JP2008180801A - Automatic answering apparatus and method - Google Patents

Abstract

[PROBLEMS] In a natural language dialogue, a dialogue that corrects errors in speech recognition and sentence analysis is incorporated in a part of a natural dialogue, and the dialogue proceeds appropriately while correcting an input error. An object is to provide an automatic answering apparatus capable of interactive processing.
An automatic answering apparatus according to the present invention includes a morpheme analysis unit that performs morpheme analysis on a speech-recognized sentence, a word extraction unit that extracts a word that satisfies a certain part of speech in the sentence subjected to the morpheme analysis, and the extraction A co-occurrence degree calculating means for obtaining the sum of co-occurrence degrees with other words for each word that has been made, and a threshold value for determining a word whose sum of co-occurrence degrees is lower than a predetermined threshold value as a speech recognition error word The determination means and correction sentence creation means for creating a correction sentence when there is a word determined to be a speech recognition error.
[Selection] Figure 2

Description

  The present invention relates to an automatic answering apparatus having an error correction function for speech recognition and sentence analysis.

Today, the use of the Internet and e-mail is rapidly spreading personal computers in homes. When operating a computer, the user interacts with the computer using a keyboard and mouse, but unlike human interaction, the user communicates a request to the computer by entering a command with the keyboard or clicking with the mouse. The personal computer processes according to the request and displays the result, so it can be said that it is still difficult for beginners who do not know anything about the interaction method. In order to interact with a computer easily and easily, anyone can interact with a computer in the natural language used by humans without using even these basic skills and knowledge to operate the computer. It is desirable to use speech recognition instead of a keyboard.
Therefore, a speech recognition method for performing an appropriate dialogue is disclosed in Japanese Patent Laid-Open No. 2003-015688.

According to the speech recognition method of the present invention, the similarity between the input word speech data and the speech pattern in the word dictionary is calculated, a plurality of speech patterns are set as word recognition candidates in descending order of similarity, and the word recognition candidates are used as speech. When the input operation is confirmed by outputting in the above, the expression content of the output voice is changed depending on whether the similarity is higher or lower than the reference.
JP 2003-015688 A

  According to the speech recognition method of the background art, the similarity between the input word speech data and the speech pattern in the word dictionary is usually calculated from zero to 9999 calculation results, and the higher the score, the higher the score. Similarity increases. Therefore, for example, it is empirically known that the probability of misrecognition increases when 3000 points or less, with 3000 points as a reference, and the possibility of misrecognition decreases when 3000 points are exceeded. By changing the content of the talkback expression based on the score, the user can understand that the device is making a proper decision, and can increase the reliability of the device.

However, the background art speech recognition method has a problem that the calculation of the similarity of speech patterns is limited to the determination of erroneous recognition within a range in the word dictionary. In order to conduct a natural language conversation using speech, the input speech is converted into text information by the speech recognition unit, the sentence is analyzed by the input analysis unit, and the response is collected by response information collection and dialogue control processing by the problem solving unit. Processing is required in which the content is determined and the response sentence generated by the utterance generation unit is output using the voice output unit. At this time, there is a possibility that a recognition error occurs in the speech recognition unit and a sentence analysis error occurs in the input analysis unit, which causes a problem in that the solution to the problem fails and an appropriate response becomes difficult.
Here, speech recognition errors and sentence analysis errors that occur when realizing a natural language conversation using speech will be described below.

[1. Speech recognition error]
When inputting characters using the keyboard, if there is a typo or grammatical error such as a typo or omission, or a kana-kanji conversion error, the user can input it with some correction before entering it. Then, what is uttered as it is thought is recognized and input as it is, so there is a high possibility of including recognition errors and grammatical errors. If there is an error in speech recognition, the error will be enlarged in the analysis based on the wrong input even in the next sentence analysis, which will be a major obstacle to solving the problem. The following are examples of speech recognition errors.
Spoken sentence: The power does not turn on.
Recognition result: The power supply does not work.
There are the following ways of speech recognition errors.

  (1) When there is no or part of the sound due to user's pronunciation or microphone malfunction as misrecognition due to voice recognition error.

  (2) The case where misrecognition occurs due to unintentional utterance or ambient noise as a misrecognition due to noise phonemes mixed in the phoneme string.

  (3) The case of misrecognition of similar phonemes due to the same vowel sound and similar phonemes.

  (4) Misrecognition due to improper separation of phoneme strings, such as when a word boundary is mistaken when converting speech to a word, or when there is a blank in the middle of the utterance of one word.

  (5) In case of conversion error of homonyms.

