JP2007265131A - Dialog information extraction device, dialog information extraction method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a dialog information extraction device capable of extracting only important information which does not overlap listening text data written by a primary operator and reducing burdens on a secondary operator. <P>SOLUTION: The dialog information extraction device 100 is provided with: a dialog voice input part 114 for inputting dialog sound data which record dialogs between a plurality of persons; a sound text conversion part 116 for applying sound recognition processing to the dialog sound data to convert them into dialog text data; a dialog sentence extraction part 118 for performing a morpheme analysis to the dialog text data to extract a group of dialog sentences to be an important group of sentences among the dialog text data; a listening text input part 122 for inputting the listening text data whose contents of the dialog have been transcribed by a listener; and a differential sentence output part 124 for outputting a group of differential sentences which do not correspond to sentences of the listening text data. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a dialog information extraction apparatus, a dialog information extraction method, and a program for extracting dialog information.

  In recent years, in addition to written documentation such as instruction manuals, inquiries regarding purchased products are directly received from users by telephone, and services corresponding to each individual are increasing. In such an inquiry, there is a wide variety of information desired by the user, and the operator carefully responds to each inquiry. There is usually one window for accepting such inquiries, and the corresponding operator is referred to herein as the primary operator.

  The primary operator responds in a dialog with the user about items that can be answered. However, if the primary operator cannot answer immediately, the inquiry is transmitted to the secondary operator and answered to the user at a later date by the secondary operator.

  When preparing such an answer, the secondary operator refers to the call memory (hereinafter referred to as listening text data) in which the primary operator describes the contents of the dialogue in text, and grasps the user's inquiry item. However, with such listening text data alone, important information from the user may be lost, and the point of inquiry is shifted and the secondary operator asks the user the same content as the primary operator, and the processing is efficient. There is a problem that is not done automatically. For this reason, the secondary operator takes measures such as re-listening all the dialogue voice data storing the dialogue between the user and the primary operator as voice, or contacting the user directly to confirm the contents of the question.

  However, it is inefficient for the secondary operator to listen to the recorded dialogue voice data every time. However, even if the speech recognition process is performed on the dialog voice data and the result converted to the dialog text data is transmitted to the secondary operator, the text is too much data like the voice, so the processing becomes complicated. End up.

  Furthermore, since the primary operator does not know the information required by the secondary operator, there is a possibility that important information may be lost from the listening text data. In addition, the primary operator can only leave a simple memo while responding to a telephone call, and since it creates formal listening text data based on the memo later, there is a problem that information tends to be lost.

Therefore, a technique for converting dialogue voice data into text and extracting a character string including a specific keyword (important keyword) prepared in advance so that important information is not lost from the voice information has been studied (for example, , Patent Document 1).
JP 2000-067064 A

  However, in the technology for extracting important information using the above keywords, it is necessary to predict matters to be asked by the user in advance and assign important keywords. Also, important matters that could not be predicted as keywords must be grasped by the secondary operator again listening to the dialog voice data.

  On the other hand, even if the speech recognition process is simply performed on the dialog voice data and the result converted to the dialog text data is transmitted to the secondary operator, the text is too much data like the voice and the processing becomes complicated. Has also occurred. In such a state, the listening text data described so that the primary operator can easily understand the secondary operator is wasted.

  The present invention has been made in view of the above-mentioned problems of conventional dialog information extraction technology, and the object of the present invention is to extract only important information that does not overlap with listening text data created by a primary operator, To provide a new and improved dialogue information extraction apparatus, dialogue information extraction method and program capable of reducing the burden on the secondary operator.

  In order to solve the above-described problem, the present invention can extract only sentences that are missing from the listening text data written by the primary operator from the dialogue text data, and the secondary operator can grasp the inquiry desired by the user without omission. A dialog information extraction device, a dialog information extraction method, and a program that can reduce the work burden of a secondary operator and contribute to user convenience.

  In order to solve the above problems, according to an aspect of the present invention, a dialog voice input unit that inputs dialog voice data; a voice text conversion unit that performs voice recognition processing on the dialog voice data and converts it into dialog text data A dialog text extracting unit that extracts a dialog text group that is an important text group in the dialog text data; a listening text input unit that inputs listening text data in which the listener has written the content of the dialog; And a differential text output unit that outputs a differential text group that does not correspond to the text of the listening text data from the dialog text group.

