JP3748429B2 - Speech input type compound noun search device and speech input type compound noun search method - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a search device and a search method for searching for a compound noun search word input by a user using a speech recognition device, searching a database, and confirming the user's input.
[0002]
[Prior art]
In a system that uses a speech recognition technology to identify a search term input by a user, there is a limit to the number of vocabularies that can be recognized in speech processing in real time, and the more words that can be recognized, It is known that it is very difficult to determine a user's voice input uniquely in a short time due to the current problem of recognition technology that the recognition accuracy decreases.
[0003]
Currently, the user waits until the recognition process is completed and the result is presented, and there is no measure other than confirming the correctness of the recognition result to the user, confirming in order of recognition score output by the recognition device, The candidate is repeatedly presented until the correct answer can be confirmed (for example, Non-Patent Document 1).
[0004]
[Non-Patent Document 1]
Kumiko Omori, Masanobu Higashida “Study on Improving the Determining Rate by Spoken Dialogue for Large Vocabulary Compound Nouns”, Proc. Of the 64th Annual Conference of Information Processing Society of Japan (2), March 2002 12th, p. 2-275,2-276
[0005]
[Problems to be solved by the invention]
The conventional system that uses a speech recognition device for user speech input and identifies the input requires more time for the recognition process and the reduction in accuracy is unavoidable as the search term to be recognized is larger. . For users, waiting time and repetition of candidate presentations lead to stress, and user satisfaction cannot be obtained when compared to operator-compatible systems, and you want to use current voice response services such as voice portals repeatedly. The current situation is that the user does not think.
[0006]
Therefore, as a technique to identify input in a short and natural flow of conversation without giving stress to the user, if the search word is composed of compound words of nouns, use the frequency of the constituent nouns. Aiming to reduce waiting time and improve accuracy by creating a search auxiliary database, using the created search auxiliary database, predefining frequently used constituent nouns as priority recognition targets, and narrowing down the recognition targets in advance. Conceivable.
[0007]
However, if there is a bias in the frequency of use of each noun that constitutes one search word, or if there are many very similar constituent nouns that are used in the same constituent order, the corresponding constituent noun is It may not be registered in the priority recognition target noun list (a list of nouns that are frequently used and nominated as a priority recognition target) included in the auxiliary search database. This leads to the result that there is no search term to be found.
[0008]
In addition, even if the corresponding constituent noun is registered in the above priority recognition target noun list, many similar words are output as recognition results, so that the correct constituent words do not appear in the output result and are actually present. As a result of the search word list creation process, it leads to a result that there is no actual search word.
[0009]
In addition, single nouns (single nouns that make up compound nouns) that can be extracted in part and that are considered to be useful information are used as additional information. A partial match search word process that extracts a search term including a single noun with a high value from the search database is performed, and in parallel with this, the time of the partial match search word process is used, and the recognition process is not completed. Forward search auxiliary database (search auxiliary database created for each single noun group that recognizes the nth single noun from the front of multiple single nouns constituting the compound noun that is the search term) and backward search auxiliary database (search Continues recognition processing for each single noun, with the search target database (recognized for each single noun group recognizing the nth single noun after the multiple nouns that make up the compound noun as words) And Each time that the presentation to use business, it is conceivable to repeat the process of merging the output at that time.
[0010]
In the voice input type search using the voice recognition device, only the voice recognition technology is used and the voice input by the user is confirmed. In addition to presenting and confirming the output result of the recognition device to the user There is no definitive method, the larger the number of words to be recognized and the more similar objects, the longer the recognition processing time and the lower the recognition accuracy, thus causing the user to wait. Moreover, it leads to the repetition of the presentation confirmation by misrecognition.
[0011]
Currently, in speech input type search using speech recognition devices, in order to obtain user satisfaction, it can be applied to limited fields and regions, and types of information that have a hierarchical structure. It is difficult to obtain speed and accuracy that are acceptable for users unless the number of recognition targets is reduced by grouping and narrowing down in order, with certainty in order starting from the lowest number It is.
[0012]
Furthermore, the ambient noise level is also very difficult to recognize by the input of an unspecified number of speakers from various places. In such a case, even in the word recognition, which is said to be the easiest, it is a practical level. It is said that the maximum number of words that can be handled in is hundreds.
[0013]
In the present invention, when a search term input by a user is searched and confirmed from a search database regardless of how large the search target is, a search is quickly and naturally performed like an operator-compatible search device. It is an object of the present invention to provide a speech input type compound noun search apparatus and a speech input type compound noun search method capable of realizing processing.
