JP2014052803A - Voice document retrieval method and voice document retrieval system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice document retrieval method and the like in which a voice section showing the content to be retrieved can be retrieved with good accuracy out of a relatively long voice document.SOLUTION: Voice data stored in a database 2 is read out and documented by using voice recognition means 3, and a voice section for each utterance unit such as 15 utterance unit, 30 utterance unit, and 60 utterance unit is extracted from the documented data. Then, a related word which is related to a query input by using this voice section is extracted by using a pseudo-compatibility feedback, and search is performed with a second voice section being an index by using the extracted related word and the original query.

Description

  The present invention relates to an audio document search method that enables an audio section of content to be searched from an audio set to be found. For example, an audio section having an optimal content is selected from a long-time audio file such as a lecture unit. The present invention relates to a voice document search method that can be found.

  Conventionally, there is a search method such as query expansion as a method for finding a desired speech section from a speech document. In this query expansion, some related terms are extracted from the initial query input by the user, and the user selects the related terms and performs a search together with the original query. As an extension of this query expansion, pseudo-conformance feedback, etc. that considers some of the top ranks from the search results of the input original query as pseudo-compliant documents and performs a search with the original query, etc. Yes (Patent Document 1 etc.).

JP 2003-242170 A

  By the way, when trying to search a speech section including a desired topic from a very long speech document such as a plurality of lectures using a method such as pseudo-adaptability feedback, the following is performed. Cause serious problems.

  In other words, pseudo relevance feedback is to extract the related terms by regarding the upper rank of the search results from the input initial query as pseudo relevance documents. May be added. In addition, when the search target is a relatively long lecture unit and includes many topics, an additional word that is not related to the query may be extracted. For this reason, if a speech segment containing a desired topic is extracted from a large number of speech documents such as lectures, is the search result suitable by reproducing the extracted speech document for each lecture unit from the beginning? There was a problem that it took a huge amount of time to work.

  Therefore, the present invention has been made paying attention to the above-mentioned problem, and provides a voice document search method and the like that can accurately search a voice section representing a content to be searched from a relatively long voice document. Objective.

  That is, in order to solve the above-described problem, the present invention separates a speech document to be searched for every predetermined time unit and extracts a speech section, and a relation related to a query in the extracted speech section. A step of extracting a word and a step of performing a search using the extracted related word are provided.

  In this way, for example, compared to the case of searching related words extracted from a very long section such as a lecture unit, the possibility of adding unrelated topics as related words is reduced, and the search accuracy is improved. Will be able to.

  Further, as the predetermined time unit, a predetermined word unit or a time unit divided by breathing unit is used.

  In this way, since the speech section for extracting the related word is divided in units of words or breathing units, it is possible to make it easier to extract the related words than in the case where the time is uniformly divided in units of seconds. become.

  Further, the speech segment separated for each second time unit different from the predetermined time unit is used as a search index to search using the extracted related terms.

  In this way, since a search can be performed for each separated second time unit, it is possible to efficiently extract a suitable speech section from a long speech document.

  In the present invention, a speech document to be searched is separated every predetermined time unit to extract a speech section, and a related word related to the query is extracted in the extracted speech section to perform a search. So, for example, compared to searching by extracting related terms from a very long section such as a lecture unit, the possibility of adding unrelated topics as related terms is reduced, and the search accuracy can be improved. Become.

Functional block diagram of a speech document search system in an embodiment of the present invention Diagram showing query expansion with pseudo-fitness feedback The figure which shows the outline | summary of the pseudo conformity feedback in one embodiment of this invention Flow chart of the system in the same form Table showing the effect of pseudo conformance feedback Example showing the impact of pseudo-conformance feedback

  Hereinafter, an audio document search system 1 according to an embodiment of the present invention will be described with reference to the drawings.

  The speech document retrieval system 1 in this embodiment documents speech data stored in a database 2 using speech recognition means 3, and provides related information related to an inputted query from among them as pseudo-fit feedback. It is used to extract and search. Characteristically, the documented text is separated into predetermined word (column) units or breathing (speech) units (for example, 15 utterances, 30 utterances, 60 utterances, etc.), and the separated speech Extract related terms using pseudo relevance feedback within the interval, and separate the documented text into a second time unit (eg, 30 utterances to the whole) that is different from the time unit. It is possible to perform a search using related terms as an index for this purpose. Hereafter, each function realization means in this Embodiment is demonstrated using the functional block diagram of FIG.

  First, voice data is stored in the database 2. Here, for example, speech data that is spoken by a speaker such as a conference lecture or a mock lecture is handled, but the present invention is not limited to this. These audio data are usually about 10 to 25 minutes, but may exceed 1 hour, and in the case of a mock lecture, there is also a speech of about 12 minutes on a daily topic by a general speaker. , Often includes a variety of topics.