[2. Sentence analysis error]
Examples of sentence analysis include morphological analysis, syntactic analysis, semantic analysis, context analysis, and discourse analysis. At present, considerable technology has been established for morphological analysis and syntax analysis, and dependency analysis is often used as Japanese syntax analysis. JUMAN, ChaSen, MeCab are widely used for morphological analyzers, and KNP, CaboCha, etc. are widely used for dependency analyzers. These are processed with considerable accuracy, but there are many errors. When trying to make meaningful conversations such as question answering, morphological analysis and syntactic analysis must be followed by semantic analysis. An error in morphological analysis affects syntactic analysis, and an error in syntactic analysis affects semantic analysis, so even though these precisions are quite high, errors that actually occur cannot be ignored. An example of an error is shown. In this example, “mail” is related to “first”, but considering the meaning of the sentence, “mail” should be related to “setting method”. In general, dependency analysis errors are likely to occur when a sentence has elements of “link modifier”, “insert phrase”, “subordinate clause”, and “parallel structure”.

  A common problem in speech recognition and sentence analysis is that even if each accuracy is increased, it will not be possible to obtain 100% accuracy by itself. This is because natural language inherently contains ambiguity, and there is a part that cannot be determined during speech recognition, morphological analysis, or syntactic analysis without considering the meaning and context. In addition, as long as humans can input freely, the input sentence itself may be grammatically incorrect. Therefore, natural language dialogue requires a mechanism for correcting these errors while conducting dialogue. In order to realize a spoken dialogue system, an error correction function for each error is required.

  The present invention has been made to solve the above-mentioned problems, and in natural language dialogue, dialogue that corrects errors in speech recognition and sentence analysis is incorporated into a part of the dialogue in a natural form. An object of the present invention is to provide an automatic answering apparatus capable of performing dialogue processing for appropriately proceeding dialogue while correcting an input error.

[Interaction Control]
As a principle of how the speaker should transmit information to the listener, the following Grice standard is generally known. (1) As an official quantity, convey information without excess or deficiency. (2) As a quality standard, state that you believe that there is grounded truth or truth. (3) Talk about the relationship between the speaker and the listener as a related standard. (4) Speak clearly and concisely in order as a standard for the mode. This is applied to the utterance in the error correction dialogue and considered as follows. (1a) Corresponding to the official amount of information, information that is judged to be necessary for correcting errors in speech recognition and sentence analysis is conveyed without excess or deficiency. (2a) Corresponding to the quality standard, the part of the input sentence that is not an error in speech recognition or sentence analysis or that can be determined to be an error is reported. (3a) Talk about errors in speech recognition and sentence analysis for input sentences in accordance with the related norms. (4a) In correspondence with the official state of the mode, speak so that it can be clearly communicated to which part of the input sentence it is judged as an error.

  In the error correction dialogue according to the present invention, an utterance is created so as to satisfy the above-mentioned conditions. For example, in response to an error in voice recognition such as “I do n’t have a power supply”, I use all of the input sentences to say “Did you say the power supply is OK?” “Is the power supply not good?” The utterances such as "does not meet the norm of quantity or quality", in this case, the utterance such as "What was the power supply?" And "What is the power supply?" Is desirable.

(1) An automatic answering apparatus according to the present invention includes a morpheme analyzing unit that performs morphological analysis on an input sentence, a word extracting unit that extracts words that satisfy a certain condition in a part of speech in the sentence subjected to the morpheme analysis, and the extracted A co-occurrence degree calculation means for obtaining the sum of co-occurrence degrees with other words for each word, and a threshold determination means for judging a word whose numerical value of the sum of co-occurrence degrees is lower than a predetermined threshold as an input error word And a corrected sentence creating means for creating a corrected sentence when there is a word determined to be an input error. Here, the “input sentence” includes, for example, a sentence that has been voice-recognized after voice input, a sentence written by a keyboard, and the like. The “input error” includes, for example, a voice recognition error, a written sentence error, and the like. Note that errors in written sentences include typographical errors and omissions, and kana-kanji conversion errors. The case where the part of speech of “the word in which the part of speech in the sentence subjected to morphological analysis satisfies a certain condition” satisfies a certain condition includes a word that is a noun, a verb, an adjective and is not independent and is not a number. The “corrected sentence” includes a rephrase request sentence when there is a recognition error, a confirmation sentence for the restatement request sentence, and an input error correction sentence in which words are replaced according to the restatement request.
This makes it possible to easily and accurately find and correct the presence of an unnatural word in a dialogue sentence in an input by speech recognition or the like. In addition, since statistical processing is performed in determining speech recognition errors, the reliability of the numerical value of the sum of co-occurrence levels increases as the number of example sentences from dialogue with the speaker increases, and The accuracy of correction is also improved.

