JP7096199B2 - Information processing equipment, information processing methods, and programs - Google Patents

Description

The present invention relates to an information processing apparatus, an information processing method, and a program.

A technique for performing speech recognition based on the certainty including the speech recognition result is known (see Patent Document 1). The certainty is set based on, for example, a simple comparison result with the corpus or a result of evaluating the similarity between the speech recognition result and the corpus.

Japanese Unexamined Patent Publication No. 2016-206487

However, with the conventional technique, it takes time to extract a suitable candidate from tens of thousands of latent word candidates for each word of the corpus, and there is a possibility that efficient speech recognition processing cannot be realized. rice field. In addition, there is a possibility that improvement in the accuracy of extracting suitable candidates from each word of the corpus has not been sufficiently examined.

The present invention has been made in consideration of such circumstances, and an object of the present invention is to provide an information processing device, an information processing method, and a program capable of performing voice recognition processing more efficiently and with high accuracy. It is one of.

One aspect of the present invention is for each of an acquisition unit for acquiring voice data, an analysis unit for analyzing the voice data and converting it into a text, and a plurality of first words included in the text of the analysis result by the analysis unit. , The first index which evaluated the frequency of the first word in the analyzed sentence included in the text and including the first word, and the rarity of the first word with respect to the sentence included in the library information. An index value derivation unit that derives a value and associates it with the analyzed sentence, a vector conversion unit that converts a sentence analyzed by the analysis unit into a vector value by distributed representation, and an index value derivation unit that is derived. The meaning is known and the vector value is the selection unit that selects a part of the sentence to be analyzed or the sentence of interest based on the first index value and the conversion result by the vector conversion unit. Information including a generation unit that generates data in which the meaning of a teacher sentence whose vector value is close to that of the selection sentence selected by the selection unit is associated with the meaning of the selection sentence among the required teacher sentences. It is a processing device.

According to one aspect of the present invention, voice recognition processing can be performed more efficiently and with high accuracy.

It is a figure which shows an example of the use environment of the information processing apparatus 100 which concerns on embodiment. It is a figure which shows typically the process of the information processing apparatus 100. It is a figure for demonstrating WFST. It is a figure for demonstrating WFST. It is a figure for demonstrating WFST. It is a block diagram of the information processing apparatus 100 which concerns on embodiment. It is a figure for demonstrating the vector conversion process by the W2V execution unit 110. It is a figure for demonstrating a sentence vector. It is a figure which shows typically the suitable candidate selection by a selection part 114. It is a figure for demonstrating a task text. It is a figure for demonstrating a representative vector. It is a figure for demonstrating the index value of the text to be extracted. It is a figure which shows an example of the tf-idf value derived by the vector conversion unit 112. It is a figure for demonstrating the tf-idf vector of a sentence vector. It is a figure for demonstrating the reliability derivation process by a reliability derivation unit 114a. It is a figure for demonstrating the similarity evaluation method. It is a flowchart which shows an example of the flow of the language model generation processing by an information processing apparatus 100. It is a flowchart which shows an example of the flow of the voice recognition processing by an information processing apparatus 100.

Hereinafter, embodiments of the information processing apparatus, information processing method, and program of the present invention will be described with reference to the drawings.

[Overview]
The information processing device is realized by one or more processors. The information processing device receives voice data recording the voice emitted by the user, performs voice recognition processing of the received input data, and performs various processing based on the recognition result (hereinafter, "terminal device"). It is a device for providing a language model for (referred to as). Various processes include controlling the IoT (Internet of Things) device according to the intention of the user who emitted the voice, and responding to the user's question.

The language model is a natural language processing model that converts input data into text in speech recognition processing, and includes the probability of the conversion result that the result of converting the input result into text is likely to be the correct answer. .. Hereinafter, the operation of the terminal device intended by the user may be referred to as a "task". The voice data may be compressed or encrypted.

FIG. 1 is a diagram showing an example of a usage environment of the information processing apparatus 100 according to the embodiment.

In the illustrated environment, the terminal device 20, the controlled device 30, and the service server 40 communicate with each other via the network NW. The network NW includes, for example, a part or all of a WAN (Wide Area Network), a LAN (Local Area Network), the Internet, a provider device, a wireless base station, a dedicated line, and the like. In the example shown in FIG. 1, the number of controlled devices 30 is N (N is an integer of 1 or more). In this specification, when the individual controlled object devices 30-1 to 30-N are not distinguished, such as when common matters are explained in the controlled object devices 30-1 to 30-N, the controlled object is simply controlled. Called device 30.

The terminal device 20 is a device that accepts a user's voice input. The terminal device 20 is a mobile phone such as a smartphone, a tablet terminal, a personal computer, a smart speaker (AI speaker), or the like.

The control target device 30 is an IoT device having a communication function and an interface for receiving control from the outside, and can be controlled in response to a command from the terminal device 20 operated by the user. The controlled device 30 is, for example, a television, a radio, a lighting fixture, a refrigerator, a microwave oven, a washing machine, a rice cooker, a self-propelled vacuum cleaner, an air conditioner, a vehicle, or the like.

The controlled device 30 may be the terminal device 20 itself. That is, the terminal device 20 may perform some kind of search processing, make a phone call, or send a message according to the processing result of the information processing device 100.

The service server 40 communicates with a web server device that provides a web page corresponding to a command from the terminal device 20 operated by the user, and the terminal device 20 in which the application is started, and exchanges various information to provide contents. Is an application server device or the like that provides.

FIG. 2 is a diagram schematically showing the processing of the information processing apparatus 100.

The information processing device 100 converts voice data input by the user via the terminal device 20 into phonemes by applying it to an acoustic model, and one or more texts to be extracted (sounds included in the voice data) based on the phonemes. Is generated, and the extracted text selected based on the comparison with the known task feature amount among the generated texts to be extracted is applied to the language model to select suitable candidates. The suitable candidate is a text that is determined to be a suitable text that is likely to reflect the intention of the user in the text to be extracted, and is a text suggesting the operation of the terminal device 20 or the controlled device. Is.

The acoustic model is a model for statistically analyzing frequency components and time changes to determine what kind of phonemes the input voice data is composed of (what is said). A phoneme is a label for specifying the smallest unit of a language such as an alphabet or a kana, and includes, for example, a vowel or a consonant. The information processing apparatus 100 obtains the text to be extracted by appropriately combining phonemes according to a language rule.