  The voice recognition means 3 reads the voice data stored in the database 2 and documents it using a voice recognition technique. As the speech recognition technology, various conventionally proposed methods are used, and these are used as a text document. At this time, since the speech recognition rate varies depending on the speech recognition technology, the recognition rate is generally about 65% to 95%.

  The first speech segment extraction means 4 extracts a speech segment having a first time length (first speech segment) from the text document recognized as described above. This voice section may be a voice section in a uniform second unit (for example, 30 seconds, 1 minute, 2 minutes, etc.), but here it is a time length in units of breathing (speech). When extracting the voice interval of this breath connection unit, it is separated into time lengths of a predetermined breath connection (speech) unit (for example, 15 speeches, 30 speeches, 60 speeches). In this way, when the time is divided in units of breathing, it is not divided at least in the middle of the utterance, so there is an advantage that related words etc. can be extracted accurately compared to the case where separation is performed uniformly in seconds. is there. Here, the first time length is divided in units of breath connection, but may be divided in “word string units”, for example. In this way, when segmented in units of word strings, separation can be performed at sentence boundaries if a typical sentence end expression is found. In addition, if a word string related to a typical topic is found, it can be separated at the boundary of the topic, so there is a merit that related words etc. can be extracted more accurately than in the case of uniformly separating in units of seconds. is there.

  The query input means 5 receives a query input by the user. As a method of accepting this input, the input may be accepted as text, or the input may be accepted as speech. Here, an example of the influence of pseudo conformity feedback will be described in an example to be described later, assuming that a query “where is Aboriginal?” Is input as an input example.

  The related word extraction / drawing means 7 extracts a related word from the input query by using the text of the first voice section extracted by the first voice section extraction means 4 as an index. Here, the concept of query expansion will be described.

In general, in a search system based on a vector space, a query is represented by a vector whose elements are word frequency. Assuming that the original query vector is q, the new index word vector q _n to be added to the query, and the weight is α, the expanded query vector q _n is expressed by Equation 1. In the vector space model to the q to q _n be said query expansion.

Generally, α <1, but since the element of q _n is a positive number, query expansion is performed with β based on Equation 2 as a positive integer. That is, the original query q is weighted again and the query is expanded as shown in Equation 2.

Next, relevance feedback will be described. Relevance feedback means that a query vector q is changed to q _n by allowing a user to input which document is “conforming document” and which document is “nonconforming document” among the obtained search results. It will be corrected (see Equation 3).

At this time, d _r and d _n are query vectors each having the appearance frequency of words included in the set of conforming documents and the set of non-conforming documents as elements. Further, ω ₀ , ω ₁ and ω ₂ are constants of 0 or more.

  Next, pseudo compatibility feedback will be described. The conformance feedback method described above requires the user to input “whether it is a conforming document” or “whether it is a nonconforming document”. This “pseudo conformance feedback” The query expansion can be performed by extracting the related words without requiring any judgment and without any interaction with the user. This technique is shown in FIG. In the pseudo relevance feedback, a search result is obtained by using an input query q with a document recognized from speech data as an index. Next, the top of the obtained search results are made a collection of conforming documents in a pseudo manner, related terms are extracted from these documents, and added to the original query to expand the query. . Here, since the non-conformance document is not used, the pseudo-conformity feedback technique is expressed by Equation 4.

  The search means 8 performs a search from the index for search using the related terms extracted in this way and the original query. Here, as an index for searching, a document having a time longer than the first time length for extracting related words is extracted by the second speech section extracting means 6, for example, 30 utterances, 60 utterances, or The total length of the document is searched using related terms and the original query. Normally, if you search by extracting related terms from the entire document, related terms related to different topics will be extracted, and the search accuracy will be reduced due to voice recognition errors. If a word is extracted and searched, the possibility of extracting such irrelevant related words is reduced, and the search accuracy can be increased.

  The output means 9 outputs the voice document searched in this way. These outputs are displayed in order from the highest score, and are selected so that they can be reproduced from the second audio section of the corresponding audio data. In this way, since it is not necessary for the user to consider whether the extracted speech data is a relevance document by reproducing the extracted speech data in units of lectures, from the speech section corresponding to the relevant relevance document in a very short time. It can be played back. In this example, the audio data is reproduced from the second audio segment of the selected audio data. However, the audio data may be reproduced from the audio segment from a predetermined length before the audio segment.

  Next, with respect to a method for searching voice data corresponding to a query from voice data stored in a database using the voice document search system configured as described above, the overview diagram of FIG. 3 and the flowchart of FIG. It explains using.

