JPH1173419A - Method and device for retrieving electronic document - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a document retrieving device capable of highly accurately retrieving an electronic document suited to a retrieving purpose by a voice input. SOLUTION: Recognition tolerance and recognition time are applied to one or plural keyword candidates recognized from an input voice. An weighting information applying part 22 calculates retrieving weight in each keyword candidate from the recongition tolerance and time elapsed from the recognition time. A word merging part 24 merges retrieving weight values as to identical keyword candidates. A word determination part 25 specifies a keyword candidate having the highest retrieving weight as a retrieving keyword and outputs the specified candidate to a document management part 30. The management part 30 calculates retrieving weighted word importance in each electronic document based on the retrieving keyword and retrieves an electronic document having the highest retrieving weighted word importance from an electronic document DB 40.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for retrieving an electronic document based on recognition results of continuously input speech, characters, and the like, and a document retrieval apparatus suitable for implementing the method.

[0002]

2. Description of the Related Art A document retrieval apparatus for displaying one or a plurality of electronic documents related to a keyword inputted through a keyboard or a pen on a display or the like is known. Such a document search device includes a file search device using a personal computer and a WWW (WoW) on the Internet.
rld Wide Web). As shown in FIG. 10, this type of document search device includes an electronic document database (hereinafter, electronic document DB) 40 storing a plurality of electronic documents to be searched, and a keyword (hereinafter, search) used by an operator for search. A keyword input unit 60 that receives an input of a keyword), and a document management unit 70 that searches for a corresponding electronic document from the electronic document DB 40 based on the search keyword. Document management unit 7
0 is a search information table 31 functioning as an index file of each electronic document stored in the electronic document DB 40, and a new external electronic document is stored in the electronic document DB 40.
And a document input unit 33 for accumulating the data in a document.

[0003]

A conventional document retrieval apparatus 2
Had to use a keyboard or pen to enter search keywords, so if a user needed to search for an electronic document while performing a manual task, the original task would be interrupted. Would.

[0004] In order to improve this point, a search method has been proposed in which keyword information is input using a voice recognition technique, thereby eliminating the need for input work using a keyboard or the like. In this search method, sentence speech, that is, speech that reads a grammatically correct sentence, or speech with equivalent content is recognized, the content is analyzed linguistically, and the analysis result is used to inquire the electronic document DB 40. The document is searched by converting it into a search expression expressed in a language. In some cases, a method of detecting the same search keyword as a word registered in advance from continuously input speech is adopted.

However, identifying a search keyword by recognizing sentence speech requires accurate recognition of all words that may be uttered, and thus poses a problem in practicality. Further, when sentence speech is continuously input, a search keyword may be specified without correctly recognizing a word in the speech, and there is a problem that search accuracy cannot be improved.

SUMMARY OF THE INVENTION It is an object of the present invention to provide an improved search method for accurately recognizing a search keyword from continuously inputted keyword information and improving the search accuracy of an electronic document. Another object of the present invention is to provide a document search device suitable for implementing the above search method.

[0007]

A search method according to the present invention for solving the above-mentioned problems is characterized in that an electronic device such as a computer includes the following processing steps. (1) a process of recognizing a plurality of words from keyword information such as sentence speech, and (2) a search for each recognized word based on a recognition likelihood indicating the certainty of the recognition and an elapsed time from the recognition time. A process of adding weights and merging search weights between words when a plurality of identical words are recognized; (3) a process of specifying a word having a relatively large search weight as a search keyword; and (4) a search target. For each of the plurality of electronic documents to be determined, a search weighted word importance is determined based on the search weight of the specified search keyword and word importance indicating how important the search keyword is, and The process of identifying the electronic document with the highest search-weighted word importance.

[0008] The search weight increases in proportion to the recognition likelihood, and decreases as the elapsed time increases.

According to another aspect of the present invention, there is provided a document retrieval apparatus for retrieving a plurality of words from electronic document storage means for storing an electronic document to be retrieved and keyword information inputted at the time of retrieval. Word recognition means for adding and outputting a recognition likelihood and a recognition time representing the likelihood of recognition to each word, and keyword specifying means for specifying at least one of the recognized words as a search keyword, A search unit that searches for the corresponding electronic document from the electronic document storage unit using the specified search keyword, wherein the keyword specifying unit is configured to execute the search based on at least one of the recognition likelihood and the elapsed time from the recognition time. A search weight according to a search purpose is created, and a word having a relatively high created search weight is preferentially used as the search keyword. It is characterized in.