As shown in FIG. 2, when the extraction target text generated as a result of phoneme conversion is “kyonotenki”, for example, “k” and “t” indicate phonemes included in the generated extraction target text. When the voice recognition process is performed on the premise of Japanese, the text to be extracted may be expressed in alphabetical notation, hiragana notation, or katakana notation. In the example shown in FIG. 2, the information processing apparatus 100 generates an extraction target text including "kyonotenki", "kyonotenkii", and "kyonodenki" based on the received voice data.

The language model generated by the information processing apparatus 100 performs morphological analysis on each of the conversion candidates including "kyonotenki", "kyonotenkii", and "kyonodenki" in the example shown in FIG. The morphological analysis is a process of determining a delimiter of words constituting the text to be extracted and deriving, for example, a part of speech of each delimited word. The morphological analysis is performed using, for example, a morphological analysis engine such as MeCAB.

For example, the language model derives three words "today (kyo)", "no (no)", and "weather (tenki)" as a result of analyzing the extraction target text "kyonotenki". Similarly, as a result of analyzing the extraction target text "kyonotenkii", "today (kyo)", "(no)", "tenkii", and as a result of analyzing the extraction target text "kyonodenki", "Kyo (kyo)" kyo) ”,“ no (no) ”,“ electricity (denki) ”are generated. In this way, when converting voice input from Hiragana to Kanji, there is a possibility that conversion candidates for a plurality of patterns will be generated.

The language model generates an evaluation value for evaluating the analysis result generated from each of one or more extraction target texts, and selects one extraction target text from a plurality of patterns of conversion candidates based on the evaluation value. More specifically, the information processing apparatus 100 evaluates the matching rate of the analysis result of the text to be extracted with the feature amount obtained from the known task voice, and is presumed to be in line with the user's intention. Select a candidate. Then, the information processing apparatus 100 outputs an instruction regarding a task for generating output information corresponding to the intention.

[WFST]
3 to 5 are diagrams for explaining a WFST (Weighted Finite-state Transducer) realized by an acoustic model and a language model. WFST is an example of a mechanism for converting input data into a "conversion candidate" and an "estimated value of the certainty of the conversion candidate".

When voice recognition using WFST is performed, the voice input received by the terminal device 20 is converted into a context-sensitive phoneme such as Triphon by the acoustic model (FIG. 3). Next, the acoustic model (or language model) converts phonemes into words (Fig. 4). Next, the language model generates text that is the result of conversion of speech input from words (Fig. 5). The language model is, for example, an N-gram language model. For example, when 3-gram is adopted as a language model, the composition of the entire text is determined based on whether or not the meaning is established for each of the three words.

The information processing apparatus 100 generates a suitable language model so that the above-mentioned morphological analysis and speech recognition using WFST can be performed at higher speed and with higher processing accuracy.

[overall structure]
FIG. 6 is a block diagram of the information processing apparatus 100. The information processing apparatus 100 includes, for example, an acquisition unit 102, an analysis unit 104, a frequency calculation unit 106, a rarity calculation unit 108, a W2V (Word2Vec) execution unit 110, a vector conversion unit 112, and a selection unit 114. A language model calculation unit 116, a command output unit 118, and a storage unit 120 are provided. These components (excluding the storage unit 120) are realized by, for example, a hardware processor such as a CPU (Central Processing Unit) executing a program (software).

In addition, some or all of these components are hardware (circuits) such as LSI (Large Scale Integration), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), and GPU (Graphics Processing Unit). It may be realized by the part; including circuitry), or it may be realized by the cooperation of software and hardware. The program may be stored in advance in a storage device (a storage device including a non-transient storage medium) such as an HDD or a flash memory of the information processing device 100, or a detachable storage such as a DVD or a CD-ROM. It is stored in a medium (non-transient storage medium), and may be installed in the HDD or flash memory of the information processing apparatus 100 by mounting the storage medium in the drive device.

The storage unit 120 is realized by, for example, a RAM (Random Access Memory), a register, a flash memory, an EEPROM (Electrically Erasable Programmable Read Only Memory), or the like. The storage unit 120 may include, for example, an acoustic model 120a, a language model 120b, a corpus analysis result 120c, a task text analysis result 120d, an extraction target text analysis result 120e, a word vector list 120f, a vector list 120g, and a language model calculation text. Information such as 120h is stored. The vector list 120g includes, for example, a task text vector list 120i and an extraction target text vector list 120j.

The acquisition unit 102 acquires character information (hereinafter referred to as “corpus I1”) used as a corpus when the information processing apparatus 100 performs voice recognition processing, and outputs the information to the analysis unit 104. The corpus I1 includes, for example, article data such as news and posted data of SNS (Social Networking Service). Corpus I1 is an example of "library information".

The corpus I1 is a colloquial text (for example, a post history in the SNS, a written sentence of the conversation history between the user and the device in the automatic response device, a text converted from the actual conversation, and a voice acquired from the terminal device 20. It is desirable that it is a written sentence of the processing history of the own device for the input).

Further, the acquisition unit 102 acquires a data set of character information (hereinafter referred to as “task text I2”) indicating a routine task set by the administrator of the information processing apparatus 100, and outputs the data set to the analysis unit 104. The task text I2 is an example of a "teacher sentence".

Further, the acquisition unit 102 acquires voice data (hereinafter referred to as “voice data I3”) input by the user of the terminal device 20 and outputs the voice data to the analysis unit 104. The voice data I3 acquired by the acquisition unit 102 may include the user's position information. The position information may be, for example, a processing result by a GNSS (Global Navigation Satellite System) receiving device included in the terminal device 20. Further, when the terminal device 20 is a device mainly used in a specific place (for example, a user's living room, a user's office, etc.), the information about the specific place corresponds to the location information.

The analysis unit 104 analyzes the information acquired by the acquisition unit 102 and converts it into text (character data). The analysis by the analysis unit 104 is, for example, a morphological analysis.

The analysis unit 104 analyzes, for example, the corpus I1 output by the acquisition unit 102. For example, the analysis unit 104 decomposes the corpus I1 output by the acquisition unit 102 into units of part of speech such as nouns, verbs, and particles. The analysis unit 104 stores the analysis result as the analysis result 120c of the corpus in the storage unit 120.

Further, the analysis unit 104 analyzes the task text I2 output from the acquisition unit 102, and stores the analysis result as the analysis result 120d of the task text in the storage unit 120.

Further, the analysis unit 104 applies the voice data I3 output by the acquisition unit 102 to the acoustic model 120a to generate one or more extraction target texts, and then analyzes each extraction target text by morphological analysis or the like. I do. Further, the analysis unit 104 stores the analysis result in the storage unit 120 as the analysis result 120e of the text to be extracted.

