JP7259349B2 - Dialogue device, dialogue method, and program - Google Patents

Description

The present invention relates to an interactive device, an interactive method, and a program for outputting a response sentence according to an input from a user.

2. Description of the Related Art Conventionally, as an interactive device for outputting a response sentence in response to an input from a user, for example, one disclosed in Japanese Patent Application Laid-Open No. 2002-200012 is known. This conventional dialogue system is applied to a vehicle navigation system and has dialogue scenario data composed of a finite automaton. A conventional dialogue device executes dialogue with a user according to dialogue scenario data in response to input from the user at that time.

JP-A-2003-329477

A dialogue apparatus having this kind of finite automaton creates a response sentence for a user based on a transition destination response rule obtained using dialogue scenario data. On the other hand, in casual conversations, the topic tends to change in various ways, so depending on the configuration of the dialogue scenario data, there is a possibility that the transition destination response rule alone may not be sufficient.

Such a problem can be avoided by configuring the above dialogue scenario data so that a wide range of topics can be acquired. It can become natural.

SUMMARY OF THE INVENTION Accordingly, an object of the present invention is to enable a natural dialogue with a user .

An example of a dialogue device includes an acquisition means for acquiring input utterance information input from a predetermined target, a set of a plurality of input words, a response sentence, an automaton state, and an automaton state to be transitioned to next. rule data according to a predetermined state including the current state of the automaton and a plurality of input words included in the input utterance information obtained by the obtaining means from a database storing a plurality of rule data associated with each other; and a response sentence output means for outputting a response sentence included in the selected rule data to the predetermined object; and a memory for sequentially storing the predetermined state corresponding to the selected rule data in a storage means. and a control means , when the input utterance information is acquired by the acquisition means after the storage control by the storage control means, the response sentence output means, if there is a plurality of rule data to be selected, the database among the plurality of rule data stored in the storage means , rule data corresponding to the same state as the predetermined state stored more recently than the plurality of predetermined states stored in the storage means are preferentially selected. .

According to the present invention , it becomes possible to interact with the user naturally.

1 is a block diagram illustrating one embodiment of an interactive device; FIG. FIG. 2 is a diagram showing an example of hardware of a computer that can implement an interactive device; FIG. 4 is a diagram showing examples of data formats of control data, input word data, and rule data; 4 is a main flowchart showing an example of interactive processing; 6 is a flowchart showing a detailed example of preprocessing; 9 is a flowchart showing a detailed example of rule search processing; 9 is a flowchart showing a detailed example of response sentence output processing; FIG. 1 is a diagram (part 1) showing an embodiment of a dialogue database based on an automaton; FIG. 2 is a diagram (part 2) showing an embodiment of a dialogue database based on an automaton; FIG. 3 is a diagram (part 3) showing an embodiment of a dialogue database based on an automaton; It is a figure which shows the operation example of a dialog apparatus.

EMBODIMENT OF THE INVENTION Hereinafter, it demonstrates in detail, referring drawings for the form for implementing this invention. FIG. 1 is a block diagram illustrating one embodiment of an interactive device. The dialogue device 100 includes a response sentence output unit 103 (response sentence output means) including a database 101 (database), an acquisition unit 106 (acquisition means), an extraction unit 107 (extraction means), and a data acquisition unit 102 (data acquisition means), and a storage unit 104 (storage means), and is configured to be able to interact with a predetermined target, for example, a user. This interactive device 100 can be used, for example, by incorporating it into a speaker with an interactive function for home use or an interactive function of a robot. In the case of a robot, the predetermined target may be, for example, another robot.

The database 101 stores a plurality of rule data 110. Each rule data 110 contains an assumed input word set (assumed input utterance sentence), a response sentence, an automaton state number, and the next transition destination of the automaton. state number (hereinafter referred to as "next transition destination state number"). This assumed input word set is composed of a plurality of input words assumed to be input by the user, and the response sentence is a response sentence to the user's utterance. The number and the next transition destination state number of the automaton are associated with each other and stored as the rule data 110 . The state number of the automaton and the next transition destination state number define which of the plurality of rule data 110 should be selected and responded during the dialogue with the user. It represents (defines) the state number of the automaton included in the rule data 110 to be selected when responding to . In this case, the automaton state number and the next transition destination state number may be the same number or different numbers. The automaton state number and the next transition destination state number of each rule data 110, and the response sentence corresponding to each of them are set so that a natural response can be made while appropriately changing the topic in the dialogue with the user. It is Here, for example, a plurality of rule data 110 including the same automaton state number may be stored in the database 101 so as to correspond to user utterances on various topics.

Acquisition unit 106 acquires user input speech information 115 via, for example, a microphone (not shown).

The extraction unit 107 digitizes the input speech information 115 through an amplifier, an A/D (analog/digital) converter, and the like, and converts it into digital voice. Next, the extraction unit 107 acquires text data of the input sentence by executing speech recognition processing on the digital voice, and executes morphological analysis on the acquired text data of the input sentence to obtain , text data of an input word set 111 consisting of a plurality of input words is extracted in a form in which the utterance is divided into words (for example, nouns, verbs, adjectives, adverbs, etc.).

The response sentence output unit 103 selects the response rule data 113 from the plurality of rule data 110 according to a predetermined state including the current state of the automaton and the input utterance information 115 acquired by the acquisition unit 106. , the response sentence 114 included in the selected response rule data 113 is output to a predetermined target.

Here, the response sentence output unit 103 may have the data acquisition unit 102 as follows. The data acquisition unit 102 acquires input utterance information 115 from a plurality of rule data less than a plurality of rule data corresponding to a predetermined state including the current state of the automaton, among the plurality of rule data 110 . Then, the candidate response rule data 112 , which are a plurality of rule data of response candidates that are candidates for the response rule data 113 , are retrieved and acquired. This search is performed, for example, as follows. First, the rule data 110 having the assumed input word set included in the acquired input word set 111 is searched as provisional candidate data of the response candidate rule data 112 . In this case, the "assumed input word set included in the input word set 111" is such that the number of words in the assumed input word set is less than or equal to the number of words in the input word set, and all the words in the assumed input word set are input. An assumed input word that matches some or all of the words in the word set. The data acquisition unit 102 searches the response candidate rule data 112 by referring to the storage unit 104 as will be described later, for one or more pieces of provisional candidate data thus retrieved.

The response sentence output unit 103 selects the response rule data 113 from the response candidate rule data 112 searched by the data acquisition unit 102, and extracts the response sentence included in the response rule data 113 (the response included in the rule data 110 in FIG. 1). sentence) is output as a response sentence 114 to the user's utterance. In response to the data (text data, for example) of the response sentence 114 output in this way, the voice corresponding to the response sentence 114 is uttered from the speech synthesis processing unit via the D/A converter, the amplifier, and the speaker. be. If the dialogue device 100 is incorporated as a dialogue function of a robot, for example, the robot causes the dialogue device 100 to utter the above-described voice while detecting the presence of the user in the surroundings with a sensor or the like. Along with this, the response sentence output unit 103 sequentially stores the automaton next transition destination state number (see FIG. 1) included in the response rule data 113 in the storage unit 104 as the stack state number NSA. Thus, the storage unit 104 functions as a stack that stacks state numbers of automatons. The plurality of stack state numbers NSA stored in the storage unit 104 indicate the history of the state numbers of the automaton that transitioned according to the topic of the user's utterance. The most recently stored stack state number NSA indicates the current state number (=topic, for example) of the transitioned automaton.

Here, the data acquisition unit 102 stores the state number (see FIG. 1) of the corresponding automaton among the one or more provisional candidate data (rule data 110) retrieved as described above in the storage unit 104. Temporary candidate data matching any of the plurality of stack state numbers NSA is retrieved as response candidate rule data 112 . Each stack state number NSA in the storage unit 104 indicates the state number (=for example, topic) of the automaton transitioned according to the topic of the user's utterance. Therefore, the data acquisition unit 102 selects the rule data 110 that includes the same state number as the state number of the automaton that has occurred so far (stack state number NSA) among the rule data 110 stored in the database 101. Answer candidate rule data 112 is selected from the rule data 110 that has become a hot topic.

Most preferably, the rule data 110 containing the most recently stored stack state number NSA in storage 104, ie, the same state number as the automaton's current state number, is retrieved. However, the state number of the automaton included in the rule data 110 corresponding to the word set 111 input by the user does not necessarily match the current state number of the automaton. In such a case, the data acquisition unit 102 includes the state number of the automaton that matches the state number of the automaton traced in the past indicated by the stack state number NSA stored in the storage unit 104 and corresponds to the input word set 111. The rule data 110 having an assumed input word set is searched as the answer candidate rule data 112 . With this configuration of the dialogue apparatus 100, it is possible to search for the answer candidate rule data 112 along the flow of dialogue with the user, for example.

In this case, the data acquisition unit 102 further collects a plurality of response candidate rule data 112 (rule data 110) whose corresponding automaton state number matches any of the plurality of stack state numbers NSA in the storage unit 104 as described above. exists, of these candidate response rule data 112, the candidate response rule data 112 corresponding to the state number matching the stack state number NSA stored more recently is preferentially selected as the response rule data 113. The score of each response candidate rule data 112 is calculated as follows.