According to another document search apparatus of the present invention, in addition to the above-described electronic document storage means, word recognition means, and keyword specifying means, the search weight of the specified search keyword and how much the search keyword is for each electronic document. A search weighted word importance is calculated for each electronic document by integrating the word importance indicating whether the word is an important word, and an electronic document having a higher calculated search weighted word importance is retrieved from the electronic document storage means. And search means for issuing the search result.

In another document search device of the present invention, in each of the above document search devices, the keyword information is a sentence voice continuously uttered, and the word recognizing means sets a silent section of the sentence voice as a delimiter. It is characterized in that it is configured to perform speech recognition repeatedly. In this case, the word recognition unit is configured to list words included in the electronic document in advance, and perform speech recognition using the words listed.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment in which the present invention is applied to a continuous speech input type document search apparatus will be described below in detail. FIG. 1 is a configuration diagram of a document search device according to an embodiment of the present invention. The document search device 1 includes a word recognition unit 10 for recognizing a sentence voice that is continuously input, a keyword specification unit 20 for specifying a search keyword based on an output result of the word recognition unit 10, and a search unit based on the search keyword. A document management unit 30 for performing a search process of an electronic document, an electronic document DB 40 storing a large amount of electronic documents to be searched, and an output unit 50 for presenting a search result to a user. Components having the same functions as those of the conventional device shown in FIG. 10 are denoted by the same reference numerals.

The word recognition section 10 detects a silent section in the input speech and extracts a voiced section,
A word list 13 in which words representing each of the documents to be searched are listed in advance, and a speech recognition unit 12 that performs speech recognition using the output result of the preprocessing unit 11 and the word list 13 Is done. The voice recognition unit 12 performs voice recognition repeatedly with a silent section as a delimiter,
Words can be continuously recognized and output for a voice (an utterance sound such as an additional description, a detailed description, and a correction) that is input after the preceding input voice.

The keyword specifying unit 20 includes the word recognizing unit 10
Acquisition unit 2 receiving the processing result output from
1. Weight information assigning unit 2 for assigning a search weight to the processing result
2. A word storage unit 23 for storing words to which search weights have been assigned by the weight information assigning unit 22, a word merging unit 24 for merging words and their search weights, and comparing the search weights assigned to each word. Thus, a word specifying unit 25 for specifying one or a plurality of search keywords is provided.

The document management section 30 receives a specified search keyword and searches for an appropriate electronic document from the electronic document DB 40 by using a search information table 31 and a search weight which is information for specifying an electronic document to be searched. It has an importance determining unit 32 for processing the attached word importance and a document input unit 33 for inputting an electronic document, and outputs an electronic document retrieved by a search engine (not shown) to an output unit 50.

Next, the processing procedure of each unit will be described with reference to FIGS. FIG. 2 is a processing procedure diagram in the word recognition unit 10. Here, it is assumed that sentence voices having the contents shown in FIGS. 4A and 4B are continuously input by being separated by a pause. First, when the sentence voice having the content shown in FIG. 4A is input to the word recognition unit 10 (step S101: Ye)
s) The pre-processing unit 11 detects a silent section in the voice (step S102) and extracts a voiced section (step S103). Then, the feature of the voice in the extracted voiced sound section is extracted (step S104). Voice recognition unit 12
Then, the extracted individual voice features are collated with the word list 13 to recognize words that may be used as search keywords (hereinafter, keyword candidates), and based on the likelihood of each keyword candidate and the recognition time point. Is calculated (steps S105 and S106).

FIG. 5A shows an example of a keyword candidate recognized from the sentence voice having the content of FIG. 4A, its recognition likelihood, and elapsed time. Here, the words "snowboard", "ski area", "high speed", and "chuo-michi" are contained in the sentence voice, but the words "snowboard" and "high speed" were correctly recognized as keyword candidates. ”, The“ ski area ”is dropped off, and the“ volume ”is recognized by an insertion error. Also, as a result of misrecognition of "Chuo Expressway", "Chubu region" is obtained. The recognition likelihood is “0.9” for “snowboard”.
6 "," 0.64 "for" volume "," 0.70 "for" high speed ", and" 0.
72 ". The elapsed time is calculated based on" snowboard ".

