JP2010123005A - Document data retrieval device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain a document (lyrics) data retrieval device for retrieving a target content and document by discriminating similarities at a high speed even when a retrieval character string that does not coincide with a retrieval original is input in a text document retrieval system for retrieving a large amount of text documents. <P>SOLUTION: Words (retrieval words) having the possibility to be retrieved are predicted and registered, and sound similar distance values to retrieval target words in a lyrics file are calculated in advance. By creating an index file 22 in which significance information is calculated by using a keyword (a retrieval word having the possibility to be included in a retrieval word string) which is acoustically similar to a word having the possibility to be retrieved in the lyrics file, a sound similar distance value between two words, and the positional information of the retrieval target word in the lyrics file as parameters, the acceleration of similarity calculation in a similar text retrieval is achieved. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a search device used in a content search system based on content-related terms and a text search system for searching a large amount of document data (text document), and a search word that does not match an original text existing in the document data. The present invention also relates to a document data search apparatus that enables a target content and document data to be searched by determining similarity at high speed even when input is made.

  In a search device for text, a complete match search of keywords is performed. However, in order to improve usability, for example, “interface” and “interface” are judged to match even if they do not match completely. It is required to perform “fuzzy search”.

  For example, in the similar text search device described in Patent Document 1, N-gram elements (N character sequences) for text are created, and the degree of coincidence of N-gram elements for a plurality of texts is calculated. In the calculation results, high-speed similar text search is realized by outputting search target candidates in descending order of coincidence.

Further, in the speech search system described in Patent Document 2, when searching for a word or phrase in speech data that has been speech-recognized, it is possible to cope with a case where a character string to be searched is erroneously recognized or an unknown word. Therefore, the input search character string is converted into a phoneme string, and a similar word or similar word that can appear in the speech recognition result of the speech data to be searched using a search algorithm used in continuous word speech recognition Searching is performed after expanding the column.
For example, if you input “Harry Potter” as a search word (character string), the speech data recognition results will pick up word candidates such as “Harry”, “Poster”, “Ha”, “Reporter”, etc. Is done. At this time, by arranging the word candidates, it expands into a word string such as “Harry” + “poster”, “ha” + “reporter”, which is acoustically close to the search character string “Harry Potter”. Since this expansion is performed so that the acoustic distance is close to the search character string, a word string that is highly likely to be a recognition result of the search character string is obtained including erroneous recognition.
JP 2003-288366 A JP 200631278 A

In a document data search device that searches for music using a part of lyrics as a search keyword, when searching with lyrics that the user has misunderstood and remembered (for example, the character string existing in the original text of the lyrics is “innumerable light”) In contrast, assume that the search character string is erroneously entered as “straight light” and is searched).
In this case, as in the similar text device of Patent Document 1, even if calculation of the edit distance of the character string and indexing of the N-gram are performed, the edit distance is different between “numberless” and “straight”. It is impossible to search for information (lyric text) of the target music.

  In addition, with respect to the device of Patent Document 2, it is necessary to calculate an acoustic distance between a search character string and a word candidate to be expanded in order to expand the search character string into a word string each time a search is performed. Especially when the number of search query strings is large (two or more non-consecutive lyrics keywords) and there are many lyrics files to be searched (more than tens of thousands for commercial search systems), online calculation time is required. The search was very slow.

  The present invention has been proposed in view of the above circumstances, and in a text document search system for searching a large number of text documents, even when a search character string that does not match the original search text is input, the similarity is determined at high speed. It is an object of the present invention to provide a document data retrieval apparatus that enables retrieval of target contents and documents.

In order to achieve the above object, the invention of claim 1 is a search device comprising an input interface through which a search character string is input, and searching for document data from a plurality of text data to be searched using the search character string. It is characterized by including the structure of.
Word extraction means. This word extracting means extracts a word constituting the document data by inputting a plurality of sentence data in advance.
Keyword registration means. The keyword registration means registers search words that may be included in the search character string as keywords in the list.
INDEX file creation means. The INDEX file creation means creates a search INDEX file by using the extracted words extracted by the word extraction means and the acoustic similarity distance value calculated from the comparison between the reading kana or phonemes decomposed for the keywords. Is.
Similarity calculation means. The similarity calculation means calculates the similarity of the search target document data based on the search word input with reference to the INDEX file, based on the search word obtained from the input search character string.
Output interface. This output interface outputs a search result of document data based on the similarity.