More specifically, the data acquisition unit 102 sets a weighting factor for each of the input words included in the input word set 111, for example, using a predetermined method such as the TF-IDF method described later. As a result, the weight coefficients of the plurality of input words are set to the same value or different values depending on their importance. The data acquisition unit 102 also indicates the degree of similarity of the assumed input word set (see the assumed input word set of the rule data 110 in FIG. 1) included in the rule data 110 as the response candidate rule data 112 to the input word set 111. A cosine similarity is calculated for each response candidate rule data 112 according to the set weighting factor. In addition, the data acquisition unit 102 calculates a score indicating an index for selecting the response rule data 113 for each response candidate rule data 112 according to the calculated cosine similarity (the cosine similarity is calculated as a score). ). Further, the data acquisition unit 102 obtains the stack state number NSA stored in the past among the plurality of stack state numbers NSA stored in the storage unit 104, among the plurality of candidate response rule data 112 (rule data 110). , the score of the answer candidate rule data 112 corresponding to the state number that matches is calculated to be a smaller value. Then, the response sentence output unit 103 selects the response candidate rule data 112 having the maximum score as the response rule data 113 based on the score calculated for each response candidate rule data 112 . With this configuration of the dialogue apparatus 100, the response sentence output unit 103 selects one or more of the candidate response rule data 112 retrieved by the data acquisition unit 102, which is closest to the current state number of the automaton (=for example, the current topic). The dialogue device 100 can preferentially select response candidate rule data 112 corresponding to a state number (=for example, a closer topic) as the response rule data 113, and can have a more natural dialogue with the user according to the topic. can be provided.

In addition to the above configuration, a response sentence storage unit 104 (response sentence storage means) may be further provided for storing a predetermined number of past response sentences 114 generated based on the selected response rule data 113 . Then, by referring to the response sentence storage unit 104, the data acquisition unit 102 determines the priority ( For example, the score described above) may be changed according to how many times in the past the response sentence 114 of the response candidate rule data 112 was generated. With this configuration of the interactive device 100, it is possible to prevent repeated output of the response sentence 114 based on the same response rule data 113. FIG.

In addition to the configuration described so far, the database 101 may include a plurality of predetermined non-assumed rules (wild card list and super wild card list, which will be described later). Then, when the content of the user's utterance is unexpected content, that is, the plurality of words of the input word set 111 do not match the plurality of words of the assumed input word set of any of the rule data 110 . If not included, among the plurality of non-assuming rules, the one corresponding to the previously selected response rule data may be acquired as the response candidate rule data.

FIG. 2 is a diagram showing an example of computer hardware that can implement the dialog apparatus 100 of FIG. This computer includes smartphones, tablet terminals, digital cameras, etc., in addition to ordinary personal computers. The computer shown in FIG. 2 includes a CPU (Central Processing Unit) CPU 201, a memory 202, an input device 203, an output device 204, an auxiliary information storage device 205, a media drive device 206 into which a portable recording medium 211 is inserted, and a network connection device. 207 , an audio input device 208 and an audio output device 209 . These components are interconnected by bus 210 . The configuration shown in FIG. 2 is an example of a computer that can implement the interactive device 100 of FIG. 2, and such a computer is not limited to this configuration.

The memory 202 is, for example, a Read Only Memory (ROM), a Random Access Memory (RAM), or a semiconductor memory such as a flash memory. Various data corresponding to FIG. 3 are stored.

A CPU (processor) 101 uses a memory 202, for example, to execute a program corresponding to processing of flow charts of FIGS. It operates as each processing block.

The input device 203 is, for example, a touch panel input device, and is used to input instructions or information from an operator or user. The output device 204 is, for example, a display device such as a liquid crystal display (LCD) integrally formed with the touch panel input device, and is used to output an inquiry to an operator or user or a processing result.

The auxiliary information storage device 205 is, for example, a semiconductor storage device, a hard disk storage device, a magnetic disk storage device, an optical disk device, a magneto-optical disk device, or the like. It operates as the storage unit 104 or the response sentence storage unit 105 in FIG. 2 stores, in the auxiliary information storage device 205, data such as programs for executing the processes of the flowcharts of FIGS. 4 to 7 and various data described later in FIG. stored, and loaded into the memory 202 for use.

The media drive device 206 drives the portable recording medium 211 and accesses the recorded contents. The portable recording medium 211 is a memory device, flexible disk, optical disk, magneto-optical disk, or the like. The portable recording medium 211 may be Compact Disk Read Only Memory (CD-ROM), Digital Versatile Disk (DVD), Universal Serial Bus (USB) memory, or the like. An operator or user can store the above-described programs and data in the portable recording medium 211 and load them into the memory 202 for use.

Thus, the computer-readable recording medium storing the above programs and data is a physical (non-temporary) recording medium such as the memory 202, the auxiliary information storage device 205, or the portable recording medium 211. is.

A network connection device 207 is a communication interface that is connected to a communication network such as a Local Area Network (LAN) and performs data conversion associated with communication. The interactive device 100 of FIG. 2 can receive the above-described programs or data from an external device via the network connection device 207, load them into the memory 202, and use them.

The voice input device 208 includes a microphone/amplifier for inputting the user's voice as an analog input voice signal, an A/D (analog/digital) converter for converting the analog input voice signal to a digital input voice signal, and a digital input voice signal. It is an interface circuit or the like for handing over to the CPU 201 as an input from the user.

The speech output device 209 is a speech synthesis processing device that synthesizes a digital speech signal corresponding to the response sentence 114 generated by the dialogue device 100 in FIG. converters, amplifiers/speakers for emitting analog audio signals to the user.

2 does not need to include all the components shown in FIG. 2, and it is possible to omit some of the components depending on the application or conditions. For example, part or all of the input device 203 and part or all of the output device 204 may be omitted if there is no need to input instructions or information from the operator or user. If the portable recording medium 211 or communication network is not used, the media drive device 206 or the network connection device 207 may be omitted.

FIG. 3 is a diagram showing an example of the format of various main data necessary for controlling the interactive device 100 of FIG. 2 realized by the computer of FIG.

FIG. 3(a) is an example of a data format of control data. The following various data are stored in the memory 202 in FIG. 2 in order from the address of the start pointer Ctrl of the control data.

transition: a pointer to the transition database. It is a pointer to the beginning of the rule data 110 in the database 101 in FIG. 2 stored in the auxiliary information storage device 205 in FIG. As described above with reference to FIG. 1, the rule data 110 includes, for each state number of the automaton, a set of words assumed to be uttered by the user (assumed input word set), a corresponding response sentence, and then a transition It manages the state number of the automaton (next transition destination state number).

inputWordCount: Input word count. It is the number of words (morphemes) in the input user utterance.

inputWords[inputWordCount]: A pointer to the input word data in FIG. 3(b). This is the entity address of the word (morpheme) group included in the input user utterance. This is an input word data array for the number of input words inputWordCount.

transCandidates: Answer candidate rule data list. This is a list for storing one or more response candidate rule data 112 (see FIG. 1) that match the input condition. Elements of each list include a pointer to each rule data transition[i] (see FIG. 3(c)), a value of the number of matching words of each candidate response rule data 112, and a score of each candidate response rule data 112. Also includes the value of

state_id: stack array of state numbers. It is a stack array for managing stack state numbers NSA (see FIG. 1) stored in the storage unit 104 .

score_coef: evaluation coefficient. It is the input side denominator component when calculating the cosine distance for evaluating each response candidate rule data 112 .

FIG. 3(b) is an example data format of input word data showing the input word set 111 of FIG. 1 detected based on the user's utterance. The head pointer of each input word data is indicated by inputWords[i] (i=0, 1, 2, . . . ). stored in

word: Input word. This is a morpheme-based text data string set by morpheme analysis processing.

weight: Weighting factor. A coefficient for indicating the importance of the word within the rule. The value of the weighting factor is, for example, assigned a large value according to the part of speech. For example, a large value is assigned to nouns and verbs, and a small value is assigned to particles.

prev: previous pointer. A pointer to the input word pronounced just before the current input word in the user's utterance.

next: next pointer. A pointer to the input word pronounced immediately after the current input word in the user's utterance.

FIG. 3(c) is an example data format of the rule data 110 stored in the database 101 of FIG. The head pointer of each rule data 110 is indicated by transition[i] (i=0, 1, 2, . It is stored in the database 101 of FIG. 1 in the information storage device 205 (or in the memory 202).

userWordCount: Expected number of input words. This is the number of assumed input words given as input to the rule data 110 .
userWords [userWordCount]: An array of text data of assumed input words given as input to the rule data 110 . It corresponds to the assumed input word set in the rule data 110 of FIG.
state_id: current state number of the automaton. This number indicates the state to which the rule data 110 belongs. It corresponds to the state number in the rule data 110 of FIG.
bot_reply: reply text. It is the text data of the response sentence that is the output of the rule data 110 . It corresponds to the response sentence in the rule data 110 of FIG.
next_state_id: Next transition destination state number. This is a number indicating the transition destination state of the automaton that transitions after the rule data 110 is selected. It corresponds to the next transition destination state number in the rule data 110 of FIG.

prev: previous pointer. A pointer to the rule data 110 connected immediately before the current rule data 110 .

next: next pointer. A pointer to the rule data 110 connected immediately after the current rule data 110 .

