JP2008009552A - Index generation device, index generation method and index generation program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To create a sufficiently useful and detailed index of entire audio content. <P>SOLUTION: When audio content (audio data) is input and an index creation command is received, an index generation device 20 generates a speech word list from words included in text data obtained through speech recognition of the audio content, calculates speech densities from the generated speech word list, extracts words whose speech density exceeds a predetermined value as topic words per unit time, generates a topic word list enumerating the extracted topic words in association with the predetermined unit time between the speech start time and speech end time of the audio content, extracts time points without any enumerated topic words on the generated topic word list as topic change points, and generates an index by sections delimited by the extracted topic change points from the topic words enumerated in each section on the topic word list. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an index generation apparatus, an index generation method, and an index generation program that generate speech data by recognizing speech content and generate an index to be added to the speech content from the text data.

  2. Description of the Related Art With the background of social broadbandization, an infrastructure for distributing a large number of video / audio contents in the Internet environment has been established. The search method on the Internet or the like for these video / audio contents is mainly a text / video content search and a search by a directory or title by document.

  In view of this, in order to search for such audio content using a document of the audio content, various techniques for recognizing the audio content (for example, audio data) and indexing (important word extraction) are disclosed. ing.

  For example, in Patent Document 1 (Japanese Patent Laid-Open No. 9-114494), “a noun that appears for the first time or a noun that has not been used for a long time since the last appearance may become a topic word of a new topic” On the other hand, if the noun that appeared in the previous utterance appears also in the next utterance, this noun is likely to be a factor in maintaining the topic. ” A technique for extracting an important word from an input document and generating an index composed of the extracted important word is disclosed.

JP-A-9-114494

  By the way, the above-described conventional technique merely creates one index for the entire audio content, and even if this index is used, a desired topic portion (a certain topic is uttered from the audio content). There was a problem that it was not possible to search for (part).

  Therefore, the present invention has been made to solve the above-described problems of the prior art, and an index generation device and an index generation capable of creating a sufficiently useful and detailed index for the entire audio content It is an object to provide a method and an index generation program.

  In order to solve the above-described problems and achieve the object, the invention according to claim 1 generates index data by recognizing speech content and generates an index to be given to the speech content from the text data. An utterance word list generation means for extracting a word included in the text data and an utterance time of the word, and generating an utterance word list in which the word and the utterance time are associated and listed; and the utterance word An utterance density indicating the number of utterances per predetermined unit time is calculated for each word included in the utterance word list generated by the list generation means, and words whose utterance density exceeds a predetermined value are used as topic words in the unit time. Topic word extracting means for extracting, and the predetermined content from the utterance start time to the utterance end time of the audio content A topic word list generating unit that generates a topic word list in which topic words extracted by the topic word extracting unit are associated and listed for each rank time, and a topic word list generated by the topic word list generating unit. And an index generating means for generating an index from the topic words thus obtained.

  Further, in the invention according to claim 2, in the above invention, in the topic word list generated by the topic word list generating means, the topic conversion that extracts a time when none of the topic words are enumerated as a topic conversion point The index generation means further generates an index from the topic words listed in each section in the topic word list for each section delimited by the topic conversion points extracted by the topic conversion point extraction means. It is characterized by doing.

  The invention according to claim 3 further comprises distance dictionary storage means for storing a distance dictionary that defines a mutual semantic distance between a plurality of words in the above invention, wherein the topic word extraction means includes the utterance When calculating the utterance density per unit time for a predetermined word included in the word list, the predetermined word and the semantic distance are within a predetermined range in the distance dictionary stored in the distance dictionary storage means. The utterance density per predetermined unit time is calculated on the assumption that the word is also the same word as the predetermined word.

  The invention according to claim 4 further comprises topic word extraction list storage means for storing a topic word extraction list in which the words to be extracted as the topic words are listed in the above invention, wherein the topic word extraction means Is extracted as a topic word in the unit time on the condition that the utterance density exceeds a predetermined value on the condition that the word is listed in the topic word extraction list stored in the topic word extraction list storage means It is characterized by doing.

  Further, in the invention according to claim 5, in the above invention, the topic turning point extracting means, when the same topic word is listed in the topic word list without a predetermined time interval, The topic turning points are extracted as the enumeration of topic words continues.

  According to a sixth aspect of the present invention, in the above invention, a distance dictionary storage unit that stores a distance dictionary that defines a semantic distance between a plurality of words, and a distance dictionary stored in the distance dictionary storage unit A word removal unit that evaluates a semantic distance between words included in the utterance word list generated by the utterance word list generation unit and removes a word with a low evaluation from the utterance word list. The topic word extraction unit calculates a utterance density per unit time for each word included in the utterance word list based on the utterance word list after the word removal by the word removal unit, and the utterance density A word having a value exceeding a predetermined value is extracted as a topic word in the unit time.

  The invention according to claim 7 is the meaning of the words listed in the utterance word list after the word removal by the word removal means using the distance dictionary stored in the distance dictionary storage means in the above invention A word having a high evaluation of a target distance is extracted from the distance dictionary, and further includes a word complementing unit that complements the extracted word in the utterance word list, and the topic word extracting unit includes a word after word completion by the word complementing unit Calculating an utterance density per predetermined unit time for each word included in the utterance word list based on the utterance word list, and extracting words having the utterance density exceeding a predetermined value as topic words in the unit time It is characterized by.

  Further, in the invention according to claim 8, in the above invention, for each section delimited by the topic turning point extracted by the topic turning point extracting means, each of the topic words listed in each section in the topic word list It further comprises an important word extraction means for calculating an utterance frequency and extracting a topic word having the utterance frequency exceeding a predetermined value as an important word, and the index generation means is configured to extract the important words for each section delimited by the topic turning points. An index listing topic words extracted as important words by the word extraction means is generated.

  The invention according to claim 9 is the above invention, wherein the index generation means is any of the utterance time, the utterance density, the utterance frequency, and the semantic distance for each section delimited by the topic turning points. The topic word is edited based on one or more to generate an index.

  The invention according to claim 10 is an index generation device that generates speech data by recognizing speech content and generates an index to be added to the speech content from the text data. Distance dictionary storage means for storing a distance dictionary that defines the semantic distance of the word, a word included in the text data and the utterance time of the word, and a utterance word list in which the word and the utterance time are associated and listed Utterance word list generation means for generating the utterance word and a distance dictionary stored in the distance dictionary storage means to evaluate a semantic distance between words included in the utterance word list generated by the utterance word list generation means. , A word removing unit that removes a word with a low evaluation from the utterance word list, and an utterance after the word removal by the word removing unit Characterized in that and a index generating means for generating an index based on the word list.

  The invention according to claim 11 is the meaning of the words listed in the utterance word list after the word removal by the word removal means using the distance dictionary stored in the distance dictionary storage means in the above invention. A word completion means for extracting a word having a high target distance evaluation from the distance dictionary and complementing the extracted word in the utterance word list, wherein the index generation means is an utterance after word completion by the word completion means An index is generated based on the word list.

  The invention according to claim 12 is an index generation method suitable for generating speech data by recognizing speech content and generating an index to be added to the speech content from the text data, the text data Utterance word list generating step of extracting a word included in the word and the utterance time of the word, generating an utterance word list in which the word and the utterance time are associated and enumerated, and the utterance generated by the utterance word list generation step A topic word extraction step of calculating an utterance density indicating the number of utterances per predetermined unit time for each word included in the word list, and extracting words whose utterance density exceeds a predetermined value as topic words in the unit time; For each predetermined unit time from the utterance start time to the utterance end time of the audio content, A topic word list generating step for generating a topic word list in which the topic words extracted by the word extracting step are associated and enumerated; and an index from the topic words listed in the topic word list generated by the topic word list generating step. And an index generation step to be generated.

  The invention according to claim 13 is an index creation for causing a computer to execute an index generation method suitable for generating text data by recognizing voice content and generating an index to be given to the voice content from the text data. An utterance word list generation procedure for extracting a word included in the text data and an utterance time of the word, and generating an utterance word list in which the word and the utterance time are associated and listed, and the utterance word An utterance density indicating the number of utterances per predetermined unit time is calculated for each word included in the utterance word list generated by the list generation procedure, and words whose utterance density exceeds a predetermined value are used as topic words in the unit time. The topic word extraction procedure to be extracted and the utterance end from the utterance start time of the audio content A topic word list generation step for generating a topic word list in which the topic words extracted by the topic word extraction procedure are associated and listed for each predetermined unit time until the time reaches, and the topic word list generation procedure An index generation procedure for generating an index from topic words listed in the generated topic word list is executed by a computer.