The frequency calculation unit 106 refers to a plurality of words (hereinafter referred to as "first word") included in one sentence included in the extracted text (hereinafter referred to as "analyzed sentence") from the analysis result 120e of the extracted text. ) Is calculated and associated with the analyzed sentence. The frequency calculation unit 106, for example, obtains a tf value (Term Frequency Value; an index value indicating frequency) for each of a plurality of words included in one sentence included in the extraction target text from the analysis result 120e of the extraction target text. Calculate and associate with the analyzed sentence.

The frequency calculation unit 106 calculates in advance the tf value for each word (hereinafter referred to as “second word”) included in one sentence included in the analysis result 120c of the corpus. The frequency calculation unit 106 refers to each of the plurality of second words included in the corpus analysis result 120c as a sentence of the corpus included in the corpus analysis result 120c and including the second word (hereinafter referred to as "attention sentence"). The tf value of the second word in (referred to as) is calculated in advance and associated with the sentence of interest.

The shortage calculation unit 108 calculates an index value indicating the rarity of each of the first words included in the analyzed sentence included in the extracted text from the analysis result 120e of the extracted text, and associates it with the analyzed sentence. .. The rarity calculation unit 108, for example, from the analysis result 120e of the text to be extracted, an idf value (Inversed Document Frequency Value; index value indicating rarity) for each of the first words included in the analyzed sentence included in the text to be extracted. ) Is calculated and associated with the analyzed sentence.

The shortage calculation unit 108 calculates in advance the idf value for each of the second words included in one sentence included in the analysis result 120c of the corpus, and associates it with the sentence of interest.

The frequency calculation unit 106 and the shortage calculation unit 108 perform at least one of the setting of the index value for the first word and the setting of the index value for the second word. The one having both the frequency calculation unit 106 and the shortage calculation unit 108 is an example of the “index value derivation unit”. The calculation result by the frequency calculation unit 106 and the rarity calculation unit 108 regarding the analysis result 120e of the extraction target text is an example of the “first index value”, and the frequency calculation unit 106 and the rarity calculation unit regarding the analysis result 120c of the corpus. The calculation result by 108 is an example of "second index value".

The W2V execution unit 110 converts each of the words included in the sentence analyzed by the analysis unit 104 into a vector value by distributed representation. The W2V execution unit 110 converts, for example, the analysis result 120c of the corpus into a vector value, and stores the conversion result in the word vector list 120f.

The vector conversion unit 112 converts the sentence analyzed by the analysis unit 104 into a vector value by distributed representation. The vector value generated by the vector conversion unit 112 is the index value for the first word or the second of the vector value converted by the W2V execution unit 110 and the calculation result by the frequency calculation unit 106 and the rarity calculation unit 108. It is based on at least one of the index values for the word.

The vector conversion unit 112 uses the analysis result 120e of the text to be extracted and the vector value of the word vector list 120f to use the vector value of the sentence unit of the text to be extracted (hereinafter, the sentence vector of the text to be extracted, or simply "sentence vector". ) Is generated.

The sentence vector is, for example, a vector value obtained by converting the analysis result 120e of the text to be extracted by the W2V execution unit 110, and a calculation result by the frequency calculation unit 106 and the rarity calculation unit 108 (hereinafter, “tf-idf value”). ) And. The vector conversion unit 112 outputs the sentence vector of the text to be extracted to the selection unit 114.

Further, the vector conversion unit 112 generates a vector value for each sentence of the task text (hereinafter referred to as “task text sentence vector”) by using the analysis result 120d of the task text and the vector value of the word vector list 120f. .. The vector conversion unit 112 outputs the sentence vector of the task text to the selection unit 114.

In the process of generating the language model 120b, the selection unit 114 selects a sentence vector that is the source of the language model 120b (reflected in the language model 120b) based on the sentence vector of the text to be extracted and the sentence vector of the task text. The text from which the sentence vector that is the source of the language model 120b is derived is an example of the "selective sentence". The selection unit 114 outputs the selection result to the language model calculation unit 116.

Further, in the process of using the language model 120b (speech recognition processing process by the information processing apparatus 100), the selection unit 114 outputs a part or all of the conversion result by the vector conversion unit 112 to the language model calculation unit 116.

The selection unit 114 includes, for example, a reliability derivation unit 114a. The priority derivation process by the reliability derivation unit 114a will be described later.

The language model calculation unit 116 performs processing related to the language model 120b.

The language model calculation unit 116 includes, for example, a language model generation unit 116a. The language model generation unit 116a generates a language model to which the selection result output by the selection unit 114 is applied in the generation process of the language model 120b, and stores it in the storage unit 120 as the language model 120b. The language model generation unit 116a generates a language model 120b based on, for example, a language model calculation text 120h preset by the administrator of the information processing apparatus 100 and conversion candidates selected by the selection unit 114.

The language model calculation text 120h is, for example, a sentence vector of a task text assumed by the administrator of the information processing device 100 or a sentence vector held as a voice recognition processing history of the past information processing device 100. The language model calculation text 120h may include a sentence vector generated from a sentence that is the same as or similar to the corpus I1, the task text I2, the voice data I3, and the like. The selection unit 114 is based on at least one of the tf-idf value of the first word or the tf-idf value of the second word by the frequency calculation unit 106 and the shortage calculation unit 108, and the conversion result by the vector conversion unit 112. Then, select a part of the sentence to be analyzed or the sentence of interest.

Further, the language model calculation unit 116 applies the selection result output by the selection unit 114 in the process of using the language model 120b (speech recognition processing process by the information processing apparatus 100) to the language model 120b, and outputs the applied result as a command. Output to unit 118.

The command output unit 118 is a recognized sentence (selected sentence to be analyzed, based on the vector value converted by the vector conversion unit 112 in the process of using the language model 120b (speech recognition processing process by the information processing apparatus 100). Or the meaning of the sentence of interest) is estimated, and the information about the command based on the estimation result (or the command itself) is output. The command output by the command output unit 118 includes an instruction of processing to be performed by the terminal device 20, information for specifying the control target device 30 of the output destination, a processing request to the control target device 30 of the output destination, and the like.

The command output unit 118 is, for example, the weather forecast provided by the service server 40 when the suitable candidate output by the language model calculation unit 116, which is the result of application to the language model 120b, is "tell me the weather today". The output information is information including a part or all of the response of the command for transmitting the request to the website of the above and transmitting to the terminal device 20.