4 to 7 are flowcharts showing examples of interactive processing executed by the computer of FIG. 2 to implement the operation of the interactive device 100 of FIG. This interactive processing is processing in which the CPU 201 of FIG. 2 executes an interactive processing program stored in the memory 202 while using various data described in FIG. .

FIG. 4 is a main flowchart showing an example of interactive processing. When a power switch (not shown) is turned on to start the system, the CPU 201 initializes various variables in the memory 202 and loads necessary data from the auxiliary information storage device 205 to the memory 202 (step S401). ).

Next, the CPU 201 pushes the state number 0 representing the initial state to the stack array state_id of the automaton state numbers on the control data in FIG. 3A (step S402).

After step S402, the CPU 201 waits for input of a user's speech (step S403) and an instruction to end the interactive device (step S404) (repetitive processing of steps S403 and S404).

When the user speaks, in the voice input device 208 of FIG. 2, the voice spoken by the user is input as an analog input voice signal through a microphone/amplifier, and the analog input voice signal is converted into a digital input voice signal in the A/D converter. It is converted and sent to CPU 201 via bus 210 in FIG. Then, when the CPU 201 detects the power of the digital input audio signal equal to or higher than the predetermined threshold, the input of the user's utterance is detected, and the determination in step S404 becomes YES. The CPU 201 sequentially executes preprocessing (step S405), rule search processing (step S406), and response sentence output processing (step S407), and then returns to standby processing in steps S403 and S404.

When the user turns off a power switch (not shown) to issue an instruction to terminate the interactive apparatus 100, the determination in step S403 becomes YES, and the CPU 201 discards the secured area on the memory 202, and the state shown in FIG. Terminates the indicated interactive processing and terminates the system.

FIG. 5 is a flow chart showing a detailed example of preprocessing in step S405 of FIG. In this preprocessing, processing for creating input word data inputWords[0], inputWords[1], . . . in FIG. 3B is mainly executed.

First, the CPU 201 initializes each data (variable) in FIGS. 3A and 3B on the memory 202 (step S501).

Next, the CPU 201 first performs speech recognition on the digital voice data based on the user's utterance input in step S404 in FIG. By executing the analysis, an input word group consisting of a plurality of words divided into morphemes is extracted (step S502). This input word group corresponds to the input word set 111 in FIG.

Next, the CPU 201 sets the initial value 1 to the score coefficient, which is a variable on the memory 202 (step S503). The score coefficient will be described later.

Next, the CPU 201 sequentially extracts morphemes from the first morpheme extracted by the morphological analysis in step S502 (step S504) until it is determined in step S505 that all morphemes have been searched (step S505). While sequentially searching, a series of processes from steps S506 to S508 below are repeatedly executed for each morpheme.

First, the CPU 201 increments the number of input words: inputWordCount of the control data in FIG. 3(a). Further, the CPU 201 generates a new entry (storage area) (for example, inputWords[i]) for the input word data shown in FIG. The text of the morpheme acquired in S509 is registered (above, step S506). The CPU 201 sets the previous pointer (prev) of the new entry of the input word data to the top address of the entry generated immediately before it, and sets the value of the next pointer (next) of the immediately preceding entry to the new entry. By setting the starting address of the input word data, a list of entries of the input word data connected in order according to the user's utterance is generated.

Next, the CPU 201 sets a weighting factor for the input word corresponding to the morpheme acquired in step S504 or S509, and stores the weighting factor in the input word data of FIG. The weighting factor of the new entry: weight is set (step S507). It can be said that the above-mentioned weighting factor indicates the importance of the input word of the corresponding morpheme in the input document. The degree of importance of a word in a document is determined by a well-known method called TF (Term Frequency), in which a word that appears more frequently in an input document is set as a higher (important) value, and by cross-cutting several documents. The words that are used can be set by the TF-IDF method in combination with another well-known method called IDF (Inverse Document Frequency), which is not set to a very large (important) value. Therefore, the weighting factors mentioned above may be set by such a TF-IDF approach. In addition, the degree of importance of a word within a document varies depending on the part of speech of the morpheme. For example, nouns and verbs are assigned large weighting coefficient values, while particles are assigned small values. Therefore, the weighting factors described above may be set from, for example, a weighting factor table for each part of speech or an IDF (Inverse Document Frequency) table held in the auxiliary information storage device 205 in FIG.

Then, the CPU 201 adds the square value of the weighting factor acquired in step S507 to the score factor in the memory 202 initialized in step S503 (step S508). Score coefficients will be described later.

The above series of processes from steps S506 to S508 are executed for all morphemes obtained by the input through repeated processes from steps S505 to S509, thereby obtaining the input word data inputWords[0], inputWords[1], . . . are created. As shown in FIG. 3B, the input word data inputWords[i] (i=0, 1, . It consists of the weight coefficient weight obtained in step S507 corresponding to the word, and pointers prev and next to the previous and subsequent entries.

When the processing for all morphemes is completed and the determination in step S505 becomes YES, the CPU 201 calculates the square root of the score coefficient finally calculated in step S508, which is a variable in the memory 202, and further calculates its reciprocal. is calculated, and the calculation result is set as the evaluation coefficient (denominator of the evaluation formula): score_coef in the control data of FIG. 3(a) (step S510). The evaluation coefficient score_coef calculated in this manner substantially obtains cos (cosine) distance=correlation coefficient.

The covariance (numerator) and the standard deviation of the rule data are calculated during the rule search process, which will be described later, but the variance corresponding to the input word set 111 on the input side is common to all rules. and is generated as a component of the denominator, first, after initial setting in step S503, the score coefficient is calculated by adding the square of the weight coefficient of each input word in step S508, and the calculated score coefficient is reciprocated in advance in step S510 to obtain the evaluation coefficient score_coef in FIG. This is to allow multiplication (rather than computationally expensive division). After that, the CPU 201 ends the pre-processing of step S405 in FIG. 4 illustrated in the flowchart in FIG.

FIG. 6 is a flowchart showing a detailed example of the rule search processing in step S406 of FIG. This rule search processing implements the processing function of the data acquisition unit 102 in FIG. As described above in the description of the data acquisition unit 102 in FIG. 1, the data acquisition unit 102 operates according to a state model called automaton while referring to the rule data 110 stored in the database 101 . Similarly, the rule search process of FIG. 6 operates according to the state model of the automaton while referring to the rule data 110 of FIG. Here, an automaton is a model that consists of a combination of states, transitions, and actions. It is the rule data 110 on the database 101 that "transitions" and defines the state. In addition, in this embodiment, as will be described later, the search of the rule data 110 and the transition of the state are also controlled according to the stack state number NSA stored in the storage unit 104 described with reference to FIG.

The processing of the flowchart of FIG. 6, which is a detailed example of the rule search processing in step S406 of FIG. 4, will be described below.

In FIG. 6, the CPU 201 first determines that all the rule searches for the input word set obtained from the user's utterance in the preprocessing of step S405 in FIG. 4 shown in FIG. Step S613: YES), a series of processes from steps S601 to S612 are repeatedly executed.

In this iterative process, the CPU 201 first reads the rule data 110 (corresponding to the rule data 110 in the database 101 in FIG. 1) of FIG. First, the database 102 is searched for the rule data 110 containing the assumed input word set included in the input word set (corresponding to the input word set 111 in FIG. 1) (step S601). As another embodiment of step S601, the CPU 201 determines that the assumed input word count (see FIG. 3C) is the input word count (FIG. 3A) of the input word set (corresponding to the input word set 111 in FIG. 1). ) and all the assumed words in the array of assumed input words (see FIG. 3(c)) match all of the input words (see FIG. 3(b)). good. Note that when the input sentence is in Japanese, the order of the words between the assumed input word set and the input word set does not matter. This is to cope with the case where words are inverted depending on the input sentence.

As a result of the search in step S601, the CPU 201 determines whether rule data has been found in the database 102 (step S602).

If no rule data is found (NO in step S602), the CPU 201 determines whether or not all rule data have been searched (step S613).

If the search for all rule data has not ended (NO in step S613), the CPU 201 returns to the search processing in step S601 and repeats the search for rule data.

If rule data is found as a result of the search in step S601 (if the determination in step S602 is YES), the CPU 201 executes a series of processes shown in steps S603 to S612 below to perform step S601. It is determined whether or not to adopt the rule data found in the above as the response candidate rule data 112 (see FIG. 1).

Specifically, the CPU 201 first sets the magnification indicated by the state number coefficient, which is a variable held in the memory 202 in FIG. 2, to the initial value of 1.0 (step S603). The state number coefficient is weight data for determining how important a state is to be given when retrieving rule data corresponding to a state that occurred in the past other than the current state.