(2) The automatic response device according to the present invention includes a morpheme analyzing unit that performs morphological analysis on an input sentence, a word extracting unit that extracts words that satisfy a certain condition in a part of speech in the sentence subjected to the morpheme analysis, and the extracted A co-occurrence degree calculation means for obtaining the sum of co-occurrence degrees with other words for each word, and a threshold determination means for judging a word whose numerical value of the sum of co-occurrence degrees is lower than a predetermined threshold as an input error word A corrected sentence creating means for creating a corrected sentence when there is a word determined to be an input error, and a corrected sentence comparison for determining whether the corrected sentence has a different word compared to the sentence before correction It is comprised from the determination means and the Yes / No determination means which determines whether there is an error in a correction sentence. In the case where the part of speech of “input sentence”, “input error”, “word in which the part of speech in the sentence subjected to morphological analysis satisfies a certain condition” satisfies a certain condition and “corrected sentence”, Same as described.
As a result, dialogues that correct input errors such as speech recognition can be incorporated into a part of the natural dialogue, and the dialogue can be appropriately advanced while correcting input errors such as speech recognition. it can.

(3) In the automatic answering apparatus according to the present invention, if necessary, the corrected sentence is a sentence before correction up to the word one word before the word determined to be input error among the words for which the co-occurrence degree is calculated. This includes means for creating a corrected sentence using a template created by paying attention to the word / phrase at the end of the answer sentence.
This clearly shows where the recognition error occurred in the input sentence. In addition, it is possible to incorporate a dialogue for correcting an input error into a part of a natural dialogue, and to appropriately advance the dialogue while correcting the input error.

(4) An automatic answering apparatus according to the present invention includes dependency analysis means for dependency analysis of an input sentence, and word extraction means for extracting one word whose part of speech in the dependency analyzed sentence satisfies a certain condition. A dependency degree calculating means for obtaining a dependency degree for the extracted set of one word and another word, a dependency relation determining means for determining a magnitude relation of the dependency degree between the pairs, It consists of dependency correction sentence creation means for creating a dependency correction sentence for a set of words having a low dependency degree. Here, the “input sentence” includes, for example, a sentence that has been voice-recognized after voice input, a sentence written by a keyboard, and the like. The case where the part-of-speech of “a word from which one part-of-speech in a sentence satisfies a certain condition” satisfies a certain condition includes nouns, verbs, adjectives that are non-independent and not numbers. Further, the “dependency correction sentence” includes a dependency confirmation sentence for the speaker, and a correction sentence that has been corrected upon confirmation from the speaker of the dependency confirmation sentence.
As a result, it is possible to easily and accurately find and correct the presence of an unnatural word dependency in a dialogue sentence. Since statistical processing is performed in determining sentence analysis errors, the reliability of dependency numerical values increases as the number of sentence examples from dialogue with the speaker increases, and the accuracy of correcting sentence analysis errors Will also improve.

(5) An automatic answering apparatus according to the present invention includes dependency analysis means for dependency analysis of an inputted sentence, word extraction means for extracting one word whose part of speech in the sentence satisfies a certain condition, A dependency calculation unit for creating a combination of a word and another word to obtain dependency, a dependency relationship determination unit for determining the magnitude relationship of the dependency between the pairs, and a set of words having a low dependency It is composed of dependency correction text creating means for creating a dependency correction text and a Yes / No determination means for determining whether or not the dependency correction text is an error. In the case where the part of speech of the “input sentence”, “the word in which the part of speech in the sentence extracts one word that satisfies a certain condition” satisfies the certain condition and “the dependency correction sentence”, It is the same.
Thereby, it is possible to incorporate a dialogue for correcting the sentence analysis errors into a part of the natural dialogue, and to appropriately advance the dialogue while correcting the input error.

(6) In the automatic answering apparatus according to the present invention, if necessary, the dependency correction sentence is a word having a larger dependency degree and a response sentence than words determined to be sentence analysis errors among words for which the dependency degree is calculated. This includes means for creating a corrected sentence using a template created by paying attention to the last word / phrase.
This clearly shows where the sentence analysis error occurred in the input sentence. Moreover, it is possible to incorporate a dialogue for correcting errors in sentence analysis into a part of a natural dialogue, and to appropriately advance the dialogue while correcting an input error.

(7) The automatic answer program according to the present invention includes a morpheme analysis procedure for morphological analysis of an input sentence, a word extraction procedure for extracting words that satisfy a certain condition in the part of speech in the sentence subjected to the morpheme analysis, and the extracted A co-occurrence degree calculation procedure for obtaining the sum of co-occurrence degrees with other words for each word, and a threshold determination procedure for determining a word whose sum of co-occurrence degrees is lower than a predetermined threshold as an input error word The computer is caused to function as a corrected sentence creation procedure for creating a corrected sentence when there is a word determined to be an input error.

(8) The automatic answering method according to the present invention includes a morpheme analysis step for morphological analysis of an input sentence, a word extraction step for extracting a word that satisfies a certain part of speech in the sentence subjected to the morpheme analysis, and the extracted A co-occurrence degree calculation step for obtaining the sum of co-occurrence degrees with other words for each word, and a threshold determination step for determining a word whose numerical value of the sum of co-occurrence degrees is lower than a predetermined threshold as an input error word And a corrected sentence creating step for creating a corrected sentence when there is a word determined to be an input error.