The invention of claim 2 is the document data retrieval apparatus of claim 1,
The acoustic similarity distance value of the INDEX file creation means is
A distance between syllables (between different syllables) can be obtained by performing DP matching in units of syllables in the two reading kana strings by converting them into reading kana strings by means of calculating the editing distance between the keyword and the extracted words. ), A penalty value at the time of syllable insertion, and a penalty value at the time of syllable drop are calculated by calculation means for giving desired numerical values.

The invention of claim 3 is the document data search device of claim 1,
The acoustic similarity distance value of the INDEX file creation means is
By converting the reading of each word into a phoneme string by means of calculating the edit distance of the two phoneme strings between the keyword and the extracted word, and performing DP matching in units of phonemes in the two phoneme strings, It is characterized in that it is calculated by a calculation means for giving desired numerical values to the distance (penalty value between different phonemes), the penalty value when a phoneme is inserted, and the penalty value when a phoneme is dropped.

According to a fourth aspect of the present invention, in the document data retrieval apparatus according to any one of the first to third aspects,
The INDEX file creation means creates a file in which importance information using the position information and appearance frequency of the extracted word extracted from the search target sentence data as parameters in the document data is registered.

The invention of claim 5 is the document data retrieval apparatus of claim 1,
The output interface includes means for calculating a fitness based on an acoustic similarity distance value and outputting a search result in descending order of the fitness with one or a plurality of search words obtained from a search character string by an INDEX file. By
The top N search target document data candidates are narrowed down in descending order of suitability with the search character string by the INDEX file, and the phoneme string of the search character string and the phoneme string of the search target document data candidate character string in the N candidates DP matching is calculated, and the output order is determined based on the calculation result as a similarity.
The goodness of fit calculated based on the acoustic similarity distance value is, for example, a value including the reciprocal of a numerical value obtained by adding an arbitrary number to the acoustic similarity distance value.

  That is, the present invention predicts a word that may be input in a document data search apparatus, calculates in advance an acoustic similarity distance between this word and a search target word in the document data, and is searched in the document data. Search index file based on the acoustic similarity distance value between two words, the position of the word candidate in the document data, and the importance information. Providing means for creating is a characteristic configuration.

According to the present invention, even when searching with a search character string that does not exactly match the search target word or is mistakenly remembered, a candidate that is acoustically similar to the search character string can be cited, so A text document (content) can be searched.
Moreover, since the acoustic similarity distance between the search keyword and the word to be searched is calculated in advance at the stage of creating the INDEX file of the present invention, it is not necessary to calculate the acoustic similarity distance at the time of searching, and the search time is reduced. Therefore, it is possible to speed up the similar text search.

Hereinafter, an embodiment of a document data retrieval apparatus of the present invention will be described with reference to the block diagram of FIG.
The document data search device 1 is an example applied to a device for searching the lyrics of a song. An input interface 2 for inputting a partial phrase or the like of a lyrics as a search character string and a search result of lyrics data (document data). An output interface 3 for outputting, a lyrics database 4 storing a plurality of lyrics data (lyric files) to be searched, and lyrics data stored in the lyrics database 4 from a search character string input to the input interface 2 It has the control part 10 which performs a search process, and is comprised.

  The control unit 10 inputs a plurality of lyrics data (sentence data) in advance and extracts a word constituting the lyrics (document data), and a search word that may be searched (included in the search character string). Keyword registration means 12 for registering a search word as a keyword in a list, and an index file for search using an acoustic similarity distance value for the extracted word (search target word) and the registered keyword INDEX file creation means 13 for creating a search result, and similarity calculation means 14 for calculating the similarity of search target document data based on a search word input with reference to the INDEX file.

In the control unit 10 of the document data search apparatus 1, an INDEX file used at the time of search is created based on information input in advance to the word extraction unit 11, the keyword registration unit 12, and the INDEX file creation unit 13.
Hereinafter, a procedure for creating an INDEX file in the control unit 10 will be described with reference to FIG.

