the Internet address in the Internet address database 31 is reused.

  On the other hand, the control device 1 recognizes and acquires a keyword (hereinafter also referred to as an utterance keyword) as a word spoken by the user from an audio signal input through the microphone 5 and the audio input unit 25. A recognition vocabulary storage unit 34 that stores in advance a plurality of recognition vocabulary that is assumed to be uttered by the user and should be recognized by the control device 1 is provided. That is, the recognized vocabulary group stored in the recognized vocabulary storage unit 34 is a recognized vocabulary database of the control device 1.

  Furthermore, the control device 1 sets the content of the utterance (hereinafter also referred to as agent utterance) output from the speaker 7 (that is, the operation content of the speaker 7) and the operation content of the device M other than the speaker 7. As a means for storing data for estimating the user's request and the mental or physical state of the user, a dialog data storage unit 35 for storing a dialog database, and request / state estimation data A request / state estimation data storage unit 36 to be stored and a user profile storage unit 37 to store personal information (hereinafter also referred to as a user profile) for a plurality of users. The dialog data storage unit 35, the request / state estimation data storage unit 36, the user profile storage unit 37, and the Internet address database 31 described above are configured by a nonvolatile memory capable of reading and writing data. ing.

  The dialog database stored in the dialog data storage unit 35, the request / state estimation data stored in the request / state estimation data storage unit 36, the user profile stored in the user profile storage unit 37, etc. See Japanese Patent Application Nos. 10-162457 and 10-184840.

  Next, an outline of processing executed by the system control unit 21 in the control system according to the present embodiment configured as described above will be described with reference to FIG. Here, in FIG. 1, the speech recognition unit performs the recognition process, and the dialogue management means performs a predetermined process on the word string candidate that is the recognition result. explain.

  When the process is started, a recognition process is performed on the input voice (S10). Then, based on the determination result in S20, it is determined that “waiting for input” and the process returns to S10, multi-stage processing (S30) is performed, or it is determined that the recognition is completed or timed out, and the process proceeds to S40. Details of the multi-stage processing in S30 will be described later.

  After the transition to the dialog management in S40, it is determined whether or not the response is an unexpected response (S50), and if it is an unexpected response, the flow proceeds to S60 to determine the type of unexpected response. Depending on the determination result, confirmation of topic change (S70), utterance after topic change (S80), or context-priority utterance (S90) is executed. Thereafter, the process returns to S10 through the speech process (S180).

  On the other hand, if it is determined in S50 that the response is not unexpected (that is, the response is in line with the expectation), the process proceeds to S100. If it is determined in the determination process in S100 that the number of questions returned from the agent has exceeded N times, a help mode process is executed (S120). In the determination process in S100, the number of questions returned from the agent is K. If the number of times exceeds (K> N), or if there is a reset request from the user, the process proceeds to S170 and returns to the initial state. In other cases, that is, when the number of questions returned from the agent is N or less, processing for presenting options on the display 8 based on the determination result in S110 (S130), normal speech strategy (S140) Then, either the voice menu mode (S150) or the inquiry (S160) is selectively executed. Details of these will be described later.

After any process of S120 to S170 is executed, the process returns to S10 through the speech process (S180).
The above is a description of the schematic flow of processing, and the detailed contents will be described subsequently. However, here, description will be made in the order of the conceptual units of the present invention: progressive hierarchical search, multi-stage processing, unexpected response processing, and misrecognition processing. However, for the convenience of explanation, the unexpected response processing and the misrecognition processing are described in [3. The misrecognition handling process] is summarized as follows.
[1. About progressive hierarchical search]
[1.1 Overview]
The progressive hierarchy search corresponds to the process of presenting the options of S130 in the flowchart of FIG. However, since each word is processed one by one, the time cycle is different from other processes, so it is necessary to set a mode in which a progressive hierarchical search is executed in advance. Therefore, S130 of FIG. 3 is executed only when the progressive hierarchical search mode is set.

[1.2 Specific example]
FIG. 4 shows a specific screen transition example of the progressive hierarchical search.
(1) The first screen G1 shown in FIG. 4 indicates that, as a result of uttering “Aichi Prefecture” in a state where the prefecture names of the four prefectures of Tokai are displayed in the default mode, the focus is on Aichi Prefecture.

  Then, when the control system recognizes the word Aichi Prefecture, it performs the “gradual hierarchy search” function, and immediately displays the utterance vocabulary expected for the user on the screen. In this case, the names of municipalities in Aichi Prefecture are displayed as shown on the screen G2. Although only four municipalities are listed on the screen G2, this is for simplifying the explanation. The screen G2 indicates that the user has focused on Okazaki as a result of speaking “Okazaki”.

  Then, when Okazaki is input, the control system side exhibits the “gradual hierarchy search” function so that the utterance vocabulary expected for the user next can be presented on the screen. The reason for "being ready to present" here is because, in practice, the next word is often spoken continuously like "I want to eat in Okazaki", so it is not presented in practice. It is. When there is a silent period of a predetermined time (for example, 1 to 2 seconds) after Okazaki, it is assumed that the user is at a loss and the utterance vocabulary is presented on the screen. That is, “Okazaki's request menu”. The content is a portion excluding “Please speak again” in the request menu of Okazaki on the screen G7 used for erroneous recognition. When the screen G3 is displayed and “Indian food is good” is spoken, the focus is on Indian food as shown in the screen G4. As shown in the screen G5, “Indian food in Okazaki City, Aichi Prefecture. When the search is completed, the search result is displayed as shown in screen G6.