The above processing is repeated for sentence voices having the contents shown in FIG. 4B (step S107: Yes). As a result, a recognition result as shown in FIG. 5B is obtained. FIG.
Referring to (b), all of the sentence voices input the second time are correctly recognized, and keyword candidates “Suwa Inter” and “Snowboard” are obtained. Here, the elapsed time is based on “snowboard” included in the first sentence voice.

When the recognition of the search keyword candidates is completed, the word recognition unit 10 adds the recognition likelihood and the recognition time to each of the keyword candidates, and adds the recognition time to the keyword identification unit 20.
(Step S108).

FIG. 3 is a processing procedure diagram in the keyword specifying unit 20. In the keyword specifying unit 20, the parameter likelihood unit 21 obtains the recognition likelihood and the elapsed time of each keyword candidate output from the word recognition unit 10 and inputs them to the weight information giving unit 22 (step S201).

The weight information assigning unit 22 calculates a search weight of each search keyword (step S202). That is, assuming that the recognition likelihood of each keyword candidate is L, a search weight SW based on the recognition likelihood L is calculated by multiplying the recognition likelihood L by a weight coefficient α determined in advance according to the search purpose. Further, based on the elapsed time given to each keyword candidate, weighting is performed so that the search weight decreases as the time elapsed from the recognition time to the present becomes longer. Here, not only the newly recognized keyword candidates but also the search weights of the previously recognized keyword candidates are weighted again. In this case, the weights are, for example, BW and AW for the search weights before and after weighting, and T
Then, it can be obtained by a calculation formula such as “AW = (β · BW) to the power of −γT (β and γ are weighting factors)” or “AW = β · BW / γT”.

FIG. 6A shows the search weight of each keyword candidate recognized from the first sentence voice, and FIG. 6B shows the search weight of each keyword candidate recognized from the second sentence voice. FIG. FIG. 7 is a diagram illustrating a result of re-weighting the search weight of the keyword candidate recognized in the past. As is apparent from FIG. 7, the search weights of the four keyword candidates recognized from the first sentence voice have a predetermined time since the keyword candidates are recognized from the second sentence voice. 6 (a) is uniformly smaller than the search weight.

The word merging unit 24 checks whether or not the same keyword candidate as the newly recognized keyword candidate exists in the past recognized keyword candidates, and if it exists (step S203: Yes), The keyword weights are merged with each other, and the search weight is recalculated (step S2).
04, S205). That is, the search weight is further increased. In this example, since “snowboard” is commonly recognized, it is merged. The search weight after merging is determined, for example, by multiplying the sum of the search weights of the same keyword candidate by a predetermined weight coefficient. In the case of this example, if the search weight of “snowboard” first recognized from the sentence speech is W1, and the search weight of the second “snowboard” is W2, the search weight MW after merging is
It can be obtained from the calculation formula of “δ (W1 + W2) (where δ is a weight coefficient)”. When the recalculation of the search weight is completed, or when there is no identical keyword candidate in step S203 (step S203: No), each keyword candidate to which the search weight is assigned is stored in the word storage unit 2.
It is stored in step 3 (step S206).

FIG. 8 shows keyword candidates stored in the case of this example and their search weights. The word specifying unit 23 evaluates the search weight of each of the stored keyword candidates (step S207), specifies a keyword candidate having a relatively large search weight as a search keyword, and outputs the keyword candidate to the document management unit 30 (step S208). ). The search keyword specified in this way is the one that is estimated to be most suitable for the search purpose.

For example, in this example, "snowboard"
Both are recognized at the time of two voice inputs, and it is presumed that the user is particularly interested. In this case, the search weight is increased to “1.57” from the first value “0.92” by the above-described merge processing for the second time. On the other hand, the "volume" and "Chubu district", which were incorrectly recognized by the word recognition unit 10, had the recognition likelihoods "0.62" and "0.72" at the time of the first speech input, but the second time Since it was not recognized at the time of voice input, the search weight decreased from “0.62” to “0.56” for “volume” and from “0.72” to “0.68” for “Chubu” It works to reduce the negative impact on search results caused by incorrect word recognition. In this way, by using the elapsed time from the recognition time and the results of multiple recognitions as parameters, the search weight is automatically learned for the search purpose, and all keyword candidates can be accurately recognized. If not, it becomes easy to achieve the required search purpose.