  First, a morphological analysis for extracting a plurality of words is performed on the document data (lyric data) of the plurality of lyrics files S1 to Si stored in the lyrics database 4 by the word extraction means 11, and an extracted word ( A list Li20 is created. The morphological analysis tool used here can cope with Japanese and English part-of-speech classifications, and can deal with fluctuations in notation in foreign words and kana English.

  Next, a keyword list (A1,..., An) 21 registered as keywords is created by the keyword registration means 12 in anticipation of words to be input for searching. The keywords (search words that may be included in the search character string) are stored in the lyrics database in order to provide flexibility in the word set of the keyword list 21 (to deal with words that are mistakenly remembered). In addition to all the words included in the lyrics, words extracted from the language corpus of the domain other than the lyrics are also registered.

  For each word in the extracted word list Li20 created for each lyric file, the acoustic similarity distance to the keyword An is calculated by the acoustic similarity distance calculation method described later, and the word Wni having the closest acoustic similarity distance (in the lyrics file Si) An INDEX file 22 is created by extracting the distance values dni (acoustic similarity distance values of words Wni and An similar to An in the lyrics file Si) respectively.

The acoustic similarity distance value in the INDEX file 22 can be calculated by the following method.
The acoustic similarity distance value of the INDEX file creation means is a means of calculating the edit distance of the reading kana between the keyword (search word) and the extracted word (search target word), or the edit distance of two phoneme strings between words Thus, the reading of the word is converted into a kana or phoneme string by means of, and DP matching is performed to obtain the similarity distance value.

  When the edit distance is calculated by comparing the reading kana strings and the acoustic similarity distance value of the INDEX file creation means is obtained, first, the word reading is converted into the reading kana string by the means for calculating the editing distance between the words. Then, DP matching is performed in syllable units in the two-reading kana string, so that the distance between kana is the distribution distance of the acoustic model representing each syllable, and the distance is 1 when syllable is inserted or dropped Calculate by means.

  Also, when calculating the edit distance by phoneme sequence comparison and obtaining the acoustic similarity distance value of the INDEX file creation means, first the word reading is converted into a phoneme sequence by means of calculating the edit distance of two phoneme sequences between words. Convert. Then, DP matching is performed in units of phonemes in the two phoneme strings, so that the distance between phonemes is the distance between the probability distributions of the acoustic model representing each phoneme, and the distance when the phoneme is inserted or dropped is 1 It is calculated by the calculation means.

That is, first, the search keyword An and the word to be searched are both converted into a reading kana or phoneme string, and DP matching is performed on the two reading kana or phoneme strings in units of characters or phonemes, and calculated by DP matching calculation. The obtained distance value becomes the acoustic similarity distance value.
DP matching is a method for calculating the similarity between series of data, and will be described below based on a specific DP matching calculation example.
In DP matching, a penalty of “one point if the character does not match” and “one point if the character is shifted by one (insertion or dropping)” is determined.

For example, when comparing between reading kana characters, comparing “Osaka” and “Otsuka” composed of four-tone characters, the characters are different at one place and the characters are the same at three places. × 1 = 1 ”, the matching part is“ 0 ”, and the total penalty is 1 + 0 = 1 point. Since the distance value is normalized to a value of 0 to 1.0, the penalty after normalization is calculated as “total penalty / character string length”, which is ¼ = 0.25. The character string length as a denominator in normalization is the longer number of characters between character strings.
In addition, regarding the comparison between the 6-character “Shin-Osaka” and the 4-character “Osaka”, the “Shin-Osaka” and “Osaka” are compared to each other, assuming that “the two characters are shifted or stagnated.” 2 parts of misalignment are “1 × 2 = 2”, 2 parts of mismatch are “1 × 2 = 2”, and the penalty after normalization is (2 + 2) /6=0.667 points become. In DP matching, the smaller the penalty value, the higher the similarity is set. Therefore, when the search word is “Osaka”, it is determined that “Otsuka” is more similar to “Osaka”.