  The dialogue sentence extraction unit extracts important sentences (dialogue sentence groups) from the dialogue text data, and the difference sentence output part extracts sentences other than sentences (character strings) already described in the listening text data from the dialogue sentence groups. Output. Therefore, the difference sentence group which is a remaining sentence becomes a sentence group which exists in the dialogue text data but does not exist in the listening text data. With such a configuration, since only sentences that have been dropped from the listening text data created by the primary operator can be extracted from the dialog text data, the secondary operator can grasp the query desired by the user without omission.

  In the present invention, an attribute dictionary storage unit that stores an attribute dictionary in which a word is associated with an attribute of the word is further provided, and the dialog sentence extraction unit selects a word that is determined to be important based on the attribute of the attribute dictionary. You may make it extract the 1 or 2 or more sentence containing. Also, the difference text output unit may determine an important word by referring to the attribute dictionary, which makes it possible to emphasize the important word when a difference document is output, which increases convenience.

  The differential sentence output unit removes the overlapping part of the listening text data and the dialog text data from the dialog text data and outputs the result. In this case, if all of the dialogue text data is left, the amount of text is too large, and the processing of the secondary operator becomes complicated. Therefore, the secondary operator is required to output only sentences containing important words. It is possible to narrow down the important sentences and to take an accurate and quick response, thereby further reducing the burden on the secondary operator.

  The dictionary storage unit may include a plurality of attribute dictionaries, and the dialog text extraction unit may use an attribute dictionary selected according to the use of the output result of the difference text output unit. The "attribute dictionary selected according to the use of the output result" here is selected to achieve the purpose in consideration of the expertise of the user (secondary operator, etc.) of the output result. Refers to the attribute dictionary.

  By setting attribute dictionaries individually for the other users above, it is possible to define important words for each user, and the dialogue text extractor is appropriate for each user's expertise. Sentences can be extracted.

  The dialog text extraction unit may extract at least any text indicating a question, an answer, or a wish. In addition, it is possible to extract a sentence indicating the question, the answer, and the hope together with the sentence including the important word as described above.

  When each sentence of the dialog text data is a question or answer of a user or a primary operator or is a user's request, the sentence is important, and the secondary operator must grasp the information accurately. In the morphological analysis of the dialog text extraction unit, when the text corresponds to such a question, answer, or hope, the text is also extracted and transmitted to the secondary operator.

  The difference sentence output unit may output a sentence other than a sentence in which all important words in the sentence are included in the listening text data from the dialogue sentence group.

  The differential sentence output unit determines a sentence already described in the listening text data. However, not all important words in sentences in dialogue text data are described in the listening text data. Therefore, the difference sentence output unit extracts sentences in which all important words in the dialog text data are included in the listening text data as duplicate sentences, and extracts sentences other than the duplicate documents as difference sentence groups.

  The difference sentence output unit may extract the dependency relationship from the listening text data and output a sentence other than the sentence having the dependency relationship from the dialogue sentence group.

  In listening text data, if an arbitrary word pair (same word in the notation, the same as a synonym) is considered to be a dependency relationship, a sentence containing a word with an equivalent dependency relationship is already included in the listening text data. Can be considered included. Therefore, sentences having such a relationship do not need to be left and are not included in the difference sentence group.

  A synonym dictionary storage unit for storing a synonym dictionary for adapting words in the text of the dialog text data to the listening text data may be further provided.

  The primary operator may generate listening text data by replacing a word being spoken with an easy-to-understand word. At this time, although the listening text data and the dialogue text data include a pair of words having the same dependency relation, the words do not match. By referring to the synonym dictionary, the differential sentence output unit can replace the word with a synonym and match the dependency relationship.

  The differential sentence output unit may exclude a sentence corresponding to a sentence corresponding to the sentence of the listening text data and a sentence having a question-answer relationship from the dialogue sentence group.

  In a sentence corresponding to the above-mentioned sentence of the listening text data in the dialogue sentence group, when only a question or only an answer is written, the corresponding answer or question is also included in the listening text data inherently. Can be considered. Therefore, it is not necessary to leave sentences having such a relationship, and they are removed by the difference sentence output unit.

  The differential text output unit may output differential voice data obtained by converting the differential text groups into speech and combining them, or may output differential text data combining the differential text groups.

  With this configuration, the secondary operator can grasp the user's inquiry using voice data or text data. In particular, with voice data, the voice quality of the user can be grasped at the same time, and the user can respond appropriately.

  It may be further provided with a dialog voice generation unit that records the contents of the dialog and generates dialog voice data.

  With this configuration, the dialog information extracting apparatus can directly record the contents of the dialog, and the primary operator does not have to handle the dialog voice data, so the processing can be simplified.

  The above dialog information extraction device is represented by an assembly of a plurality of components, but is not limited to the case where each component or functional module belongs to a single device, and each component is a separate device or a separate device. It may be configured to be incorporated in the apparatus.