  According to the invention of claim 1 or 10, 11, 12, and 13, a word included in text data and an utterance time of the word are extracted, and an utterance word list in which the word and the utterance time are associated and listed is generated. And calculating an utterance density indicating the number of utterances per predetermined unit time for each word included in the generated utterance word list, and extracting a word whose utterance density exceeds a predetermined value as a topic word in the unit time, For each predetermined unit time from the utterance start time of the audio content to the utterance end time, a topic word list in which the extracted topic words are associated and listed is generated, and in the generated topic word list, the topic word In each case, the time when enumeration does not continue is extracted as a topic turning point, and for each section delimited by the extracted topic turning point, the stories listed in each section in the topic word list Because it generates an index from the word, it is possible to create a sufficiently useful and detailed index for the entire audio content.

  According to the invention of claim 2, in the generated topic word list, the time when none of the topic words are enumerated is extracted as a topic turning point, and for each section delimited by the extracted topic turning point, the topic word Since the index is generated from the topic words listed in each section in the list, multiple topics included in the audio content can be extracted, and multiple indexes can be created for each topic. It is possible to create a more useful and detailed index.

  According to the invention of claim 3, a distance dictionary that defines a semantic distance between a plurality of words is stored, and an utterance density per predetermined unit time is calculated for a predetermined word included in the utterance word list. In the stored distance dictionary, the utterance density per predetermined unit time is calculated assuming that the word whose semantic distance is within the predetermined range in the stored distance dictionary is the same as the predetermined word. Therefore, it is possible to create a more sufficiently useful and detailed index for the entire audio content.

  According to the invention of claim 4, a topic word extraction list in which words to be extracted as topic words are listed is stored, and on the condition that the words are listed in the stored topic word extraction list, Since words whose utterance density exceeds a predetermined value are extracted as topic words in unit time, topic turning points can be extracted more accurately.

  According to the invention of claim 5, when the same topic word is enumerated without a predetermined time interval in the topic word list, it is assumed that the enumeration of the topic word has continued and the topic turning point Therefore, it is possible to prevent unnecessary information from being extracted from the entire index by extracting topic words in detail.

  According to the invention of claim 6, a distance dictionary that defines a mutual semantic distance between a plurality of words is stored, and the words between words included in the generated utterance word list are stored using the stored distance dictionary. The utterance density per unit time is determined for each word included in the utterance word list based on the utterance word list after the word is removed. Since the word whose utterance density exceeds a predetermined value is extracted as a topic word in unit time, compared to the case where a topic word is extracted from extracted word data in which words that are semantically far are not deleted, It is possible to accurately identify the topic word.

  Further, according to the invention of claim 7, using the stored distance dictionary, a word having a high semantic distance evaluation with the words listed in the utterance word list after the word removal is extracted from the distance dictionary, The extracted words are supplemented to the utterance word list, and the utterance density per unit time is calculated for each word included in the utterance word list based on the utterance word list after the word completion, Since words exceeding the value are extracted as topic words in unit time, it is possible to specify the topic words more accurately than in the case where words that are semantically close are not complemented.

  Further, according to the invention of claim 8, for each section delimited by the extracted topic turning points, the utterance frequency of each topic word listed in each section in the topic word list is calculated, and the utterance frequency is predetermined. Topic words that exceed the value are extracted as important words, and an index that lists the topic words extracted as important words is generated for each section delimited by topic turning points. Therefore, useful words are created as indexes for each topic. It is possible.

  According to the invention of claim 9, the topic word is edited on the basis of one or more of the utterance time, the utterance density, the utterance frequency, and the semantic distance for each section delimited by the topic turning point. Since an index is generated, it is possible to display an index according to the purpose of use, such as searching the Internet or searching for agenda items such as conference recording data.

  Exemplary embodiments of an index generation device, an index generation method, and an index generation program according to the present invention will be described below in detail with reference to the accompanying drawings. In the following, description of main terms used in the embodiment, outline and features of the index generation device according to the first embodiment, configuration and processing flow of the index generation device, and effects of the first embodiment will be described.

[Explanation of terms]
First, main terms used in the following examples will be described. “Audio content (audio data)” used in the embodiment is not a conventional text data but a set of voices, music, and videos spoken by a person expressed in digital data. It's being used. Specifically, data as shown in FIG. 3 corresponds to this.

  “Voice recognition” is a technique for converting the audio content described above into text. Specifically, the spoken language spoken by a person is analyzed by a computer, and the content being spoken is extracted as character data. For example, when the audio content (audio data) shown in FIG. 3 is recognized by voice, the text data as shown in FIG. 4 can be extracted. Thus, by using text data obtained by converting audio content into text, it is possible to search for audio content over the Internet or the like.

  The “index” is data that gives an index to a specific item so that the specific item can be referred to quickly in the data handled by the computer. As an example of the index to be given, for example, “strategic conference” is given to the audio content recorded in the conference, and “song name” is given to the audio content of music.

[Outline and Features of Index Generation Device (Example 1)]
Next, the outline and characteristics of the index device according to the first embodiment will be described with reference to FIG. FIG. 1 is a diagram for explaining an overview and characteristics of the index generation device according to the first embodiment.

  As illustrated in FIG. 1, when audio content (audio data) is input and an index creation instruction is received, the index generation device 20 performs speech recognition of the input audio content. To give a specific example, when the audio content as shown in FIG. 3 is input, the index generating device 20 performs the text data shown in FIG. 4 (more specifically, the utterance as shown in FIG. 5). Text data in which data is associated with each time). Note that the audio content may be input from another computer or the like via a network, or may be input from a recording medium such as an FD or a CD.

  Then, the index generation device 20 according to the first embodiment outlines the generation of an index to be added to the audio content from the text data obtained by recognizing the audio content in this manner, and in particular, the audio content. The main feature is that it is possible to generate a sufficiently useful and detailed index for the whole.

  Specifically, the main feature is that the index generation device 20 extracts a word included in the text data and the utterance time of the word, and generates an utterance word list in which the word and the utterance time are associated and listed. (Refer to (1) in FIG. 1). To give a specific example, the words appearing in the text data (actually spoken words) and the appearing times (spoken time) are associated with each other by “10:06:00, environment, education”. Or “10:06:04, library” or the like is extracted to generate an utterance word list (see FIG. 6).

  Then, the index generation device 20 calculates an utterance density indicating the number of utterances per predetermined unit time for each word included in the generated utterance word list, and selects a word whose utterance density exceeds a predetermined value in the unit time. It is extracted as a topic word (see (2) in FIG. 1). Specifically, an utterance density (Δh / Δt = X) indicating the number of utterances (Δh: number of utterances) per predetermined unit time (Δt) is calculated for each word included in the generated utterance word list, Words whose utterance density (X) exceeds a predetermined value (for example, 20) are extracted as topic words in unit time (see FIG. 7). For example, here, the indexing apparatus 20 extracts “word A, word B, word C” as topic words.

  Subsequently, the index device 20 generates a topic word list in which the extracted topic words are associated and listed for each predetermined unit time from the utterance start time to the utterance end time of the audio content (FIG. 1). (See (3)). More specifically, the index generation device 20 uses “9:00: 00, word A”, “9:10:06, word A, word B as“ speech (appearance time), word name ”. ", 9:17:00, word A, word B" and "9:20:12, word C" are created as a topic word list.

  And the index production | generation apparatus 20 extracts the time when none of enumeration of a topic word continues as a topic turning point in the produced | generated topic word list (refer (4) of FIG. 1). Specifically, in the above example, the index generation device 20 plots the appearance density for each word included in the extracted topic word list in the order of appearance (utterance) as shown in (4) of FIG. To do. Then, the index device 20 extracts the time when the sections do not overlap (discontinuous), that is, the appearance time (T = 9: 20: 12) of the “word C” as the topic turning point.

  Subsequently, the index generation device 20 generates an index from the topic words listed in each section in the topic word list for each section delimited by the extracted topic turning points (see (5) in FIG. 1). More specifically, in the above example, the index generation device 20 performs “word A” and “word B” at time “T0-T” before the extracted topic turning point “appearance time of word C: T”. "Topic section 1" which is a topic about "Topic section", and "Topic section 2" which is a topic about "Word C" at time "T-T1" after "Appearance time: T". The index “9: 00: 00: 00 to 9:20:11, word A, word B”, “9:20:12 to 9:30:30, word C”, and the like are generated.

  In this way, it is possible to identify a topic in a certain time zone for audio content, give an index for identifying the topic between the topics, and assign another index between other topics, etc. As a result of extracting multiple topics included in the audio content and creating multiple indexes for each topic, a sufficiently useful and detailed index is created for the entire audio content as described above. It is possible.

[Configuration of Indexing Device (Example 1)]
Next, the configuration of the index generation device 20 shown in FIG. 1 will be described with reference to FIGS. FIG. 2 is a block diagram showing the configuration of the index generation device 20, FIG. 3 is a diagram showing an example of the configuration of information stored in audio content, and FIGS. 4 and 5 are stored in text data. FIG. 6 is a diagram illustrating a configuration example of information stored in the utterance word list, and FIG. 7 is a diagram illustrating an example of an utterance density calculation method according to the first embodiment. FIG. 8 is a diagram illustrating an example of the configuration of information stored in the topic word list, FIG. 9 is a diagram illustrating an example of a topic turning point calculation method according to the first embodiment, and FIG. It is a figure which shows the structural example of the information memorize | stored in an index.