Next, the CPU 201 stores a stack array of state numbers held (stored) as control data in the memory 202 (storage unit 104): state_id (see FIG. 3A). The first stack state number NSA, ie, the latest stored state number, is read out of the plurality of stack state numbers NSA (step S604).

Next, the CPU 201 determines whether the state number: state_id (see FIG. 3C) in the rule data found in step S601 matches the current stack state number NSA selected in step S604 or S611, which will be described later. It is determined whether or not (step S605).

If the determination in step S605 is YES, the CPU 201 stores a pointer to the rule data found in step S601 corresponding to the rule data in response candidate rule data list transCandidates (see FIG. 3A): transition (see FIG. 3 ( By registering the value of c)), the rule data is added to the response candidate rule data list as new response candidate rule data 112 (step S606).

Next, the CPU 201 adds each assumption in the assumed input word set of the rule data 110 found in step S601 of FIG. The square value of each weight coefficient: weight (see FIG. 3B) of each input word corresponding to the input word is accumulated for all input words (step S607). Note that the initial value of the score of the answer candidate rule data 112 is set to a predetermined value. Any value can be adopted as the predetermined value, for example, the value 0 may be used.

Subsequently, the CPU 201 accumulates the square value of the state number coefficient, which is a variable held in the memory 202, to the score of the response candidate rule data 112 added this time, which is referred to from the response candidate rule data list transCandidates. (step S608). As described above, the state number coefficient is weight data for determining how important the state number is when adopting the rule data 110 including the state number other than the current state number of the automaton. .

Next, the CPU 201 converts the score value of the rule data 110 as the response candidate rule data 112 added in step S606, which is referenced from the response candidate rule data list transCandidates, into the cosine similarity according to the following equation (1). Convert (step S609). "sqrt()" denotes an operation that computes the square root. "Ctrl.score_coef" is an evaluation coefficient referenced from the start address Ctrl of the control data in FIG. Also, “transCandidates->score” indicates the score variable value of the response candidate rule data 112 added this time on the memory 202 .

transCandidates->score
=sqrt(transCandidates->score)
x Ctrl. score_coef (1)

Since the covariance and the standard deviation corresponding to the rule data 110 as the new response candidate rule data 112 are calculated by the above equation (1), the standard deviation corresponding to the above-mentioned input word set is calculated. The correlation coefficient can be calculated by multiplying by the reciprocal evaluation coefficient.

Next, the CPU 201 returns from the process of step S613 to step S601, and repeats the search process for the next rule data.

On the other hand, if the determination in step S605 is NO, that is, if the state number: state_id of the rule data 110 found in step S601 does not match the current stack state number NSA selected in step S604 or S611, the CPU 201 It is determined whether or not the search for the stack state number NSA in the array has ended (step S610). When the determination in step S610 is NO, the stack state number NSA stored one before the stack state number NSA being read at that time is read out (step S611), and the state number coefficient is attenuated by a predetermined value. A coefficient (1.0>attenuation coefficient>0) is multiplied (step S612), and the process returns to step S605.

On the other hand, when the determination in step S610 is YES and the search for the stack state number NSA in the stack array of state numbers is completed, steps S613 and subsequent steps are executed.

On the other hand, when the determination in step S613 is YES and the search for all rule data 110 is completed, wild card list search processing and super wild card list search processing are executed in steps S614 and S615, respectively, as will be described later. , the process ends.

As described above, in the rule search process shown in FIG. 6, when the state number of the rule data 110 found in step S601 matches the first stack state number NSA read out in step S604, that is, the current state of the automaton ( In step S605: YES), the state number coefficient set to 1.0 in step S603 is used as it is, and is accumulated in step S608.

On the other hand, if the state number of the rule data found in step 601 does not match the first stack state number NSA (step S605: NO), the stack state numbers NSA stored in the stack array are read out from the newest one. (Step S611), the state number coefficient is multiplied by the attenuation coefficient (Step S612), and step S605 is executed again. A match is determined. Steps S611 and S612 are repeatedly executed until step S605 becomes YES unless the retrieval of all stack state numbers NSA is completed (step S610: NO).

As a result, the older the stack state number NSA read in step S611 is, the more times the state number coefficient is multiplied by an attenuation factor smaller than 1.0. , is set to a smaller value. In this case, for example, when the damping coefficient is 0.9, the state number coefficient is changed from 1.0, which is the initial value set in step S603, to 0.9→0 each time the damping coefficient is multiplied. .81, . . . Then, if the determination in step S605 becomes YES, the score of the response candidate rule data 112 is calculated using the attenuated state number coefficient (step S608).

FIG. 7 is a flowchart showing a detailed example of response sentence output processing in step S407 of FIG. First, the CPU 201 converts the maximum likelihood (highest score) candidate rule data 110 to the response rule data list of FIG. It is determined as data 113 (step S701).

Subsequently, the CPU 201 outputs the response sentence: bot_reply (see FIG. 3C) included in the maximum likelihood candidate rule data 110 in step S701 to the voice output device 209 in FIG. 2 (step S702). The voice output device 209 synthesizes a digital voice response signal corresponding to the response sentence: bot_reply, converts the digital voice response signal to an analog voice response signal with a built-in D/A converter, and converts the analog voice response signal to A sound is emitted toward the user via an amplifier and a speaker. The text data of the response sentence: bot_reply may be displayed on the display of the output device 204 without outputting the voice signal of the response sentence: bot_reply.

Subsequently, the CPU 201 acquires the transition destination state number: next_state_id (see FIG. 3C) of the maximum likelihood candidate rule data 110 (step S703).

Then, the CPU 201 prevents the transition destination state numbers obtained in step S702 from being consecutive on the stack array of state numbers: state_id (= stack state number NSA in FIG. It is pushed to the stack array as NSA (step S704). In this case, when this process is executed for the first time, that is, when no stack state number NSA is stacked in the state number stack array, the current state number and transition destination state number are pushed to the stack array in this order. .

8, 9, and 10 show operation examples of the processing described above. FIGS. 8B, 9, and 10 show an operation example when the state number of the automaton transitions from state number 0 to state number 4. FIG. Also, FIG. 8(a) shows a legend regarding the operation examples shown in FIGS. 8(b), 9, and 10. FIG. Numbers (X=0, 1, 2, 3, 4) enclosed in thick solid circles indicate state numbers of the automaton. A broken-line frame #Xi with a pound sign indicates the i-th rule data 110 (see FIG. 1) when the state number of the automaton is the state number X. FIG. In this rule data 110, dark-colored frames indicate "conditions" for selecting the rule data 110. FIG. Also, the text in the balloon facing left is the input word group (input word group 111 ) indicates an assumed input word set (see FIG. 1) to be matched. The text in the balloon facing to the right indicates the text of the response sentence (see FIG. 1).

Here, the method of acquiring the sentence input by the user and the input word set is as described above. The text data of this input word set corresponds to the input word set 111 in FIG. 1, but hereinafter simply referred to as "input word set" corresponds to the input word set 111 in FIG. .

For example, when the state number of the automaton shown in FIG. 8(b) is 0, it is in a state in which a rule data group for introducing a general speech at the start of a conversation is associated. In this state number 0, the rule data 110 [#0-0] is such that, for example, when an interrogative sentence including the word "I like" is given as an input word set corresponding to the input sentence by speaking of the user, This is rule data for outputting a response sentence corresponding to the text "I don't understand." Further, in this operation example, after the response sentence based on the rule data 110[#0-0] is output (step S702 in FIG. 7), the solid line output from the dashed frame of the rule data 110[#0-0] As indicated by the arrow, the state number of the automaton maintains the current state number 0. pushed onto the stack). This solid arrow corresponds to the next transition destination state number in the rule data 110 of FIG. The same applies to the rule data 110 [#0-1] and [#0-3] in state number 0 shown in FIG. 8(b).

On the other hand, the rule data 110 [#0-2] in state number 0 shown in FIG. 8B includes two words "animal" and "talk" corresponding to the interrogative sentence of the input sentence "animal story". This is a rule for outputting a response sentence corresponding to the text "Yes, let's talk about animals" when given an input word set. Also, in this operation example, after the response sentence based on this rule data 110[#0-2] is output (step S702), the solid line arrow output from the dashed frame of the rule data 110[#0-2] is changed to a thick line. As shown by adding "1" surrounded by a dashed circle, the state number of the automaton transitions from the current state number 0 to the state number 1 shown in FIG. The transition destination state number=1 is acquired from the data 110 [#0-2] and pushed onto the stack in step S704).

When the state number of the automaton shown in FIG. 9 is state number 1, it is in a state in which a group of rule data for talking about animals is associated. For example, the rule data 110 [#1-0] is such that when an input word set generated from an input sentence includes the word "cat", the response sentence "Cat's eyes are big, isn't it?" as a general topic about cats. It is a rule for output. In addition, the rule data 110 [#1-1] is such that when an input word set generated from an input sentence includes the word "eyes", "cat's eyes are big, aren't they?" as a general topic about cat's eyes. This is a rule for outputting a response sentence. After these rule data 110 [#1-0] and [#1-1] are output (step S702 in FIG. 7), each of the rule data 110 [#1-0] and [#1-1] The state numbers of the automaton are shown in FIG. 9 through the current state number 1 shown in FIG. Transition to state number 2 (transition destination state number=2 is acquired from rule data 110 [#1-0] or [#1-1] in step S703 of FIG. 7, and pushed to the stack in step S704).