(9) An automatic answer program according to the present invention includes a dependency analysis procedure for dependency analysis of an input sentence, and a word extraction procedure for extracting one word in which the part of speech in the dependency analyzed sentence satisfies a certain condition. A dependency degree calculation procedure for obtaining a dependency degree for the extracted set of one word and another word, a dependency relationship determination procedure for determining a magnitude relation of the dependency degree between the pairs, The computer is caused to function as a dependency correction sentence creation procedure for creating a dependency correction sentence for a set of words having a low dependency degree.

(10) An automatic answering method according to the present invention includes a dependency analysis step of performing dependency analysis on an input sentence, and a word extraction step of extracting one word in which the part of speech in the dependency analyzed sentence satisfies a certain condition. A dependency level calculating step for determining a dependency level for the pair of the extracted one word and another word, a dependency relationship determination step for determining a magnitude relationship of the dependency level between the sets, This is composed of a dependency correction sentence creation step for creating a dependency correction sentence for a set of words having a low dependency degree.
These outlines of the invention do not enumerate the features essential to the present invention, and a sub-combination of these features can also be an invention.

(First embodiment of the present invention)
Here, the present invention can be implemented in many different forms. Therefore, it should not be interpreted only by the description of the following embodiment.
In the embodiment, the apparatus will be mainly described. However, as is apparent to those skilled in the art, the present invention can also be implemented as a program usable on a computer. Further, the present invention can be implemented in hardware, software, or software and hardware embodiments. The program can be recorded on any computer-readable medium such as a hard disk, CD-ROM, DVD-ROM, optical storage device, or magnetic storage device. Furthermore, the program can be recorded on another computer via a network.

  In the first embodiment according to the present invention, as an example of a dialogue domain, a question and answer at a call center that provides personal computer technical support is taken as an example, and about 35,000 cases totaled by about 3 years of practice at a mail call center are collected. It applies to interactive processing that performs error correction interactively by applying domain knowledge obtained from email data of questions and answers.

[1. Hardware configuration diagram.]
FIG. 1 shows a hardware configuration diagram of an automatic answering apparatus according to the first embodiment of the present invention. The computer 1 includes, for example, a CPU (Central Processing Unit) 2, a main memory 3, an HDD (Hard Disk Drive) 4, a video card 5, a mouse 6, a keyboard 7, an optical disk 8, and the like. Note that a microphone as an input device necessary for voice recognition and a speaker as an output device can be externally connected. As the speech recognition engine, commercially available speech recognition software for PC can be used. As the interactive method, a multimodal interactive method that uses not only voice but also images can be used.

[2. Block diagram]
FIG. 2 is a block diagram of the automatic answering apparatus according to the first embodiment of the present invention. The voice data of the speaker is input from the voice input unit 10. The input voice data is stored in the voice storage unit 20. Here, the voice storage unit 20 is operated by the main memory 3, the HDD 4, and the like. Note that recording can also be performed by magnetic storage, optical storage, semiconductor storage, or the like. Then, the voice data read from the voice storage unit 20 is recognized by the voice recognition unit 30 and is analyzed by the input analysis unit 50. The voice data is sent to the problem solving unit 70 when there is no voice recognition error or sentence analysis error. The problem solving unit 70 selects an appropriate answer sentence for the question sentence from the speaker. The voice data of an appropriate answer sentence is processed in the utterance generation unit 80 and the voice output unit 90. Note that not only voice data but also sentence data written with a keyboard or the like can be processed.

[2.1 Correction of speech recognition errors]
If there is an error in speech recognition, the error recognition correction unit 40 performs processing to correct the speech recognition error. A dialogue that corrects an error in speech recognition uses word co-occurrence to determine whether or not there is an error. If there is an error, a dialogue is requested as a response based on the dialogue control. In addition to voice recognition, errors to written sentences can be similarly corrected.

  First, in the misrecognition determination unit 41, the speech data is sent to the morphological analysis unit 411, the misrecognition word extraction unit 412, the co-occurrence sum calculation unit 413, and the threshold calculation unit 414, and whether or not there is an error in speech recognition. Determine.