  On the other hand, when a vocabulary other than the vocabulary that is expected to be spoken next is spoken when Okazaki is spoken on screen G2, the screen moves to screen G7 and says "Please speak again" along with the "Okazaki request menu". Display the content with guidance. Here, if “meal” is spoken, the screen shifts to a screen G3. If “What is a department store?” Is spoken, a corresponding department store is displayed in a list as shown in a screen G8. Then, when any of them is designated and “map of ΔΔΔ” is uttered, as shown on the screen G9, the map is displayed together with the surrounding maps so that the position of the department store is clear.

  (2) FIG. 5 is also an example of a progressive hierarchical search. Screens G1 to G3 in FIG. 5 are the same as the screen contents shown in FIG. 4, but when “Tokyo” is spoken in the default mode in which the prefecture names of the four prefectures of Tokai are displayed, the screen shifts to screen G11. . In this case, since the term “Tokyo” does not necessarily refer to only Tokyo, the area around Tokyo may be used.

  Then, when the word “Tokyo” is recognized, the control system performs the “gradual hierarchy search” function, and immediately presents the utterance vocabulary expected for the user to the screen. In this case, as shown on the screen G12, the names of cities, towns and villages in Tokyo are displayed on the screen. And when I speak with Ginza, I focus on Ginza.

  Then, when the control system recognizes the word Ginza, it performs a “gradual hierarchy search” function so that the utterance vocabulary expected for the user next can be presented on the screen. In this case, “to be able to present” is the same as the above-described transition from the screen G2 to G3. In fact, the next word such as “Ginza, XXX Building” continues. This is because it is not presented in practice. Presented is a case where there is a silent section for a predetermined time (for example, 1 to 2 seconds) after Ginza. The utterance vocabulary presented on the screen is place name related information in Ginza. For example, the place name itself may be used, or in terms of usability, it may be a facility name such as XX Building or □□ Department Store.

Then, when the user utters, for example, XX building, as shown on the screen G13, it is displayed together with the surrounding map so that the position of the XX building becomes clear.
[1.3 Effect]
For example, when a user utters “I want to eat in Okazaki, Indian food is good”, the output result of the speech recognition method by word spotting is the word string “Okazaki, meal, Indian food”. In the conventional speech recognition method, the processing on the system side corresponding to the recognition is disclosed at the stage when these three are gathered, but according to this method, when “Okazaki” is input, Since the utterance vocabulary expected to be input can be presented immediately, the user can speak utterly. As a result, it is possible to prevent the user from uttering a vocabulary outside the recognition dictionary that causes misrecognition.
[2. About multi-step processing]
[2.1 Overview]
(1) With the current word spotting technique, the number of recognized vocabulary per process is about 100 words, and it is difficult to increase the scale like a dictionary for continuous word recognition. On the other hand, from a practical viewpoint, when focusing on a specific topic (for example, “going to a meal”) that is expected to occur in the passenger compartment, for example, about 100 words can be handled. Therefore, it is only necessary to accurately recognize the topic, detect the discontinuity of the topic in the case of misrecognition, and switch the word spotting vocabulary accordingly. Therefore, multistage processing is performed.

  FIG. 6 shows an example of multi-stage processing that is performed when an utterance “Yokohama no Chugai ga I want to go to” is made. In this utterance example, it can be classified into the following word attributes.

Yokohamano Chugai I would like to have a good time (place) (facility name) (request target) (request keyword)
Therefore, first, in order to grasp what is the request, in the first stage, the request keyword “I want to go” is spotted to determine the topic.

  In the second stage, the vocabulary is limited from the topics determined in the first stage, and the dictionary is switched. That is, in this case, most of the 100 words are composed of a restaurant name based on the destination and a dish name related to the restaurant name. This is because the request keyword “Kaitai” is understood to be a meal request, and therefore, it can be predicted that other words constituting the word string are restaurant names and dish names. This makes it possible to construct a dictionary in which “Chukagai”, “Syumai”, and the like are easy to hit as recognition vocabulary.

  (2) In FIG. 6, the topic is determined by spotting the request keyword as the first stage process, but the topic may be determined by spotting other facility names or places or the request target. . However, in reality, the number of vocabularies for word spotting is currently about 100 words, so that it is preferable to determine the topic with the requested keyword in view of the availability of such vocabularies.

  (3) In the specific example shown in FIG. 6, the n-dimensional axis is set from the viewpoint of the attributes of the words constituting one word string. However, the relevance along the time axis may be considered. . In other words, the topic is determined in consideration of the context.

[2.2 Basic structure of user utterance]
In the case of most content searches, the user's utterance is considered to consist of “location”, “facility name”, “request target”, “request keyword”, or a rearrangement or abbreviated form of the basic form. When the word order changes, it is unclear where the requested keyword exists in the audio signal. For example, use a method for estimating an appropriate word string as proposed in Japanese Patent Application No. 11-20349 by the applicant of the present application. Thus, a priority order can be set for a plurality of candidates based on syntactic constraints.