Further, for example, when the suitable candidate is "lower the volume of music", the command output unit 118 identifies the controlled target device 30 during music reproduction and outputs a command to lower the volume. When the output destination generates the output information of the control target device 30, the command output unit 118 also generates the output information notifying the terminal device 20 that the output information has been output to the control target device 30. You may.

[W2V vector conversion]
FIG. 7 is a diagram for explaining the vector conversion process by the W2V execution unit 110.

The W2V execution unit 110 generates a word vector by vector-expressing (distributed representation) the meaning of each word included in the analysis result 120c of the corpus, for example. In the example of FIG. 7, the W2V execution unit 110 generates a word vector of “volume”.

The W2V execution unit 110 generates a word vector so that the distance (cosine similarity) between word vectors is close between words having similar meanings such as "sound" and "volume", and "music" and "music". do. The W2V execution unit 110 stores the generated vector value in the storage unit 120 as a word vector list 120f in the storage unit.

Further, when the task text or the extraction target text contains a word that is not stored in the word vector list 120f, the W2V execution unit 110 sets the analysis result 120d of the task text or the analysis result 120e of the extraction target text into, for example, a corpus. By adding them, they may be analyzed in the same manner and their vector values may be generated. This vector value may or may not be reflected in the word vector list 120f each time the processing is performed by the W2V execution unit 110.

[Sentence vector]
FIG. 8 is a diagram for explaining a sentence vector.

For example, when the vector conversion unit 112 generates a sentence vector of "lowering the volume", the vector conversion unit 112 performs predetermined operations on the word vectors of "volume", "o", and "lowering" (for example, respectively). (By adding the word vectors of) to generate a sentence vector.

As a result, for the sentence vector that is the sum of the word vectors of the words that make up the sentence, the distance between the sentence vectors of sentences that have similar meanings such as "make the sound of music quieter" and "lower the volume". Will be closer.

Further, the vector conversion unit 112 generates a sentence vector of the task text using the analysis result 120d of the task text and the word vector output by the W2V execution unit 110, and stores it in the storage unit 120 as the task text vector list 120i. .. The task text is a text known to include the intention of the user, and is set in advance by, for example, the administrator of the information processing apparatus 100.

[Candidate selection]
The selection unit 114 evaluates the extraction target text output by the language model calculation unit 116 based on the evaluation value, and selects a suitable candidate having a high possibility of reflecting the input intention of the user. The selection unit 114 outputs a suitable candidate as a selection result to the language model calculation unit 116.

FIG. 9 is a diagram schematically showing suitable candidate selection by the selection unit 114.

The language model is a model for generating suitable candidates from the text to be extracted. For example, from the similarity between the sentence vector of the candidate vector and the sentence vector of the task text, the selection unit 114 gives a higher evaluation value as it is closer to the task text, and further, the score regarding the sequence of words is higher by using the language model. Suitable candidates are selected by comprehensively evaluating these evaluation values, which give higher evaluation values. The language model may be evaluated in consideration of the surrounding environment of the user.

[Task text]
The task text will be described below. The administrator of the information processing apparatus 100 refers to the selection unit 114 as an evaluation criterion in the process of generating the language model 120b based on, for example, the past voice input history of the terminal apparatus 20 and the processing history of the information processing apparatus 100. The task text I2 to be processed is extracted.

FIG. 10 is a diagram for explaining the task text.

The left figure of FIG. 10 shows the voice recognition results R1 to R7 of the past voice input history of the terminal device 20. The voice recognition result includes a result reflecting the input intention of the user of the terminal device 20 and a voice recognition result having no input intention by the user.

For example, when the administrator of the information processing apparatus 100 determines that the voice recognition result R4 is a text close to the task, the administrator sets a high priority as shown in the upper right figure of FIG. "Close to a task" means that text that is likely to reflect the input intention of the user is included, and that text that has a high meaning of operation for the terminal device 20 or the controlled device 30 is included. ..

Further, when the administrator of the information processing apparatus 100 determines that the voice recognition result R6 is a text far from the task, the administrator sets the priority low as shown in the lower right figure of FIG.

Further, the administrator of the information processing apparatus 100 determines that the voice recognition results R1, R2, R3, R5, and R7 are texts far from the task, and sets the priority low. The priority of the task text is registered in the task text vector list 120i together with the sentence vector value of the task text, for example.

The task text vector list 120i may have a cluster structure of about 10, and in that case, the task texts having similar meanings and contents of the tasks are collected as a cluster. The cluster is configured by, for example, the k-means clustering method or the like. Similar evaluation of meaning and content will be described later.

Further, in the task text vector list 120i, a representative vector may be set for each cluster for the purpose of improving the search efficiency, and the representative vector may be stored. The representative vector may be, for example, the average of the sentence vectors of the task texts constituting the cluster, or may be the weighted average of the priority of the task texts and the sentence vectors.

When position information is given to the text to be extracted, the selection unit 114 estimates the input environment of the user from the position information and determines whether or not the text includes the execution intention of the user of the text to be extracted. Then, subsequent processing may be performed based on the determination result.

For example, when the user is estimated to be in the living room at home from the position information of the text to be extracted, the selection unit 114 sets a high matching rate of the task related to the controlled device 30 used in the living room, and at the same time, the office. By setting the matching rate of the task related to the controlled device 30 used in the above to a low value, the accuracy (high matching rate) of selecting the corresponding task may be changed.

For example, in the example of FIG. 10, when the voice data I3 has the position information corresponding to the user's home living room, the text including the execution intention of the task "I want a light bulb" rather than "I want an annual holiday". Is likely to be recognized, so the precision rate of "I want a light bulb" is set high. On the other hand, when the voice data I3 has the location information corresponding to the user's office, there is a high possibility that the text including the execution intention of the task "I want an annual holiday" rather than "I want a light bulb" is recognized. In the case (when it is unlikely to accept the voice data I3 saying "I want a light bulb"), the matching rate different from the example shown in the figure (for example, the matching rate of "I want a light bulb" and "I want an annual holiday" is reversed. Etc.) may be set.

FIG. 11 is a diagram for explaining a representative vector.

For example, when selecting a task text, the selection unit 114 first compares the representative vector with the sentence vector of the text to be extracted to select a cluster, and then selects a suitable task text from the selected clusters. select.

[Index value of text to be extracted]
The selection unit 114 uses the tf-idf value as a determination factor as to whether or not the text is "close to the task" as described above.

FIG. 12 is a diagram for explaining an index value of the text to be extracted.