  As illustrated in FIG. 2, the index generation device 20 includes a communication control I / F unit 21, a storage unit 22, and a control unit 23. The communication control I / F unit 21 controls communication related to various types of information exchanged with other devices connected to the index generation device 20 via a network or the like. To give a specific example, audio content input to the index generation device 20 is received from another device.

  In addition to storing (storing) data and programs necessary for various processes by the control unit 23, the storage unit 22 is particularly closely related to the present invention and includes text data 22a, an utterance word list 22b, topic words A list 22c and an index 22d are provided.

  The text data 22a stores text data obtained by voice recognition of input voice content. More specifically, the index generation device 20 recognizes the voice when the voice content as shown in FIG. 3 is input, and stores FIG. 4 and FIG. To do.

  The utterance word list 22b stores data generated by an utterance word list generation unit 23b described later. More specifically, as shown in FIG. 6, “10: 6: 00, environment, education” is made by associating the “uttered word” with “the time when the spoken word was spoken”. Or “10:06:02, elderly, environment, education, disabled”.

  The topic word list 22c stores topic words generated by a topic word list generation unit 23d described later. As a specific example, as shown in FIG. 8, each topic word extracted by the topic word extraction unit 23c is stored in association with each appearing time axis.

  The index 22d stores data generated by an index generation unit 23f described later. More specifically, as shown in FIG. 10, “a topic section” delimited by a topic turning point extracted by a topic word turning point extracting unit 23e described later and an index generating unit 23f are used. “10: 00: 10-10: 06: 00, Future Library, Construction, Central”, “10: 06: 01-10: 17: 20, Library, Future Library, “Land” or “10: 17: 21-10: 30, Library, Request, Self-Defense Force”.

  The control unit 23 has a control program such as an OS (Operating System), a program that defines various processing procedures, and an internal memory for storing necessary data, and in particular, is closely related to the present invention. A speech recognition unit 23a, an utterance word list generation unit 23b, a topic word extraction unit 23c, a topic word list generation unit 23d, a topic word turning point extraction unit 23e, and an index generation unit 23f are provided. Execute the process. Note that the utterance word list generation unit 23b corresponds to the “utterance word list generation unit” described in the claims, and similarly, the topic word extraction unit 23c corresponds to the “topic word extraction unit”, and the topic word The list generation unit 23d corresponds to the “topic word list generation unit”, the topic word turning point extraction unit 23e corresponds to the “topic word turning point extraction unit”, and the index generation unit 23f corresponds to the “index generation unit”. Correspond.

  The voice recognition unit 23 a generates text data 22 a from the input voice content and stores it in the storage unit 22. Specifically, when an audio content is input and an index generation instruction is received, the audio recognition unit 23a analyzes the input audio content (see FIG. 3) and generates text data (FIG. 4, FIG. 4). 5) is generated and stored in the storage unit 22. The audio content handled by the voice recognition unit 23a may be data input in real time from a microphone or the like, or may be data recorded once and recorded on a recording medium such as a CD or a DVD. When real-time input speech data is recognized as speech, the speech recognition processing time is used as time information. When the stored speech data is recognized as speech, the time obtained by correcting the speech recognition processing time with the recording start time is handled as time information.

  The utterance word list generation unit 23b extracts a word included in the text data 22a and the utterance time of the word, and generates an utterance word list 22b that lists the word and the utterance time in association with each other. More specifically, the text data 22a generated by the speech recognition unit 23a and stored in the storage unit 22 is read out, and the word included in the text data 22a and the utterance time of the word are extracted. The word and the utterance time are associated with each other and extracted as “10:06:00, environment, education”, “10:06:04, library”, etc. to generate the utterance word list 22b (see FIG. 6). And stored in the storage unit 22.

  The topic word extraction unit 23c calculates an utterance density indicating the number of utterances per predetermined unit time for each word included in the utterance word list 22b generated by the utterance word list generation unit 23b, and the utterance density is a predetermined value. Words that exceed are extracted as topic words in unit time. Specifically, the utterance word list 22b generated by the utterance word list generation unit 23b and stored in the storage unit 22 is read, and as shown in (1) and (2) of FIG. An utterance density indicating the number of utterances (Δh) per predetermined unit time (Δt) is calculated for each word included in the word list 22b, and words whose utterance density (X) exceeds a predetermined value (for example, 20) are calculated. Extract as topic words in unit time.

  To explain with an example, the utterance word list of “10:06:00 to 10:06:57” shown in FIG. 6 with the unit time being “1 minute” and the threshold value “20” is the above-mentioned “Δh”. When the utterance density (X) is calculated by “/ Δt”, as shown in (3) of FIG. 7, “environment, 15”, “education, 4”, “ Elderly, 7 ”“ Disabled, 1 ”“ Library, 21 ”. Among these, “library” exceeding the threshold “20” is extracted as a topic word in “10:06:00 to 10:06:57”.

  The topic word list generation unit 23d associates and lists the topic words extracted by the topic word extraction unit 23c for each predetermined unit time from the utterance start time to the utterance end time of the audio content. Is generated. Specifically, as shown in FIG. 8 and FIG. 9, the topic words “question, reform, parking lot, library, Yokosuka, shopping street, etc.” extracted by the topic word extraction unit 23 c, respectively, A topic word list 22 c associated with the time when the topic word of the person speaks (appears) is generated and stored in the storage unit 22. Specifically, in the example described above, when the input audio content is 10 minutes, “time at 1 minute interval” and “topic word” in 10 minutes from the utterance start time to the utterance end time of the audio content. , “10: 06: 00-10: 06: 57, library”, “10: 07: 00-10: 08: 57, library, environment” “10: 09: 00-10: 09: 57 , “Education” and the like, a topic word list listing topic words every other minute is generated and stored in the storage unit 22.

  The topic word turning point extraction unit 23e extracts, as the topic turning point, a time when none of the topic words are enumerated in the topic word list 22c generated by the topic word list generation unit 23d. Specifically, as shown in FIG. 10, “word A” “word B” included in the topic word list 22 c created by the topic word list generation unit 23 d and stored in the storage unit 22 as shown in FIG. “Word C” is read, and the appearance density for each word is plotted in the order of appearance (utterance). Then, when “word A” and “word B” are plotted, they become continuous areas (with overlapping appearance times) (topic section 1), and when “word C” is plotted, “word A” and “word B” are not. Since it is a continuous area (topic section 2), the time when none of the topic words are enumerated, that is, the appearance time “T” of “word C” where the area is discontinuous is extracted as a topic turning point.

  The index generation unit 23f generates an index from the topic words listed in each section in the topic word list for each section delimited by the topic conversion points extracted by the topic word conversion point extraction unit 23e. Specifically, the topic turning points extracted by the topic word turning point extraction unit 23e are “10:06:00, Future Library”, “10:06:01, Library”, “10:17:21”. When “Self-Defense Force” is extracted, as shown in FIG. 10, the index generation unit 23f creates an index from the topic words listed in each section in the topic word list for each section delimited by the extracted topic turning points. "10: 00: 00-10: 06: 00, Future Library, Construction, Central", "10: 06: 01-10: 17: 20, Library, Future Library, Land" or "10: 17: 21- 10:30, library, request, self-defense force "and the like are stored in the storage unit 22.

[Processing by Index Generation Device (Example 1)]
Next, processing by the index generation device will be described with reference to FIG. FIG. 11 is a flowchart showing the flow of index generation processing.

  As shown in FIG. 11, when audio content (audio data) is input and an index generation instruction is accepted (Yes in step S1101), the audio recognition unit 23a of the index generation device 20 detects the input audio content from audio recognition. Text data 22a is generated and stored in the storage unit 22 (step S1102).

  Subsequently, the utterance word list generation unit 23b of the index generation device 20 reads the generated and stored text data 22a from the storage unit 22, extracts the words included in the text data 22a and the utterance time of the words, An utterance word list 22b in which the word and the utterance time are associated and listed is generated and stored in the storage unit 22 (step S1103).

  Then, the topic word extraction unit 23c of the index generation device 20 reads the utterance word list 22b generated and stored by the utterance word list generation unit 23b from the storage unit 22, and performs predetermined processing for each word included in the utterance word list 22b. Utterance density indicating the number of utterances per unit time is calculated, and words whose utterance density exceeds a predetermined value are extracted as topic words in unit time (step S1104).

  Subsequently, the topic word list generation unit 23d associates the topic words extracted by the topic word extraction unit 23c in association with each predetermined unit time from the utterance start time to the utterance end time of the audio content. A word list 22c is generated and stored in the storage unit 22 (step S1105).

  Then, the topic word turning point extraction unit 23e reads the topic word list 22c generated by the topic word list generation unit 23d from the storage unit 22, and extracts a time when no enumeration of topic words continues as a topic turning point. (Step S1106).