[2.3 Configuration of recognition dictionary]
It is assumed that the recognition dictionary can be dynamically constructed from a recognition vocabulary database stored in the recognition vocabulary storage unit 34 shown in FIG. The recognized vocabulary database is divided into vocabulary clusters for each category in advance corresponding to actual topics handled by the system (see FIG. 7). The vocabulary cluster for each category is called a cluster dictionary. The categories include the following types, for example.

(1) Various commands ・ Navigation command ・ Schedule book ・ Address book ・ Telephone (2) Request keyword (request KW)
(3) Facility name 1) Restaurant name ・ Cooking name ・ Atmosphere ・ Price 2) Ski resort name 3) Golf course name 4) Department store name 5) Amusement park name 6) Park name 7) Movie theater name 8) Hot spring (4) Event Name (5) Search result (6) Place name (7) Railway station name (8) Basic dialogue vocabulary-Affirmation, denial-Inquiry-Explanation, status notification, confirmation, ...
These constituent vocabularies include not only proper nouns that are elements of the database, but also synonyms (dialogues, meals, etc) in dialogue. From each of these, the vocabulary of one word spotting, that is, the name of the restaurant based on the destination and the name of the dish associated with the name (here also refer to the user profile) constitutes the majority of 100 words. Based on this, "Chukagai" and "Syumai" are hit as recognition vocabulary.

[2.4 Network between words]
The basic structure of the dictionary is defined according to the hierarchical expression described above, but other network relationships such as semantic relationships are provided from users and database suppliers as needed. For example, user utterances are positioned in a dialogue database (FIG. 8), which is a collection of attribute sets (called dialogue vectors) as described below.

(Topic, time / position, environment / situation, state / request, user utterance, agent utterance, control output)
There are an infinite number of interaction vectors, but they do not take all vector values, but can be clustered into meaningful finite groups as a unit of practical communication between humans and machines. It uses the semantic classification of words, grammatical constraints, topic continuity, physical / common sense constraints, and event continuity. Therefore,
(A) Range of vocabulary used for a word sequence that constitutes a certain user utterance (B) Vocabulary restrictions from the current utterance to the next utterance are the main factors that constitute the space spanned by the conversation vector (topic, Time / location, environment / situation, condition / request).

Therefore, what kind of network structure a word in a certain user utterance has with respect to other words is determined by the following factors.
(1) Relationship between cluster dictionaries (2) Relationship between words (3) Relationship between topics (4) Continuity of context (5) User characteristics and situation (6) Between applications Words based on the above relationships Another word W2 is activated from W1 and added to the vocabulary of the recognition dictionary for the next user utterance. Further, the parameters of the speech recognition module are adjusted so as to increase the likelihood value LFK accompanying the recognition result. Here, (1) relationship between cluster dictionaries, (2) relationship between words, (3) relationship between topics, (4) continuity of context, (5) user characteristics and situations will be supplementarily explained.

[2.4.1 Relationship between cluster dictionaries]
Basically, it conforms to the relationship described in [2.3] Dictionary configuration.
(Example) Facility → Ski resort → Ontake ski resort [2.4.2 Semantic relationship between words]
[2.4.2.1 Inclusion Relationship]
(Example) Chinese food → Shumai, Ramen, Gyoza ...
(Example) Sports → Tennis, skiing, swimming, jogging, ...
[2.4.2.2 Association relationship]
(1) When associating objects of the same classification (Example) Udon → Noodles + Ramen (2) When associating scene components (Example) Slope → Ski → Lift, Snowboard, Goggles, ...
(Example) Golf-> Golf course-> Hall, Caddy, Fairway, Club ...
(Example) seaside → bathing → swimsuit, beach umbrella, blue sky, white clouds, ...
(3) When reminiscent of the object of interest related to the scene (Example) Ski → slope, snow quality, lift ...
(Example) Golf-> weather, route, cost, score ...
(4) When reminiscent of a typical scene from the season (Example) Summer → Pool, bathing, shaved ice, cicada, cooler, ...
(5) Associations based on requested keywords (Example) Harahata → Restaurant [2.4.3 Relationship between topics]
A recognition vocabulary can be set by activating a topic keyword associated with the current topic. The category of association includes means, accompanying actions, and common incidental events.

(Example) Shopping → Parking lot (means), restaurant (accompanying action), bargain (incidental event) ……
[2.4.4 Continuity of context]
[2.4.4.1 Topic continuity]
As can be seen in normal natural dialogue, it is conceivable that the dialogue continues in a range of recognized vocabulary closed on a certain topic (for example, shopping). The recognition vocabulary can be set under such a restriction of topic continuity.

[2.4.4.2 Validity of utterance-response]
In the dialogue environment represented by the passenger compartment, a certain utterance content (by a user or an agent) can be classified as {call, declaration, notification, teaching, commentary, instruction, request, warning, reminder, inquiry}. be able to. On the other hand, the response to this utterance can be classified into {response, confirmation, hold, judgment, answer, other response}. This combination of utterance and response is called an utterance pair or dialog unit. Based on this dialogue unit, the logical continuity of context can be defined regardless of the topic content. In FIG. 9, a portion indicated by “◯” indicates an utterance-response combination established as a dialogue unit. Based on this dialog unit, the agent can predict and set the recognition vocabulary included in the next user utterance.