On the other hand, in state 1 of FIG. 9, the rule data 110 [#1-2], for example, defines a general topic about foxes as a general topic about foxes when an input word set generated from an input sentence includes the keyword "foxes." This is a rule for outputting a response sentence such as "I have an image of you acting at night." In addition, the rule data 110 [#1-3] indicates that when an input word set generated from an input sentence includes the word "eyes", a general topic related to fox eyes is "A fox might have eyes like a cat." This is a rule for outputting the response sentence. After these rule data 110 [#1-2] and [#1-3] are output (step S702 in FIG. 7), each of the rule data 110 [#1-2] and [#1-3] The state numbers of the automaton are shown in FIG. 9 through the current state number 1 shown in FIG. Transition to state number 4 (transition destination state number=4 is acquired from rule data 110 [#1-2] or [#1-3] in step S703 of FIG. 7, and pushed to the stack in step S704).

Note that both the rule data 110 [#1-1] regarding the topic of cats and the rule data 110 [#1-3] regarding the topic of foxes are generated when the input word set generated from the input sentence includes the word "eyes". However, which one is selected may change depending on the conditions for calculating the degree of similarity (cosine similarity) between the input word set and the assumed input word set of each rule data 110 (FIGS. 5 and 6). (steps S607, S608, S609, step S701 in FIG. 7). Such a configuration avoids the dialogue from becoming uniform.

The state of state number 2 in FIG. 10 that transitions after rule data 110 [#1-0] or rule data 110 [#1-1] in state number 1 in FIG. 9 is selected, and the state that transitions from state number 2 State number 3 corresponds to the group of rule data 110 relating to deeper topics about cats. On the other hand, the state of state number 4 in FIG. 9, which transitions after rule data 110 [#1-2] or rule data 110 [#1-3] in state number 1 of FIG. etc.).

The rule data 110 [#1-4] in state number 1 in FIG. This is a rule for returning similar ambiguous response sentences such as After the rule data 110 [#1-4] is output (step S702 in FIG. 7), the solid-line arrow that emerges from the dashed-line frame of the rule data 110 [#1-4] is surrounded by a thick solid-line circle with "1 , the state number of the automaton maintains the state number 1 shown in FIG. The number=1 is obtained, and 1 remains stored at the top of the stack in step S704).

FIG. 11 is a diagram showing a specific operation example of the rule search processing in FIG. 6 corresponding to the automaton illustrated in FIGS. 8 to 9. In FIG. First, in the state number 0 of the automaton, the user speaks, for example, and the input sentence In[0] "Let's talk about animals" is input. Suppose that an input word set consisting of is generated by morphological analysis. 3C stored in the memory 202 or the auxiliary information storage device 205, the assumed input word set userWords[0], userWords [1] (see FIG. 3(c)) is included in (here, "matches") the input word set "animal" and "story", and the rule data 110 is retrieved (step S601 in FIG. 6). . As a result, the current state number state_id (see FIG. 3(c)) matches the automaton state number 0 in FIG. )) is included in (or matches with) the input word set “animal” and “talk” is searched for one rule data 110 [#0-2]. Also, although not explained in step S601, it is determined that the input sentence In[0] is an interrogative sentence of the "proposal" type as the "type" and has "affirmative" as the "affirmative/negative" item. , the rule data 110 [#0-2] is retrieved and set as the response candidate rule data 112 (the first determination in step S605 in FIG. 6 is YES, and step S606 fart).

Next, as shown in FIG. 11, this one response candidate rule data 112[#0-2] is selected as the response rule data 113 corresponding to the input sentence In[0] (step S701 in FIG. 7). ). As a result, as shown in FIG. 11, the response text bot_reply (see FIG. 3(c)) of the rule data 110 [#0-2] selected as the response rule data 113 becomes the response text Out[0]=“ Yes, let's talk about animals" (see FIG. 11), that is, is generated and output as the response sentence 114 (see FIG. 1) (step S702). Along with this, the rule data 110 [#0-2] selected as the response rule data 113 is referred to, and the rule data 110 [#0-2] contains the value 1 as the next transition destination state number next_state_id (Fig. 7 step S703), the state number of the automaton changes from state number 0 to state number 1 (step S704 in FIG. 7).

At this time, if the response rule data 113 is selected for the first time like the rule data 110 [#0-2] selected as the response rule data 113, the current rule data 110 [#0-2] The value 0 of the state number state_id (see FIG. 3(c)) is first pushed and stored as the stack state number NSA in the state number stack array: state_id (see FIG. 3(a)). The value 1 of the next transition destination state number next_state_id of the data 110 [#0-2] is further pushed to the state number stack array: state_id (see FIG. 3A) and stored as the next stack state number NSA. (step S704 in FIG. 7). In this case, the state number 0 is stored as the previous past stack state number NSA, and the next transition destination state number 1 is stored as the latest stack state number NSA.

Also, at this time, although not shown in the flowchart of FIG. ), the response sentence Out[0]=“Yes, let's talk about animals” of the rule data #0-2 selected as the response rule data 113 is stored.

Next, in state number 1 of the post-transition automaton, the input sentence In[1] "I'm thinking of getting a cat" is input, and the input word set 111 including the word "cat" is generated by morphological analysis of this. Suppose it was On the other hand, among the plurality of rule data 110 in FIG. 3C, the assumed input word set userWords[0] (see FIG. ” is searched for rule data 110 that is included in (or matches with) (step S601 in FIG. 6). As a result, the current state number state_id (see FIG. 3(c)) matches the state number 1 of the automaton shown in FIG. is included in (or matches with) the input word set "cat". Also, the input sentence In[1] is determined to be an interrogative sentence of the "other" type as the "type" and has "affirmative" as the "affirmative/negative" item, which corresponds to the input sentence In[1] in FIG. Then, the rule data 110 [#1-0] is searched as the answer candidate rule data 112 (the first judgment in step S605 in FIG. 6 becomes YES, and the process proceeds to step S606).

Next, as shown in FIG. 11, this one response candidate rule data 112[#1-0] is selected as the response rule data 113 corresponding to the input sentence In[1] (step S701 in FIG. 7). ). As a result, as shown in FIG. 11, the response text bot_reply (see FIG. 3(c)) of the rule data 110[#1-0] selected as the response rule data 113 becomes the response text Out[1]=“ Cat's eyes are big, aren't they?" (see FIG. 11), that is, is generated and output as response sentence 114 (see FIG. 1) (step S702). Along with this, the rule data 110 [#1-0] selected as the response rule data 113 is referred to, and this rule data 110 [#1-0] contains the value 2 as the next transition destination state number next_state_id (see FIG. 7 step S703), the state number of the automaton transitions from the previous state number 1 to state number 2 (step S704 in FIG. 7).

At this time, if the rule data 110 [#1-0] selected as the response rule data 113 is the response rule data 113 selected after the second time, the rule data 110 [#1-0] ] is pushed to the state number stack array: state_id (see FIG. 3A) and stored as the latest stack state number NSA (step S704).

Also, at this time, although not shown in the flowchart of FIG. Cats have big eyes." is remembered.

Next, assume that in state number 2 of the automaton after the transition, the input sentence In[2] is "That's right, your eyes are big and cute, isn't it?" . On the other hand, among the plurality of rule data 110 in FIG. 3C, the assumed input word set userWords[0] (see FIG. ” is retrieved (or matches) (step S601).

As a result, the assumed input word set userWords[0] (see FIG. 3C) is included in the input word set "me" (or matched ), one rule data 110 [#2-0] whose current state number state_id (see FIG. 3(c)) matches the state number 2 of the automaton, and the current state number state_id (see FIG. 3 The rule data 110 [#1-1] and [#1-3] in which (see (c)) matches the state number 1 are sequentially searched, and the determination in step S602 becomes YES. Subsequently, in subsequent steps S603 to S612, for the rule data 110 [#2-0], the state number 2 of the automaton is added to the top of the state number stack array: state_id (see FIG. 3(a)). Matches the latest pushed stack state number NSA (=2), and as a result of the first determination of YES in step S605, the state number coefficient=1 is accumulated in the score (step S608). ). Similarly, for the rule data 110 [#1-1] and [#1-3], the state numbers of these automatons are the stack array of state numbers: state_id (see FIG. 3(a)). The second pushed stack state number NSA (=1) is matched, and as a result, step S612 is executed once each by making a second determination of YES in step S605. , state number coefficient=1×0.9=0.9 are accumulated (steps S612, S608).

From the above, rule data 110 [#2-0] whose score is multiplied by 1.0, and rule data 110 [#1-1] and [#1-3] whose respective scores are multiplied by 0.9 are acquired as response candidate rule data 112 in the response candidate rule data list transCandidates.

