JP6374771B2 - Retrieval device, program, and method for retrieving vague record of user's memory - Google Patents

Description

  The present invention relates to a technique for retrieving a record (voice data or text) previously spoken or described by a user.

  2. Description of the Related Art Conventionally, there is a technique for extracting a conversation at an arbitrary time from recorded voice data using a conversation start time and an end time as search keys for an apparatus that constantly records conversation (see, for example, Patent Document 1). According to this technology, the used storage capacity is minimized, but it is necessary to manually specify the start time of the conversation based on the voice data to be stored, and the more voice data the more time is required. It becomes.

  There is also a technique for recognizing the degree of emotion from a voice message and assigning importance (priority) to the voice message based on the degree of emotion for the recording function of the answering machine (see, for example, Patent Document 2). The user can make the playback order earlier as the voice message has a higher priority. According to this technique, priority is automatically assigned to audio data recorded in one segment.

JP 2005-114898 A JP 2007-49657 A

The inventors of the present application have thought that “can an ambiguous memory be searched for a user?” As a new problem (service) that has not existed before. For example, the following cases that can be provided to the user are assumed.
(Example 1) The case of remembering the memory of a trip When there are many stories about a trip, as a result, the memory of which trip was a memory becomes ambiguous. For example, there are cases such as “What was the trip to the memory of the ryokan that stayed in a room with a very nice view?” “How many years ago was it to see the beautiful Mount Fuji?”.
(Example 2) Case of recalling conversations at work Since we are doing almost the same work continuously, such as reception work and sales, it becomes ambiguous to remember who and what kind of conversation. For example, there are cases such as "I will hear your hobbies and work contents again in a few days".

A record in which a user's memory is ambiguous has the following characteristics.
(1) There are many similar experiences (2) The past is far from the current date and time

  Therefore, an object of the present invention is to provide a search device, a program, and a method capable of searching for a record (voice data or text) in which a user's memory is ambiguous as a new service that has not existed before.

According to the present invention, there is provided a search device for searching a record in which a user's memory is estimated to be ambiguous with respect to a query text from a record storage unit that stores a plurality of records including text,
The record has a recording date and time,
Category classification means for classifying a plurality of records into a predetermined number of categories using the record storage means;
Similar record search means for searching for records that are similar in terms of text or more in comparison with query text using record storage means;
Ambiguity calculation means for calculating the ambiguity by multiplying the number of records included in the category into which the record is classified for each searched record and the time difference from the recording date and time of the record to the current date and time;
Search result output means for outputting the query text as a search result in descending order of the ambiguity.

According to another embodiment of the search device of the present invention,
The record storage means preferably stores the ambiguity calculated by using the ambiguity calculation means in association with each record.

According to another embodiment of the search device of the present invention,
It is also preferable to further include voice recognition means for converting the voice data acquired by the microphone into text, generating a record in which the text is divided into silent sections of a predetermined time or longer, and storing the record in the record storage means.

According to another embodiment of the search device of the present invention,
It is also preferable that the microphone and the voice recognition unit are always activated, and the voice data as a log in the daily conversation of the user is always collected and stored in the record storage unit.

According to another embodiment of the search device of the present invention,
The category classification means preferably classifies a plurality of records so as to belong to each category using an unsupervised classification algorithm.

According to another embodiment of the search device of the present invention,
Each time new text is acquired as a new record, the category classification means compares the new text with the text of a record belonging to each category, and classifies the new text so as to belong to any one category. preferable.

According to another embodiment of the search device of the present invention,
It is also preferable to delete records whose ambiguity is greater than or less than a predetermined value from the record storage means.

According to the present invention, a computer mounted in an apparatus is caused to function so as to search a record in which a user's memory is estimated to be ambiguous with respect to a query text from a record storage unit that stores a plurality of records including text. A search program,
The record has a recording date and time,
Category classification means for classifying a plurality of records into a predetermined number of categories using the record storage means;
Similar record search means for searching for records that are similar in terms of text or more in comparison with query text using record storage means;
Ambiguity calculation means for calculating the ambiguity by multiplying the number of records included in the category into which the record is classified for each searched record and the time difference from the recording date and time of the record to the current date and time;
The computer is caused to function as a search result output means for sequentially outputting records with a high degree of ambiguity as search results for the query text.