Hereinafter, the utterance content and a specific example of a response to the utterance will be described. Note that the user's utterance is indicated by "...", and the agent's utterance is indicated by "...".
(1) Calls This includes general calls and greetings.

(Example)
Call: “Hey, XYZ.”
Answer: “Yes, what is it? ]
(Example)
Call: "Good morning, today is a nice weather. ]
Answer: “Oh, good morning.”
(2) Report (Example)
Declaration: “Today ’s drive with my family.”
Confirmation: “It ’s a drive with your family. ]
(3) Notification (example)
Notice: “It is about 1km ahead and there is traffic. ]
No response: “” or Confirmation: “Okay.”
(Example)
Notice: “My name is XYZ. ]
Confirmation: "OK."
(4) Teaching (Example)
Teaching: “It's raining now.”
Confirmation: “The message“ It is raining now ”was confirmed. ]
(5) Explanation (Example)
Explanation: “If you don't know how to operate, please say help. ]
No reply:""
(6) Instruction (Example)
Instructions: “Speak user password. ]
Confirmation: “Okay. Xxxxxxx”
(7) Request (example)
Request: “Soon, gasoline is running low. Gasoline at the next intersection
Could you refuel at the stand? ]
Hold: “No, let's do it later”
(8) Warning (Example)
Warning: “Accident occurred in XX tunnel 10km ahead. Get off at the next interchange
Please. ]
Judgment: “Okay, let ’s do it.”
(9) Reminder (Example)
Dunning: “User name has not been registered yet. Register immediately. ]
Confirmation: “I understand.”
(10) Inquiry There are the following four types of inquiries.

1) Request for agreement (example): “The telephone connection will be canceled because it will not be issued. Is it OK? ]
2) Selection request (Example) Inquiry: “A? Is it B? ]
Answer: “A.”
3) Inquiry A unit for inquiring specific data such as location, time, and information.

(Example) Inquiry: “What is the snow cover information on the XX ski resort?”
(Example) Inquiry: “Where are you going to go? ]
4) Confirmation of topic Confirm with the user that sudden topic transition out of context has occurred.

(Example)
: “Go to Okazaki at 12:00.” “Nemui.”
Inquiries: I heard "Nemu" but what about Okazaki's story?
Did you do that? ]
[2.4.4.3 Connectivity between dialogue units]
In addition to the above-mentioned responsiveness within the dialog unit, restrictions on the validity of connections between dialog units (including topic transitions, calls, and terminations) can be taken into account in setting the recognition vocabulary.

[2.4.5 User characteristics and status]
Based on the user environment / situation / request / state associated with the user utterance and the user profile, a recognition dictionary that can respond to the next user utterance is set. In this case, the continuity of the above-mentioned context is not always maintained.

(1) Natural request estimation Requests stored in the request / state estimation data storage unit 36 shown in FIG. 2 as well as based on the request estimation device proposed by the applicant of the present application in Japanese Patent Application No. 10-184840, for example. -Refer to the state estimation data and limit the next recognition vocabulary from the user's environment / situation / request / state and user profile.

(2) Sudden changes in driving conditions A recognized vocabulary to pause the continuity of context and interrupt the topics required for situations in the event of discontinuous unpredictable emergencies or warnings Set up a dictionary.

(Example)
Agent: What are you going to do now? ]
User: “Shopping in front of XX station”
Agent: “The arrival time is around 11:00. Where is the parking lot? ]
(Because the preceding car suddenly stopped, the brakes were applied suddenly.)
User: “Oh, it was dangerous.”
Agent: “It was dangerous, please try to drive safely. Exhaustion
Do you take a rest? ]
(3) System function transition When the function is switched, the system speaks and sets the recognition vocabulary to start the necessary dialogue.

[Specific example of 2.5 multi-stage processing]
A specific example of multi-stage processing will be described with reference to the flowcharts of FIGS.
Here, it is first determined whether or not local information that is frequently used is prioritized (S210), and the recognition processing is divided into two. Note that other branching conditions can be considered and depend on the dialogue strategy.

  In the case of local priority (S210: YES), an N-word dictionary is constructed with representative locations, facility names, request keywords, and request-related attributes, word spotting is applied (S220), and the likelihood of the request keywords is evaluated. (S230). Then, the likelihood is corrected by weighting the keyword attributes (S240), and the word strings are ordered (S250). After that, the speech section that has been recognized is removed from the next recognition target (S360). If all speech sections have not been recognized (S370: NO), the vocabulary of the missing attribute is added to prepare for dictionary update. (S380), the process returns to S210.

  In this case, S220 to S250 are processes when local priority is given, and this process is not a “multi-stage process”. According to this local priority processing, all speech segments can be processed with one recognition and the recognition time is short, but the search target is limited to be within 100 words, for example. In the case of an utterance that deviates from the demand, the probability of misrecognition increases.

  On the other hand, when the local priority is not given, that is, when the request is given priority (S210: NO), first, it is determined whether or not the first recognition (S260), and if the first recognition (S260: YES), the recognition is performed. The attribute is set as a request keyword (S270), and the process proceeds to S280. If it is the second or later recognition (S260: NO), the process of S270 is not executed and the process proceeds to S280. In S280, a word set including many recognition attribute keywords is formed, and word spotting is executed using the formed word set (S290).