  The index generation unit 23f generates an index from the topic words listed in each section in the topic word list for each section delimited by the topic conversion points extracted by the topic word conversion point extraction unit 23e (step S1107).

[Effects of Example 1]
As described above, according to the first embodiment, a word included in text data and an utterance time of the word are extracted, an utterance word list in which the word and the utterance time are associated with each other is generated, and the generated utterance is generated. Calculates the utterance density indicating the number of utterances per unit time for each word included in the word list, extracts words whose utterance density exceeds a predetermined value as topic words in the unit time, and starts utterance of audio content For each predetermined unit time from the time to the utterance end time, a topic word list in which the extracted topic words are associated and enumerated is generated, and in the generated topic word list, all of the topic words are enumerated. Time is extracted as a topic turning point, and an index is generated from the topic words listed in each section in the topic word list for each section delimited by the topic turning point Therefore, it is possible to identify a topic in a certain time zone for audio content, assign an index for identifying the topic between the topics, and assign another index between other topics. It is possible to create a sufficiently useful and detailed index for the entire content.

  Further, according to the first embodiment, in the generated topic word list, a time when none of the topic words are enumerated is extracted as a topic turning point, and for each section divided by the extracted topic turning point, the topic word Since the index is generated from the topic words listed in each section in the list, multiple topics included in the audio content can be extracted, and multiple indexes can be created for each topic. It is possible to create a more useful and detailed index.

  The embodiment of the present invention has been described in the first embodiment. However, the present invention may be implemented in various different forms other than the above-described embodiment. Therefore, using FIG. 12 to FIG. 14, (1) utterance density calculation for words that are semantically similar, (2) topic word extraction from a topic word list prepared in advance, (3) slightly discontinuous time zone Are considered to be continuous, and will be described separately.

  FIG. 12 is a diagram illustrating an example of a distance dictionary according to the second embodiment, FIG. 13 is a diagram illustrating a configuration example of information stored in an utterance word list prepared in advance, and FIG. It is a figure for demonstrating that the discontinuous time slot | zone of the topic word which concerns on is considered as the same time slot | zone.

(1) Calculation of utterance density assuming that words that are semantically similar are the same For example, in the first embodiment, the case has been described where the utterance density of words included in the generated utterance word list is calculated and topic words are extracted. The invention is not limited to this, and a topic word may be extracted by calculating the utterance density of words that are semantically close to words included in the utterance word list.

  Specifically, the index generation device stores a distance dictionary that prescribes a mutual semantic distance between a plurality of words in advance, and utterance density per unit time for a predetermined word included in the utterance word list In the stored distance dictionary, the utterance density per unit time is assumed to be the same word as the predetermined word even if the semantic distance from the predetermined word is within the predetermined range. Is calculated.

  For example, as shown in FIG. 12, the index generation device, when a word close to “library” (collection, dictionary, IT, location, user, book, etc.) is included in the utterance word list, actually Assuming that a “library” that has not been uttered and is not included in the utterance word list has been uttered, the utterance density per unit time is also calculated for this “library”, and the topic word list Is created, topic turning points are extracted, and an index is generated.

  Thereby, when storing a distance dictionary that defines the mutual semantic distance between a plurality of words, and calculating the utterance density per unit time for a predetermined word included in the utterance word list, the stored distance Since the utterance density per predetermined unit time is calculated assuming that a word whose semantic distance is within a predetermined range in the dictionary is also the same word as the predetermined word, for example, audio content “Museum”, “Citizen”, etc. are determined as topic words from the words included in the word, and “construction plan” and “construction place” that were not included in the audio content are semantically close to “Museum” and “Citizen” By calculating the utterance density based on the word, it is possible to identify the topic words “Museum” and “Citizen” more. It is possible to create a sufficiently useful and detailed index.

(2) Extracting Topic Words from Preliminary Topic Word Lists In Example 1, a case where words whose utterance density calculated for each word included in the utterance word list exceeds a predetermined value is extracted as topic words in unit time. As described above, the present invention is not limited to this, and a word whose calculated utterance density exceeds a predetermined value is extracted as a topic word in a unit time on condition that it is included in a pre-stored topic word list. You may make it do.

  Specifically, the index generation device stores the topic word extraction list as shown in FIG. 13 and is the utterance density on the condition that the words are listed in the stored topic word extraction list. Words that exceed a predetermined value as topic words in unit time

  For example, for words whose utterance density calculated for each word included in the utterance word list exceeds a predetermined value, a word list as shown in FIG. 13 (for example, decentralization, home, traffic safety, tax, library) If it is included, it is extracted as a topic word.

  As a result, a topic word extraction list in which the words to be extracted as written topic words are stored is stored, and the utterance density is set to a predetermined value on condition that the words are listed in the stored topic word extraction list. Since more words are extracted as topic words in unit time, for example, specific words such as “employment”, “beach”, “volunteer” are prepared in advance, and whether or not the extracted words are these specific words As a result of more accurately specifying the topic word, it is possible to extract the topic turning point more accurately.

(3) Slightly discontinuous time zones are regarded as continuous. In the first embodiment, the case where a time when enumeration of all topic words does not continue is extracted as a topic turning point in the generated topic word list has been described. However, the present invention is not limited to this, and when the same topic word is enumerated without a predetermined time interval, the enumeration of the topic word may be continued. Good.

  Specifically, as shown in FIG. 14, when the same topic word is listed without a predetermined time interval ((1) to (3) in FIG. 14), as shown in FIG. Then, a topic turning point is extracted as the enumeration of the topic words (land) continues.

  Thereby, when the same topic word is enumerated without leaving a predetermined time interval in the topic word list, the topic turning point is extracted as the enumeration of the topic word continues, so for example, discrete If a topic word is extracted, the sections in which the topic word is detected can be merged to extract the topic as one section. As a result, unnecessary information is extracted from the entire index by extracting the topic words in detail. Can be prevented from being extracted.

  In the embodiment described above, a case has been described in which topic words are extracted from an utterance word list created from text data. However, the present invention is not limited to this, and word removal and words are added to the created utterance word list. You may make it extract a topic word using the result of processing, such as a complement.

  Therefore, in the following, as Example 3, an example in which topic words are extracted using the results of processing such as word removal and word completion on the created utterance word list will be described. In the following, the outline and features of the index generation device according to the third embodiment, the configuration of the index generation device, and the effects of the third embodiment will be described.

[Outline and Features of Index Generation Device (Example 3)]
Next, the outline and characteristics of the index device according to the third embodiment will be described with reference to FIG. FIG. 15 is a diagram illustrating an overview and characteristics of the index generation device according to the third embodiment.

  As illustrated in FIG. 15, the index generation device 40 stores a distance dictionary that defines a semantic distance between a plurality of words in the storage unit 22. Then, as in the first embodiment, when audio content (audio data) is input and an index creation instruction is received, the index generation device 40 performs speech recognition of the input audio content.

  Subsequently, the index device 40 extracts a word included in the text data created by speech recognition and the utterance time of the word, and generates an utterance word list in which the word and the utterance time are associated and listed. To give a specific example, the index generation device 40 associates a word and an utterance time, and “9:00: 00, word A”, “9:06:10, word B”, “9:20:00, An utterance word list such as “word C”, “9:25:00, word D”, “9:30, 0, word E” is created.

  And the index production | generation apparatus 40 evaluates the semantic distance between the words contained in the produced | generated utterance word list | wrist using the distance dictionary (refer FIGS. 17-19) memorize | stored in the memory | storage part 42, The said evaluation Remove low words from the spoken word list. Specifically, the index generation device 40 includes “9:00: 00, word A”, “9:06:10, word B”, “9:20:” included in the created utterance word list. “00, word C”, “9:25:00, word D”, “9:30, 00, word E”, and using the distance dictionary stored in the storage unit 22 (see FIGS. 17 to 19), "9:06:10, word B" which is far away is deleted. As a result, in the utterance word list, “9:00: 00, word A”, “9: 20:00, word C”, “9: 25:00, word D”, “9:30: 00, word E” are included. It will be memorized.

  More specifically, for example, when the utterance word list includes “art museum”, “art”, and “bipolarization”, as shown in FIG. 19, the distance is semantically close to “art museum” and “art”. “Bipolarization” deletes “bipolarization” from the utterance word list on the basis that both “art museum” and “art” are semantically distant. Note that “semantically close” or “semanticly far” means that if they are directly connected or are adjacent to each other in FIG. In this case, it is determined that “it is semantically far”.

  Subsequently, the index generation device 40 uses the distance dictionary (see FIGS. 17 to 19) stored in the storage unit 42 to evaluate the semantic distance from the words listed in the utterance word list after word removal. A high word is extracted from the distance dictionary, and the extracted word is complemented to the utterance word list. More specifically, the index generation device 40 uses the “word” that is semantically close to “word A” and “word C” included in the utterance word list from which “word B” has been removed. F ”is extracted using the distance dictionary stored in the storage unit 22 and added to the spoken word list.