After all, in this example, as shown in FIG.
Since the search weight of “Snowboard” is “1.57”, which is the relatively highest search weight, this “Snowboard” is output to the document management unit 30 as a search keyword.

FIG. 4 is a processing procedure diagram in the document management unit 30. In the document management unit 30, the keyword specifying unit 20
"Snowboard" identified as a search keyword
(Step S301), and this "snowboard"
Search for electronic documents that match. At this time, the importance determining unit 32 obtains a search-weighted word importance for each of the plurality of electronic documents to be searched.

That is, the first electronic document is extracted from the electronic document DB 40 through the search information table 31 (step S302), and the word importance indicating how important the search keyword is to the electronic document is calculated (step S303). ). The word importance can be determined by the frequency of appearance of words in the electronic document. After that, the word importance and the keyword specifying unit 20
Is added to the search weight of the search keyword assigned in the step (a) to obtain a search-weighted word importance (step S304). This search weighted word importance WWI is
Assuming that the word importance of the search keyword is WI and the search weight given to the search keyword is SW, for example, “e (WI
SW) ”. Here, “e” is a weight coefficient.

After this processing is repeated for all the electronic documents to be searched (step S305), each search-weighted word importance is evaluated (step S30).
6). Then, an electronic document having the highest search-weighted word importance is searched for and output to the output unit 50 (step S307). As described above, by performing the search processing based on the search-weighted word importance, it is possible to easily search for an electronic document with a high degree of attainment of the purpose.

As described above, the document search device 1 of the present embodiment is an electronic document DB storing a large amount of electronic documents.
From 40, the user automatically recognizes the sentence voice uttered continuously as if speaking to a human to explain the necessary electronic document and automatically searches for the target electronic document. It is not necessary to consider an effective search keyword and input it from a keyboard or the like, and the search can be performed while performing other operations.

Further, according to the contents of the input sentence voice,
Since the search of the document and the presentation of the search result are performed every moment, when the search result is disappointing and the search is newly started, the past recognition result can be obtained simply by continuously inputting the explanation voice of the necessary information. You will be able to get reflected search results.

Further, it is possible to use a dialogue between humans as an input to the document search apparatus 1 to search for an electronic document related to the contents of the dialogue and automatically present it to the other party. This can be effective in facilitating the dialogue. Such a use form can be applied to automatic help output in a face-to-face or online consultation operation using a voice input / output device.

In the present embodiment, an example has been described in which a plurality of keyword candidates are recognized from continuously input sentence voices. However, the present invention uses not only sentence voices but also ordinary character recognition and pattern recognition. The same applies to the case of specifying keyword candidates.

[0034]

As is apparent from the above description, according to the present invention, even if the first keyword information cannot be accurately recognized, only the subsequent related keyword information is merely input. Since a search keyword effective for achieving the search purpose is accurately specified, there is an effect that the search operation of the electronic document becomes easy.

In addition, for each of a plurality of electronic documents to be searched, the importance of a search-weighted word is evaluated, and the document having the highest evaluation result is searched for, so that the search accuracy of the electronic document is improved. There is.

Further, since the sentence voice continuously uttered can be used as the keyword information in the document search apparatus of the present invention, the application scene can be used by a service system using a voice input / output device or a communication means. It can be extended to network services. Furthermore, since the speech recognition is repeatedly performed using the word list at intervals of the silent section of the sentence speech, there is also an effect that detection errors are suppressed.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of a voice input type document search device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a processing procedure of a word recognition unit in the embodiment.

FIG. 3 is an explanatory diagram of a processing procedure of a keyword specifying unit in the embodiment.

FIG. 4 is a processing procedure diagram of a document management unit in the embodiment.

FIG. 5A is an explanatory diagram showing the content of a first input voice, and FIG. 5B is a diagram showing the content of a second input voice;

FIG. 6A is a chart showing a recognition result of a first input voice, and FIG. 6B is a chart showing a recognition result of a second input voice;

FIG. 7 is a chart showing search weights of the first recognition result (past keyword candidates) recalculated when calculating the search weight for the second recognition result.