  In a specific example, a question sentence is extracted from about 16,500 mail data, morphological analysis is performed on 46,224 sentences using a morphological analyzer MeCab, and parts of speech are nouns, verbs, and adjectives. Examining the co-occurrence degree C with other words in the same sentence of words that are not independent words but not numbers, the co-occurrence degree for 13,942 words was obtained. Here, in one sentence, the co-occurrence degree C (a, b) of a word a with another word b is the co-occurrence frequency cf (a, b) in all question sentences and the word co-occurring with the word a. Using the number cn (a), it is defined as:

この値は語ａの他の語ｂとの同じ文中での出現しやすさを表しているので、入力文中の各語ｘについて、文中の他の語Ｗｉとの共起度の和

Since this value represents the ease of appearance in the same sentence with another word b from the word a, the sum of co-occurrence degrees with other words Wi in the sentence for each word x in the input sentence

を求めると、Ｃ（ｘ）が低いものは、音声認識誤りによって発話とは違う語が認識された可能性が高く、逆にＣ（ｘ）が高いものは、正確に認識できている可能性が高いと推測できる。

When C (x) is low, there is a high possibility that a word different from the utterance is recognized due to a speech recognition error, and conversely, a case where C (x) is high may be correctly recognized. Can be estimated.

  Therefore, the morphological analysis unit 411 analyzes the voice-recognized question sentence, and the misrecognized word extraction unit 412 extracts words that are nouns, verbs, and adjectives and are not independent and are not numbers. For each word extracted by the co-occurrence degree calculation unit 413, the sum C (x) of the co-occurrence degree with other words is calculated by the above-described method, and the threshold value calculation unit 414 calculates the sum of the co-occurrence degrees from the threshold value. A low word is determined as a recognition error word. Although the threshold is set to 0.01 for convenience, an empirical value in statistical processing can also be applied.

As an example, for the input sentence “Personal computer is very slow”, the result of the morphological analysis is as shown in FIG. 3, and the part of speech is a noun, verb, adjective and is not independent, and is not a number. “Operation” and “Slow” are extracted. When calculating the sum of the co-occurrence of the “computer” with the other two words,
C (PC) = C (PC, operation) + C (PC, slow)
= 0.139954790
Similarly, when calculating “motion” and “slow”,
C (operation) = 0.08925836809999173
C (slow) = 0.07646221645888845
Result. In this case, since any value is equal to or greater than the threshold value, it can be determined that the word is correctly recognized.

Since the voice recognition error is judged statistically, the reliability of the sum of the co-occurrence levels increases as the number of example sentences from the dialogue with the speaker increases. Accuracy is also improved.
Processing performed in the corrected sentence creation unit 42, the corrected sentence comparison determination unit 43, and the erroneous recognition Y / N determination unit 44 when the erroneous recognition determination unit 41 determines that there is a recognition error will be described below.

  If there is a word determined to be a speech recognition error, the corrected sentence creation unit 42 creates a rephrase request sentence. In other words, in order to avoid unnecessary speech recognition errors, it is better to eliminate the re-input of correctly recognized words, include an error part, and make the sentence as short as possible before the previous input. Therefore, the request sentence to be created needs to clearly indicate where the recognition error has occurred in the question. So, using the input sentence up to the word before the word judged to be incorrect among the words whose co-occurrence degree has been calculated, add "What?" To do. In addition, when the part of speech of the word immediately before the word determined to be an error is a noun, if there is a particle immediately after that, it is added afterwards so that the rephrasing request sentence becomes a natural form.

  For example, if the utterance “Tell me how to delete the search word history” is misrecognized as “Tell me how to crop the search word history”, it is morphologically analyzed as shown in FIG. Crop) = 0.0 and the noun “Crop” is judged to be wrong. ”Is used to create a rephrasing request sentence“ What is in the search word history? ”. This restatement request sentence has various expressions that are determined according to the control of the dialogue, such as “How did the search word history change?”. The corrected sentence comparison / determination unit 43 compares the input sentence re-phrased according to the re-phrase request sentence, and if the word determined to be wrong is different, the corrected sentence preparation unit 42 selects the word. Replace it to create a correction and proceed with the dialogue. If they are the same, the corrected sentence creation unit 42 creates a confirmation sentence that can be answered with Yes / No, using only that word and adding "?" Or "Did you say?" Here, the sentence end expression in the rephrasing request sentence and the confirmation sentence can be made into a template. If the answer is YES in the misrecognition Y / N determination unit 44, the conversation is continued as it is, and if it is No, the restatement request sentence is re-executed.

FIG. 5 shows an example of a speech recognition error correction dialogue of the automatic answering apparatus according to the first embodiment of the present invention.
First, the user (speaker) inputs a voice as “a character suddenly jumps to a complicated place while inputting characters”. However, the system (computer 1) mistakenly recognizes that “the character suddenly jumps to the wrong place in the middle of inputting the character”. Therefore, the system (computer 1) creates a request sentence by rephrasing “What is the character suddenly in the middle of inputting the character?”. A user (speaker) inputs a corrected sentence by voice, “Jump to a strange place”. As a result, the system (computer 1) can correct the voice input sentence “The character suddenly jumps to a strange place in the middle of inputting the character”.

[2.2 Correction of sentence analysis errors]
If there is an error in the sentence analysis, the error analysis correction unit 60 performs processing to correct the sentence analysis error. For correction of sentence analysis errors, use the dependency level of words to find a dependency relationship that may be erroneous in the dependency analysis result of the input sentence, and if found, check the dependency. . The method will be described in detail below.