  More specifically, for example, when “art museum” and “art” are included in the utterance word list, “activity” is meaningful for each of “art museum” and “art” as shown in FIG. "Activity" is added to the utterance word list on the basis of being close to.

  Then, the index generation device 40 calculates an utterance density indicating the number of utterances per predetermined unit time for each word included in the generated utterance word list, and selects words whose utterance density exceeds a predetermined value in the unit time. A topic word list that is extracted as topic words and enumerated in association with the extracted topic words is generated for each predetermined unit time from the speech start time to the speech end time of the audio content. Here, specific processing such as utterance density calculation is the same as that in the first embodiment, and thus detailed description thereof is omitted. However, in the above example, “word A” “word C” “word D” It is assumed that “word E” and “word F” are extracted as topic words.

  And the index production | generation apparatus 40 extracts the time when none of a topic word enumeration continues as a topic turning point in the produced | generated topic word list. Note that topic turning points are extracted by plotting topic words as shown in FIG. 4 as described in the first embodiment. Here, it is assumed that the appearance time of “word C” is extracted as a topic turning point.

  Then, the index generation device 40 calculates the utterance frequency of each topic word listed in each section in the topic word list for each section delimited by the extracted topic turning points, and the utterance frequency exceeds a predetermined value. Topic words are extracted as important words, and an index listing the topic words extracted as important words is generated for each section delimited by topic turning points. As a specific example, words whose utterance frequency is less than “1” and words included in the unnecessary word list are deleted and extracted.

  Specifically, in the above example, “word A” and “word F” in the section “section 1” divided by the extracted topic turning point “word C” have an utterance frequency of “1” or more, and “ Assuming that “word C”, “word D”, “word E”, and “word F” in section 2 also have an utterance frequency of “1” or higher and “word D” is listed in the unnecessary word list, an index is generated. The device 40 generates an index in which “word A”, “word F” in “section 1” and “word C”, “word E”, and “word F” in “section 2” are extracted and listed as topic words.

  Subsequently, the index generating device 40 edits the topic word based on one or more of the utterance time, the utterance density, the utterance frequency, and the semantic distance for each section delimited by the topic turning point. Generate. More specifically, for example, the index generation device 40 displays the index generated by extracting topic words by “displaying from the left in order of speaking time from the left” or “changing the font size in descending order of speaking density”. The index is edited by combining any one or more of utterance time, utterance density, utterance frequency, and semantic distance (see FIGS. 38 and 39).

  For this reason, according to the third embodiment, it is possible to specify a topic word more accurately than in the case where a topic word is extracted from extracted word data in which words that are semantically distant are not deleted. . In addition, it is possible to specify a topic word more accurately than in the case where words that are semantically close are not complemented. Furthermore, it is possible to create a useful word as an index for each topic, and it is possible to display an index according to the purpose of use, such as searching the Internet or searching for an agenda such as conference recording data.

[Configuration of Indexing Device (Example 3)]
Next, the configuration of the index generation device 40 shown in FIG. 15 will be described with reference to FIGS. 16 is a block diagram showing the configuration of the index generation device 40, FIGS. 17 to 19 are diagrams showing examples of distance dictionaries, and FIG. 20 is a diagram showing examples of utterance word extraction conditions, 21 is a diagram illustrating an example of the utterance word extraction result, FIG. 22 is a diagram illustrating an example of the shortest distance calculation condition, and FIG. 23 is a diagram illustrating an example of the shortest distance calculation result. FIG. 25 is a diagram showing an example of keyword extraction conditions, FIG. 25 is a diagram showing an example of evaluation based on the cumulative number by shortest distance between words, and FIG. 26 is a word extraction result (utterance word list) after word deletion FIG. 27 is a diagram showing an example of keyword extraction conditions, and FIG. 28 is a diagram showing an example of evaluation based on average distance and shortening distance.

  FIG. 29 is a diagram illustrating an example of a word extraction result (utterance word list) after word deletion, FIG. 30 is a diagram illustrating an example of complementary word extraction conditions, and FIG. 31 is a diagram after word completion. FIG. 32 is a diagram illustrating an example of a word extraction result (utterance word list), FIG. 32 is a diagram illustrating an example of an utterance density count condition, and FIG. 33 is a diagram illustrating an example of an utterance density count result. FIG. 35 is a diagram illustrating an example of important word extraction conditions, FIG. 35 is a diagram illustrating an example of a word list that should not be extracted as an index, and FIG. 36 is a word extraction after deleting a word that should not be extracted as an index. FIG. 37 is a diagram illustrating an example of a result, FIG. 37 is a diagram illustrating an example of an index generation condition, and FIGS. 38 and 39 are diagrams illustrating an example of index generation.

  As shown in the figure, the index generation device 40 includes a communication control I / F unit 41, a storage unit 42, and a control unit 43. The storage unit 42 includes text data 42a, an utterance word list 42b, a topic word list 42c, and an index 42d. The control unit 43 includes a voice recognition unit 43a, an utterance word list generation unit 43b, and the like. The topic word extracting unit 43c, the topic word list generating unit 43d, the topic word turning point extracting unit 43e, the index generating unit 43f, the word removing unit 51, the word complementing unit 52, and the keyword extracting unit 53 Prepare. In the following, description will be given along the processing flow of the above-mentioned “overall configuration of the index generation device”.

  The utterance word list generation unit 43b corresponds to the “utterance word list generation unit” described in the claims, and similarly, the topic word extraction unit 43c corresponds to the “topic word extraction unit”. The list generation unit 43d corresponds to the “topic word list generation unit”, the topic word turning point extraction unit 43e corresponds to the “topic word turning point extraction unit”, and the index generation unit 43f corresponds to the “index generation unit”. Correspondingly, the word removal unit 51 corresponds to the “word removal unit”, the word complementing unit 52 corresponds to the “word complementing unit”, and the important word extraction unit 53 corresponds to the “important word extracting unit”.

  The speech recognition unit 43a, the utterance word list generation unit 43b, the text data 42a, and the topic word list 42c are the speech recognition unit 23a, the utterance word list generation unit 23b, the text data 22a described in the first embodiment, Since it has the same function as the topic word list 22c, its detailed description is omitted here.

  The utterance word list 42b stores data in which the words included in the text data and the utterance times of the words are extracted by the utterance word list generation unit 43b and associated with the words and the utterance times. As a specific example, as shown in FIG. 20, the utterance word list 42 b includes “processing execution unit (step), 1” as “item indicating setting item,“ value indicating setting value ””, and the like. The word extraction condition of is retained. Then, based on the stored word extraction conditions, as shown in FIG. 21, “extracted word node number indicating the extracted word node number”, “extracted word” indicating the extracted word, and the time when the word appears As “time information” to be shown, “1, city, 16:29:00”, “2, opinion, 16:29:10” and the like are stored. Note that the word extraction conditions shown in FIG. 20 can be arbitrarily changed. Here, since “value” is set to “1”, word extraction is executed step by step (one word at a time). .

  The distance dictionary storage unit 50 stores a distance dictionary that defines a semantic distance between a plurality of words. Specifically, as shown in FIG. 17, “a node number that uniquely identifies a word, a“ node name ”that indicates the name of the word, a“ part of speech ”that indicates the part of speech of the word, and evaluation information As “the average distance of all nodes” to be used, “1, city, noun, 2.971”, “2, opinion, noun, 3.429”, etc. are stored. Further, as shown in FIG. 18, “1, 9”, “2, 4”, etc. are stored as “link source node number, link destination node number indicating the semantic connection relationship between words”. As shown in FIG. 19, a knowledge network or the like illustrating the semantic distance between words is stored based on FIG. 17 and FIG.

  The word removal unit 51 uses the distance dictionary stored in the distance dictionary storage unit 50 to evaluate the semantic distance between words included in the utterance word list 42b generated by the utterance word list generation unit 43b, and performs the evaluation. Remove low words from the spoken word list. Specifically, a semantic distance is evaluated by either “an evaluation based on the cumulative number of words by the shortest distance between words” or “an evaluation based on an average distance and a shortened distance of words”, and a word with a low evaluation is selected from the utterance word list. Remove.

(Calculate shortest distance between words)
First, “shortest distance calculation between words” will be described. As shown in FIG. 22, the word removal unit 51 reads “a distance calculation window width indicating the maximum number of words held in a memory (not shown) for the shortest distance calculation result”. “5, 4” and the like as the “calculation target shortest distance” indicating the maximum value for calculating the shortest distance between words. When the number of input words exceeds the “distance calculation window width”, the oldest one is overwritten, and when the shortest distance calculation result exceeds the “calculation target shortest distance”, “N” is input. Note that it is not always necessary to hold this, and it may be set manually for each process execution.