In the same example, when comparing “Osaka” and “Otsuka” with phoneme sequences, “O”, “O”, “S”, “A”, “K”, “A”, and “O”, “O”, “TS”. This is a comparison with “U”, “K”, and “A”. Considering “O”, “O”, “S”, “A”, “K”, “A” and “O”, “O”, “TS”, “U”, “K”, and “A”, there are 0 deviations: 2 × 1 = 2. Two differences: 2 × 1 = 2, and the penalty after normalization is (0 + 2) /8=0.25 points. The denominator for normalization is the phoneme number of the longer phoneme string.
Further, when comparing phonemes between “Osaka” and “Shin Osaka”, “O” “O” “S” “A” “K” “A” and “S” “I” “N” “O” “O” “S” “A” “K” “A”. “O” “O” “O” “O” “O” “S” “A” “K” “A” and “S” “I” “N” “O” “O” “S” “A” “ Considering “K” and “A”, the deviation is 3 × 3 = 3, the difference is 3 × 3 = 3, and the penalty after normalization is (3 + 3) /9=0.667 points. At the phoneme level, when the search word is “Osaka”, it is determined that “Otsuka” is more similar than “Shin Osaka”.

In the acoustic similarity distance calculation method described above, the penalty when inserting or dropping a character or phoneme in the case of a reading character is “1”, and the distance value between syllables or phonemes (penalty value between different phonemes). Is set to “1”, but may be calculated using other penalty values.
For example, for the distance value between phonemes, the inter-model distance value of the phoneme acoustic model used at the time of recognition may be used. In this case, the distance value between phonemes can be defined by the Mahalanobis distance between the probability distributions of the acoustic model representing each phoneme. When the acoustic model of each phoneme is modeled with one state and a single Gaussian distribution, the Mahalanobis distance ADij (phonemes i and j) between the two models is expressed by the following equation (1).

In the above equation 1,
K is the number of dimensions of the MFCC (Mel-Frequency Cepstrum Coefficient) vector (K = 12),
μik and σik are the elements of the kth dimension of the mean and variance MFCC vectors of phoneme i, respectively.

Further, the distance between phonemes may be calculated based on the interphoneme confusion matrix. The inter-phoneme confusion matrix is obtained by a recognition experiment or the like, and represents the elements of the matrix as probabilities. A calculation example of the interphoneme distance using this interphoneme confusion matrix will be described with reference to Table 1.
Table 1 shows the probability that the input phonemes a, i, u, e, o,... Can be heard as a, i, u, e, o,. For example, the probability that the phoneme a can be heard as a is 0.9, the probability that it can be heard as i is 0.2, the probability that it can be heard as u is 0.3, and so on. The reciprocal of the probability of the phoneme confusion matrix can be defined as the distance between phonemes.

  In the INDEX file created by the INDEX file creation means 13, information (ranking element value) that is used as a criterion for selecting lyrics data for a search word is registered in addition to the acoustic similarity distance value described above. It may be a thing. As such information, for example, position information in the lyric data (document data) of a word similar to the search word (search target word existing in the lyrics) and importance information for calculating the appearance frequency as a parameter can be considered. .

The position information is stored in the INDEX file as an appearance position pni of similar word candidates (where pni is an appearance position of the similar word candidate Wni in the lyric file Si: a numerical value indicating what number of words appears). If there are a plurality of appearance positions pni, all of them are registered.
The importance level information is defined as a feature quantity tf ・ idf that evaluates how important the search target word has in the lyric data, and tf is the frequency of occurrence of the search target word in the lyric data and idf is log The feature quantity tf · idf is calculated by tf · id f which is the product of these, where (number of all search target documents / number of documents including search target words).
Further, in order to normalize the value of the feature value tf · idf to a value of 0 to 1.0, the following calculation is performed. The tf · idf value of the word Wni in the lyric file Si is tf · idfni, and Σtf · idfi is the sum of the tf · idf values of all the words of Si. Normalized Wni tf · idf'ni = tf · idfni / Σtf · idfi.

  For all the words of the keywords A1 to An registered in the keyword list 21, the word Wni having the closest acoustic similarity distance, the distance value dni by the acoustic similarity distance calculation, the appearance position pni, the normalized feature amount tf · id f 'Create a table with registered values. The table of the INDEX file 22 is created as shown in Table 2 in the lyrics file S10, for example.