According to the present invention, there is provided a search method of an apparatus for searching a record in which a user's memory is estimated to be ambiguous with respect to a query text from a record storage unit that stores a plurality of records including text,
The record has a recording date and time,
A first step of classifying a plurality of records into a predetermined number of categories using a record storage unit;
A second step of using a record storage unit to search for records that are similar in terms of text or more compared to the query text;
A third step of calculating, for each retrieved record, an ambiguity obtained by multiplying the number of records included in the category into which the record is classified and the time difference from the recording date and time of the record to the current date and time;
And a fourth step of outputting the query text in order from the record with the higher ambiguity as a search result.

  According to the search device, program, and method of the present invention, it is possible to search for a record in which a user's memory is ambiguous.

It is a functional block diagram of the search device in this invention. It is explanatory drawing showing the category classification | category and ambiguity calculation in this invention. It is explanatory drawing showing the similar record search and search result output in this invention. It is a system configuration figure consisting of a terminal and a search server in the present invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a functional configuration diagram of a search device according to the present invention.

  The search apparatus 1 in FIG. 1 inputs a query text as a past event, and searches for a record (voice data or text) in which a user's memory is estimated to be ambiguous. The record to be searched is a user's utterance or a written log, and is always collected in the whole life. The search device 1 is mounted on a terminal such as a smartphone or a wearable device that is always possessed by the user in order to acquire a log from the user.

  According to FIG. 1, the terminal 1 includes a microphone unit 101, a touch panel display 102, and a speaker unit 103 as hardware. The microphone unit 101 is for collecting user's voice at all times, and is a device such as a pin microphone, for example. The touch panel display 102 is a user interface that inputs query text from a user and displays search results. The speaker unit 103 reproduces audio data as a search result. When the voice from the character displayed on the touch panel display 102 is uttered from the speaker unit 103, the user can recognize the search result as if the character agent is answering.

  Further, according to FIG. 1, the terminal 1 includes a voice storage unit 11, a voice recognition unit 12, a record storage unit 13, a category classification unit 14, an ambiguity calculation unit 15, a similar record search unit 16, And a search result output unit 17. These functional components can be realized by executing a program that causes a computer installed in the terminal to function.

[Voice storage unit 11]
The audio storage unit 11 stores the audio data acquired by the microphone unit 101. Here, the recording unit of the audio data is stored in association with the record stored in the record storage unit 13. That is, the record as the search result is associated with the audio data.

[Voice recognition unit 12]
The voice recognition unit 12 converts the voice data acquired by the microphone unit 101 into text. Here, the text is generated for each record divided into silent sections of a predetermined time or longer. For example, when the silent section continues for n seconds (for example, 100 seconds) or longer, the text may be separated as a record. In addition, the speaker recognition technology may be used to divide the record as a record when the utterance of only the user having the terminal 1 continues for a predetermined silent period.

  The microphone unit 101 and the voice recognition unit 12 may be constantly activated, and may always collect voice data as a daily conversation log of the user and output it to the record storage unit 13.

[Record accumulation unit 13]
The record storage unit 13 stores a plurality of records including text. According to the present invention, “recording date and time” and “ambiguity” are stored in association with each record (text). The record storage unit 13 preferably stores not only the text (conversation log) output from the voice recognition unit 12 but also the text (handwriting log) input from the touch panel display 102.

  FIG. 2 is an explanatory diagram showing category classification and ambiguity calculation in the present invention.

[Category classification unit 14]
The category classification unit 14 classifies a large number of records accumulated in the record accumulation unit 13 into a predetermined number of categories (see the category classification unit 14 in FIG. 2). For this classification, an “unsupervised classification algorithm” may be used. For example, LDA (Latent Dirichlet allocation) or SVM (Support Vector Machine) can be used.

  “LDA” refers to a technique for calculating the probability (topic ratio) of each record belonging to each topic group and classifying the records into topic groups. Topic classification is a probabilistic model of the document generation process expressed in bag-of-words in order to analyze records (text). LDA introduces a probabilistic framework based on fluctuations in a feature vector of a word to a technique (LSI (latent Semantic Indexin)) for dimensional compression of a word document matrix. The compressed set of dimensions is called a topic.