[0014]
[Means for Solving the Problems]
The present invention has a recognition likelihood that exceeds the threshold when the recognition likelihood of only some of the single nouns exceeds a predetermined threshold among a plurality of single nouns constituting a search term input by a user. A partial match search word candidate extraction means for extracting partial match search word candidates from the search database, the partial match search word candidates having the same number of single nouns as the number of single nouns constituting the search word. Re-recognition means for re-recognizing the input speech for each single noun constituting the extracted partial match search word candidate, and re-recognition The speech interactive compound noun search device includes integrated likelihood calculating means for calculating the integrated likelihood by calculating the recognition likelihood as a result of re-recognition by the means by a predetermined calculation method.
[0015]
BEST MODE FOR CARRYING OUT THE INVENTION
FIG. 1 is a block diagram showing a spoken dialogue compound noun search apparatus 1 according to an embodiment of the present invention.
[0016]
The speech interactive compound noun search device 1 includes a speech input unit 2, a speech recognition unit 3, speech recognition software 3S, a speech recognition result output unit 4, a search word candidate list creation unit 5, and a dialog control unit. 6, an audio output unit 7, audio output software 7 </ b> S, and a system database 8.
[0017]
The search word candidate list creation unit 5 includes an actual search word list creation unit 51 and a partial match search word list creation unit 52.
[0018]
The dialogue control unit 6 includes a search word candidate selection dialogue unit 61 and a partial match search word candidate selection dialogue unit 62. The speech recognition unit 3 uses the system database 8 when performing speech recognition processing on the input speech.
[0019]
The system database 8 includes a search database 81, a search auxiliary database 82, and a YES / NO database 83.
[0020]
The search database 81 is a database in which compound nouns composed of a plurality of single nouns are registered as search words, and the search words are registered by being divided into single nouns.
[0021]
The search auxiliary database 82 calls a group of single nouns written in the n-th (n is an integer value) of each registered compound noun as a “n-th notation set of single nouns”, and A single noun set group of notation is registered in the order of the frequency, and the single nouns registered in the order of the frequency are grouped into a plurality of sub-groups in order from the highest frequency. A database in which a set group is formed.
[0022]
The voice recognition unit 3 uses a voice recognition device, and the voice output unit 7 uses a voice output device for voice output.
[0023]
FIG. 2 is a diagram illustrating an example of the overall image of the search database 81.
[0024]
FIG. 3 is a diagram illustrating an example of the overall image of the search assistance database 82.
[0025]
The search database 81 holds a number of search words that cannot be subjected to speech recognition processing within the real time of interactive processing, and each search word is recorded in a state of being divided into constituent nouns as shown in FIG. The search auxiliary database 82 calculates the frequency of use of the single nouns in the n-th notation group of nouns for all search words in the search database 81, and the single nouns are calculated in the calculated usage frequency order. Stored in a state of being arranged (ordered state).
[0026]
The search auxiliary database 82 includes a search auxiliary database for a single noun of the first notation to a search auxiliary database for a single noun search of the pth notation, where p is the maximum number of single nouns constituting the search word.
[0027]
The number of single nouns (constituent nouns) held in each search auxiliary database 82 is expected to be smaller than the total number of search words held in the search database 81, and cannot be processed within the interactive processing real time. Suppose that it is a large number.
[0028]
Each search auxiliary database is stored in a form in which a plurality of single nouns are subdivided in order from the single noun with the highest frequency of use for each number of words that can be processed within the real time of the interactive processing preset by the search device 1. Among a plurality of single nouns stored in the nth search auxiliary database, a group of single nouns having the highest use frequency is referred to as a “first sub-set group”.
[0029]
The speech interactive compound noun search device 1 sends the user's voice input via the voice input unit 2 to the voice recognition unit 3.
[0030]
The voice recognition unit 3 performs a recognition process on the user input voice using a voice recognition device. If the speech recognition unit 3 does not have a function of dividing a compound noun into single nouns (constitutive nouns) and outputting a recognition result for each constituent noun, the user is required to divide and input each single noun. To instruct. When the speech recognition unit 3 has a function of dividing a speech signal subjected to speech power into constituent noun units (single noun units) and outputting a recognition result for each single noun, the user wants to search Input a series of nouns by voice.
[0031]
In addition, the speech recognition unit 3 does not have a function for the speech recognition unit 3 to divide speech-powered speech signals into single noun units (constant noun units) and output a recognition result for each single noun. Then, the user Pc is instructed to input speech by dividing into single nouns (every word).
[0032]
The voice recognition unit 3 selects a database to be recognized from the system database 8 in accordance with the processing status of the search device 1. When a search term is input, the search auxiliary database 82 is referred to, and when a response to the correctness confirmation to the user is recognized, the YES / NO database 83 is referred to.
[0033]
The speech recognition result output unit 4 sends the recognition result to the search word candidate list creation unit 5.