Next, although not shown in the flow chart of FIG. 6, the respective response sentences 114 for the rule data 110 [#2-0] and the rule data 110 [#1-1] and [#1-3] are: It is determined whether or not it is registered in the above-described response sentence storage unit. As a result, the response sentence of the rule data 110[#2-0] (FIG. 10) and the rule data 110[#1-1] (FIG. 9)=“Cat's eyes are big, aren't they?” ] and registered in the response sentence storage unit. As a result, rule data 110 [#2-0] (FIG. 10) and rule data 110 [#1-1] (FIG. 9) are selected in order to avoid outputting the same response sentence 114 in succession. First, the rule data 110[#1-3] is selected as the response rule data 113 corresponding to the input sentence In[2], as shown in FIG.

As a result, as shown in FIG. 11, the response text bot_reply (see FIG. 3(c)) of the rule data 110 [#1-3] selected as the response rule data 113 becomes the response text Out[2]=“ A fox might have eyes like a cat' (see FIG. 11), that is, it is generated and output as the response sentence 114 (see FIG. 1) (step S702). Along with this, the rule data 110 [#1-3] selected as the response rule data 113 is referred to, and the rule data 110 [#1-3] contains the value 2 as the next transition destination state number next_state_id (Fig. 7 step S703), the state number of the automaton representing the rule data 110 to be selected next maintains the previous state number 2 (step S704 in FIG. 7).

At this time, since the rule data 110 [#1-3] selected as the response rule data 113 is the response rule data 113 selected after the second time, the next rule data included in the rule data 110 [#1-3] The value 2 of the transition destination state number next_state_id is pushed to the state number stack array: state_id (see FIG. 3A) and stored as the latest stack state number NSA (step S704).

Also, at this time, although not shown in the flowchart of FIG. A fox might have eyes like a cat' is remembered.

Next, in state number 2 of the post-transition automaton, assume that the input sentence In[3] is "Suddenly you're talking about a fox, isn't it?", and an input word set 111 including the word "fox" is generated. . On the other hand, among the plurality of rule data 110 in FIG. 3(c), the assumed input word set userWords[0] (see FIG. 3(c)) is included in the input word set "fox" (or , matching) rule data 110 is retrieved (step S601).

As a result, in the loop processing of steps S601→S602→S613→S601 in FIG. 6, the rule data 110 including the current state number 2 of the automaton in FIG. , “fox” in the assumed input word array userWords[] (see FIG. 3(c)) does not exist, but as illustrated in FIG. 3(c)), the rule data 110 [#1-2] including "fox" in the assumed input word array userWords[] (see FIG. 3(c)) is searched. , the determination in step S602 is YES. Subsequently, in subsequent steps S603 to S612, for the rule data 110 [#1-2], the state number 1 of the automaton is pushed before the latest state number 2 of the state number stack array: state_id. and the past stack state number NSA (=1), which results in a YES determination in step S605 for the second time. 0.9=0.9 is multiplied (steps S612, S608).

As described above, only rule data 110 [#1-2] is acquired as response candidate rule data 112 in the response candidate rule data list transCandidates.

Next, as shown in FIG. 11, the rule data 110 [#1-2] acquired as the answer candidate rule data 112 is selected as the answer rule data 113 corresponding to the input sentence In[3] (step S701). As a result, as shown in FIG. 11, the response text bot_reply (see FIG. 3(c)) of the rule data 110 [#1-2] selected as the response rule data 113 becomes the response text Out[3]=“ There is an image that foxes act at night.” (see FIG. 11), that is, is generated and output as response sentence 114 (see FIG. 1) (step S702). After that, the rule data 110 [#1-2] selected as the response rule is referred to, and since this rule data 110 [#1-2] contains the value 4 as the next transition destination state number next_state_id (step S703), the state number of the automaton transitions from state number 2 to state number 4 (step S704 in FIG. 7).

At this time, since the rule data 110 [#1-2] selected as the response rule data 113 is the response rule data 113 selected after the second time, the next rule data included in the rule data 110 [#1-2] The value 4 of the transition destination state number next_state_id is pushed to the state number stack array: state_id and stored as the latest stack state number NSA (step S704).

Also, at this time, the response sentence Out[1] of the rule data 110 [#1-2] selected as the response rule data 113 is stored in the response sentence storage unit. remembered.

Further, in state number 1 after the transition, it is assumed that the input sentence In[4] is "that's right, you know well" and that an input word set 111 including the word "that's right" is generated. On the other hand, among the plurality of rule data 110 in FIG. 3C, the assumed input word group userWords[0] (see FIG. The rule data 110 included in (or matching with) is searched (step S601).

As a result, in the loop processing of steps S601→S602→S613→S601 in FIG. 6, the rule data 110 including the current state number 4 of the automaton in FIG. , "Sona na" in the assumed input word array userWords[] (see FIG. 3(c)) is not found, but as illustrated in FIG. The rule data 110 [#1-4] includes "that's right" in the array userWords[] of assumed input words (see FIG. 3(c)) in the rule data 110 included as the number state_id (see FIG. 3(c)). is retrieved, and the determination in step S602 becomes YES. Subsequently, in subsequent steps S603 to S612, for the rule data 110 [#1-4], the state number 1 of the automaton is higher than the latest stack state number NSA (=4) of the state number stack array: state_id. Matches the past stack state number NSA (=1) that was pushed earlier, and as a result of the determination of YES in step S605 for the third time, step S612 is executed twice, and the state number is added to the score. A coefficient=1×0.9×0.9=0.81 is multiplied (steps S612 and S608).

As described above, the rule data 110 [#1-4] is acquired as the response candidate rule data 112 in the response candidate rule data list transCandidates.

Next, as shown in FIG. 11, for rule data 110 [#1-4] acquired as response candidate rule data 112, the response sentence bot_reply (see FIG. 3(c)) is the response sentence Out[4]. ]="I think so" is checked whether or not it is registered in the above-described response sentence storage unit. As a result, it is determined that the response sentence Out[4]=“I think so” is not registered in the response sentence storage unit. As a result, the rule data 110 [#1-4] including the state number 1 of the automaton is selected as the response rule data 113 corresponding to the input sentence In [4], as shown in FIG. 11 (step S701). . Then, as shown in FIG. 11, the response sentence bot_reply (see FIG. 3(c)) of the rule data 110 [#1-4] selected as the rule data 113 is the response sentence Out[4]=“That's right. I think" (see FIG. 11), that is, is generated as a response sentence 114 (see FIG. 1) and output (step S702). Along with that, the rule data 110 [#1-4] selected as the response rule data 113 is referred to, and since this rule data 110 [#1-4] contains the value 1 as the next transition destination state number next_state_id ( Step S703 in FIG. 7), the state number of the automaton representing the rule data 110 to be selected next maintains the previous state number 1 (step S704 in FIG. 7).

At this time, since the rule data 110 [#1-4] selected as the response rule data 113 is the response rule data 113 selected after the second time, the next rule data included in the rule data 110 [#1-4] The value 1 of the transition destination state number next_state_id is pushed to the state number stack array: state_id (see FIG. 3A) and stored as the latest stack state number NSA (step S704).

At this time, although not shown in the flow chart of FIG. of” (response sentence 114) is stored.

In the rule data 110 [#0-3] including the state number 0 of the automaton shown in FIG. 8(b), the item of the input statement is "*". This is a rule with a word matching condition of "treat any word (morpheme) as applicable". Then, for example, as an input sentence, there is a keyword (and there is no condition such as a question sentence) that is not set in any of the other rule data 110 [#0-0] to [#0-2] including the state number 0. It is a rule for outputting a response sentence such as "What did you say?" when given. This "*" is called a wild card. A solid-line arrow emerging from the dashed frame of the rule data 110[#0-3] indicates that after the rule data 110[#0-3] is selected, the same state number 0 as before the selection is maintained. (That is, the next transition destination state number is 0). By setting such wildcard rule data, ambiguous dialogue can be realized.

In the rule data 110 [#3-2] including the state number 3 of the automaton shown in FIG. 10, the input statement item is "#". This is called a super wild card. The super-wildcard has the same word matching condition as the wildcard, ``treat any word (morpheme) as corresponding'', but the rule data 110 is the state number that matches the current state number of the automaton. Only those that include are added as answer candidate rule data 112, and when this rule data 110 is added to the candidates, rule data 110 in other states are deleted from the answer candidate rule data list. Similar to the wildcard described above, a keyword (and a condition such as an interrogative sentence) that is not set in any other rule data 110 [#3-0] to [#3-1] including state number 3 is used as an input sentence. This is a rule for outputting a response sentence such as "I would like to touch it" when given "No"). Here, for example, after the rule data 110 [#3-3] is selected by the solid-line arrow emerging from the dashed frame of the rule data 110 [#3-3], a different state number (for example, state number 2 ). In this way, when an appropriate transition destination is described in the super wildcard rule data 110, the behavior of transitioning to another state number can be realized. By setting such super wild card rule data 110, it is possible to realize a dialogue that greatly changes the topic when the dialogue gets stuck. Alternatively, for example, it is possible to implement behavior such as staying in the same state until a satisfactory answer is obtained systematically, that is, repeating the same question.