  In addition, there are provided a program for causing a computer to function as the dialog information extraction device and a storage medium storing the program.

  In order to solve the above problems, according to another aspect of the present invention, a dialog voice input step for inputting dialog voice data; a voice text conversion for converting the dialog voice data into dialog text data by performing a voice recognition process; A dialog text extraction step for extracting a dialog text group that is an important text group in the dialog text data; and a listening text input step for inputting listening text data in which a listener has made the text of the dialog text; And a differential text output step for outputting a differential text group that does not correspond to the text of the listening text data from the dialog text group.

  The components corresponding to the dependent claims in the dialog information extraction apparatus and the description thereof can be applied to the dialog information extraction method, program, and storage medium.

  As described above, according to the present invention, only important information that does not overlap with the listening text data created by the primary operator is extracted, and the secondary operator can accurately and accurately refer to the listening text data and this important information. In addition to being able to respond quickly, the efficiency of the processing in this way reduces the overall burden on the primary and secondary operators and also contributes to user convenience.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the present specification and drawings, components having substantially the same functional configuration are denoted by the same reference numerals and description thereof is omitted.

(First embodiment: Dialog information extracting apparatus 100)
The dialogue information extraction apparatus 100 according to the first embodiment extracts matters that are not described in the listening text data from the correspondence between the listening text data and the dialogue text data, and extracts only that portion. The extracted information is transmitted to the secondary operator.

  FIG. 1 is a functional block diagram showing a schematic configuration of a dialogue information extraction apparatus 100 according to the first embodiment. The dialogue information extraction apparatus 100 includes a central control unit 110, a dialogue voice generation unit 112, a dialogue voice input unit 114, a voice text conversion unit 116, a dialogue sentence extraction unit 118, a dictionary storage unit 120, and a listening text. The input unit 122, the difference text output unit 124, and the attribute dictionary selection unit 180 are configured. The interactive information extraction apparatus 100 itself may be configured by an electronic device such as a server, a computer, a PDA (Personal Digital Assistant), a mobile phone, and a game device.

  The central control unit 110 manages and controls the entire dialogue information extraction apparatus 100 by a semiconductor integrated circuit including a central processing unit (CPU).

  The dialog voice generation unit 112 records the contents (voice) of the dialog made by the user and the primary operator who has received an inquiry from the user, and generates dialog voice data 150. The dialogue information extraction apparatus 100 according to the present embodiment can directly acquire the content of the dialogue.

  The dialog voice input unit 114 inputs the dialog voice data 150 recorded by the dialog voice generation unit 112 or by a separate recording device 140. When the dialog voice data 150 is generated by the dialog voice generation unit 112, the dialog voice data 150 is automatically transmitted to the voice text conversion unit 116 by the dialog voice input unit 114.

  The voice text conversion unit 116 performs voice recognition processing on the dialog voice data 150 input to the dialog voice input unit 114 to convert it into text, and generates dialog text data 152.

  The dialog text extraction unit 118 performs morphological analysis on the dialog text data 152 and extracts a dialog text group which is an important text group in the dialog text data 152.

The dictionary storage unit 120 stores one or more attribute dictionaries 160 that associate a word with an attribute of the word. Here, the dictionary storage unit 120 may be configured by a storage medium such as a RAM, an E ² PROM, a nonvolatile RAM, a flash memory, a card memory, a USB memory, and an HDD (Hard Disk Drive).

  The dialogue sentence extraction unit 118 extracts one or more sentences (character strings) including words that are determined to be important based on the attribute of the attribute dictionary 160 in the dictionary storage unit 120 and forms a dialogue sentence group.

  The differential sentence output unit 124 described later removes the overlapping part of the listening text data 170 and the dialog text data 152 from the dialog text data 152 and outputs the result. However, if all of the dialogue text data 152 is left, the amount of text is too large, and a load is imposed on the secondary operator. Thus, by setting only sentences including important words as output targets, important sentences required by the secondary operator are narrowed down, and the burden on the secondary operator is further reduced. As a result, since the secondary operator can take an accurate and prompt response to the user, the overall efficiency is improved.

  The dictionary storage unit 120 may also include a synonym dictionary 190 described in the second embodiment. Here, the dictionary storage unit 120 stores the attribute dictionary 160 and the synonym dictionary 120 together. In the present embodiment, the dictionary storage unit 120 has the functions of both the attribute dictionary storage unit and the synonym dictionary storage unit. Have. In the present invention, each storage unit can be stored in a separate storage unit. In this case, the storage unit storing the attribute dictionary 160 is the attribute dictionary storage unit, and the storage unit storing the synonym dictionary 190 is the synonym dictionary. Each can be a storage unit.