[0034]
At this time, if there is a sub-set group for which the recognition process has not been completed yet in each of the search auxiliary databases 82, the speech recognition unit 32 restarts the recognition process for the sub-set group, and the speech recognition result output unit 43 repeats the process of outputting the recognition result and sending it to the search word candidate list creation unit 5.
[0035]
In the search word candidate list creation unit 5, the actual search word candidate list creation unit 51 selects a possible constituent noun candidate from the recognition results, creates a combination of possible constituent noun candidates, and refers to the search database 81. Then, an actual search word candidate list obtained by extracting actual search words is created, and the created real search word candidate list is sent to the dialogue control unit 6.
[0036]
When extracting a real search word, all combinations of influential constituent nouns are created, and the search database 81 is referenced to check whether or not it exists.
[0037]
In addition, each time a candidate for a prominent constituent noun is combined one by one, the search database 81 is used to check whether or not a search word having a connection with the relevant prominent constituent noun actually exists. It is also possible to check the search database 81 again by combining influential constituent nouns in other constituent orders only with respect to the possible constituent noun candidate combinations.
[0038]
Furthermore, as an order of combining the prominent constituent nouns, they may be combined in order from the top, or the combination may be selected so that the search time is minimized.
[0039]
While creating the actual search word list, the partial match search word list creation unit 52 sets a search word partially including the prominent constituent noun candidate selected from the recognition result as a “partial match search word candidate”. Extracted from the search database 81, “partial match likelihood” is calculated, and recognition processing is performed for a partial match search word having a partial match likelihood equal to or greater than a predetermined threshold.
[0040]
When the search term is divided into single nouns and the user inputs a voice, only the search words having the same number of constituent nouns are selected from the search database 81, and the first input voice for the entire partially matched search word is selected. At the same time as using the recognition process, a list of nouns to be recognized is created only for the places where powerful constituent nouns cannot be extracted. To process. That is, speech recognition is performed again for each single noun.
[0041]
When the speech recognition apparatus has a function of dividing and recognizing a speech signal input for each single noun and recognizing the speech signal, the speech signal input by the user is recognized using a partial match search word candidate. The result of this recognition processing is sent to the dialogue control unit 6.
[0042]
The dialogue control unit 6 executes a detection candidate selection dialogue by using the actual search word candidates and the partial match search word candidates. At this time, the partial match likelihood is calculated for each search word candidate using the result for the partial match search word sent from the partial match search word list creation unit 52.
[0043]
When recognition processing is performed for the entire partial match search word, the calculated recognition likelihood may be used as the partial match likelihood without using the partial match likelihood so far. It is also possible to obtain the partial match likelihood by adding the sum of the recognition likelihoods for each leading constituent noun and the calculated recognition likelihood and dividing by “number of partially matching constituent nouns + 1”. Alternatively, the partial match likelihood may be obtained simply by taking a sum or product.
[0044]
In addition, when a single noun is recognized only for a portion where no powerful constituent noun is output, it is added or multiplied by the likelihood of each powerful constituent noun calculated in advance and divided by the number of constituent nouns. Alternatively, the products of the respective likelihoods may be calculated (the respective recognition likelihoods are multiplied with each other), and this may be divided by the number of constituent nouns.
[0045]
If the search word candidate in the actual search word candidate list satisfies the specified conditions of the search device that can specify the search word only by the confirmation process with the user, the confirmation process guidance output command is sent to the voice output unit 7. send.
[0046]
If the search word candidate in the actual search word candidate list does not satisfy the conditions for specifying the search word only by the confirmation process with the user, or if the actual search word candidate is not extracted, the partial match search word processing As a result, when the calculated new partial match likelihood satisfies the specified conditions of the search device that can specify the search word only by the confirmation process with the user, an output instruction of the confirmation process guidance is sent to the voice output unit 7. .
[0047]
If the search word candidate that can be specified only by the confirmation process is not extracted from the actual search word candidate or the partial match search word process using the partial match search word candidate, the recognition process is terminated at this point.
[0048]
Using the powerful constituent noun candidates obtained as a result of the recognition process for the next highest subset set, update the actual search word candidate list and the partial match search word list, and the search word candidate selection dialogue, The partial match search word processing is repeated.
[0049]
When the confirmation processing guidance output command is sent, the voice output unit 7 outputs a guidance for performing correct / incorrect confirmation between the search word candidate and the designated candidate to the user.
[0050]
When a response from the user to the confirmation processing guidance is input again from the voice input unit 2, the voice recognition unit 3 refers to the YES / NO database 83 to recognize the user's response, and the voice recognition result When the output unit 4 outputs the recognition result and a response indicating affirmation is obtained from the user, the dialogue control unit 6 sends a command for guiding the user to the effect that the search term specification is completed, to the voice output unit 7. send.
[0051]
The search word candidate list is updated, the search word candidate selection dialog, and the partial match search word processing are repeated until a search word can be specified.