  Then, the word removal unit 51 calculates the “shortest distance between words” for each word stored in the utterance word list 42 b using the distance dictionary stored in the distance dictionary storage unit 50 and the shortest distance calculation condition held. And stored in the memory as “shortest distance data”. Specifically, as shown in FIG. 23, the shortest distance between words for each word stored in the utterance word list 42b is calculated as “region, node name, node number, time information, shortest distance”. For example, the word removal unit 51 searches from both sides of the node number of the start point and the node number of the end point based on the two node numbers to be calculated, using FIG. Let the number be the shortest distance.

  The above example will be described more specifically. In FIG. 19, the semantic distance between “city” and “opinion” is “3”, and the semantic distance between “opinion” and “cholera” is “4”. It can be seen that it is. As described above, the word removing unit 51 calculates the semantic distance (shortest distance) between the words using FIG. 19, and calculates the distance between the words for each word stored in the utterance word list 42 b illustrated in FIG. 23. The shortest distance is calculated, and the “shortest distance data” is stored in the memory. The shortest distance calculation may be performed using a known Dijkstra method or a Warsal Floyd method.

(Evaluation based on the cumulative number of shortest distances between words)
Next, the evaluation based on the cumulative number of shortest distances between words will be described. The word removal unit 51 reads the calculated “shortest distance data” from the memory, and calculates the cumulative number of shortest distances between words included in the utterance word list 42b. A value counted for each distance is used as an evaluation for each word, and a high evaluation is extracted as a keyword and stored in the utterance word list 42b. That is, the word removal unit 51 deletes words that have not been extracted as keywords from the utterance word list 42b on the basis of low evaluation.

  More specifically, as shown in FIG. 24, the word removal unit 51 performs “an“ evaluation window width ”indicating the maximum number of words to be subjected to keyword extraction evaluation, and“ cumulative distance 0 ”as a keyword extraction evaluation target. “Number of times”, “cumulative number of distance 1” as a keyword extraction evaluation object, “accumulation number of distance 2” as a keyword extraction evaluation object, and “cumulative number of distance 3” as a keyword extraction evaluation object ” The keyword extraction condition such as “1, 1, 3” is held. Note that it is not always necessary to hold this, and it may be set manually for each process execution.

  Under such a configuration, the word removal unit 51 uses the keyword extraction condition to be held and the distance dictionary storage unit 50 based on the distance dictionary (FIGS. 17 to 19) to shorten the shortest distance between words included in the utterance word list 42 b. Evaluate the cumulative number of distances by distance. The result is shown in FIG. From this result, if any of the cumulative number of distances for each word is equal to or greater than the cumulative number of extraction targets for each distance, the word removal unit 51 extracts the corresponding word as a keyword.

  For example, if keyword extraction is performed based on the above keyword extraction conditions (shortest distance 0 = “1”, shortest distance 1 = “1”, shortest distance 2 = “1”, shortest distance 3 = “3”), From (1) to (3) in FIG. 25, since the cumulative number by distance of each word (city, opinion, cholera) does not exceed the cumulative number of extraction targets by distance, any word is extracted as a keyword. Not. Then, in (4) of FIG. 25, the word removal unit 51 sets “3” as the “shortest distance cumulative count” of the area 1 “city” to be equal to or greater than the keyword extraction condition (shortest distance 3 = “3”). To extract “city” as a keyword. Similarly, the word removal unit 51 sets “2, opinion” in the areas 2 and 4 to “1” for “the shortest distance accumulation count 1”, which is equal to or greater than the keyword extraction condition (shortest distance 1 = “1”). , Each is extracted as a keyword. Thus, the result extracted as a keyword is shown in FIG.

  As can be seen by comparing FIG. 25 before removing words that are semantically far away from FIG. 26 after removal, the word removal unit 51 extracts “city”, “opinion”, “citizen”, and “museum” as keywords. However, “cholera” and “disappointment” are not extracted as keywords. That is, the word removal unit 51 deletes “cholera” and “disappointment” from the spoken word list 42b.

(Evaluation based on average word distance and shortened distance)
Next, the evaluation based on the average word distance and the shortened distance will be described. The word removing unit 51 reads the calculated “shortest distance data” from the memory, and calculates the average distance between the words included in the utterance word list 42b and the shortened distance. Are evaluated and extracted as keywords, and stored in the spoken word list 42b.

  More specifically, as shown in FIG. 27, the word removing unit 51 “the“ evaluation window width ”indicating the maximum number of words to be subjected to keyword extraction evaluation, and the distance calculation result used for calculating the average distance. “Number of average target distance calculation results” indicating the maximum number, “Extraction target average distance” indicating the average distance for keyword extraction, The keyword extraction condition such as “shortened distance” ”“ 5, 3, 3.0, 1.0 ”is held.

  Based on such a configuration, the word removal unit 51 uses the keyword extraction condition to be held and the distance dictionary storage unit 50 based on the distance dictionary (FIGS. 17 to 19) and the average distance of words included in the utterance word list 42b. And evaluation by shortening distance. The result is shown in FIG. From this result, when the “average distance value is equal to or smaller than the extraction target average distance setting value” or “the shortening distance value is equal to or larger than the extraction target shortening distance setting value” for each word, the word removing unit 51 Is extracted as a keyword.

  Here, when calculating the shortest distance average, the average target distance calculation result number of the keyword extraction condition is read, and if there are more shortest distance calculation results than this set value in the evaluation window, the upper (short distance) average target The number of distance calculation results is used as an average calculation target. In addition, in the shortest distance average calculation, a calculation result with a distance calculation result “0” is an average object, and a calculation result with a distance calculation result “N” is not an average calculation object. Thus, the result extracted as a keyword is shown in FIG.

  As can be seen by comparing FIG. 28 before removing semantically distant words and FIG. 29 after removing words, the word removal unit 51 extracts “city”, “opinion”, “citizen”, and “museum” as keywords. , “Cholera” and “Disappointment” are not extracted as keywords. That is, the word removal unit 51 deletes “cholera” and “disappointment” from the spoken word list 42b.

  In the above-mentioned “evaluation based on the cumulative number of shortest distances between words” and “evaluation based on average distance and shortening distance of words”, general network analysis such as average distance of all nodes, number of links, distance centrality, and mediation centrality The index value used may be used as a keyword extraction condition. Further, an upper limit may be set on the number of keywords to be extracted, and the keywords may be extracted in order of the index values used in the general network analysis described above.

  Returning to the description of FIG. 16, the word complementing unit 52 uses the distance dictionary stored in the storage unit 42 to select words having a high semantic distance evaluation with the words listed in the spoken word list 42 b after the word removal. Extract from the distance dictionary and complement the extracted word to the spoken word list. Specifically, words that are close in semantic distance to the words included in the utterance word list 42b extracted by the word removal unit 51 shown in FIG. 29 are extracted using the distance dictionary shown in FIGS. The utterance word list 42b is complemented.

  For example, as shown in FIG. 30, the word complementing unit 52 outputs “the number of extraction target distances indicating the maximum distance from the keyword to be extracted” and the distance from the keyword “1”. The “distance 1 output upper limit number” indicating the maximum number and the “distance 2 output upper limit number” indicating the maximum output number with the distance from the keyword “2” are held as extraction conditions such as “2, 1, 0”.

  With such a configuration, the word complementing unit 52 searches the distance dictionary for words having a high semantic distance evaluation with the words listed in the utterance word list 42b extracted by the word removing unit 51 shown in FIG. The extracted word is complemented to the utterance word list. The complemented results are shown in FIG.

  As can be seen by comparing FIG. 29 before supplementing words that are close in meaning and FIG. 31 after completion, the word completion unit 52 is “activity” as a word having a meaning similar to “museum, opinion, citizen”. It is understood that “art” is complemented. The words complemented in this way are stored in the utterance word list 42b.

  Returning to the description of FIG. 16, the topic word extraction unit 43 c is included in the utterance word list 42 b in which words having a long semantic distance are deleted by the word removal unit 51 and words having a short semantic distance are complemented by the word completion unit 52. For each word, an utterance density indicating the number of utterances per predetermined unit time is calculated, and words whose utterance density exceeds a predetermined value are extracted as topic words in the unit time.

  Subsequently, the topic word turning point extraction unit 43e selects words having an utterance density exceeding a predetermined value on the condition that the words are listed in the topic word extraction list stored in the topic word list 42c. Extracted as a topic word. Note that the topic word extraction unit 43c and the topic word turning point extraction unit 43e have the same functions as the topic word extraction unit 23c and the topic word turning point extraction unit 23e described in the first embodiment. Description is omitted.

  The keyword extraction unit 53 determines the utterance frequency of each topic word listed in each section in the topic word list 42c for each section divided by the topic word (topic conversion point) extracted by the topic word conversion point extraction section 43e. A topic word whose utterance frequency exceeds a predetermined value is extracted as an important word.