  Thereafter, it is determined whether the likelihood of the request keyword is a predetermined value (here, 0.6 as an example) or more (S300), and if the request keyword has a likelihood of 0.6 or more (S300: YES), the process proceeds to S310. In S310, when there are a plurality of words having the same attribute, the syntax position is given priority. That is, the request based on the requested keyword is confirmed, and in the subsequent S320, the attribute corresponding to the hit keyword is set in the next recognized vocabulary. Thereafter, the recognized speech segment is removed from the next recognition target (S360). If all speech segments have not been recognized (S370: NO), the vocabulary of the missing attribute is added to prepare for dictionary update. (S380), the process returns to S210.

  On the other hand, when the likelihood of the request keyword is less than 0.6 (S300: NO), the process proceeds to S330, and it is determined whether or not there is a word having a likelihood of 0.6 or more in other attributes. If there is a word having a likelihood of 0.6 or more (S330: YES), the word with the highest likelihood is set as a recognition attribute, and a word set including a large number of keywords of the recognition attribute is formed to perform word spotting. Is executed (S340). However, if there is no word with a likelihood of 0.6 or more (S330: NO), an N-word dictionary is constructed with the location, facility name, and request-related attributes, and word spotting is executed (S350). After the processing of S340 and S350, the speech segment that has been recognized is removed from the next recognition target (S360). As described above, if the recognition of all speech segments has not been completed (S370: NO), the vocabulary of the missing attribute is After adding and preparing to update the dictionary (S380), the process returns to S210.

  On the other hand, if the entire speech section has been recognized (S370: YES), the word string is estimated (S390 in FIG. 11), echoed back (S400), and the utterance for inquiring about the shortage condition is performed (S410). Thereafter, it is determined whether the user's response is a negative utterance (for example, different or not), and if it is not a negative utterance (S420: NO), the vocabulary of the missing attribute is added to update the dictionary. After preparation (S430), the process returns to S210 in FIG. On the other hand, if it is a negative utterance (S420: YES), the misrecognition response process of S440-S470 will be performed.

  Specifically, branching is performed based on the mode setting in S440, and switching to continuous word recognition is performed to instruct the user to speak (S450), or the user is instructed to enter the requested keyword as a help mode (S460), Either a vocabulary that can be recognized as a progressive hierarchical search is displayed (S470). After that, the vocabulary having insufficient attributes is added to prepare for dictionary update (S480), and the process returns to S210 in FIG.

  A case will be described in which such processing is performed on a user's uttered voice, for example, “I want to pick up a good ramen”. First, a request keyword (Kitai, etc.) is preferentially detected from the voice section. For this purpose, the recognition vocabulary also includes a large number of request keywords (enriched with synonyms, synonyms, and associative words for request types and the same request). However, in the case of request priority, recognition is not completed by one word spotting. Therefore, as shown in FIG. 6, after the request keyword is fixed (meal-related), the next facility name and request target are each recognized corresponding to one word spotting, and recognition of all speech sections is completed. To drive word string estimation to narrow down word string recognition results. For details of narrowing down the word string recognition results, refer to, for example, Japanese Patent Application No. 11-20349.

  Based on this, the system utters an echo-back such as “I'll eat an angelo ramen” (S400), and utters an inquiry about insufficient conditions by driving a logical utterance strategy (S410). "Where do you want to see?" On the other hand, if the user utters a completely different utterance “I want to laugh at XX”, the user utters a negative response (such as “all different”) to the above echo back result ( S420: YES). Accordingly, in this case, the erroneous recognition handling process (S440 to S470) is driven.

  On the other hand, when a negative response is not made (S420: NO), the agent interprets that the dialogue with the user is correctly performed, and then the keyword group expected to be spoken by the user (Ikenken), “sushi” ”In the recognized vocabulary dictionary (S430), and waits for the user's utterance.

[3. Misrecognition handling process]
Likelihood judgment and vocabulary limitation of recognition vocabulary output from speech recognition unit, help request from user, simplification of input based on screen display support (combined with switch operation), judgment of insufficient condition, organic service content Helps, option presentation, initialization, and questioning are driven based on relevant associations.

[3.1 Detection of false recognition]
The situation in which the agent cannot recognize the user utterance is misrecognition, and there are the following cases.

1) A user utters a word that is not in any recognition dictionary 2) A user utters a word that is in another dictionary but not in the dictionary currently in use 3) Responds to the utterance content of a speaker other than the user, and the user Mode transitions against the intention of the user Such a situation is detected when the user utters “No”, “I don't know”, or “None at all” in response to the agent. In this case, one of the following dialogue contents is selected according to the user's situation.

(1) Voice menu mode (2) Gives an option Note that, for example, when an utterance such as “No” is repeated K or more times from the user, the state returns to the initial state. For example, K is 5.

[3.1.1 Detection of unexpected response]
Based on the threshold processing of likelihood associated with the recognized word, if the likelihood is low, the speech recognition reliability is low, and it is determined that there is an utterance other than the recognized vocabulary, and there is a possibility of misrecognition. .

[3.2 Detection of topic change]
Whether or not the user utterance deviates from the agent's expectation is determined by whether or not the user utterance is in line with the previous context, that is, whether the utterance is expected in a certain topic. An expected utterance is derived based on the relationship between words shown in “2.4 Network between Words”, and the corresponding vocabulary is set in the recognition dictionary. This is called Case A for convenience. Otherwise, it is an unexpected utterance, and the words constituting it can be classified as follows.