  First, in the misanalysis determination unit 61, the speech data is processed by the dependency analysis unit 611, the misanalysis word extraction unit 612, the dependency degree calculation unit 613, and the dependency relationship determination unit 614, so that there is an error in sentence analysis. Determine.

  As a specific example, dependency analysis was performed on the question sentences 46 and 224 using the dependency analyzer CaboCha. From the analysis result, when a word a is related to the word b for a word that is part of a noun, a verb, or an adjective and is not independent and is not a number, it is easy to relate to the word b of the word a. Was determined as the degree of dependency D (a, b). The dependency degree D (a, b) is a frequency df (a, b) in which the words a and b appear as a dependency relationship in all questions, and the number dn (a) of the words in the dependency relationship with a. And defined as:

このとき語ａの品詞が名詞で、その直後に助詞ｃがあるときは、語ｂに助詞ｃを加えて考え、助詞が違う係り受け関係は別のものとして扱った。Ｄ（ａ，ｂ）は語ａ（まはた語ａと助詞ｃ）の語ｂとの係りやすさを表しているので、入力文を係り受け解析したときに、Ｄ（ａ，ｂ）が低いものは、係り受け解析が誤っている可能性があり、逆にＤ（ａ，ｂ）が高いものは、正確に係り受け解析ができている可能性が高いと推測できる。

At this time, when the part of speech of the word a is a noun, and there is a particle c immediately thereafter, the particle c is added to the word b, and the dependency relationship with different particles is treated as different. Since D (a, b) represents the ease with which the word a (or the word a and the particle c) is related to the word b, when the input sentence is subjected to dependency analysis, D (a, b) is A low value may indicate that the dependency analysis is erroneous, and conversely, a high value of D (a, b) can be estimated to have a high possibility that the dependency analysis is correctly performed.

  As described above, dependency analysis may be wrong, so this value cannot be overly trusted, but since the frequency is examined by analyzing a large number of sentences, the accuracy of dependency analysis In other words, it expects that a large number of errors will not appear, and it is used as a decision material for determining which words are subject to correction of analysis errors.

  The dependency analysis unit 611 performs dependency analysis on the input sentence, and the misanalyzed word extraction unit 612 extracts, for each clause of the input sentence, a non-independent, non-numerical word whose part of speech is a noun, verb, or adjective. When nouns are continuous in a phrase, they are grouped together as nouns, and those that have the verb “s” after the nouns are combined as verbs. If there is a particle after the noun, the particle is taken out. In addition, the nouns, verbs, and adjectives are similarly extracted from the clauses in which the clause is related, and a dependency set is created. The dependency level calculation unit 613 obtains the above-described dependency level D (a, b) for the dependency set, and the dependency level determination unit 614 further D with another word Wi appearing in the sentence of the word a. When (a, Wi) is present, an utterance is performed to confirm the dependency of the words a and b.

For example, for an input of “I do not know the initial setting of mail”, the result of dependency analysis shown in FIG. 6 is obtained, and D (first of mail), D (first, setting), D (setting is , Understand) is calculated. For "email", D (email setting) and D (email understanding) are calculated,
D (first of mail) = 0.0
D (setting of mail) = 0.01777993527508091
D (I understand mail) = 0.0
Thus, it can be seen that the degree of dependency is higher when “Mail” is related to “Setting”.

  Since sentence analysis errors are statistically processed, as the number of sentence examples from conversations with speakers increases, the reliability of dependency-related numerical values also increases, and the accuracy of sentence analysis errors increases. Will also improve. The contents processed in the dependency sentence creation unit 62 and the error analysis Y / N determination unit 63 when the error analysis determination unit 41 determines that there is an analysis error are shown below.

  Since the dependency level is higher when “Mail” is “Setting” than the dependency analysis described above, the dependency statement creation unit 62 sends a confirmation message that can be answered with Yes / No, “Is this mail first?” create. If the answer is yes, the error analysis Y / N judgment unit 63 will go directly to the next dialogue. If the answer is No, if the answer contains a correction such as “Email setting”, change the “mail” destination to “setting”. To the next dialogue. If the reply does not include a correction, the dependency sentence creation unit 62 creates a confirmation sentence “Is it a mail setting?” Using “setting” that has a high degree of mail dependency.

FIG. 7 shows an example of a dialogue for correcting a sentence analysis error in the automatic answering apparatus according to the first embodiment of the present invention.
First, the user (speaker) inputs a voice saying that the application that the child has installed has been deleted. However, the system (computer 1) gives an analysis result that “children” relate to “install” in the sentence analysis after speech recognition. Therefore, the system (computer 1) creates a dependency confirmation message, "Is the child installed?" The user (speaker) inputs the confirmation sentence “Sound has been deleted” by voice. As a result, the system (computer 1) corrects the “children” relationship to “delete”.