  Here, the dialogue sentence extraction unit 118 may extract only a sentence indicating at least one of a question, an answer, and a wish. In addition, it is possible to extract a sentence indicating the question, answer, and hope together with a sentence including an important word using the attribute dictionary 160.

  When each sentence of the dialog text data 152 is a question and answer of a user or a primary operator or a user's wish, the sentence is important, and the secondary operator must grasp the information. In the morphological analysis of the dialogue sentence extraction unit 118, when the sentence corresponds to such a question, answer, or hope, the sentence is also extracted and transmitted to the secondary operator.

  The listening text input unit 122 is a listening text in which the content of the dialogue with the user is written by the primary operator who is the listener, or written by a third party listening to the dialogue between the user and the primary operator. The text data 170 is input.

  The differential text output unit 124 performs morphological analysis on the listening text data 170 and outputs a differential text group that does not correspond to the text of the listening text data 170 from the dialog text group extracted by the dialog text extraction unit 118. To do. That is, the remaining sentences are extracted when sentences that can be regarded as overlapping of the listening text data 170 and the conversation text group are excluded from the dialog text data 152.

  The difference sentence output unit 124 removes one or more sentences in which all important words in the sentence are included in the listening text data 170 from the conversation sentence group. The differential sentence output unit 124 extracts sentences already described in the listening text data 170 from the dialogue text data 152, and all important words in the sentences in the dialogue text data 152 are described in the listening text data 170. Not necessarily. Therefore, the difference sentence output unit 124 sets a sentence in which all important words in the dialogue text data 152 are included in the listening text data 170 as a duplicate sentence, and a sentence that does not contain all important words is a difference sentence group. Extract as

  The remaining difference sentence group is present in the dialog text data 152 but not in the listening text data 170 and is an important sentence. With this configuration, it is possible to extract only the missing text in the listening text data 170 created by the primary operator from the dialogue text data 152, and the secondary operator can grasp the inquiry desired by the user without omission. .

  The remaining difference sentence group may be output from the difference sentence output unit 124 as difference voice data 162 converted into speech and combined, or may be output as simply combined difference text data 164. In this way, the secondary operator can grasp the user's inquiry from the voice data or text data. In particular, with voice data, the voice quality of the user can be grasped at the same time, and the user can respond appropriately.

  Further, the difference sentence output unit 124 may exclude a sentence corresponding to the sentence of the listening text data 170 and a sentence having a question-answer relationship from the dialogue sentence group. In the listening text data 170, when only a question or only an answer is described, it can be considered that the corresponding answer or question is also included inherently. Accordingly, it is not necessary to leave sentences having such a relationship, and the difference sentence output unit 124 removes them.

  With the above-described configuration, the dialog information extracting apparatus 100 extracts only important information that does not overlap with the listening text data 170 described by the primary operator, and the secondary operator refers to the listening text data 170 and this important information. This makes it possible to increase the processing efficiency and reduce the overall burden.

  Also, the differential voice data 162 and differential text data 164 output in this way can be used not only by the secondary operator but also for pointing out omissions to the primary operator and for evaluating the description of the primary operator. .

(Program, storage medium)
Also provided are a program that causes the computer to operate as the dialogue information extraction apparatus 100, and a storage medium that stores the program.

(Dialogue information extraction method)
Next, a dialog information extraction method for extracting dialog information using the above-described dialog information extraction apparatus 100 will be described in detail.

  FIG. 2 is a flowchart showing a process flow of the dialog information extraction method according to the first embodiment. Hereinafter, the process flow will be described together with the configuration of FIG.

  First, the dialog voice generation unit 112 of the dialog information extraction apparatus 100 records the dialog contents (voice) between the user and the primary operator, and generates dialog voice data 150 (S200). The generated dialog voice data 150 is subjected to voice recognition processing by the voice text conversion unit 116 and converted into dialog text data 152 (S202).

  FIG. 3 is an explanatory diagram showing an example of the dialog text data 152. In the dialogue text data 152, a continuous identifier is assigned to each sentence, and the speaker, here, the user and the primary operator are also distinguished, and the content of the dialogue between the user and the primary operator is converted into text. In addition, the utterance time during which the dialogue was made is also described.

  The dialog text extraction unit 118 performs morphological analysis on the dialog text data 152 converted into text, and extracts dialog text groups that are important text groups in the dialog text data 152 (S204). The extraction of the dialogue text group is largely performed by two processes. One is a process of extracting a sentence including a word that is determined to be important based on the attribute by the attribute dictionary 160, and the other is a process of extracting a sentence indicating a question, an answer, and a wish. Hereinafter, the two processes will be described with reference to FIGS.