  “SVM” is applied to pattern recognition, and comprises two classes of pattern discriminators using linear input elements, and calculates distances from feature values using learning samples, Learn the parameters of a linear input element. The identification engine of the support vector machine learns the extracted plurality of learning feature quantities as teacher data. It is determined whether or not the target feature amount (the feature amount of the newly input log text) belongs to the learning feature amount of the specific category (the feature amount of the category). According to the identification engine of the support vector machine, a clear rule is not generated, and it looks like a black box, but a learning feature amount for each specific category is generated as a support vector.

The category classification unit 14 classifies the category with respect to the record storage unit 13 at the following timing, for example.
(1) Every predetermined period (for example, every hour or every two weeks)
(2) Every time a predetermined number (for example, 100) or more records are accumulated in the record accumulation unit 13, an assigned category ID (IDentifier) is assigned to each record. For example, when it is classified into 300 categories, the category ID is a numerical value of 1 to 300.

[Ambiguity calculator 15]
The ambiguity calculation unit 15 calculates, for each searched record, an ambiguity obtained by multiplying the number of records included in the category into which the record is classified and the time difference from the recording date and time of the record to the current date and time (see FIG. 2 ambiguity calculation unit 15).
Ambiguity = number of records in category x (current date-recording date)

The ambiguity is calculated as follows, for example.
Number of records in category = 100 Current date and time-Recording date and time = 24,000 hours Ambiguity = 100 x 24,000 = 2,400,000
In this case, the greater the number of records in the category, the higher the ambiguity level, and the farther the record recording is in the past, the higher the ambiguity level.

  FIG. 3 is an explanatory diagram showing similar record search and search result output in the present invention.

[Similar record search unit 16]
The similar record search unit 16 uses the record storage unit 13 to search for a record that is similar in terms of text as compared to the query text (see the similar record search unit 16 in FIG. 3).

The similarity between the query text and the record text is, for example, calculating a vector that analyzes the part of speech or meaning between words of both texts, and calculating the cosine similarity (based on Bag of words) of these vectors. Also good. Of course, the similarity may be calculated only for the important word by using the important word dictionary storing the important words (nouns) having high appearance frequency. The cosine distance is calculated by the following conceptual expression, for example.
Query text word feature vector: D
Record text word feature vector: E
Similarity between two sentences: sim (D, E)
sim (D, E) = cos θ = (D · E) / (| D || E |)
When the cosine distance is the same word, the similarity weight is set to 1.

[Search result output unit 17]
The search result output unit 17 outputs the query text as a search result in descending order of the ambiguity (see the search result output unit 17 in FIG. 3). Thereby, it is possible to present to the user in order from the record in which the user's memory is ambiguous.

  The search result output unit 17 may have a character agent function for the user. For this purpose, the search result output unit 17 has a voice conversion function for converting text to be output as a search result into voice data, and a character display function for displaying a character linked to the voice data. Then, control is performed so that the voice from the character displayed on the touch panel display 102 is uttered from the speaker unit 103. Thereby, the user can recognize the search result as if the character agent is answering.

<Reduction in storage amount in the record storage unit 13>
Accumulating all the large number of records collected as logs in the past by the record accumulating unit 13 only uses the storage capacity wastefully and is expensive. Further, when searching, it is necessary to search for unnecessary data, and the processing efficiency of the search also decreases. For this reason, it is preferable to accumulate only the ambiguous records of the user as much as possible.

The record storage unit 13 may delete a record whose degree of ambiguity is greater than or equal to a predetermined value (shifts to a secondary storage capacity).
Ambiguity is greater than or equal to a predetermined value: Past with a lot of similar experiences and far from the current date and time Ambiguity is less than or equal to a predetermined value: Past with few similar experiences and near from the current date and time Ambiguity is greater than or equal to a predetermined value Or, as will be described below, any record can be searched. As a result, it is possible to reduce unnecessary data that will not be searched in the future in the record storage unit 13 and prevent a decrease in search processing efficiency.

  FIG. 4 is a system configuration diagram including a terminal and a search server in the present invention.