  First, when a user inputs a query using the query input means 5 (step S1), the voice document search system 1 extracts voice data stored in the database 2 (step S2), and the voice recognition means 3 Use for voice recognition (step S3). In this example, the text is converted into text using the voice recognition means 3 after the query is input, but the voice data in the database 2 may be converted into text in advance.

  Then, the voice document converted into text in this way is subjected to morphological analysis (step S4) and separated into first voice sections, for example, 15 speech units (step S5). Then, a search is performed for each first speech section using the input query (step S6), ranking is performed, and the related words are extracted by regarding the top several search results as pseudo matching documents (step S6). S7) The search is performed again by pseudo conformity feedback using the original query (step S9).

  When this re-search is performed, the second voice section is extracted from the text converted in step S3 (step S8), and a matching voice document is searched from the voice section (step S9).

  Then, the user selects a desired voice document from the search results, reads the voice data associated with the voice document from the database, and outputs the voice data of the corresponding voice section (step S10).

  As described above, according to the above embodiment, the speech document to be searched is separated for each predetermined utterance unit, the speech section is extracted, and the related terms related to the query are extracted in the extracted speech section. For example, compared with the case of searching by extracting related words from a very long section such as a lecture unit, the possibility of adding irrelevant topics as related words is reduced, and the search accuracy is reduced. Can be increased.

  Next, experimental results in the present invention will be shown. Here, the experiment is performed using the test collection used in NTCIR-9 Spoken Doc. This searches for a total of 2702 lectures, including 987 academic lectures and 1715 mock lectures in the Japanese Spoken Language Corpus (hereinafter referred to as CSJ). Speech recognition results for each CSJ lecture speech are required, but we used those included in the speech document search test collection (recognition rate 65% -95%). In this experiment, 15 utterances, 30 utterances, and 60 utterances are adopted as passage search targets. The 15 utterance units are divided into 15 utterances in order from the beginning of the lecture, and each section becomes a document to be searched. Similarly to 15 utterances, 30 and 60 utterances are divided into 30 utterances and 60 utterances in order from the beginning of the lecture.

  Queries are 125 texts written in natural language sentences. Relevance information as an answer to each query is given in units from which utterance to which utterance of which lecture. Partial fit (P) exists. Here, the section with the matching label (R) is treated as the correct answer to the query.

  The evaluation of the search performance of the voice document search system 1 in the present embodiment generally uses the recall and accuracy. Here, 11-point average accuracy, which is an evaluation scale combining these, is used as the evaluation scale. This is obtained by Equation 5, and means that the closer the value is to 1, the more suitable it is.

1. Effects of pseudo-conformance feedback in lecture document search (1) Baseline search results Here, a document search system based on a vector space model was used. SMART was used for the similarity between vectors. Here, the baseline search system was designed using the index word as the basic form of morphemes (only nouns and verbs) and the general-purpose associative engine GETA as the search engine. When query q is given, similarity SMART (q, Di) with q is calculated for all documents Di, and all documents with similarity greater than 0 are output in descending order.

  The results of this baseline system are shown in FIG. 5, and the 11-point average accuracy of the search for the lecture unit, 60 speech units, 30 speech units, and 15 speech units was 0.531, 0.294, 0.241, and 0.183, respectively.

(2) Search result of pseudo suitability feedback for each utterance unit to be searched There are three parameters in pseudo suitability feedback. The first is how many of the top search results obtained by the original query q are treated as pseudo matching documents, and the second is how many words are extracted when extracting related words from the matching documents using Equation 5. 3 The second is the weight β for the original query in Equation 4. In this embodiment, leave-one-out cross-validation was performed using 125 queries. Specifically, each parameter was estimated using 124 queries so that the search accuracy was maximized, and the estimated parameters were used for the remaining one search experiment. At that time, the number of documents regarded as compliant documents was counted from 1 to 5 in increments of 1; related words were searched from 10 to 50 words in increments of 10; and β was searched from 1 to 10 in increments of 1. This was done for every split set. Here, based on a preliminary experiment, it is assumed that a low-frequency word contains a speech recognition error word when a related word is extracted, and a word with a document appearance frequency of 1 (DF = 1) is extracted from the related word. I decided to exclude it.

  The result of query expansion with pseudo-fitness feedback is shown in the middle column of FIG. Here, the utterance unit to be searched and the extracted utterance unit of the related word are the same. Search accuracy decreased from 0.531 to 0.470 and from 0.294 to 0.262 for the lecture unit and 60 utterance units, respectively. Search accuracy improved from 0.241 to 0.253 and from 0.183 to 0.210 for 30 utterance units and 15 utterance units, respectively. These indicate that the smaller the document to be searched is, the greater the effect of pseudo conformity feedback is.