[0052]
When the recognition process is completed for all single noun sets in the auxiliary search database 82 and the actual search word candidate list cannot be updated any more, the dialogue control unit 6 uses the single noun unit (constituent noun unit). The user is requested to input again, and based on the input recognition result of the single nouns in the composition order, the search term including the corresponding constituent noun is extracted from the search database 81 and the speech recognition is repeated again.
[0053]
While the actual search word candidate list is updated and the search word candidate selection dialogue is performed, the search device 1 is configured to include at least one set of constituent noun sets other than the priority recognition target noun group (one of a set of specified numbers). Min), the number of constituent nouns is defined so that the recognition process is completed.
[0054]
Next, the operation of the above embodiment will be described more specifically.
[0055]
In the following description, a case where a corporate company name included in a telephone book composed of compound nouns is searched will be described as an example.
[0056]
There are 22 million corporate names in the phone book across Japan. It is known that if the current speech recognition technology is applied as it is, it is impossible to recognize and process 22 million cases within the real time of interactive processing, and the accuracy is very low.
[0057]
FIG. 4 shows that 22 million corporate name data registered in the telephone directory are delimited by single nouns (constituent nouns) constituting the data, and corporate name data delimited by the single nouns is the search database 81. It is a figure which shows the example registered into.
[0058]
For the corporate names registered in the search database 81, the frequency of each single noun constituting the corporate name is checked for each single noun and registered in the search auxiliary database 82 in descending order of frequency.
[0059]
Among compound nouns constituting 22 million corporate names, when the maximum number of constituent words is 7, the total number of single nouns in the first notation is about 3.6 million, and the number of single nouns in the second notation The total number is about 2.5 million, the total number of single nouns in the third notation is about 2.7 million, the total number of single nouns in the fourth notation is about 1 million, and so on. Within the time, recognition processing is impossible even for each single noun, and it is expected that the accuracy is low.
[0060]
FIG. 5 is a diagram showing an example of the search assistance database 82 in the above embodiment.
[0061]
The search device 1 arranges the single nouns in the search auxiliary database 82 of each configuration order in order of frequency of use for each single noun, and subdivides (sorts) the q nouns in descending order of frequency of use. Construct a single noun set group (1 ≦ n ≦ 7).
[0062]
The above q is defined according to the performance of the speech recognition apparatus to be used, and in this embodiment, it is assumed that q = 500. Then, a group of 500 single nouns having the highest use frequency in the n-th notation single noun set group is defined as a “first sub-set group”.
[0063]
Further, the recognition device used in the present embodiment has a function of dividing an input single noun, performing recognition processing on a single noun basis, calculating a recognition likelihood for a single noun for each configuration order, and outputting it. The case is not assumed. Therefore, the user is requested to speak the corporate name (compound noun), which is the purpose of the search, divided into single nouns.
[0064]
In addition, this embodiment is applied to the case where the input single noun is divided, subjected to recognition processing in units of single nouns, and has a function of calculating and outputting the recognition likelihood of single nouns for each configuration order. May be.
[0065]
Here, a case where the user Pc inputs a voice using “Tokyo / Wholesale / Market” as a search term will be considered.
[0066]
“Tokyo” inputted by voice is recognized from single nouns belonging to the first sub-set group Gc1-1 of the first noun set group Gc1 of the first notation, and “wholesale” inputted by voice is indicated by the second notation. Recognized from the single nouns belonging to the first sub-set group Gc2-1 of the single noun set group Gc2, and for the “market” inputted by speech, the first sub-set group Gc3- of the third-stated single noun set group Gc3 Recognize from single nouns belonging to 1.
[0067]
Each of the forward search auxiliary database 82 and the backward search auxiliary database 82 is subjected to recognition processing, and the results of the recognition processing are merged. Then, the recognition results are output in descending order of the recognition score output by the recognition device.
[0068]
FIG. 6 is a diagram showing a state in which single nouns are recognized using the forward search auxiliary database 82 and the backward search auxiliary database 82 and the results of recognition are merged in the above embodiment.
[0069]
In the merge result shown in FIG. 6, “Tokyo”, which is the first noun single noun, is included in the first sub-set group Gc1-1 in the first noun single noun set group Gc1. “Market” as a noun is included in the first sub-set group Gc3-1 in the single noun set group Gc3 of the third notation, but “wholesale” as the single noun of the second notation is the second It is not included in the first sub-set group Gc2-1 in the single noun set group Gc2.
[0070]
In this embodiment, a single noun that outputs a recognition likelihood of 80 or more is defined as a “possible influential noun candidate”, and in FIG. 6, influential constituent noun candidates are indicated by halftone dots.