(B1) A word that is not registered in the recognized vocabulary dictionary (B2) A word that is registered in the recognized vocabulary dictionary but has a different topic (B2a) A word that is included in the current recognized vocabulary range (B2b) A current recognized vocabulary Words not included in the range Of these, (B1) and (B2b) cannot be recognized by normal word spotting, so they are considered as non-terminable terms or other recognizable vocabularies that are considered close in signal processing. Replaced and output. These are dealt with by the misidentification process described later.

On the other hand, (A) and (B2a) correspond to the following three processing forms.
(1) Context priority processing [(A)]
When there is not much difference in likelihood between the plurality of output recognition candidates (lattices) (the variance is small), recognition candidates according to the context are preferentially selected.

(2) Sudden topic transition (topic change) [(B2a)]
If there is one recognition candidate that has been output and the likelihood is equal to or greater than a certain threshold, it is recognized that the topic has suddenly changed.

(3) Confirmation of sudden topic transition [in the case of (B2a)]
If the output recognition candidate is one but has a likelihood not exceeding a certain threshold value, the user is inquired to confirm whether or not the topic has suddenly changed.

[3.2.1 Detection of topic change]
If a request keyword different from the current topic is recognized, it is considered that the topic may have changed.

[3.2.2 Confirmation of topic change]
Based on the above, the system generates a query utterance that confirms to the user that the topic has changed.

(Example)
Agent “□□ There are three ramen shops in front of the station. ]
User "Call XXX in Tokyo"
Agent "Do you want to call me? ]
User “Yes. To Mr. XX in Tokyo”
Agent "Calls Mr. XX in Tokyo"
[3.2.3 Utterance change utterance]
Generate an utterance declaring that the topic has changed (see example above).

[3.3 Context-first utterance]
Assuming that the topic T (n) (n: serial number of the utterance pair) continues, the next agent utterance is also generated based on the topic. Therefore, even if the interpretation result U (n) of the user utterance this time is a vocabulary that is not related to the topic T (n), the processing is limited to the utterance content according to the context from the topic T (n) without immediately responding to it. Add

(Example)
Agent “There are three ramen shops in front of the station. ]
・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ {T (n) = meal}
User “Oh, I have to make a phone call.”
・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ {"Phone" is recognized
Then Tnew = phone but don't update topic}
Agent “Which ramen shop would you like to have? ]
・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ {T (n + 1) = T (n) = meal}
User “Uh, □□”
Displays the route to the agent “□□”. ]
The topic change described above and the context priority explained here are contradictory responses. For example, the condition judgment is U (n) likelihood information LFK or the number of consecutive occurrences Ntnew of the topic Tnew indicated by U (n). Etc. are used. That is, if Ntnew> 2 and LFK> 0.4, the topic shifts from T (n-1) to T (n) = Tnew, otherwise Tnew obtained from U (n) is rejected, A conditional branch such as T (n) = T (n-1) is used.

[3.4 Help mode]
It displays choices such as examples of misrecognition that tend to occur, typical request keywords, and utters speech.

(Example)
・ To make a call, use "Call phone" to display the phone number screen and enter the phone number of the other party. In the case of a registered partner (for example, Mr. XX), “Call phone, XX” is fine.

・ Which of the following? Please speak again. {Meals (restaurant, rice, hungry), schedule book, address book (address book, phone book), map (route guidance, drive assistant)}
・ Please speak again with a synonym in parentheses.

・ In the case of a map display, it is better to select a road map from the route guidance.
・ The destination is more accurate when speaking with a municipality. (Example: Kariya → Kariyashi)
[3.5 Presentation of options]
Choice presentation corresponds to the progressive hierarchical search already described.

[3.6 Normal speech strategy]
If no misrecognition is detected based on the recognition result (recognized word, likelihood), a normal speech strategy is applied. Since the contents of this normal speech strategy are not the main points of the present invention, details are not mentioned here.

[3.7 Adaptation of dialog management]
[3.7.1 Adaptation to the environment and situation]
Time (season, date and time, time [morning, day and night]), space (vehicle position, area [prefecture, municipality]), environment (road environment [highway, general road, tunnel, etc.], road condition [road surface freezing, slippery , Etc.], traffic environment [expressway, speed limit, etc.], geographical environment [close to the sea, in the mountains, in the city, in front of the station, etc.]), outside conditions (weather, traffic conditions [such as traffic congestion], It is possible to perform topic selection, dialogue management, transmission of appropriate messages, etc., depending on the situation of the overtaking car, etc.)) and in-vehicle conditions (driving state, riding state, purpose of movement, topic). These are also reflected in the display system.

[3.7.2 Adaptation to schedule]
(1) Drive schedule Create a drive schedule based on the destinations and waypoints set by the user, determine topics and manage conversations based on the meaning of each event (meal, shopping, sightseeing, etc.) and location / time. You can even make suggestions for blank parts of your drive schedule.

(2) Personal schedule It is possible to download personal schedule data on a PDA or PC via an interface for the PDA or PC, and based on that, determine topics, manage dialogues, and further propose blank portions of the personal schedule.

[3.7.3 When the user cannot understand the meaning of the response]
When the user utters “What do you mean” or “I don't know well”, the user judges that he / she does not understand the meaning of the agent response, and performs one of the following processes.