  FIG. 4 shows that the present invention is configured as a system including a terminal and a search server, as compared with FIG. According to FIG. 4, the terminal 1 includes a microphone unit 101 as a user interface, a touch panel display 102 and a speaker unit 103, and a voice recognition unit 12, and further includes a voice transmission unit. The voice transmission unit transmits the voice data acquired by the microphone 101 unit to the search server 2.

  Further, according to FIG. 4, the search server 2 further includes a voice receiving unit, a record receiving unit, and a query text input unit. The voice receiving unit receives the voice data itself from the terminal 1 and stores it in the voice storage unit 11. The record receiving unit receives the record text recognized by the terminal 1 and outputs it to the record storage unit 13. The query text input unit receives the query text input by the user from the terminal 1 and outputs it to the similar record search unit 16. The search result output unit 17 transmits voice data and / or text as a search result to the terminal 1.

  As described above in detail, according to the search device, program, and method of the present invention, it is possible to search for a record (voice data or text) in which the user's memory is ambiguous.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Search apparatus, terminal, smart phone 101 Microphone part 102 Touch panel display 103 Speaker part 11 Voice storage part 12 Voice recognition part 13 Record storage part 14 Category classification part 15 Ambiguity calculation part 16 Similar record search part 17 Search result output part 2 Search server

Claims (9)

A search device for searching for a record in which a user's memory is estimated to be ambiguous with respect to a query text from a record storage unit that stores a plurality of records including text,
The record is given a recording date and time,
Category classification means for classifying a plurality of records into a predetermined number of categories using the record accumulation means;
Similar record search means for searching records similar in terms of text or more in comparison with the query text using the record storage means;
Ambiguity calculation means for calculating the ambiguity obtained by multiplying the number of records included in the category into which the record is classified and the time difference from the recording date and time of the record to the current date and time for each of the retrieved records;
A search apparatus comprising: search result output means for outputting the query text as a search result in order from the record with the highest ambiguity. The search apparatus according to claim 1, wherein the record storage unit stores the ambiguity calculated by using the ambiguity calculation unit in association with each record.   Voice recognition means for converting voice data acquired by a microphone into text, generating a record in which the text is divided into silent sections of a predetermined time or longer, and storing the record in the record storage means is further provided. The search device according to claim 1 or 2. 4. The microphone and the voice recognition unit are always activated, and voice data as a log in a daily conversation of a user is always collected and stored in the record storage unit. The search device described in 1.   The search device according to any one of claims 1 to 4, wherein the category classification means classifies a plurality of records so as to belong to each category by using an unsupervised classification algorithm. Each time the new text is acquired as a new record, the category classification means compares the new text with the text of a record belonging to each category, and classifies the new text so as to belong to any one category. The search device according to any one of claims 1 to 5, wherein: The search apparatus according to any one of claims 1 to 6, wherein a record having an ambiguity greater than or less than a predetermined value is deleted from the record storage means. A search program for causing a computer mounted on a device to function so as to search a record in which a user's memory is estimated to be ambiguous with respect to a query text from a record storage unit that stores a plurality of records including text,
The record is given a recording date and time,
Category classification means for classifying a plurality of records into a predetermined number of categories using the record accumulation means;
Similar record search means for searching records similar in terms of text or more in comparison with the query text using the record storage means;
Ambiguity calculation means for calculating the ambiguity obtained by multiplying the number of records included in the category into which the record is classified and the time difference from the recording date and time of the record to the current date and time for each of the retrieved records;
A search program for causing a computer to function as search result output means for outputting the query text as a search result in order from the record with the higher ambiguity. A search method of an apparatus for searching a record in which a user's memory is estimated to be ambiguous with respect to a query text from a record storage unit that stores a plurality of records including text,
The record is given a recording date and time,
A first step of classifying a plurality of records into a predetermined number of categories using the record storage unit;
A second step of searching for a record that is textually similar to or more than a predetermined condition using the record storage unit as compared with the query text;
A third step of calculating, for each retrieved record, an ambiguity obtained by multiplying the number of records included in the category into which the record is classified and a time difference from the recording date and time of the record to the current date and time;
And a fourth step of outputting the query text as a search result in order from the record with the higher ambiguity.

Images