[0071]
In the merge result shown in FIG. 6, two candidates “Tokyo” and “Tokyu” are selected as probable constituent noun candidates for “Tokyo” inputted by voice, and the prominent constituent noun candidate is inputted for “wholesale” inputted by voice. For the “market” that is not selected and is voice-inputted, one candidate “estate” is selected as a potential constituent noun candidate.
[0072]
Based on the merge result shown in FIG. 6, the actual search word candidate list creation process is performed. However, for “wholesale” inputted by voice, since no prominent constituent noun candidate is selected, there is no actual search word, Therefore, as soon as the recognition process for the first sub-set group of each noun single noun set group is completed, the search device 1 sets the next most frequently used set group among the remaining set groups as recognition targets ( The second sub-set group (a set group made up of 500 single nouns) is recognized. Then, after this recognition process, a partial match search process is performed.
[0073]
A search word including the leading single noun candidate constituent nouns “Tokyo” and “Tokyu” and a search word including the third single noun leading constituent noun candidate “terrestrial” are extracted from the search database 81. To do.
[0074]
In the present embodiment, the sum of the recognition likelihoods of the leading constituent noun candidates is set as the partial match likelihood.
[0075]
FIG. 7 is a diagram illustrating a result of the partial match search word process and an example of a search word candidate selection dialogue using the result in the above embodiment.
[0076]
FIG. 8 is a diagram showing the relationship between the partial match likelihood and the partial match search word candidate in the above embodiment.
[0077]
As shown in FIG. 8, the partial match likelihood of a search term including “Tokyo” as the single noun of the first notation and including “the estate” as the single noun of the third notation is maximized. The entire partial match search word extracted as the partial match search word processing is set as a recognition target, and the voice of the user spoken first is recognized again for each single noun.
[0078]
Thus, since speech recognition is performed again for each single noun, the recognition likelihood shown in FIG. 7 and the recognition likelihood shown in FIG. 8 are different for the single noun “Tokyo”. Similarly, the recognition likelihoods shown in FIG. 7 and the recognition likelihoods shown in FIG. 8 are different for each of the single nouns “Tokyu”, “Place”, and “Market”. Moreover, since voice recognition is performed for each single noun, recognition accuracy is increased.
[0079]
Create a partial match search word extracted at the same time, and create a list of single nouns in the first notation, a list of single nouns in the second notation, and a list of single nouns in the third notation, respectively. Re-recognize single nouns uttered by, and fit the recognition results output for each order into the single nouns of the second notation that could not extract potential constituent noun candidates for partial match search words. The degree is calculated again, and only the actual candidates are left as shown in FIG.
[0080]
In this case, when calculating the partial match search word likelihood, the partial match search word likelihood may be calculated by taking the sum of the recognition likelihood of each single noun, or by taking the product. The partial match search word likelihood may be calculated, and the partial match search word likelihood is calculated by dividing the sum of the recognition likelihood of each single noun by the number of single nouns constituting the compound noun. You may make it do.
[0081]
In this embodiment, the partial match search word likelihood is calculated by taking the sum of the recognition likelihood of each single noun, and the recognition result for the entire partial match search word for which the recognition process has been completed at this point And the result of the real search word candidate creation process using the recognition result of the nth sub-set group having the next highest frequency of use, which is considered to have been recognized at this point. Merge.
[0082]
As a result of the merging, as shown in FIG. 8 (2), the recognition likelihood for “Tokyo / Wholesale / Market” exceeds the predetermined threshold 240, so “Tokyo / Wholesale / Market” remains as a real candidate.
[0083]
Further, in the present embodiment, since the telephone directory database is used as the search database 81, if the address is known in advance, it is possible to specify the corporate company name or the individual that actually exists at the address.
[0084]
In this embodiment, focusing on the fact that the search word is a compound noun, the frequency of use is checked for each single noun, and recognition processing is performed for each single noun, so the search target word is a large vocabulary that cannot be processed in real time. In addition, it is possible to perform a search process quickly and accurately for the user.
[0085]
In the present embodiment, when there is a bias in the frequency of single nouns constituting a search word, even when there are very similar single nouns, a plurality of nouns that constitute a speech-input compound noun with respect to single noun recognition results Among the simple nouns, a search word in which a partial single noun is a probable constituent noun candidate is extracted from the search database 81 as a partial match search word candidate, and the extracted partial match search word candidate is used as a target. Since speech recognition is performed for each noun, it is possible to quickly and accurately narrow down potential candidates from large vocabulary search words that are all subject to speech recognition and cannot be recognized correctly in real time.
[0086]
Thus, it is considered that the search process can be performed without causing the user to wait.
[0087]
In the above embodiment, when the search database is a personal first and last name such as a telephone directory database, the last name and the first name are the constituent units of the compound noun, and the search auxiliary database has the first notation. The last name may be registered by registering the last name as a noun and registering the name as a single noun of the second notation.