One sentence S1 (hereinafter referred to as "extraction target text S1") included in the extraction target text is "next week / no / Saturday / hot spring / ni / want to go / n / but / good / hot spring / ha / aru / no" ( In the case of 14 words (/: word delimiter position), the vector conversion unit 112 determines the "importance" in the text for each word based on the calculation results by the frequency calculation unit 106 and the rarity calculation unit 108. Generate a sentence vector that is the judgment source information.

In the following description, corpus I1 contains 200,000 sentences, corpus I1 contains 150 sentences containing the word "hot spring", and corpus I1 contains 30,000 sentences containing the word "no". Explain as if the sentence is included.

In the example of FIG. 12, the extraction target text S1 is an example of the “analyzed sentence”. Further, the double underlined word "hot spring" included in the extraction target text S1 is an example of the "first word". Further, the underlined word "no" included in the extraction target text S1 and the underlined word "no" included in the extraction target text S2 are examples of the "first word" to be focused on in the following description. Is.

Further, the task text S3 and the task text S4 are examples of the task text included in the task text I2 of FIG. The task text S3 is an example of a "sentence of interest", and the task text S4 is an example of a "sentence other than the sentence of interest". The double underlined word "hot spring" included in the task text S3 is an example of the "second word". Further, the underlined word "no" included in the task text S3 and the task text S4 is an example of the "second word".

In the example of FIG. 12, the frequency calculation unit 106 sets the tf value of the word “hot spring” included in the extraction target text S1 to 2/14 (occupies 2 words out of the 14 words constituting the extraction target text S1). Calculate that there is. Similarly, the frequency calculation unit 106 calculates that the tf value of the word "no" included in the extraction target text S1 is 2/14.

The shortage calculation unit 108 calculates the idf value of the word "hot spring" included in the extraction target text S1 as log (200000/150). Similarly, the shortage calculation unit 108 calculates the idf value of the word "no" included in the extraction target text S1 to be log (200000/30000).

Next, the vector conversion unit 112 multiplies the calculation results by the frequency calculation unit 106 and the rarity calculation unit 108 of the words included in the extraction target text S1, and each of the words included in the extraction target text S1. The tf-idf value is derived.

For example, the vector conversion unit 112 derives that the tf-idf value of the word “hot spring” included in the extraction target text S1 is 2/14 × log (200000/150) ≈0.446. Similarly, the vector conversion unit 112 derives the tf-idf value of the word “no” included in the extraction target text S1 as 2/14 × log (200,000 / 30,000) ≈0.118.

A word having a larger tf-idf value derived by the vector conversion unit 112 is a word having a higher "importance" in the extraction target text S1. That is, in the extraction target text S1 of FIG. 12, the word "hot spring" is a more important word when evaluated based on the tf-idf value derived by the vector conversion unit 112.

The vector conversion unit 112 derives a tf-idf value for the sentence S2 “next week / of / weather / of / information” included in the extraction target text, as in the extraction target text S1.

Further, the vector conversion unit 112 includes a sentence S3 "near / no / hot spring / check / want" included in the task text I2 and a sentence S4 "tomorrow / no / Tokyo / no / weather" included in the task text I2. The tf value and the idf value of the second word included in the task text are derived for each of the above, and the tf-idf value is derived.

[Sentence vector (tf-idf vector)]
FIG. 13 is a diagram showing an example of the tf-idf value derived by the vector conversion unit 112.

When the extraction target text is "today / no / weather / tell / tell", the vector conversion unit 112 derives the tf-idf value of each word included in the extraction target text. It is assumed that the vector conversion unit 112 derives, for example, that the tf-idf value of the word "today" is 0.5 and the tf-idf value of the word "no" is 0.02.

FIG. 14 is a diagram for explaining the tf-idf vector of the sentence vector.

As shown in FIG. 12, the vector conversion unit 112 generates a tf-idf vector using the derivation result of each tf-idf value of the word included in the extraction target text. For example, when the vector conversion unit 112 generates a tf-idf vector from the text “today / no / weather / tell / ga”, it can be represented by a vector represented by a distributed expression as shown in FIG. The word "ga" included in the text is an example of a word not included in the text to be extracted. The tf-idf vector value of the word not included in the extracted text is 0.

Similarly, the vector conversion unit 112 also performs a process of deriving the tf-idf vector value for the analysis result 120c of the corpus. By doing so, it becomes possible to refer to the tf-idf vector value in the selection process by the selection unit 114, so that a sentence vector suitable for generating the language model 120b can be selected, and the sentence vector can be selected with high accuracy. The generation of the language model 120b is realized by the language model generation unit 116a.

[Degree of reliability]
Hereinafter, the reliability derivation process of the reliability derivation unit 114a will be described more specifically. The reliability is a numerical value between 0 and 1.0 indicating the degree of evaluation of the reliability of the speech recognition result, and is a measure of how reliable the recognition result can be.

The reliability derivation unit 114a sets the reliability to 1.0, for example, when the reliability of the text is high, that is, when there is no other text that is a candidate for competition. The reliability is derived, for example, using the posterior probabilities of words obtained as search results of a large vocabulary continuous speech recognition engine. In addition, p * (similarity of tf-idf vector value) may be used for deriving the reliability.

FIG. 15 is a diagram for explaining the reliability derivation process by the reliability derivation unit 114a.

The reliability derivation unit 114a derives, for example, the reliability of each of the extraction target texts E1 to E4. For example, the selection unit 114 preferentially selects the extraction target texts E1 and E4 whose reliability derived by the reliability derivation unit 114a is a threshold value (for example, about 0.8) or more as task text. When a plurality of task texts can be selected, the selection unit 114 may preferentially select a highly reliable task text.

Further, the reliability derivation unit 114a may reset the reliability at a predetermined cycle. In that case, the reliability derivation unit 114a has a high possibility that the extraction target texts E1 to E4 are erroneous (such as erroneous conversion being included or none matching the task text). By setting a lower reliability for the texts E2 and E3, the processing accuracy by the selection unit 114 may be improved.

The selection unit 114 selects the sentence vector corresponding to the sentence to be analyzed based on the reliability derived by the reliability derivation unit 114a. The selection unit 114 preferentially selects, for example, the analyzed sentence obtained from the analysis result whose reliability derived by the reliability derivation unit 114a is equal to or higher than the threshold value. By setting the reliability by the reliability deriving unit 114a, it is possible to prevent the erroneous recognized sentence from being reflected in the language model 120b.