(1) Help function (2) Menu selection (3) Voice menu [3.7.4 When agent cannot respond to user request]
Even if the user utterance can be recognized normally, the agent may not be able to respond to the user's request in the following cases. About these, the situation is conveyed by a voice message.

1) No search results (no corresponding database)
2) There is no corresponding device, it is out of order, or it is not ready (the telephone is not connected, etc.)
3) When a control command exceeding the operating range of the device to be operated is issued [3.8 Voice Menu Mode]
For example, the present applicant shifts to the device operation mode as proposed in Japanese Patent Application No. 10-177667. Commands required in each operation mode are always preferentially added to the recognition vocabulary dictionary to increase the likelihood of recognition.

[3.9 Questions]
There are a number of cases where the question is answered as follows. When the question is asked, a recognition dictionary limited to the vocabulary according to the content is set.

    (1) If the agent correctly recognizes the topic but the user utterance is not made in accordance with the topic, the subsequent misrecognition is avoided by issuing a query for confirming the topic.

(Example)
Agent “□□ There are three ramen shops in front of the station. ]
User "Call XXX in Tokyo"
Agent "Do you want to call me? ]
(2) When an agent recognizes a topic by mistake and inquires about a deficiency condition under a different topic, the correct topic is obtained by a question from the user or a negative remark (such as “No” or “No”). Asks what is. This stops subsequent misrecognition.

(Example)
User “Calls Mr. XX in Tokyo”
Agent “□□ Displays restaurants near the station. Which would you like? ]
User "I don't care, make a call"
Agent: Is it a phone call or a meal? ]
User “Denwa”
This is a case where a restaurant search request has already been issued, and the user makes an interruption request to make a call during the search. This is an example of asking the priorities of the two requirements (phone or meal).

    (3) After the agent correctly recognizes the topic and confirms that the necessary condition in context is insufficient, the agent asks the question. This develops the right dialogue in context and promotes problem solving.

(Example 1)
User “Calls Mr. XX in Tokyo”
Agent “You are calling XXX in Tokyo. That is ○○ ◇◇
Or is it Mr. ○○ △△? ]
User “○○ ◇◇”
Agent “Call XX ◇◇”
(Example 2)
User "Call XXX in Tokyo"
Agent “□□ Displays restaurants near the station. Which would you like? "
User “Later. Call first.”
I'll call the agent "Tokyo, Japan."It's all right. "
User “Yes, then”
The case described here (example 1) is an example in which an agent has inquired about an insufficient condition for solving the problem of correctly recognizing and making a call.

  Also, in the case of (Example 2), if the user's intention to make an electric bill is known from the progress and schedule of the conversation so far, it is necessary for the telephone requirement (meal with Mr. XX) It becomes possible to search for information in advance. In other words, it can be said that the restaurant search was voluntarily performed and the questions were asked back.

[3.10 Initialization]
When the number of questions from the user exceeds K times (K> N), the screen returns to the initial state or the top menu screen.

  As described above, according to the control system of the present embodiment, the progressive hierarchical search, the multi-stage process, the unexpected response process, and the misrecognition process are performed.

First, in the case of progressive hierarchical search, taking the utterance of “I want to eat in Okazaki, I like Indian food” as an example, the output result of the speech recognition method by word spotting is “Okazaki, Meals, India. In the case of the word sequence “cooking”, the conventional speech recognition method discloses the processing on the system side corresponding to the recognition at the stage when these three are gathered, whereas according to this method, “Okazaki” Then, the utterance vocabulary expected to be input next can be immediately presented. Therefore, the user can speak without any trouble. As a result, it is possible to prevent the user from uttering a vocabulary outside the recognition dictionary that causes misrecognition.

  Also, in the case of multi-stage processing, a recognition dictionary is dynamically and smallly constructed using semantic constraints between words constituting the word string, with the request keyword constituting the word string as an axis. The appropriate recognition is performed. As a result, erroneous recognition unintended by the user can be prevented.

  Also, in the unexpected response process, if the utterance from the user during the conversation is not the expected content according to the context, a question to confirm whether the topic has been changed is checked. One of processing, topic conversion declaration processing for declaring that the topic has been changed, and context priority processing for performing correspondence in accordance with the context on the assumption that the previous topic has continued. On the other hand, in the misrecognition handling process, if the utterance from the user during the conversation has a predetermined negative content, it responds to the question-reply processing to confirm what the correct topic is, and the requested content A help mode process for presenting an instruction for including the keyword so that the user can recognize visually or auditorily, an initialization process for returning to the initial state, and the like are executed. By doing so, it is possible to prevent the inconvenience that the user is at a loss due to misrecognition.