[0088]
In addition, when the recognition processing for all of the constituent word set groups in the search auxiliary database is completed, each of the actual search word candidate list and the partial match search word candidate list cannot be updated any more, and the candidate cannot be specified, A search term may be specified by specifically requesting the user to input the first constituent noun or the last constituent noun, and acquiring and combining information input in response to the input request.
[0089]
In other words, in the above embodiment, a compound noun composed of a plurality of single nouns is registered as a search word, and the search word is registered by being divided into each single noun, and the above registered A group of single nouns written in the nth (n is an integer value) of each compound noun is called an nth single noun set group, and the nth single noun set group has a high frequency. A search auxiliary database in which single nouns that are registered in order and that are registered in order from the highest frequency are grouped into a predetermined number in order from the highest frequency, and a plurality of sub-set groups are formed, When the user inputs the search word by speech for each single noun, the n-th single noun of the compound noun is recognized by the n-th single noun set, and the n-th single noun set. The most frequent of the group A recognition result list creating means for creating a recognition result list by performing recognition processing within the range of the first sub-set group including a single noun and corresponding recognition likelihood, and a constituent noun that is the single noun subjected to the recognition processing A recognition result list in which pairs of candidates and recognition likelihoods for the constituent noun candidates are arranged in the order of recognition likelihood is created for each of the single nouns that are input by speech, and are listed in the recognition result list. Among the constituent noun candidates, the constituent noun candidate selecting means for selecting the constituent noun candidate having the recognition likelihood exceeding the predetermined first threshold as the dominant constituent noun candidate, and a plurality of constituent nouns constituting the search word If the recognition likelihood of only some of the single nouns exceeds the first predetermined threshold, the single noun having the recognition likelihood exceeding the first threshold is provided, and the search word is Composed of single nouns A partial match search word candidate extraction means for extracting partial match search word candidates having the same number of single nouns from the search database, and voice recognition for the extracted partial match search word candidates, For each single noun constituting the extracted partial match search word candidate, a re-recognizing means for re-recognizing the input speech, and a recognition likelihood as a result of re-recognizing the re-recognizing means, a predetermined calculation method This is an example of a speech interactive compound noun search device having an integrated likelihood calculating means for calculating an integrated likelihood.
[0090]
In the above embodiment, a compound noun composed of a plurality of single nouns is registered as a search word, and the search word is registered by being divided into each single noun, and the above registration is performed. A group of single nouns written in the nth (n is an integer value) of each compound noun is called an nth single noun set group, and the nth single noun set group has a high frequency. A search auxiliary database in which single nouns that are registered in order and that are registered in order from the highest frequency are grouped into a predetermined number in order from the highest frequency, and a plurality of sub-set groups are formed, When a user inputs a series of voices without separating compound nouns as search terms into single nouns, search term separation means for separating compound nouns as speech input words into single nouns and the use If the user is the above search term When the nouns are input in a series of speech, the n noun single noun of the compound noun is recognized by the n noun single noun set group, and most frequently among the n noun single noun set groups. A recognition result list creating means for creating a recognition result list by performing recognition processing within the range of the first sub-set group including a single noun having a high level, and corresponding to the recognition likelihood, and a configuration that is a single noun subjected to the recognition processing A recognition result list in which pairs of noun candidates and recognition likelihoods for the constituent noun candidates are arranged in the order of recognition likelihood is created for each of the single nouns that are input by speech and listed in the recognition result list. Among the constituent noun candidates, a constituent noun candidate selecting means for selecting a constituent noun candidate having a recognition likelihood exceeding a predetermined first threshold as a potential constituent noun candidate, and a plurality of constituent noun candidates constituting the search word If the recognition likelihood of only some of the single nouns exceeds the first predetermined threshold, the single noun having the recognition likelihood exceeding the first threshold is provided, and the search word The partial match search word candidate extraction means for extracting from the search database partial match search word candidates having the same number of single nouns as the number of single nouns constituting the target, and the extracted partial match search word candidates as targets And re-recognizing means for re-recognizing the input speech for each single noun constituting the extracted partial match search word candidate, and the likelihood of the result of re-recognizing by the re-recognizing means This is an example of a speech interactive compound noun search device having an integrated likelihood calculating means for calculating an integrated likelihood by calculating with a predetermined calculation method.
[0091]
【The invention's effect】
According to the present invention, when searching for and confirming a search word input by a user from a search database, such as a corporate company name or a four-character idiom in a dictionary, regardless of how large the search target is, an operator There is an effect that the search processing can be realized quickly and naturally like a corresponding search device.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a voice interactive compound noun search apparatus 1 according to an embodiment of the present invention.
FIG. 2 is a diagram showing an example of an overall image of a search database 81. FIG.
FIG. 3 is a diagram showing an example of an overall view of a search assistance database 82. FIG.