  Looking at the detailed results for each query, the queries that showed improved search accuracy were 28.8% (36/125) of the talk units, 32.0% (40/125) of the 60 utterance units, 30 utterance units accounted for 54.4% (68 out of 125 cases), and 15 utterance units accounted for 66.4% (83 out of 125 cases). An example of the effect of pseudo-compatibility feedback is shown in FIG. In the lecture unit, it can be seen that there are many words selected from large topics including the search intention of the query in the additional words. In contrast, in the unit of 15 utterances, it is understood that there are several words related to the search intention of the query in the additional words. This is because the search by lecture unit is performed well, but the document obtained as a search result includes a plurality of small topics centering on a certain topic. This is probably because unrelated small topic words are extracted as additional words. The search in units of 15 utterances is considered to be because words that are greatly related to the query were extracted at the time of pseudo relevance feedback because the document obtained as a search result is small and difficult to include a plurality of topics.

2. Use of pseudo-adaptability feedback using related word extraction index (1) How to use pseudo-adaptability feedback using related word extraction index What is the retrieved utterance unit for related word extraction for pseudo-adaptability feedback? A method for performing pseudo-adaptability feedback from independent utterance units is described.

  Here, related words are extracted in units of 15 utterances, and search is performed using these additional words for pseudo-adaptability feedback in units of lectures, 60 utterances, and 30 utterances. This is intended to prevent a word not related to the query from being added to the query in a search of a large utterance unit that may include a plurality of topics.

(2) Experimental results
Using the additional words obtained in the pseudo-adaptation feedback in the search of 15 utterance units, we performed pseudo-adaptation feedback in lecture units, 60 utterance units, and 30 utterance units. The results are shown in the rightmost column of FIG. Overall, the search accuracy was improved. 11 points average accuracy 0.500 was obtained in the lecture unit, and 11 points average accuracy 0.313 was obtained in 60 utterance units. In 30 utterance units, the average accuracy of 11 points was 0.262. These indicate that the smaller the document to be searched, the higher the increase rate of search accuracy. Looking at the detailed results for each query, 33.6% (42/125) of queries that showed improved search accuracy compared to baseline search results, 59.2% (42 of 125) 74 cases out of 125 cases) and 60.0% (75 out of 125 cases) in 30 utterance units. Compared to the results obtained with normal pseudo-conformance feedback, queries with improved search accuracy were 56.0% (70/125) for the lecture unit and 69.6% (125/125) for the 60 utterance unit. 85) and 30 utterances, 55.2% (69 out of 125). In terms of lecture units, the search accuracy remained unchanged for 4.0% (5/125) of queries. These results show the effect of performing pseudo-adaptive feedback of lecture units, 60 utterance units, and 30 utterance units using additional words extracted in units of 15 utterances. This is thought to be because the correct utterance intervals are distributed more than 15 utterance intervals, especially from 1 utterance interval to 3 utterance intervals. Future topics include extracting additional words in shorter utterance units and performing pseudo-adaptability feedback.

  In a search task for CSJ lectures, the effect of pseudo-conformance feedback was clarified by using a sufficient amount of search queries. We also proposed a method using pseudo-adaptability feedback using an index for extracting related words, and found that it is effective to use additional words obtained by searching short utterance units for pseudo-adaptability feedback. It was.

DESCRIPTION OF SYMBOLS 1 ... Voice document search system 2 ... Database 3 ... Voice recognition means 4 ... First voice section extraction means 5 ... Query input means 6 ... Related word extraction means 7 ... No. Two voice segment extraction means 8 ... search means 9 ... output means

Claims (6)

Separating a voice document to be searched for every predetermined time unit and extracting a voice section;
Extracting related terms related to the query within the extracted speech segment;
Performing a search using the extracted related terms;
A voice document search method characterized by comprising: The speech document search method according to claim 1, wherein the predetermined time unit is a predetermined word unit or a time unit divided by a breathing unit. The step of performing a search using the extracted related word uses the extracted related word as an index for search using a speech segment separated for each second time unit different from the predetermined time unit. The speech document search method according to claim 1, wherein the search is performed using the search. A first voice segment extracting means for separating a voice document to be searched every predetermined time unit and extracting a voice segment;
Related word extraction means for extracting related words related to the query in the extracted speech section using pseudo-adaptability feedback;
Search means for performing a search using the extracted related terms;
A voice document search system characterized by comprising: 5. The speech document search system according to claim 4, wherein the predetermined time unit is a predetermined word unit or a time unit divided by a breathing unit. The search means performs a search using the extracted related terms by using a speech segment separated for each second time unit different from the predetermined time unit as an index for search. 4. The voice document search system according to 4.

Images