  The similarity calculation means 14 of the control unit 10 calculates the similarity from the word Wni having the closest acoustic similarity distance registered in the table, the distance value dni by the acoustic similarity distance calculation, the appearance position pni, and the feature value tf · idf ′ value. Is calculated.

  The output interface 3 outputs the search result of the document data based on the calculation result in the similarity calculation means 14 of the control unit 10, and has a high similarity (or goodness of fit) with the search character string by the INDEX file. The search results are output in order.

Next, a case where a lyric file is searched by a search character string using a document data search apparatus in which an INDEX file is registered will be described with reference to the flowchart of FIG. 3 and FIG.
First, a character string (lyric sentence or phrase) that is a set of words familiar to the user is input to the input interface 2 as a search character string (step 101).
The input search character string is subjected to the same morphological analysis as that used in the creation of the INDEX file in the control unit 10 (step 102), and the words in the search query word list Q1,. A column is extracted (step 103).
Next, the search query word list is collated with the keyword list of the INDEX file 22 (step 104), and the matched keywords A1 to An are extracted. For example, if there are Q1, Q2, and Q3 in the search query word list, Q1 matches A5, Q2 matches A7, and Q3 matches A9, A5, A7, and A9 are extracted.

With respect to the extracted keyword word (or word string), the index file 22 is referred to, and in each lyrics file, a word candidate that is acoustically most similar to the keyword word (or word string), and its acoustic similarity distance The value and position information of similar word candidates are obtained and the similarity is calculated (step 105), and the lyrics file is selected (step 106).
Since Q1 matches A5, Q2 matches A7, and Q3 matches A9, the acoustic similarity distance Di of the search queries Q1, Q2, Q3 for the lyrics file Si is d5i + d7i + d9i, and the similar word candidate string Wi for the lyrics file Si is W5i. , W7i, W9i. Also, the appearance position information Pi in the lyrics file Si of the similar word candidate strings W5i, W7i, W9i is p5i, p7i, p9i.

Next, the degree of matching between the search character string and the lyrics file hit in the search is calculated, and ranking is performed accordingly (step 107).
The calculation of the fitness can be performed by a simple method in which the fitness of the lyrics file Si is 1 / (Di + 0.1) (Di is the acoustic similarity distance value of the lyrics file Si). Actually, the degree of conformity of the lyrics file is defined as 1 / (Di + β), and β is set to 0.1.
Further, after the appearance position information Pi of the similar word candidate string Wi with respect to the keyword word string is read, the adjacent relationship between the search word strings in the similar word candidate string is checked. Then, the adjacent weight αi between the word strings is set to the value of the acoustic similarity distance Di (the higher the adjacent weight αi value is set for the stronger adjacent relationship). As a calculation example, Q1 coincides with A5, Q2 coincides with A7, Q3 coincides with A9, and the adjacent weight αi for the lyric file Si can be calculated by the following equation (2). Further, when there is only one word, αi = 1 in Equation 2.

(Equation 2)
αi = 1 / [(p7i-p5i) ^ 2 + (p9i-p7i) ^ 2]

The ranking suitability for ranking the lyrics file is calculated by associating the acoustic similarity distance value Di with the adjacent weight αi and the feature quantity tf · idf. As a calculation example, the ranking suitability is a value calculated by αi / (Di + 0.1) + tf · idf ′. In addition, only when the value of Di is 0, the value of tf · idf ′ may be added to the fitness.
Then, when ranking the lyric files, the degree of fitness is determined among all the lyric file data stored in the lyric database.
After ranking the lyric files according to the ranking suitability values, the lyric files that are ranked in the top N in order of the suitability and their music information are output from the output interface 3 as search results (step 108).

In the above-described example, when the INDEX file 22 is created, the acoustic similarity distance with the keyword An is calculated for each word in the word list Li20 created for each lyrics file, and the word Wni () having the closest acoustic similarity distance is calculated. The words in the lyric file Si that are acoustically most similar to An are extracted, but the top M word candidates that are closest in acoustic similarity (the top M words that are acoustically most similar to An in the lyric file Si). May be extracted, and the acoustic similarity distance value Di and the adjacent weight αi may be stored.
In this case, in the lyric file Si, a plurality of similar words are candidates for the keyword that is the search word constituting the search word string. When calculating the ranking suitability, the candidate αi / (Di + 0. 1) The highest value among the values of + tf · idf ′ may be used as the ranking suitability for the lyrics file Si.