  More specifically, as shown in FIG. 32, the keyword extraction unit 53 reads “a totaling period indicating a data summing period, a“ totaling time width ”indicating a time width before and after the period time, an association “Associative word count necessity setting” indicating whether or not to extract words is stored as a count condition such as “2 (minutes), 1 (minutes), 1”. An associative word indicates a word that is semantically close to a certain word. When “associative word count necessity setting” is “0”, only keywords (important words) are extracted, and when “1”, both keywords and associative words are extracted. Further, the important word extraction unit 53 is a word list (unnecessary word list) that should not be extracted in advance as an index as shown in FIG. 35, for example, words unsuitable for specifying an index such as “self” or “people”. Hold.

  Further, as shown in FIG. 34, the important word extraction unit 53 displays “an index word extraction target frequency indicating the minimum cumulative frequency extracted as an index word” and “non-extraction target word exclusion setting” indicating the non-extraction target word exclusion necessity setting. Key word extraction conditions such as “2, 1, 1” are held for “word exclusion necessity setting” and “neighboring word aggregation necessity setting” indicating association word extraction conditions. Further, when the “exclusion target word exclusion necessity setting” is “0”, the unnecessary word list is not considered, and when it is “1”, the unnecessary word list is considered. Further, when “neighboring word aggregation necessity setting” is “0”, neighboring words are not extracted, when “1”, only the neighboring words with the distance “1” are extracted, and when “2”, the distance “2”. ”Is extracted, and in the case of“ 3 ”, only the neighborhood word of distance“ 3 ”is extracted. Note that the neighborhood word has the same meaning as the association word described above.

  In such a configuration, when the topic turning point is “16:30,” the keyword extraction unit 53 is divided by the topic turning point based on the totaling condition held (see FIG. 32). If the utterance frequency of each topic word listed in each section in the topic word list 42c is calculated for each section ("16: 28: 0 to 16:30:30"), a result as shown in FIG. 33 is obtained. In FIG. 33, “time information, node number, total word, cumulative frequency”, “16:28:10, 4, citizen, 7,” or “16:29:40, 58, awareness” is obtained. .

  Subsequently, the keyword extraction unit 53 determines the topic whose utterance frequency exceeds a predetermined value (for example, 1) from FIG. 33 based on the keyword extraction condition (see FIG. 34) and the unnecessary word list in FIG. When a word is extracted as an important word, the result shown in FIG. 36 is obtained. 36, the words whose utterance frequency is less than “1” and the words included in the unnecessary word list of FIG. 35 are deleted and extracted from FIG.

  Returning to the description of FIG. 16, the index generation unit 43f edits the topic word based on one or more of the utterance time, the utterance density, the utterance frequency, and the semantic distance for each section delimited by the topic turning points. To generate an index.

  More specifically, the index generation unit 43f generates indexes such as “2, 1” as “generation order setting indicating index generation order” and “font setting” indicating display font as shown in FIG. The conditions are held, and indexes such as FIG. 38 and FIG. 39 are generated according to the conditions. When “Generation order setting” is “0”, the items are displayed in ascending order of appearance times. When “1”, the items are displayed in descending order of the cumulative frequency. When “2”, the appearance time is early and the cumulative frequency is displayed. Display in descending order. When “Font setting” is “0”, the fonts are generated with the same size, and when “1”, they are generated with a font size proportional to the cumulative frequency.

  When adjacent words that are semantically close are extracted, for example, two indexes having a dependency relationship, such as merging into “opposite”, “opinion”, and “opposite opinion”, are integrated into one word. Chunk processing may be performed.

  Returning to the description of FIG. 16, the index 42d stores the index created by the index generation unit 43f. To give a specific example, as shown in FIG. 38, “time zone (topic section) and index” “16: 27-30, art museum, construction plan, citizen, approval, disagreement, opinion, construction location , Activity, art, regular meeting, problem, talk, plan, etc., and the index generated by the index generation unit 43f changing the index generation order and display font. Also, as shown in FIG. 39, the entire audio content is divided into a plurality of topics (sections), and indexes are stored corresponding to each.

[Effects of Example 3]
As described above, according to the third embodiment, a distance dictionary that defines a mutual semantic distance between a plurality of words is stored, and words between words included in the generated utterance word list are stored using the stored distance dictionary. Utterance per predetermined unit time for each word included in the utterance word list based on the utterance word list after the word is removed. Since the density is calculated and words whose utterance density exceeds a predetermined value are extracted as topic words in unit time, for example, extraction is performed by deleting “cholera” that is semantically far from “museum”, “cholera”, “art”, etc. As a result of being able to create word data and identifying topics from this extracted word data, compared to extracting topic words from extracted word data that does not delete words that are semantically far It is possible to more accurately identify the topic words.

  Further, according to the third embodiment, using the stored distance dictionary, a word having a high semantic distance evaluation with words listed in the utterance word list after word removal is extracted from the distance dictionary and extracted. A word is complemented to the utterance word list, and an utterance density per unit time is calculated for each word included in the utterance word list based on the utterance word list after word completion, and the utterance density is a predetermined value. For example, if “art museum” or “construction place” is extracted from the audio content, but the word “construction plan” is not extracted, “construction plan” is extracted. “Museum” and “construction site” are presumed to be semantically close to each other, so this “construction plan” was not extracted from the audio content, such as newly extracted, but semantically The results can be complemented words have, as compared with the case that did not complement the semantically close words, it is possible to identify more accurately topic words.

  Further, according to the third embodiment, for each section delimited by the extracted topic turning points, the utterance frequency of each topic word listed in each section in the topic word list is calculated, and the utterance frequency is set to a predetermined value. The topic words that are extracted as important words are extracted as important words, and an index that lists the topic words extracted as important words is generated for each section delimited by topic turning points. As a result of extracting words that are important for specifying a topic in a certain section, it is possible to create more useful words as indexes for each topic.

  Further, according to the third embodiment, for each section delimited by the topic turning point, the index is edited by editing the topic word based on one or more of the utterance time, the utterance density, the utterance frequency, and the semantic distance. Since it is generated, for example, it is possible to display the index according to the purpose of use such as searching the Internet and agenda searching such as conference recording data, such as displaying the index by increasing the display font in order of appearance time and appearance frequency It is.

  Although the embodiments of the present invention have been described so far, the present invention may be implemented in various different forms other than the embodiments described above. Therefore, another embodiment included in the present invention will be described below as a fourth embodiment.

(1) Processing order For example, in Example 3, after the utterance word list was created, the word removal processing and the word complementing processing were performed, and then the topic word extraction processing was performed. However, the present invention is limited to this. Instead, only one of the processes may be performed. For example, after creating an utterance word list, only word removal processing may be performed, and topic word extraction is performed based on the utterance word list after performing one or both of word removal processing and word completion processing. An index may be created without performing processing, and the processing order can be changed as needed.

(2) System configuration etc. In addition, among the processes described in the present embodiment, all or part of the processes described as being performed automatically can be performed manually or manually. All or a part of the processing described as a thing can also be automatically performed by a known method. In addition, the processing procedure, control procedure, specific name, information including various data and parameters shown in the above document and drawings (setting conditions described in the first and second embodiments, various lists to be extracted, indexes) About can be changed arbitrarily unless otherwise specified.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. The topic word extraction unit and the topic word list generation unit may be integrated. Further, all or any part of each processing function performed in each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.

  In the above-described embodiments, the index generation apparatus that implements the present invention has been described in terms of functions. However, each function of the index generation apparatus may be realized by causing a computer such as a personal computer or a workstation to execute a program. it can. That is, the various processing procedures described in the above embodiments can be realized by executing a program prepared in advance on a computer. These programs can be distributed via a network such as the Internet. Further, these programs can be recorded on a computer-readable recording medium such as a hard disk, a flexible disk (FD), a CD-ROM, an MO, and a DVD, and can be executed by being read from the recording medium by the computer. In other words, for example, a CD-ROM (which may be a separate CD-ROM for each apparatus) storing the index generation program as shown in the embodiment is distributed and stored in this CD-ROM. The program may be read and executed by each computer.