FIG. 4 shows 22 million corporate name data registered in the telephone book, separated by single nouns (constituent nouns) constituting the data, and the corporate name data delimited by the single nouns is the search database 81. It is a figure which shows the example registered into.
FIG. 5 is a diagram showing an example of a search auxiliary database 82 in the embodiment.
FIG. 6 is a diagram showing a state in which single nouns are recognized using the forward search auxiliary database 82 and the backward search auxiliary database 82 and the results of the recognition are merged in the embodiment.
FIG. 7 is a diagram showing a result of partial match search word processing and an example of a search word candidate selection dialogue using the result in the embodiment.
FIG. 8 is a diagram illustrating a relationship between a partial match likelihood and a partial match search word candidate in the embodiment.
[Explanation of symbols]
1 ... Spoken dialogue compound noun search device,
2 ... Voice input part,
3 ... voice recognition unit,
4 ... voice recognition result output unit,
5 ... Search word candidate list creation part,
6 ... Dialogue control unit,
7 ... Audio output unit,
8 ... System database,
81 ... Search database,
82 ... Search auxiliary database.

Claims (6)

A search database in which compound nouns composed of a plurality of single nouns are registered as search terms, and the search terms are registered by being divided into single nouns;
A group of single nouns described in the nth (n is an integer value) of each registered compound noun is referred to as an nth single noun set group, and the single noun set group of the nth notation is A search in which single nouns registered in descending order of frequency are grouped by a predetermined number in order from the highest frequency to form a plurality of sub-set groups. With an auxiliary database;
When the user inputs the search word by speech for each single noun, the n-th single noun of the compound noun is recognized by the n-th single noun set, and the n-th single noun set. A recognition result list creation means for creating a recognition result list by performing recognition processing within the range of the first subset group including the single noun having the highest frequency among the groups, and by associating the recognition likelihoods;
A recognition result list in which a combination of a constituent noun candidate that is the recognition-nominated single noun and a recognition likelihood for the constituent noun candidate is arranged in the order of recognition likelihood is obtained for each of the single nouns that are input by speech. Among the constituent noun candidates that are created and described in the recognition result list, the constituent noun candidates that select the constituent noun candidates that have the recognition likelihood exceeding the predetermined first threshold as the dominant constituent noun candidates are selected. Means;
When the recognition likelihood of only some of the single nouns constituting the search word exceeds the first predetermined threshold, the single noun having a recognition likelihood exceeding the first threshold is used. Partial match search word candidate extraction means for extracting from the search database partial match search word candidates having nouns and having the same number of single nouns as the number of single nouns constituting the search word;
Re-recognizing means for recognizing the extracted partial match search word candidate as a target and re-recognizing the input voice for each single noun constituting the extracted partial match search word candidate;
An integrated likelihood calculating means for calculating an integrated likelihood by calculating a recognition likelihood as a result of re-recognition by the re-recognizing means by a predetermined calculating method;
A spoken dialogue compound noun search device characterized by comprising: A search database in which compound nouns composed of a plurality of single nouns are registered as search terms, and the search terms are registered by being divided into single nouns;
A group of single nouns described in the nth (n is an integer value) of each registered compound noun is referred to as an nth single noun set group, and the single noun set group of the nth notation is A search in which single nouns registered in descending order of frequency are grouped by a predetermined number in order from the highest frequency to form a plurality of sub-set groups. With an auxiliary database;
A search term delimiting means for delimiting the compound noun that is the voice input search term for each single noun when the user inputs the compound noun that is the search term for each single noun, without separating the compound noun for each single noun;
When the user inputs a series of compound nouns that are the search terms, the n-th notation of the compound noun is recognized by the n-th notation set of nouns, and the n-th notation A recognition result list creation means for creating a recognition result list by performing recognition processing within the range of the first sub-set group including the single noun having the highest frequency among the single noun set group;
A recognition result list in which a combination of a constituent noun candidate that is the recognition-nominated single noun and a recognition likelihood for the constituent noun candidate is arranged in the order of recognition likelihood is obtained for each of the single nouns that are input by speech. Among the constituent noun candidates that are created and described in the recognition result list, the constituent noun candidates that select the constituent noun candidates that have the recognition likelihood exceeding the predetermined first threshold as the dominant constituent noun candidates are selected. Means;
When the likelihood of only some of the single nouns constituting the search word exceeds the first predetermined threshold, the single noun having the likelihood exceeding the first threshold is determined. A partial match search word candidate extraction unit that extracts partial match search word candidates having the same number of single nouns as the number of single nouns constituting the search word from the search database;
Re-recognizing means for recognizing the extracted partial match search word candidate as a target and re-recognizing the input voice for each single noun constituting the extracted partial match search word candidate;
An integrated likelihood calculating means for calculating an integrated likelihood by calculating a recognition likelihood as a result of re-recognition by the re-recognizing means by a predetermined calculating method;