Further, when the selection unit 114 finds a sentence vector whose reliability is equal to or higher than the threshold value derived by the reliability derivation unit 114a, the selection unit 114 interrupts the selection process even if the selection process is in progress, thereby causing the language. The processing time required for the generation processing of the model 120b may be shortened.

[Similar evaluation of the meaning and content of text]
Hereinafter, a method for evaluating the similarity of the meaning and content of the text will be described.

The language model calculation unit 116 applies, for example, to a mathematical formula for obtaining the cosine similarity between the sentence vector of the text to be extracted (hereinafter referred to as “vector vi”) and the representative vector V of each cluster, thereby text. Evaluate the similarity of the meaning and content of. The formula for obtaining the cosine similarity is, for example, a formula for dividing the product of an arbitrary sentence vector v1 and an arbitrary sentence vector v2 by the product of the absolute value of the sentence vector v1 and the absolute value of the sentence vector v2. If it is close to 1, it is an expression indicating that the sentence vector v1 and the sentence vector v2 are similar.

If the derived cosine similarity is equal to or higher than the threshold value, the language model calculation unit 116 is similar to the sentence vector v1 and the sentence vector v2, that is, the text of the derivation source of the sentence vector v1 and the derivation source of the sentence vector v2. It is determined that the texts have the same or similar meanings.

FIG. 16 is a diagram for explaining a similarity evaluation method.

The language model calculation unit 116 derives, for example, a vector vi of the text to be extracted "How is the weather today?". The language model calculation unit 116 is a representative vector of cluster C1 including sentence vectors such as "tell me the weather today", "tell me the weather tomorrow", and "tell me if the weather is sunny" (hereinafter, "cluster representative"). The cosine similarity between the representative vector of cluster C2 (hereinafter referred to as "cluster representative vector CV2") including sentence vectors such as "vector CV1" and "make the sound of music quieter" and the vector vi. Evaluate the similarity of the meaning and content of the text by applying it to the desired formula.

Each of the task texts included in the cluster C1 is an example of a "teacher sentence".

For example, as shown in the figure, when the similarity between the vector vi and the cluster representative vector CV1 is 0.75 and the similarity between the vector vi and the cluster representative vector CV2 is 0.1, the language model calculation unit 116 is twisted. It is determined that the cluster C1 which is the derivation source of the cluster representative vector CV1 having a high degree of similarity has the same or similar meaning as the vector vi of the extraction target text.

Further, the language model calculation unit 116 selects a task text having the same or similar meaning as the vector vi of the extraction target text from the task texts included in the cluster C1.

The language model generation unit 116a generates a language model 120b that generates data in which the meaning of the task sentence selected by the language model generation unit 116a is associated with the meaning of the text S1 to be extracted.

In the illustrated example, for example, the vector values of the extraction target text "How is the weather today" and the task text "Tell me the weather today" in the cluster C1 have high similarity (the closest meaning). ) Is determined. In that case, when the extraction target text "How is the weather today?" Is input in the language model 120b, the result of the vector similarity evaluation as described above is reflected in the estimation, and the extraction target text is the task text ". I presume that it has the same or similar meaning as "Tell me about the weather today."

The command output unit 118 outputs a command based on the task text "Tell me the weather today", which is the estimation result, to the terminal device 20. As a result, the terminal device 20 obtains information about today's weather via the network NW, or the like, based on the processing result of the information processing device 100, based on the task text "Tell me the weather today". ) Is executed.

The similarity evaluation of the meaning and content of the text may be evaluated by a method other than the cosine similarity, and even if the text comparison evaluation based on the Levenshtein distance or the text comparison evaluation based on the Jaro-Winkler distance is performed. good.

[Language model generation process flow]
Hereinafter, the process of generating the language model 120b by the information processing apparatus 100 will be described. The information processing apparatus 100 generates, for example, a language model 120b for each type of corpus I1. Further, the administrator of the information processing apparatus 100 may periodically change / update the language model calculation text 120h. For example, the language model 120b may be changed / updated at the timing of the change / update of the language model calculation text 120h. Regeneration is done.

FIG. 17 is a flowchart showing an example of the flow of the generation processing of the language model 120b by the information processing apparatus 100.

First, the acquisition unit 102 acquires character information (corpus I1) to be used as a corpus (S100). Next, the analysis unit 104 analyzes the corpus I1 by an analysis method such as morphological analysis realized by applying the corpus I1 to the acoustic model 120a, and stores the analysis result as the corpus analysis result 120c in the storage unit 120 (S102). .. Next, the W2V execution unit 110 generates a vector value (word vector) of each of the words constituting the character information included in the analysis result 120c of the corpus (S104), and stores the word vector list 120f in the storage unit 120. (S106).

Next, the acquisition unit 102 acquires the task text I2 (S106). Next, the analysis unit 104 analyzes the task text I2 in the same manner as the corpus I1 (S108), and stores the analysis result as the analysis result 120d of the task text in the storage unit 120 (S110).

Next, the acquisition unit 102 acquires the voice data I3, which is the original information of the text to be extracted (S112). Next, the analysis unit 104 analyzes the voice data I3 in the same manner as the corpus I1 and the task text I2, and stores the analysis result in the storage unit 120 as the analysis result 120e of the text to be extracted (S114).

Next, the vector conversion unit 112 refers to the analysis result 120d of the task text and the word vector list 120f, generates a sentence vector of the task text, and stores it in the storage unit 120 as the task text vector list 120i (S114). Next, the vector conversion unit 112 generates a sentence vector of the text to be extracted (S116).

Next, the selection unit 114 selects a sentence vector (reflected in the language model 120b) that is the source of the language model 120b based on the sentence vector of the text to be extracted and the sentence vector of the task text (S118). Next, the language model generation unit 116a generates the language model 120b based on the selection result by the selection unit 114 (S120). This is the end of the description of the processing of this flowchart.

[Voice recognition processing]
FIG. 18 is a flowchart showing an example of the flow of voice recognition processing by the information processing apparatus 100.

First, the acquisition unit 102 acquires the voice data I2 from the terminal device 20 (S200). Next, the analysis unit 104 applies the voice data I2 output by the acquisition unit 102 to the acoustic model 120a to generate the text to be extracted (S202).

Next, the language model calculation unit 116 applies the extraction target text output by the analysis unit 104 to the language model 120b (S204). Next, the selection unit 114 selects a suitable candidate from the application result output by the language model calculation unit 116 (S206).

Next, the language model generation unit 116a generates output information based on suitable candidates (S208). Next, the command output unit 118 outputs the output information to the terminal device 20 or the like (S210). This is the end of the description of the processing of this flowchart.