  In addition, in order to increase the calculation accuracy of similarity, the top N search target lyric file candidates are narrowed down in descending order of the fitness value for ranking by the INDEX file, and the character string of each of the N candidate lyric files is input. The search character string (the inputted phrase before the morpheme analysis described above) may be converted into a phoneme string and a process of calculating DP matching may be performed. By calculating the fitness based on the distance value that is the calculation result and ranking in descending order, ranking display with higher similarity accuracy can be performed.

According to the above configuration, if a word similar to or the same as the search word constituting the input search character string exists in the word list An (search target word) in the INDEX file 22, the numerical values such as the similarity distance have already been calculated. Since it is registered as data, the similarity text search can be speeded up by calculating the similarity based on this numerical value.
In the above example, the document data search apparatus for searching for lyric data has been described. However, the present invention is not limited to lyric data, and can be applied to similar text search for text such as document data.

  If the input word is not in the keyword list, the word (or word string) is converted to a phoneme string in real time, DP matching with the phoneme string of each search target lyrics file is calculated, and the resulting distance Search for values as similarities.

It is a functional block diagram which shows an example of embodiment of the document data search apparatus of this invention. It is explanatory drawing for demonstrating the INDEX file creation process in the document data search apparatus of this invention. It is a flowchart which shows an example of the search process in the document data search apparatus of this invention.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Document data search device, 2 ... Input interface, 3 ... Output interface, 4 ... Lyrics database, 10 ... Control part, 11 ... Word extraction means, 12 ... Keyword registration means, 13 ... INDEX file creation means, 14 ... Similarity degree Arithmetic means, 20 ... extracted word list, 21 ... keyword list, 22 ... INDEX file.

Claims (5)

In a search device comprising an input interface for inputting a search character string, and searching for document data from a plurality of search target text data by the search character string,
Word extraction means for inputting a plurality of sentence data in advance and extracting words constituting the document data;
Keyword registration means for registering a search word that may be included in the search character string as a keyword in a list;
INDEX file creation means for creating an index file for search using an acoustic similarity distance value calculated from a comparison between reading kana or phonemes decomposed for extracted words and keywords extracted by the word extraction means;
Based on a search word obtained from the input search character string, similarity calculation means for calculating the similarity of search target document data with reference to the INDEX file,
A document data search apparatus comprising: an output interface that outputs a search result of document data based on the similarity. The acoustic similarity distance value of the INDEX file creation means is
By converting each of the keyword and the extracted word into a reading kana string by means of calculating the edit distance, and performing DP matching in syllable units in the two reading kana strings, the distance between syllables (between different syllables) 2. The document data retrieval apparatus according to claim 1, wherein the calculation means calculates a desired value for each of the penalty value at the time of syllable insertion, the penalty value at the time of syllable dropping, and the penalty value at the time of syllable dropping. The acoustic similarity distance value of the INDEX file creation means is
By converting each reading into a phoneme string by means of calculating the edit distance of the two phoneme strings between the keyword and the extracted word, and performing DP matching in units of phonemes in the two phoneme strings, 2. The document data retrieval apparatus according to claim 1, wherein calculation is performed by calculation means for giving desired numerical values to distances (penalty values between different phonemes), penalty values at the time of phoneme insertion, and penalty values at the time of phoneme dropout.   The INDEX file creation means creates a search INDEX file in which importance information using the position information and appearance frequency of the extracted word extracted from the search target sentence data as parameters in the document data is registered. The document data search device according to claim 3. The output interface includes means for calculating a fitness based on an acoustic similarity distance value and outputting a search result in descending order of the fitness with one or a plurality of search words obtained from a search character string by an INDEX file. By
The top N search target document data candidates are narrowed down in descending order of suitability with the search character string by the INDEX file, and the phoneme string of the search character string and the phoneme string of the search target document data candidate character string in the N candidates The document data search apparatus according to claim 1, wherein DP matching is calculated and an output order is determined based on a similarity as a result of the calculation.

Images