  As described above, the index generation device, the index generation method, and the index generation program according to the present invention generate speech data by recognizing speech content, and generate an index to be given to the speech content from the text data. In particular, it is suitable for creating a sufficiently useful and detailed index for the entire audio content.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a diagram for explaining an overview and characteristics of an index generation device according to Embodiment 1; It is a block diagram which shows the structure of the index apparatus which concerns on Example 1. FIG. It is a figure which shows the structural example of the information memorize | stored in an audio | voice content. It is a figure which shows the structural example of the information memorize | stored in text data. It is a figure which shows the structural example of the information memorize | stored in text data. It is a figure which shows the structural example of the information memorize | stored in an utterance word list. It is a figure which shows the example of the calculation method of the speech density which concerns on Example 1. FIG. It is a figure which shows the structural example of the information memorize | stored in a topic word list. It is a figure which shows the example of the calculation method of a topic turning point. It is a figure which shows the structural example of the information memorize | stored in an index. 7 is a flowchart illustrating a flow of processing in the index generation device according to the first embodiment. It is a figure which shows the example of the distance dictionary which concerns on Example 2. FIG. It is a figure which shows the structural example of the information memorize | stored in the utterance word list prepared beforehand. It is a figure for demonstrating that the discontinuous time slot | zone of the topic word which concerns on Example 2 is considered as the same time slot | zone. It is a figure for demonstrating the outline | summary and the characteristic of the index production | generation apparatus which concern on Example 3. FIG. It is a block diagram which shows the structure of the index apparatus which concerns on Example 3. FIG. It is a figure which shows the example of a distance dictionary. It is a figure which shows the example of a distance dictionary. It is a figure which shows the example of a distance dictionary (knowledge network). It is a figure which shows the example of the utterance word extraction conditions. It is a figure which shows the example of the utterance word extraction result. It is a figure which shows the example of the shortest distance calculation conditions. It is a figure which shows the example of the shortest distance calculation result. It is a figure which shows the example of keyword extraction conditions. It is a figure which shows the example of the evaluation by the frequency | count of accumulation according to the shortest distance between words. It is a figure which shows the example of the word extraction result (utterance word list) after word deletion. It is a figure which shows the example of keyword extraction conditions. It is a figure which shows the example of evaluation by an average distance and shortening distance. It is a figure which shows the example of the word extraction result (utterance word list) after word deletion. It is a figure which shows the example of a complementary word extraction condition. It is a figure which shows the example of the word extraction result (utterance word list) after word complementation. It is a figure which shows the example of speech density totalization conditions. It is a figure which shows the example of an utterance density total result. It is a figure which shows the example of an important word extraction condition. It is a figure which shows the example of the word list which should not be extracted as an index. It is a figure which shows the example of the word extraction result after the word deletion which should not be extracted as an index. It is a figure which shows the example of an index production | generation condition. It is a figure which shows the example of an index production | generation. It is a figure which shows the example of an index production | generation.

Explanation of symbols

DESCRIPTION OF SYMBOLS 20 Index production | generation apparatus 21 Communication control I / F part 22 Storage part 22a Text data 22b Utterance word list 22c Topic word list 22d Index 23 Control part 23a Speech recognition part 23b Utterance word list generation part 23c Topic word extraction part 23d Topic word list generation Unit 23e topic word turning point extraction unit 23f index generation unit 40 index generation device 41 communication control I / F unit 42 storage unit 42a text data 42b utterance word list 42c topic word list 42d index 43 control unit 43a speech recognition unit 43b utterance word list Generating unit 43c topic word extracting unit 43d topic word list generating unit 43e topic word turning point extracting unit 43f index generating unit 50 distance dictionary storage unit 51 word removing unit 52 word complementing unit 53 important word extracting unit

Claims (13)

An index generation device that generates speech data by recognizing speech content and generates an index to be given to the speech content from the text data,
An utterance word list generating means for extracting a word included in the text data and an utterance time of the word, and generating an utterance word list in which the word and the utterance time are associated and listed;
An utterance density indicating the number of utterances per predetermined unit time is calculated for each word included in the utterance word list generated by the utterance word list generation means, and words whose utterance density exceeds a predetermined value are calculated in the unit time. Topic word extraction means for extracting as a topic word;
Topic word list generation for generating a topic word list in which the topic words extracted by the topic word extraction means are associated and listed for each predetermined unit time from the utterance start time to the utterance end time of the audio content. Means,
Index generating means for generating an index from the topic words listed in the topic word list generated by the topic word list generating means;
An index generation device comprising: In the topic word list generated by the topic word list generating means, the topic word list further includes a topic turning point extracting means for extracting a time when none of the topic words are enumerated as a topic turning point,
The index generating means generates an index from the topic words listed in each section in the topic word list for each section delimited by the topic turning points extracted by the topic turning point extracting means. Generator. A distance dictionary storage means for storing a distance dictionary that defines a semantic distance between the plurality of words;
The topic word extracting means, when calculating the utterance density per predetermined unit time for a predetermined word included in the utterance word list, means the predetermined word and the meaning in the distance dictionary stored in the distance dictionary storage means. The index according to claim 1 or 2, wherein words having a target distance within a predetermined range are also considered to be the same word as the predetermined word, and the utterance density per predetermined unit time is calculated. Generator. A topic word extraction list storage means for storing a topic word extraction list that lists the words to be extracted as the topic words;
The topic word extraction means, on the condition that it is a word listed in the topic word extraction list stored in the topic word extraction list storage means, the word whose utterance density exceeds a predetermined value as the unit time The index generation device according to claim 1, 2 or 3, wherein the index generation device extracts the topic word.   When the same topic word is listed without a predetermined time interval in the topic word list, the topic turning point extraction means determines that the topic turning point has been enumerated, and the topic turning point is extracted. The index generation apparatus according to claim 1, wherein the index generation apparatus extracts the index generation apparatus. A distance dictionary storage means for storing a distance dictionary that defines a semantic distance between a plurality of words;
Using the distance dictionary stored in the distance dictionary storage means, a semantic distance between words included in the utterance word list generated by the utterance word list generation means is evaluated, and a word with a low evaluation is selected as the utterance word Word removal means to be removed from the list;
Further comprising
The topic word extraction unit calculates an utterance density per unit time for each word included in the utterance word list based on the utterance word list after the word removal by the word removal unit, and the utterance density is predetermined. The index generation apparatus according to claim 1, wherein words exceeding the value of the index are extracted as topic words in the unit time. Using the distance dictionary stored in the distance dictionary storage means, extract words from the distance dictionary having a high semantic distance evaluation with words listed in the utterance word list after the word removal by the word removal means, Further comprising word complementing means for complementing the extracted word in the utterance word list,
The topic word extraction unit calculates an utterance density per unit time for each word included in the utterance word list based on the utterance word list after word completion by the word completion unit, and the utterance density is predetermined. The index generation apparatus according to claim 6, wherein a word exceeding the value is extracted as a topic word in the unit time. For each section delimited by topic turning points extracted by the topic turning point extraction means, the utterance frequency of each topic word listed in each section in the topic word list is calculated, and the utterance frequency exceeds a predetermined value It further includes an important word extraction means for extracting a topic word as an important word,
The index generation means generates an index listing the topic words extracted as important words by the important word extraction means for each section delimited by the topic turning points. The index generation device according to any one of the above.   The index generation means edits the topic word based on one or more of the utterance time, the utterance density, the utterance frequency, and the semantic distance for each section delimited by the topic turning point. The index generating apparatus according to claim 1, wherein the index is generated. An index generation device that generates speech data by recognizing speech content and generates an index to be given to the speech content from the text data,
A distance dictionary storage means for storing a distance dictionary that defines a semantic distance between a plurality of words;
An utterance word list generating means for extracting a word included in the text data and an utterance time of the word, and generating an utterance word list in which the word and the utterance time are associated and listed;
Using the distance dictionary stored in the distance dictionary storage means, a semantic distance between words included in the utterance word list generated by the utterance word list generation means is evaluated, and a word with a low evaluation is selected as the utterance word Word removal means to be removed from the list;
Index generating means for generating an index based on the spoken word list after word removal by the word removing means;
An index generation device comprising: Using the distance dictionary stored in the distance dictionary storage means, extract words from the distance dictionary having a high semantic distance evaluation with words listed in the utterance word list after the word removal by the word removal means, Further comprising word complementing means for complementing the extracted word in the utterance word list,
The index generation device according to claim 10, wherein the index generation means generates an index based on the utterance word list after word completion by the word completion means. An index generation method suitable for generating speech data by recognizing speech content and generating an index to be given to the speech content from the text data,
An utterance word list generating step of extracting a word included in the text data and an utterance time of the word and generating an utterance word list in which the word and the utterance time are associated and listed;
An utterance density indicating the number of utterances per predetermined unit time is calculated for each word included in the utterance word list generated by the utterance word list generation step, and words whose utterance density exceeds a predetermined value are calculated in the unit time. A topic word extraction process for extracting as a topic word;
Topic word list generation that generates a topic word list in which the topic words extracted by the topic word extraction step are associated and listed for each predetermined unit time from the utterance start time to the utterance end time of the audio content. Process,
An index generation step of generating an index from the topic words listed in the topic word list generated by the topic word list generation step;
The index generation method characterized by including. An index creation program for causing a computer to execute an index generation method suitable for generating speech data by speech recognition of speech content and generating an index to be added to the speech content from the text data,
An utterance word list generation procedure for extracting a word included in the text data and an utterance time of the word and generating an utterance word list in which the word and the utterance time are associated and listed;
An utterance density indicating the number of utterances per predetermined unit time is calculated for each word included in the utterance word list generated by the utterance word list generation procedure, and words whose utterance density exceeds a predetermined value are calculated in the unit time. A topic word extraction procedure to extract as a topic word;
Topic word list generation that generates a topic word list in which the topic words extracted by the topic word extraction procedure are associated and listed for each predetermined unit time from the utterance start time to the utterance end time of the audio content. Process,
An index generation procedure for generating an index from the topic words listed in the topic word list generated by the topic word list generation procedure;
An index generation program for causing a computer to execute

Images