According to the information processing apparatus 100 of the embodiment described above, the acquisition unit 102 that acquires voice data, the analysis unit 104 that analyzes the voice data acquired by the acquisition unit 102 and converts it into text, and the analysis unit 104. For each of the plurality of first words included in the text of the analysis result by, the frequency of the first word in the analyzed sentence (voice data I3) included in the text and including the first word, and the library information. Derivation of tf value and idf value (or tf-idf vector), which are the first index values that evaluate the rarity of the first word for the contained sentence, and associating them with the analyzed sentence, corpus I1, task text For each of the plurality of second words included in the library information such as I2 and the text 120h for language model calculation, the frequency of the second word in the note of interest included in the library information and including the second word, and the library. The tf value and the idf value (or tf-idf vector), which are the second index values for evaluating the rarity of the second word for sentences other than the sentence of interest included in the information, are derived and associated with the sentence of interest. Of these, the frequency calculation unit 106 and the rarity calculation unit 108 that perform at least one of them, the vector conversion unit 112 that converts the sentence analyzed by the analysis unit 104 into the vector value by the distributed representation, the frequency calculation unit 106 and the rarity calculation. A selection unit that selects a part of the sentence to be analyzed or the sentence of interest based on at least one of the first index value or the second index value derived by unit 108 and the vector conversion result by the vector conversion unit 112. Data in which 114 and a teacher sentence whose meaning is known and whose sentence vector is required are associated with the meaning of a teacher sentence whose sentence vector is close to that of the selected sentence selected by the selection unit 114 as the meaning of the selected sentence. By providing a language model generation unit 116a for generating data, it is possible to perform voice recognition processing more efficiently and with high accuracy.

[Modification example]
The language model 120b generated by the language model generation unit 116a may be a language model specialized for a fixed word. "Specialized in fixed words" means, for example, words that are always fixed to the language entered (such as "weather", "hot springs", "baseball" in the above example), or the same as or similar to fixed words. Words are included, and only processing related to fixed words is assumed.

In that case, when the frequency calculation unit 106 and the shortage calculation unit 108 generate a language model based on the text to be extracted, the first word is fixed and processed, and the language model is generated based on the corpus. , The second word is fixed and processing is performed. Further, when the frequency calculation unit 106 and the shortage calculation unit 108 generate a language model based on both the corpus and the text to be extracted, the first word and the second word are fixed to the same word for processing. Thereby, for example, a language model 120b specialized for the word "hot spring" and a language model 120b specialized for the word "weather" can be generated.

Although the embodiments for carrying out the present invention have been described above using the embodiments, the present invention is not limited to these embodiments, and various modifications and substitutions are made without departing from the gist of the present invention. Can be added.

20 ... Terminal device, 30 ... Control target device, 40 ... Service server, 100 ... Information processing device, 102 ... Acquisition unit, 104 ... Analysis unit, 106 ... Frequent calculation unit, 108 ... Rarity calculation unit, 110 ... W2V execution Unit, 112 ... Vector conversion unit, 114 ... Selection unit, 114a ... Reliability derivation unit, 116 ... Language model calculation unit, 116a ... Language model generation unit, 118 ... Command output unit, 120b ... Language model

Claims (11)