A spoken dialogue compound noun search device characterized by comprising: In claim 1 or claim 2,
When the recognition processing for all of the constituent word set groups in the search auxiliary database is completed and each of the actual search word candidate list and the partial match search word candidate list cannot be updated any more and the candidate cannot be specified, Spoken dialogue type characterized by specifying a search term by specifically requesting a user to input a constituent noun or a constituent noun at the end, and acquiring and combining information input in response to the input request Compound noun search device. In claim 1 or claim 2,
If the search database is a personal first and last name, the last name and first name are the constituent units of the compound noun, and the last name is registered in the search auxiliary database as the first notation of the first notation. A compound noun search device, wherein an individual first name is confirmed by registering the name as a single noun. A compound noun composed of a plurality of single nouns is registered as a search word, and the search database in which the search word is divided into each single noun is registered, and the nth of each compound noun registered above A group of single nouns represented by (n is an integer value) is called an n-th group of single nouns, and the above-mentioned n-th group of single nouns is registered in descending order of frequency. Search for compound nouns using a search auxiliary database in which single nouns registered in descending order of frequency are grouped by number in order from the highest frequency and a plurality of sub-set groups are formed. In the way to
When the user inputs the search word by speech for each single noun, the n-th single noun of the compound noun is recognized by the n-th single noun set, and the n-th single noun set. A recognition result list creation stage in which recognition processing is performed within the range of the first subset group including a single noun having the highest frequency among the groups, and a recognition result list is created by associating the recognition likelihood;
A recognition result list in which a combination of a constituent noun candidate that is the recognition-nominated single noun and a recognition likelihood for the constituent noun candidate is arranged in the order of recognition likelihood is obtained for each of the single nouns that are input by speech. Among the constituent noun candidates that are created and described in the recognition result list, the constituent noun candidates that select the constituent noun candidates that have the recognition likelihood exceeding the predetermined first threshold as the dominant constituent noun candidates are selected. Stages;
When the recognition likelihood of only some of the single nouns constituting the search word exceeds the first predetermined threshold, the single noun having a recognition likelihood exceeding the first threshold is used. A partial match search word candidate extraction step of extracting from the search database partial match search word candidates having nouns and having the same number of single nouns as the number of single nouns constituting the search word;
A re-recognition step of recognizing the extracted partial match search word candidate as a target and re-recognizing the input voice for each single noun constituting the extracted partial match search word candidate;
An integrated likelihood calculating step of calculating an integrated likelihood by calculating the likelihood of the result of speech recognition in the re-recognition step by a predetermined calculation method;
A spoken dialogue type compound noun search method characterized by comprising: A compound noun composed of a plurality of single nouns is registered as a search word, and the search database in which the search word is divided into each single noun is registered, and the nth of each compound noun registered above A group of single nouns represented by (n is an integer value) is called an n-th group of single nouns, and the above-mentioned n-th group of single nouns is registered in descending order of frequency. Search for compound nouns using a search auxiliary database in which single nouns registered in descending order of frequency are grouped by number in order from the highest frequency and a plurality of sub-set groups are formed. In the way to
A search term separation step of separating the compound noun, which is the speech-inputted search word, for each single noun when the user inputs a series of voices without separating the compound noun, which is the search word, for each single noun;
When the user inputs a series of compound nouns that are the search terms, the n-th notation of the compound noun is recognized by the n-th notation set of nouns, and the n-th notation A recognition result list creation step of creating a recognition result list by performing recognition processing within the range of the first sub-set group including the single noun having the highest frequency among the single noun set groups;
A recognition result list in which a combination of a constituent noun candidate that is the recognition-nominated single noun and a recognition likelihood for the constituent noun candidate is arranged in the order of recognition likelihood is obtained for each of the single nouns that are input by speech. Among the constituent noun candidates that are created and described in the recognition result list, the constituent noun candidates that select the constituent noun candidates that have the recognition likelihood exceeding the predetermined first threshold as the dominant constituent noun candidates are selected. Stages;
The single noun having a Nishiki likelihood exceeding the first threshold when the likelihood of only some of the single nouns constituting the search word exceeds the first predetermined threshold A partial match search word candidate extraction stage that extracts partial match search word candidates having the same number of single nouns as the number of single nouns constituting the search word from the search database;
A re-recognition step of recognizing the extracted partial match search word candidate as a target and re-recognizing the input voice for each single noun constituting the extracted partial match search word candidate;
An integrated likelihood calculating step of calculating the integrated likelihood by calculating the recognition likelihood of the result of speech recognition in the re-recognition step by a predetermined calculation method;
A spoken dialogue type compound noun search method characterized by comprising:

Images