In the rule search process of FIG. 6, after the search of all rule data for the current input word set is completed and the determination in step S613 becomes YES, the CPU 201 also performs the above-described wild card and super wild card: The same search processing as in the processing from steps S601 to S613 described above is executed. Its details are omitted.

As described above, after the rule data 110 corresponding to the super wild card is selected as the response rule data 113 (see FIG. 1), the state number of the automaton transitions to a state number different from that before selection. Therefore, the state number included in the rule data 110 is the same regardless of whether the stack array of state numbers: state_id (=stack state number NSA in FIG. 1) includes the stack state number NSA of the same number. , a rule that does not match the current state number may be controlled so as not to be added to the response candidate rule data list transCandidates. Also, when the super wild card rule is added to this list, the list may be controlled to exclude rules having state numbers other than the current state number.

As described above, in the present embodiment, the rule data 110 corresponding to the input utterance information 115 among the plurality of rule data 110 stored in the database 102, for example, the set of assumed input words corresponding to the set of input words in the input utterance information 115 is set as a candidate for response rule data 113 (response candidate rule data 112). A rule indicating a predetermined state including the current state of the automaton among the plurality of rule data 110, for example, a state included in a plurality of states stored in the storage unit 104 for sequentially storing the current state or the state of the automaton. Data 110 is selected as a candidate for response rule data 113 (response candidate rule data 112). Then, the response sentence 114 included in the response rule data 113 selected as such is output. As a result, in this embodiment, it is possible to output the response sentence 114 based on the response rule data 113 along the flow of conversation with the user, for example. At this time, in the present embodiment, among the plurality of rule data 110, the state of the corresponding automaton responds with rule data indicating the same state as the state stored more recently among the plurality of states stored in the storage unit 104. It can be preferentially selected as the rule data 113 . As a result, it is possible to output a response sentence 114 based on the response rule data 113 that better corresponds to the current topic.

Unlike the case of this embodiment, in the conventional interactive device, for example, appropriate rule data including the current state number of the automaton is not set. If it was not selected as the data 113, for example, rule data containing other state numbers had to be searched at random, and as a result, the topic suddenly changed. On the other hand, in the present embodiment, since the response rule data 113 can be selected as described above, the conversation apparatus 100 can suppress abrupt switching of the topic and can have a natural conversation with the user. can be provided.

Further, in the present embodiment, a score indicating an index for selecting the response rule data 113 is calculated for each rule data 110. The score is calculated so that the rule data 110 corresponding to the state of the automaton showing the same state as the previously stored state is less likely to be selected as the response rule data. For example, in each rule data 110, the state number coefficient of the rule data 110 is multiplied by an attenuation coefficient whose value is attenuated as the state indicated by the rule data 110 is stored in the storage unit 104 in the past, and The multiplied state number factor is accumulated in the score of that rule data 110 . Then, the rule data 110 having the maximum score among the multiple rule data 110 is selected as the response rule data 113 . Therefore, it is possible to provide the dialogue device 100 that enables natural dialogue based on past topics with the user and in line with more recent topics.

Further, in this embodiment, a variable weight is set for each input word constituting the input sentence according to the context of the input sentence (input utterance information 115). The weight of each input word corresponding to each assumed input word in 112 is accumulated to calculate a similarity parameter indicating the similarity of the response candidate rule data 112 to the input sentence. A score is calculated for the candidate rule data 112 . Then, the candidate response rule data 112 having the maximum score among the score values of each candidate response rule data 112 is selected as the response rule data 113 . Therefore, it is possible to select correct response rule data 113 according to the context of the input sentence.

Furthermore, in this embodiment, based on the results of comparison between the assumed input words included in each of the plurality of rule data 110 and the plurality of input words in the input word set 111 extracted by morphological analysis from the input sentence, Response candidate rule data 112 is retrieved from a plurality of rule data 110 in the database 101 . In this way, in this embodiment, since the response candidate rule data 112 is searched by comparing words, it is possible to determine dialogue rules based on appropriate words included in the topic.

In addition, in this embodiment, when selecting response rule data 113 from response candidate rule data 112, the same response sentence 114 is not repeatedly output by referring to the response sentence storage unit 105, which stores a predetermined number of past response sentences. It is possible to prevent the dialogue from becoming monotonous.

In this embodiment, a weighting factor is set for each input word, but the same weighting factor may be set uniformly for all input words.

In this embodiment, the cosine similarity is calculated as the score of each response candidate rule data, and the response rule data 113 is selected from the plurality of response candidate rule data 112 according to the magnitude of the cosine similarity. A similarity operation may be applied.

In the present embodiment, a damping coefficient is set such that the value of the answer candidate rule data 112 is attenuated as the stack state number NSA stored in the past among the plurality of stack state numbers NSA stored in the storage unit 104 is attenuated. The state number coefficient is multiplied, the state number coefficient is accumulated in the score of the answer candidate rule data 112, and the answer rule data 113 is selected from the plurality of answer candidate rule data 112 according to the score. The candidate response rule data 112 corresponding to a specific past status may be preferentially selected as the response rule data 113 based on another algorithm from the status history.

In this embodiment, the assumed input word set set for each rule data 110 is compared with the input word set 111 to select the response candidate rule data 112. However, instead of the assumed input word set, for example A word set obtained by morphological analysis from a response sentence in the rule data 110 may be compared with the input word set 111 . Alternatively, variously set word sets or sentences may be compared with the input word set.

In this embodiment, when the content of the user's utterance is unexpected content (when a plurality of words in the input word set 111 does not include a plurality of words in any of the assumed input word sets of the rule data 110) ) are made to refer to non-expected rules (wildcard list, super wildcard list), but if the contents are unexpected, they are not referred to, and dialogue is carried out according to the predetermined rules. , or may not speak at all.

In this embodiment, when selecting response rule data 113 from response candidate rule data 112, the same response sentence 114 is prevented from being repeatedly output by referring to the response sentence storage unit 105 that stores a predetermined number of past response sentences. However, the same response sentence may be output repeatedly according to a predetermined algorithm.

In this embodiment, the candidate response rule data 112 corresponding to the input word set 111 is selected from a plurality of rule data 110 in the database 101, and finally the response rule data 113 is selected from the plurality of candidate response rule data 112. 6 and 7 is shown as a method for determining the number of stacks corresponding to the most recently stored stack state number NSA among the plurality of stack state numbers NSA stored in the storage unit 104. Various algorithmic techniques can be employed under the condition that the response candidate rule data 112 is preferentially selected as the response rule data 113 . For example, from the stage of searching the answer candidate rule data 112 from the rule data 110 in the database 102, each state in the plurality of stack state numbers NSA stored in the storage unit 104 is compared with each state in the rule data 112. The search may be performed while

In addition to the configuration of this embodiment described above, when rule data containing the same state number as the current state number is found, the search of the answer candidate rule data list 112 is terminated at that point, and the stack state number stored in the storage unit 104 is retrieved. An approach that does not consider NSA may be adopted.

In the present embodiment described above, the process of step S601 of searching for the rule data 110 including the assumed input word set included in the input word set is repeatedly executed until all rule searches in FIG. 6 are completed. On the other hand, even if the rule data 110 including the state number matching either the current state number of the automaton or the stack state number NSA included in the stack array is searched according to the input word set, good.

Furthermore, in the above-described embodiment, the cosine similarity reflecting the attenuation coefficient is calculated as the score. A process may be performed that causes the Further, although the attenuation coefficient is used as a multiplication term, the score calculation formula may be set so that the score is calculated using the attenuation coefficient as a subtraction term.

In the above-described embodiment, a microphone may be further provided, and the acquisition unit 106 may acquire the input utterance information 115 based on the voice of a predetermined target, such as a user, input via the microphone.
Further, in the above-described embodiment, a speaker may be further provided, and the response sentence output unit 103 may output a voice corresponding to the response sentence 114 to a predetermined target, such as a user, via the speaker. .
With these configurations, for example, the interactive device 100 according to this embodiment can be realized as an interactive application for a robot or a smartphone.

In this embodiment, it is also possible to provide the interactive device 100 as a computer program executed by a computer having the hardware configuration example of FIG.

In this embodiment, the input sentence from the user is given as voice data, and text data of the input sentence is given by executing voice recognition on it. The text data of the input sentence may be given directly via the system, various messaging systems, SNS systems, or the like.

While the disclosed embodiments and their advantages have been described in detail above, those skilled in the art can make various modifications, additions, and omissions without departing from the scope of the invention, which is clearly defined in the appended claims. .

In addition, the present invention is not limited to the above-described embodiments, and can be modified in various ways without departing from the gist of the present invention. Also, the functions executed in the above-described embodiments may be combined as appropriate as possible. Various steps are included in the above-described embodiments, and various inventions can be extracted by appropriately combining the disclosed multiple constituent elements. For example, even if some constituent elements are deleted from all the constituent elements shown in the embodiments, if an effect can be obtained, a configuration in which these constituent elements are deleted can be extracted as an invention.