[3. Flow sheet]
8 and 9 show a flow sheet of the automatic answering apparatus according to the first embodiment of the present invention. The speaker inputs voice from a microphone or the like (S100). The voice storage unit 20 such as the main memory 3 or the HDD 4 stores voice (S200). The CPU 2 extracts one sentence from the stored voice (S300). The voice recognition unit 30 recognizes voice (S400). The misrecognition determination unit 41 determines whether there is an error in speech recognition (S500). If there is no error in speech recognition, the input analysis unit 50 performs input analysis (S600).

  When there is an error in the speech recognition, the corrected sentence creation unit 42 creates a rephrase request sentence (S501). The voice output unit 90 outputs the rephrasing request sentence as a voice (S502). In response to the restatement request sentence, the speaker inputs the restatement sentence by voice (S503). The speech recognition unit 30 performs speech recognition of the restated sentence (S504). The corrected sentence comparison / determination unit 43 performs comparison with the original input sentence (S505), and determines whether the words are different (S506). If the words are different, the corrected sentence creation unit replaces the words and creates a corrected sentence (S507). If the words are not different, the corrected sentence creation unit 42 creates a confirmation sentence (S508). The audio output unit 90 outputs audio (S509). The speaker inputs “Yes” or “No” to the confirmation sentence (S510). The voice recognition unit 30 performs voice recognition of Yes or No (S511). The misrecognition Y / N determination unit 44 determines Yes or No (S512). In the case of No, the corrected sentence creation unit 42 creates a rephrase request sentence again (S501). In the case of Yes, the input analysis unit 50 performs input analysis (S600).

  The error analysis determination unit 61 determines whether there is an error in the dependency (S700). When there is an error in the dependency, the dependency sentence creating unit 62 creates a dependency confirmation sentence (S701). The voice output unit 90 outputs the dependency confirmation sentence by voice (S702). The speaker inputs Yes or No to the dependency confirmation sentence (S703). The voice recognition unit 30 performs voice recognition of Yes or No (S704). The misanalysis Y / N determination unit 63 determines Yes or No (S705). In the case of No, the dependency sentence creating unit 62 newly creates a dependency confirmation sentence (S701). In the case of Yes, the dependency sentence creation unit 62 creates a dependency correction sentence (S706). Then, the input analysis unit 50 performs input analysis again (S600).

If there is no error in the dependency, the input analysis unit 60 determines whether there is a next sentence (S800). If there is a next sentence, the voice recognition unit 30 extracts one sentence (S300). If there is no next sentence, the problem solving unit 70 extracts a response pattern (S900). The sound output unit 90 outputs sound (S1000).
It should be noted that the portion corresponding to the voice described above is not limited to the voice, but includes a sentence written with a keyboard or the like.

  Although the present invention has been described with the above embodiments, the technical scope of the present invention is not limited to the scope described in the embodiments, and various modifications or improvements can be added to these embodiments. . And embodiment which added such a change or improvement is also contained in the technical scope of the present invention. This is apparent from the claims and the means for solving the problems.

It is a hardware block diagram of the automatic answering apparatus which concerns on the 1st Embodiment of this invention. It is a block block diagram of the automatic answering apparatus which concerns on the 1st Embodiment of this invention. It is a result (1) of the morphological analysis of the automatic answering apparatus which concerns on the 1st Embodiment of this invention. It is a result (2) of the morphological analysis of the automatic answering apparatus which concerns on the 1st Embodiment of this invention. It is an example of a dialogue for correcting a voice recognition error in the automatic answering apparatus according to the first embodiment of the present invention. It is a result of the dependency analysis of the automatic answering apparatus which concerns on the 1st Embodiment of this invention. It is an example of a correction dialogue of a sentence analysis error of the automatic answering apparatus according to the first embodiment of the present invention. It is a flow sheet (1) of the automatic answering device concerning a 1st embodiment of the present invention. It is a flow sheet (2) of the automatic reply apparatus which concerns on the 1st Embodiment of this invention. It is an example of an analysis error in CaboCha.

Explanation of symbols

1 Computer 2 CPU
3 Main memory 4 HDD
5 Video Card 6 Mouse 7 Keyboard 8 Optical Disk 10 Audio Input Unit 20 Audio Storage Unit 30 Speech Recognition Unit 40 Error Recognition Correction Unit 41 Error Recognition Determination Unit 42 Correction Sentence Generation Unit 43 Correction Sentence Comparison Determination Unit 44 Error Recognition Y / N Determination Unit 50 input analysis unit 60 error analysis correction unit 61 error analysis determination unit 62 dependency sentence creation unit 63 error analysis Y / N determination unit 70 problem solving unit 80 utterance generation unit 90 speech output unit 411 morphological analysis unit 412 error recognition word extraction Unit 413 Co-occurrence degree sum calculation unit 414 Threshold value calculation unit 611 Dependency analysis unit 612 Error analysis word extraction unit 613 Dependency calculation unit 614 Dependency relationship determination unit