The acquisition unit that acquires audio data,
An analysis unit that analyzes the voice data and converts it into text,
For each of the plurality of first words included in the text of the analysis result by the analysis unit, the frequency of the first word in the analyzed sentence included in the text and including the first word, and library information. An index value deriving unit that derives a first index value that evaluates the rarity of the first word with respect to the sentence included in the sentence and associates it with the analyzed sentence.
A vector conversion unit that converts the sentence analyzed by the analysis unit into a vector value by distributed representation, and
A selection unit that selects a part of the sentence to be analyzed based on the first index value derived by the index value derivation unit and the conversion result by the vector conversion unit.
Among the teacher sentences whose meanings are known and whose vector value is required, the data in which the meanings of the teacher sentences whose vector values are close to those of the selection sentences selected by the selection unit are associated with the meanings of the selection sentences. The generator to generate and
Equipped with
For each of the plurality of second words included in the library information, the index value deriving unit determines the frequency of the second word in the sentence of interest included in the library information and including the second word. A second index value that evaluates the rarity of the second word for a sentence other than the note of interest included in the library information is derived and associated with the sentence of interest.
The selection unit is the analyzed sentence or the attention sentence based on at least one of the first index value or the second index value derived by the index value derivation unit and the conversion result by the vector conversion unit. Select some sentences from
The index value derivation unit is
When deriving only the first index value, the first word is fixed and processed.
When deriving only the second index value, the second word is fixed and processed.
When deriving the first index value and the second index value, the first word and the second word are fixed to the same word for processing.
The generation unit generates the associated data for each fixed word.
Information processing equipment. The acquisition unit that acquires audio data,
An analysis unit that analyzes the voice data and converts it into text,
For each of the plurality of first words included in the text of the analysis result by the analysis unit, the frequency of the first word in the analyzed sentence included in the text and including the first word, and library information. An index value deriving unit that derives a first index value that evaluates the rarity of the first word with respect to the sentence included in the sentence and associates it with the analyzed sentence.
A vector conversion unit that converts the sentence analyzed by the analysis unit into a vector value by distributed representation, and
A selection unit that selects a part of the sentence to be analyzed based on the first index value derived by the index value derivation unit and the conversion result by the vector conversion unit.
Among the teacher sentences whose meanings are known and whose vector value is required, the data in which the meanings of the teacher sentences whose vector values are close to those of the selection sentences selected by the selection unit are associated with the meanings of the selection sentences. The generator to generate and
Equipped with
For each of the plurality of second words included in the library information, the index value deriving unit determines the frequency of the second word in the sentence of interest included in the library information and including the second word. A second index value that evaluates the rarity of the second word for a sentence other than the note of interest included in the library information is derived and associated with the sentence of interest.
The selection unit is the analyzed sentence or the attention sentence based on at least one of the first index value or the second index value derived by the index value derivation unit and the conversion result by the vector conversion unit. Select some sentences from
A command output unit that estimates the meaning of the recognized sentence based on the vector value converted by the vector conversion unit and outputs a command based on the estimation result is further provided.
The vector conversion unit converts the recognized sentence included in the text of the analysis result by the analysis unit into a vector value by distributed representation.
The command output unit estimates the meaning of the recognized sentence based on the similarity of the vector value with the sentence included in the associated data, and outputs a command based on the estimation result.
The selection unit determines whether or not the voice data includes the user's task execution intention based on the position information given to the voice data.
Information processing equipment. At least one of the first index value and the second index value is a tf-idf value.
The information processing apparatus according to claim 1 or 2 . Further equipped with a reliability derivation unit for deriving the reliability of the analysis result,
The selection unit selects the sentence to be analyzed based on the reliability.
The information processing apparatus according to any one of claims 1 to 3 . The selection unit preferentially selects the analysis target sentence obtained from the analysis result whose reliability is equal to or higher than the threshold value.
The information processing apparatus according to claim 4 . When the selected sentence selected from the analysis result whose reliability is equal to or higher than the threshold value, the selection unit ends the selection process.
The information processing apparatus according to claim 5 . The selection unit changes the accuracy with which the corresponding task is selected according to the input environment of the voice data estimated based on the position information.
The information processing apparatus according to claim 2 . The computer
Get voice data,
The voice data is analyzed and converted into text, and then
For each of the plurality of first words included in the analysis result text, the frequency of the first word in the analyzed sentence included in the text and including the first word, and the sentence included in the library information. A first index value that evaluates the rarity of the first word is derived and associated with the analyzed sentence.
Convert the parsed sentence into a vector value by distributed representation,
Based on the first index value and the vector conversion result, a part of the sentences to be analyzed is selected.
Among the teacher sentences whose meanings are known and whose vector value is required, data is generated in which the meanings of the selected selection sentences and the teacher sentences whose vector values are close to each other are associated with the meanings of the selection sentences.
For each of the plurality of second words included in the library information, the frequency of the second word in the sentence of interest included in the library information and including the second word, and the frequency of the second word included in the library information. A second index value that evaluates the rarity of the second word for a sentence other than the sentence of interest is derived and associated with the sentence of interest.
In the selection process, a part of the sentence to be analyzed or the sentence of interest is selected based on at least one of the first index value or the second index value and the vector conversion result.
When deriving the first index value or the second index value,
When deriving only the first index value, the first word is fixed and processed.
When deriving only the second index value, the second word is fixed and processed.
When deriving the first index value and the second index value, the first word and the second word are fixed to the same word for processing.
Generate the associated data for each fixed word.
Information processing method. The computer
Get voice data,
The voice data is analyzed and converted into text, and then
For each of the plurality of first words included in the analysis result text, the frequency of the first word in the analyzed sentence included in the text and including the first word, and the sentence included in the library information. A first index value that evaluates the rarity of the first word is derived and associated with the analyzed sentence.
Convert the parsed sentence into a vector value by distributed representation,
Based on the first index value and the vector conversion result, a part of the sentences to be analyzed is selected.
Among the teacher sentences whose meanings are known and whose vector value is required, data is generated in which the meanings of the selected selection sentences and the teacher sentences whose vector values are close to each other are associated with the meanings of the selection sentences.
For each of the plurality of second words included in the library information, the frequency of the second word in the sentence of interest included in the library information and including the second word, and the frequency of the second word included in the library information. A second index value that evaluates the rarity of the second word for a sentence other than the sentence of interest is derived and associated with the sentence of interest.
In the selection process, a part of the sentence to be analyzed or the sentence of interest is selected based on at least one of the first index value or the second index value and the vector conversion result.
Based on the converted vector value, the meaning of the recognized sentence is estimated, and the command based on the estimation result is output.
The recognized sentence included in the text of the analysis result is converted into a vector value by distributed representation, and is converted into a vector value.
Based on the similarity of the vector value with the sentence included in the associated data, the meaning of the recognized sentence is estimated, and the command based on the estimation result is output.
Based on the position information given to the voice data, it is determined whether or not the voice data includes the execution intention of the user's task.
Information processing method. On the computer
Get voice data,
The voice data is analyzed and converted into text, and then
For each of the plurality of first words included in the analysis result text, the frequency of the first word in the analyzed sentence included in the text and including the first word, and the sentence included in the library information. A first index value that evaluates the rarity of the first word is derived and associated with the analyzed sentence.
Convert the parsed sentence into a vector value by distributed representation,
Based on the first index value and the vector conversion result, a part of the sentences to be analyzed is selected.
Among the teacher sentences whose meanings are known and whose vector value is required, data is generated in which the meanings of the selected selection sentences and the teacher sentences whose vector values are close to each other are associated with the meanings of the selection sentences.
For each of the plurality of second words included in the library information, the frequency of the second word in the sentence of interest included in the library information and including the second word, and the frequency of the second word included in the library information. A second index value that evaluates the rarity of the second word for a sentence other than the sentence of interest is derived and associated with the sentence of interest.
In the selection process, a part of the sentence to be analyzed or the sentence of interest is selected based on at least one of the first index value or the second index value and the vector conversion result.
When deriving the first index value or the second index value,
When deriving only the first index value, the first word is fixed and processed.
When deriving only the second index value, the second word is fixed and processed.
When deriving the first index value and the second index value, the first word and the second word are fixed to the same word for processing.
Generate the associated data for each fixed word.
A program that lets you do things . On the computer
Get voice data,
The voice data is analyzed and converted into text, and then
For each of the plurality of first words included in the analysis result text, the frequency of the first word in the analyzed sentence included in the text and including the first word, and the sentence included in the library information. A first index value that evaluates the rarity of the first word is derived and associated with the analyzed sentence.
Convert the parsed sentence into a vector value by distributed representation,
Based on the first index value and the vector conversion result, a part of the sentences to be analyzed is selected.
Among the teacher sentences whose meanings are known and whose vector value is required, data is generated in which the meanings of the selected selection sentences and the teacher sentences whose vector values are close to each other are associated with the meanings of the selection sentences.
For each of the plurality of second words included in the library information, the frequency of the second word in the sentence of interest included in the library information and including the second word, and the frequency of the second word included in the library information. A second index value that evaluates the rarity of the second word for a sentence other than the sentence of interest is derived and associated with the sentence of interest.
In the selection process, a part of the sentence to be analyzed or the sentence of interest is selected based on at least one of the first index value or the second index value and the vector conversion result.
Based on the converted vector value, the meaning of the recognized sentence is estimated, and the command based on the estimation result is output.
The recognized sentence included in the text of the analysis result is converted into a vector value by distributed representation, and is converted into a vector value.
Based on the similarity of the vector value with the sentence included in the associated data, the meaning of the recognized sentence is estimated, and the command based on the estimation result is output.
Based on the position information given to the voice data, it is determined whether or not the voice data includes the execution intention of the user's task.
A program that lets you do things .

Images