The following notes are further disclosed with respect to the above embodiments.
(Appendix 1)
a database storing a plurality of rule data each including a response sentence and associated with mutually different states of the automaton, and defining transition destination states of the states of the automaton;
Acquisition means for acquiring input utterance information input from a predetermined target;
selecting response rule data from the plurality of rule data according to a predetermined state including the current state of the automaton and the obtained input utterance information, and the response sentence included in the selected response rule data; response sentence output means for outputting to the predetermined target;
storage means for sequentially storing states of the automaton,
The predetermined state includes a plurality of states stored in the storage means,
The response sentence output means outputs rule data indicating that the state of the corresponding automaton, among the plurality of rule data, is the same as the state stored more recently among the plurality of states stored in the storage means. , preferentially selected as the response rule data,
interactive device.
(Appendix 2)
The response sentence output means is
According to the obtained input utterance information, from a plurality of rule data less than the plurality of rule data corresponding to a predetermined state including the current state of the automaton, out of the plurality of rule data, the response data acquisition means for retrieving and acquiring a plurality of rule data of response candidates that are candidates for rule data;
response candidate rule data indicating that the state of the corresponding automaton, among the plurality of rule data of the acquired response candidates, is the same as the state stored more recently among the plurality of states stored in the storage means; , preferentially selected as the response rule data,
1. An interactive device according to Appendix 1.
(Appendix 3)
The response sentence output means is
A score indicating an index for selecting the response rule data is calculated for each of the rule data, and the score stored earlier than the plurality of states stored in the storage means among the plurality of rule data is calculated. calculating the score so that the rule data corresponding to the state of the automaton indicating the same state as the state of the response is less likely to be selected as the response rule data;
3. A dialogue device according to appendix 1 or 2.
(Appendix 4)
each of the plurality of rule data includes an assumed input utterance sentence assumed to be input from the predetermined object, associated with the state of the automaton and the response sentence;
extraction means for extracting an input word included in the predetermined target utterance sentence based on the acquired input utterance information;
weight setting means for setting a weight for each extracted input word;
a similarity parameter calculating means for calculating a similarity parameter indicating a similarity of the assumed input utterance sentence to the predetermined target utterance containing the input word according to the weight set for each input word; further prepared,
The response sentence output means calculates the score according to the set similarity parameter.
3. An interactive device according to appendix 3.
(Appendix 5)
further comprising response sentence storage means for storing a predetermined number of past response sentences generated based on the selected response rule data;
The response sentence output means selects the response rule data from the response candidate rule data by referring to the response sentence storage means so that the same response sentence is not repeatedly output. A dialogue device according to any one of the preceding claims.
(Appendix 6)
Equipped with an additional microphone,
The acquisition means acquires the input utterance information based on the predetermined target voice input via the microphone.
6. A dialogue device according to any one of appendices 1 to 5.
(Appendix 7)
Equipped with additional speakers,
The response sentence output means outputs a voice corresponding to the response sentence to the predetermined target via the speaker.
7. A dialogue device according to any one of appendices 1 to 6.
(Appendix 8)
a process of acquiring input utterance information input from a predetermined target;
Using a database that stores a plurality of rule data each including a response sentence and associated with mutually different states of the automaton, and defining transition destination states of the states of the automaton, from the plurality of rule data, the automaton selecting response rule data according to a predetermined state including the current state of and the obtained input utterance information, and outputting the response sentence included in the selected response rule data to the predetermined target and output processing to
a process of sequentially storing the state of the automaton in a storage means,
The predetermined state includes a plurality of states stored in the storage means,
In the output processing, rule data indicating the same state of the corresponding automaton among the plurality of rule data as the state stored more recently among the plurality of states stored in the storage means is output as the rule data. Including the process of preferentially selecting as response rule data,
how to interact.
(Appendix 9)
A program for causing a computer to execute the interaction method according to appendix 8.

100 dialogue device 101 database 102 data acquisition unit 103 response sentence output unit 104 storage unit 105 response sentence storage unit 106 acquisition unit 107 extraction unit 110 rule data 111 input word set 112 response candidate rule data 113 response rule data 114 response sentence 115 input utterance Information NSA Stack State Number 201 CPU
202 Memory 203 Input Device 204 Output Device 205 Auxiliary Information Storage Device 206 Media Drive Device 207 Network Connection Device 208 Audio Input Device 209 Audio Output Device 210 Bus 211 Portable Recording Medium

Claims (9)

Acquisition means for acquiring input utterance information input from a predetermined target;
A predetermined state including the current state of the automaton and a predetermined state including the current state of the automaton from a database that stores a plurality of rule data in which a plurality of sets of input words, response sentences, states of the automaton, and states of the automaton to be transitioned to next are associated with each other. a response sentence for selecting rule data in accordance with a plurality of input words contained in the input utterance information acquired by said acquisition means, and outputting a response sentence contained in the selected rule data to said predetermined target; an output means;
storage control means for sequentially storing the predetermined states corresponding to the selected rule data in a storage means ;
After the storage control by the storage control means, when the input utterance information is acquired by the acquisition means, the response sentence output means , if there is a plurality of rule data to be selected, among the data , preferentially selecting rule data corresponding to the same state as the predetermined state stored more recently among the plurality of predetermined states stored in the storage means;
A dialogue device characterized by: According to the obtained input utterance information, candidate and further comprising data acquisition means for retrieving and acquiring a plurality of rule data,
The response sentence output means stores a newer predetermined state than the plurality of predetermined states in which the state of the corresponding automaton is stored in the storage means among the acquired plurality of candidate rule data. preferentially select the rule data showing the same state as
2. The interactive device according to claim 1, characterized by: The response sentence output means is
A score indicating an index for selecting the rule data is calculated for each of the rule data, and a state stored earlier than the plurality of states stored in the storage means among the plurality of rule data is calculated. calculating the score so that the rule data corresponding to the predetermined state indicating the same state is less likely to be selected;
3. A dialogue device according to claim 1 or 2, characterized in that: each of the plurality of rule data includes an assumed input utterance sentence assumed to be input from the predetermined object, associated with the predetermined state and the response sentence;
extraction means for extracting an input word included in the predetermined target utterance sentence based on the acquired input utterance information;
weight setting means for setting a weight for each extracted input word;
a similarity parameter calculating means for calculating a similarity parameter indicating a similarity of the assumed input utterance sentence to the predetermined target utterance containing the input word according to the weight set for each input word; further prepared,
The response sentence output means calculates the score according to the set similarity parameter.
4. The interactive device according to claim 3, characterized by: further comprising response sentence storage means for storing a predetermined number of past response sentences generated based on the selected rule data;
5. The method according to any one of claims 1 to 4, wherein said response sentence output means selects said rule data by referring to said response sentence storage means so that the same response sentence is not repeatedly output. Dialog device as described. Equipped with an additional microphone,
The acquisition means acquires the input utterance information based on the predetermined target voice input via the microphone.
6. A dialogue device according to any one of claims 1 to 5, characterized in that: Equipped with additional speakers,
The response sentence output means outputs a voice corresponding to the response sentence to the predetermined target via the speaker.
7. A dialogue device according to any one of claims 1 to 6, characterized in that: A dialogue method executed by a dialogue device,
an acquisition step of acquiring input utterance information input from a predetermined target;
A predetermined state including the current state of the automaton and a predetermined state including the current state of the automaton from a database that stores a plurality of rule data in which a plurality of sets of input words, response sentences, states of the automaton, and states of the automaton to be transitioned to next are associated with each other. , a response sentence for selecting rule data in accordance with a plurality of input words included in the input utterance information acquired in the acquiring step , and outputting a response sentence included in the selected rule data to the predetermined target. an output step;
a storage control step for sequentially storing the predetermined states corresponding to the selected rule data in a storage means ;
After the storage control by the storage control step, when the input utterance information is acquired in the acquisition step, the response sentence output step, if there is a plurality of rule data to be selected, stores a plurality of rule data in the database. among the rule data , preferentially selecting rule data corresponding to the same state as the predetermined state stored more recently than the plurality of predetermined states stored in the storage means;
A dialogue method characterized by: to the computer of the interactive device,
an acquisition step of acquiring input utterance information input from a predetermined target;
A predetermined state including the current state of the automaton and a predetermined state including the current state of the automaton from a database that stores a plurality of rule data in which a plurality of sets of input words, response sentences, states of the automaton, and states of the automaton to be transitioned to next are associated with each other. , a response sentence for selecting rule data in accordance with a plurality of input words included in the input utterance information acquired in the acquiring step, and outputting a response sentence included in the selected rule data to the predetermined target. output step,
executing a storage control step of sequentially storing the predetermined states corresponding to the selected rule data in a storage means;
After the storage control by the storage control step, when the input utterance information is acquired in the acquisition step, the response sentence output step, if there is a plurality of rule data to be selected, stores a plurality of rule data in the database. A program for preferentially selecting, among rule data, rule data corresponding to the same state as the predetermined state stored more recently than the plurality of predetermined states stored in the storage means.

Images