Claims (10)

Morphological analysis means for morphological analysis of the input sentence;
Word extraction means for extracting words satisfying a certain condition in parts of speech in the sentence subjected to morphological analysis;
A co-occurrence degree sum calculating means for obtaining a sum of co-occurrence degrees with other words for each of the extracted words;
Threshold value determination means for determining a word whose sum of co-occurrence degrees is lower than a predetermined threshold value as an input error word;
An automatic answering device comprising corrected sentence creating means for creating a corrected sentence when there is a word determined to be an input error. Morphological analysis means for morphological analysis of the input sentence;
Word extraction means for extracting words satisfying a certain condition in parts of speech in the sentence subjected to morphological analysis;
A co-occurrence degree sum calculating means for obtaining a sum of co-occurrence degrees with other words for each of the extracted words;
Threshold value determination means for determining a word whose sum of co-occurrence degrees is lower than a predetermined threshold value as an input error word;
A corrected sentence creating means for creating a corrected sentence when there is a word determined to be an input error;
A corrected sentence comparison judging means for judging whether or not the corrected sentence has a different word compared to the sentence before correction;
An automatic answering device comprising Yes / No determining means for determining whether or not there is an error in a corrected sentence.   The corrected sentence uses a template created by paying attention to the pre-correction sentence and the word / phrase at the end of the answer sentence up to the word before the word determined to be input error among the words whose co-occurrence degree is calculated. The automatic answering apparatus according to claim 2, further comprising means for creating a corrected sentence. Dependency analysis means for dependency analysis of input sentences;
Word extracting means for extracting one word whose part of speech in the dependency analyzed sentence satisfies a certain condition;
A dependency degree calculating means for obtaining a dependency degree for the set of one extracted word and another word;
Dependency relationship determination means for determining the magnitude relationship of the dependency level between the groups;
An automatic answering device comprising dependency correction sentence creating means for creating a dependency correction sentence for a set of words having a low dependency degree. Dependency analysis means for dependency analysis of input sentences;
Word extraction means for extracting one word whose part of speech in the sentence satisfies a certain condition;
A dependency calculation means for making a set of the one word and another word to obtain a dependency;
Dependency relationship determination means for determining the magnitude relationship of the dependency level between the groups;
A dependency correction sentence creating means for creating a dependency correction sentence for a set of words having low dependency;
An automatic answering device comprising Yes / No determination means for determining whether or not the dependency correction text is an error.   The dependency correction sentence is corrected using a template created by paying attention to a word having a dependency degree larger than a word determined to be a sentence analysis error and a word / phrase at the end of the answer sentence among words whose dependency degree is calculated. 6. The automatic answering apparatus according to claim 5, further comprising means for creating a sentence. A morphological analysis procedure for morphological analysis of the input sentence;
A word extraction procedure for extracting words satisfying a certain condition in parts of speech in the sentence subjected to morphological analysis;
A co-occurrence degree sum calculation procedure for obtaining a sum of co-occurrence degrees with other words for each extracted word;
A threshold value determination procedure for determining a word whose sum of co-occurrence degrees is lower than a predetermined threshold value as an input error word;
An automatic answer program that causes a computer to function as a corrected sentence creation procedure for creating a corrected sentence when there is a word determined to be an input error. A morphological analysis step for morphological analysis of the input sentence;
A word extraction step for extracting words satisfying a certain condition for part of speech in the sentence subjected to morphological analysis;
A co-occurrence degree sum calculating step for obtaining a sum of co-occurrence degrees with other words for each extracted word;
A threshold value determining step for determining a word whose sum of co-occurrence degrees is lower than a predetermined threshold value as an input error word;
An automatic answering method comprising a corrected sentence creating step for creating a corrected sentence when there is a word determined to be an input error. A dependency analysis procedure for dependency analysis of input sentences;
A word extraction procedure for extracting one word in which the part of speech in the dependency analyzed sentence satisfies a certain condition;
A dependency degree calculation procedure for obtaining a dependency degree for the set of one extracted word and another word;
A dependency relationship determination procedure for determining the magnitude relationship of the dependency between the groups;
An automatic answer program for causing a computer to function as a dependency correction sentence creation procedure for creating a dependency correction sentence for a group of words having a low dependency degree. A dependency analysis step for dependency analysis of the input sentence;
A word extracting step of extracting one word in which the part of speech in the dependency analyzed condition satisfies a certain condition;
A dependency degree calculating step for obtaining a dependency degree for the set of the extracted one word and another word;
A dependency relationship determination step for determining the magnitude relationship of the dependency between the groups;
An automatic answering method comprising a dependency correction sentence creation step of creating a dependency correction sentence for a set of words having a low dependency degree.

Images