FIG. 8 is a table showing keyword candidates stored in a word storage unit and their search weights.

FIG. 9 is a processing procedure diagram of a document management unit in the embodiment.

FIG. 10 is a functional block diagram of a conventional document search device.

[Explanation of symbols]

 1 Speech input type document search device according to the present embodiment 2 Conventional document search device 10 Word recognition unit 11 Preprocessing unit 12 Speech recognition unit 13 Word list 20 Keyword specification unit 21 Parameter acquisition unit 22 Weight information addition unit 23 Word storage unit 24 Word merge unit 25 Word determination unit 30, 70 Document management unit 31 Search information table 32 Importance determination unit 33 Document input unit 40 Electronic document database (DB) 50 Output unit 60 Keyword input unit

──────────────────────────────────────────────────の Continued on the front page (51) Int.Cl. ⁶ Identification code FIG06F 15/40 370A 15/403 310Z

Claims (10)

[Claims] 1. A process of recognizing a plurality of words from keyword information, and adding a search weight to each recognized word based on a recognition likelihood indicating a certainty of the recognition and an elapsed time from a recognition time. A process of merging search weights between words when the same word is recognized a plurality of times; a process of specifying a word whose search weight is relatively large as a search keyword; With respect to the search weight of the specified search keyword and the word importance indicating how important the search keyword is, a word weight with a search weight is determined. The process of identifying high electronic documents;
How to search for electronic documents containing. 2. The search method according to claim 1, wherein the search weight increases in proportion to the recognition likelihood. 3. The search method according to claim 1, wherein the search weight decreases as the elapsed time increases. 4. The retrieval method according to claim 1, wherein the keyword information is sentence voice continuously uttered by a user. 5. An electronic document storing means for storing an electronic document to be searched, a plurality of words are recognized from keyword information input at the time of searching, and a recognition likelihood and a recognition indicating a certainty of recognition of each word. A word recognizing means for adding a time and outputting; a keyword specifying means for specifying at least one of the plurality of recognized words as a search keyword; and a corresponding one of the electronic document storage means using the specified search keyword. A search unit for searching for an electronic document, wherein the keyword specifying unit creates a search weight according to a search purpose based on at least one of the recognition likelihood and an elapsed time from a recognition time, and the created search weight A document search apparatus that is configured to preferentially use a word having a relatively large as the search keyword. 6. A parameter acquisition unit for acquiring the recognition likelihood and the recognition time from the word recognition unit for each word, and at least one of the acquired recognition likelihood and an elapsed time from a recognition time. A weight information assigning unit that assigns a search weight to each word based on one of the words, a word identifying unit that identifies a predetermined number of words having a relatively large search weight, and a word that stores the word assigned the search weight A storage unit, a value obtained by multiplying a sum of search weights given to each word by a predetermined coefficient when the same word as the word recognized with the input of new keyword information is present in the word storage unit; The document search device according to claim 5, further comprising: a word merging unit that sets a new search weight of the word. 7. The method according to claim 1, wherein the weight information assigning unit, when assigning the search weight to the newly recognized word, re-assigns the search weight to the search weight of the previously recognized word. 7. The document search device according to claim 6, wherein: 8. An electronic document storage means for storing an electronic document to be searched, and a plurality of words are recognized from keyword information inputted at the time of search, and a recognition likelihood indicating a certainty of recognition is added to each word. Word recognition means for outputting as a search keyword a keyword specifying means for specifying at least one of the plurality of recognized words as a search keyword; and a search weight of the specified search keyword and a description of the search keyword for each electronic document. Only the word importance representing whether the word is important or not is integrated to calculate a search weighted word importance for each electronic document, and the calculated search weighted word importance is used to obtain a higher electronic document from the electronic document storage means. A document search device comprising: a search means for searching; 9. The method according to claim 8, wherein the keyword information is sentence speech uttered continuously, and the word recognizing means is configured to perform speech recognition repeatedly with a silent section of the sentence speech as a delimiter. 9. The document search device according to claim 5, wherein 10. The apparatus according to claim 1, wherein the word recognition unit lists words included in the electronic document in advance, and performs voice recognition using the words listed. Item 9. The document search device according to Item 9.