Furthermore, in the case of the present embodiment, these progressive hierarchical search, multi-stage processing, unexpected time response processing, and misrecognition response processing are combined, which is more effective.
By the way, in this embodiment, for example, as a system that is mounted on a car and controls various devices mounted on the vehicle while talking with a vehicle occupant (mainly a driver) as a user. As described above, as shown in the conceptual diagram of FIG. 1, the same processing can be performed for word string candidates as a result of recognition by other signal system recognition units. For example, it may be a character recognition device that compares a handwritten character string input by a person to be recognized with dictionary data and outputs a plurality of word string candidates having a high degree of coincidence. Since handwritten characters are also likely to be erroneously recognized, it is highly effective to estimate an appropriate word string. Moreover, it is not limited to the content of the word string directly at the time of input to the recognition device, such as voice recognition or character recognition, but may be an image recognition device. That is, any recognition device that recognizes an image that captures a recognition target as a scene and converts the scene into a natural language can be realized. Specifically, for example, a sign language pattern is recognized from an image in which the person to be recognized is sign language, and a word string candidate indicating a natural language meaning represented by the sign language pattern is output. As for the sign language pattern, because the words expressed by subtle fingering are different, misrecognition still occurs due to individual differences among persons who sign language. Therefore, in this case as well, in the sign language pattern recognition apparatus that is likely to include misrecognition, it is possible to prevent misrecognition by performing the above-described various processes, or to perform appropriate processing after misrecognition occurs. It is possible to prevent the inconvenience that the user is at a loss due to misrecognition.

As mentioned above, although one Embodiment of this invention was described, it cannot be overemphasized that this invention can take a various form, without being limited to the said embodiment.
For example, in the flowchart showing the process outline of FIG. 3, the multi-stage process (S30) is performed first, and the progressive hierarchical search (corresponding to S130) is performed in the subsequent misrecognition handling process. The order is not limited.

  In the case of the above embodiment, the progressive hierarchical search, multi-stage processing, unexpected time response processing, and misrecognition response processing are combined. However, these are effective even if they are executed alone. However, as described above, it is more effective if implemented in combination. In that case, it is not necessary to combine all four processes, and it is sufficient to combine two or more processes.

It is the block diagram which showed notionally the function about the word string recognition apparatus of embodiment, and was shown notionally. It is a block diagram which shows the structure at the time of applying a word sequence recognition apparatus to a vehicle-mounted control system. It is a flowchart which shows the outline | summary of the process performed by a system control part. It is explanatory drawing which shows the specific example of a screen transition of a progressive hierarchy search. It is explanatory drawing which shows the specific example of a screen transition of a progressive hierarchy search. It is explanatory drawing which shows the outline | summary of a multistep process. It is explanatory drawing which shows the outline | summary of the dynamic structure of a dictionary. It is explanatory drawing which shows the structure of a dialogue database table | surface. It is explanatory drawing which shows the combination of the utterance-response which is materialized as a dialogue unit. It is a flowchart which shows the first half of a multistage process. It is a flowchart which shows the second half of multistage processing.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Control apparatus, 3 ... Input device, 5 ... Microphone, 7 ... Speaker, 8 ... Display, 9 ... Navigation apparatus, 11 ... Display apparatus, 13 ... Air-conditioner apparatus, 15 ... Audio apparatus, 17 ... Communication apparatus, 21 ... System Control unit, 25 ... Voice input unit, 27 ... Speech synthesis unit, 28 ... Screen control unit, 29 ... Device control I / F, 31 ... Internet address database, 33 ... Search control unit, 34 ... Recognition vocabulary storage unit, 35 ... Dialog data storage unit, 36... Request / state estimation data storage unit, 37... User profile storage unit.

Claims (5)

A word string output means for inputting information reflecting the action content of a human being recognition object and outputting a word string candidate having a high degree of matching compared to the dictionary data for recognition;
A request indicating what the request in the syntax is whether the word string output from the word string output means has the expected content in accordance with the context, that is, whether it is an expected utterance in a certain topic. Judgment based on keywords, and if it is determined that the request keyword of a topic different from the current topic has been recognized as being disappointing, ask at least whether the topic has changed Unexpected to perform either topic conversion confirmation processing, topic conversion declaration processing that declares that the topic has changed, or context-first response processing that responds according to the context assuming that the previous topic has continued With time handling means,
The unexpected time response means executes the topic change declaration process if the word string output from the word string output means after execution of the topic change confirmation process is a content in line with the changed topic. A word string recognition device characterized by the above. The word string recognition device according to claim 1,
Even when the topic is changed, the unexpected time response means executes the context-first response processing immediately after the conversion, and only when the converted topic continues after that. A word string recognition apparatus characterized by executing the topic change confirmation process . A word string output means for inputting information reflecting the action content of a human being recognition object and outputting a word string candidate having a high degree of matching compared to the dictionary data for recognition;
A request indicating what the request in the syntax is whether the word string output from the word string output means has the expected content in accordance with the context, that is, whether it is an expected utterance in a certain topic. Judgment based on keywords, and if it is determined that the request keyword of a topic different from the current topic has been recognized as being disappointing, ask at least whether the topic has changed Unexpected to perform either topic conversion confirmation processing, topic conversion declaration processing that declares that the topic has changed, or context-first response processing that responds according to the context assuming that the previous topic has continued With time handling means,
Even when the topic is changed, the unexpected time response means executes the context-first response processing immediately after the conversion, and only when the converted topic continues after that. A word string recognition apparatus characterized by executing the topic change confirmation process. In the word string recognition device according to any one of claims 1 to 3,
The word string output device outputs a word string candidate having a high degree of coincidence obtained by comparing speech inputted by a person who is the recognition target with recognition dictionary data. In the word string recognition device according to any one of claims 1 to 4,
The word string output means compares a handwritten character string input by a person who is the recognition target with recognition dictionary data, and outputs a plurality of word string candidates having a high degree of coincidence. .

Images