  FIG. 4 is an explanatory diagram showing an example of the attribute dictionary 160. In the attribute dictionary 160, words are listed on the upper side of the drawing, and the attributes of the words are associated with the lower side of the drawing. Here, the attribute “phenomenon / situation” indicates a word indicating a situation where a trouble or an inquiry occurs, the attribute “product name” indicates the product name of each product, and the attribute “part / accessory name” The part name and accessory name of the product are shown. The attribute “date / time / frequency” indicates the date / time and frequency of occurrence of the event. Further, the attribute “knowledge state” indicates a word indicating whether the user understands the situation of the knowledge, for example, the content. In the attribute “knowledge state” in FIG. 4, “I understand + no” represents a combination of a negative verb such as “I don't know” or “I don't know” and an auxiliary verb.

  Here, attributes are associated with words, for example, the attribute of the word “MP200” is “product name”. Based on the attribute of the attribute dictionary 160, the attribute of the word of the dialog text data is analyzed.

  FIG. 5 is an explanatory diagram in which the attribute of each word in the dialog text data 152 shown in FIG. 3 is analyzed based on the attribute dictionary 160 shown in FIG. Here, the dialog text extraction unit 118 analyzes the dialog text data 152 to decompose the text into word levels, searches each word from the attribute dictionary 160, and associates the attribute with the hit word. The underline of such a word indicates a morpheme break. For example, “product name” is associated with “MP200” of identifier 3 user utterance, “part / accessory name” is associated with “paper”, and “phenomenon / situation” is associated with “jammed”. In this way, attributes are assigned to the words in the text of the dialog text data 152.

  The dialogue sentence extraction unit 118 grasps important sentences according to the associated attributes. In this embodiment, it can be determined that the response is a technical response among the attributes “phenomenon / situation”, “product name”, “part / accessory name”, “date / frequency”, “acquisition method”, and “knowledge state”. The attributes “phenomenon / situation”, “product name”, “part / accessory name”, “date / frequency”, “knowledge state” are important words, and sentences including the words are important sentences. Therefore, identifiers 3, 4, 7, 8, and 9 correspond to important sentences. Here, the attributes “phenomenon / situation”, “product name”, “part / accessory name”, “date / frequency”, and “knowledge state” are important words. Various combinations can be applied depending on the purpose.

  Here, the word whose attribute corresponds to “knowledge state” is also regarded as an important word. The content that the user says “I don't know” is important information for the secondary operator, although it is highly likely that it will fall out of the listening text data. It is because it may become.

  Subsequently, the dialogue sentence extraction unit 118 analyzes whether or not the sentence in the dialogue text data 152 corresponds to an important utterance type. Here, the utterance type refers to, for example, questions, answers, hopes, and the like, and the important utterance types indicate questions, answers, and hopes therein. First, the dialogue sentence extraction unit 118 pays attention to the sentence end expression and question words. For example, the identifiers 8 and 12 in FIG. 5 have the final particle ending with “ka”, and the identifier 6 includes the “when” question word. Accordingly, the identifiers 6, 8, and 12 are the utterance type “question”. Then, the sentence immediately after the “question” sentence is determined to be the utterance type “answer”. In FIG. 5, the identifiers 7, 9, and 13 are “answers”.

  Since the identifiers 4 and 10 include the auxiliary verb “tai” indicating hope in the sentence, it is determined that the speech type is “hope”. In this way, utterance types (question, answer, hope, etc.) as shown on the right side of FIG. 5 are associated.

  In addition to the sentence including the important word described above, the dialog sentence extracting unit 118 extracts the sentence including the important utterance type, in this embodiment, “question”, “answer”, and “hope” as the important sentence. Accordingly, the identifiers 4, 6, 7, 8, 9, 10, 12, and 13 correspond to important sentences here. However, words such as “name” and “address” as shown in identifiers 12 and 13 are stop words, and questions and answers including those words are not important sentences. This is because the user's name and address are not important information for the secondary operator, but are considered to be specified when the primary operator handles the telephone. Accordingly, the identifiers 12 and 13 are excluded from important sentences.

  When the two processes executed in this way are combined, it can be seen that identifiers 3, 4, 6, 7, 8, 9, and 10 are extracted.

  As the next stage, in parallel with the dialog voice recording step (S200) in FIG. 2, or after the fact, the primary operator generates the listening text data 170 by converting the contents of the dialog with the user into sentences. Then, after the dialogue is finished, the listening text data 170 is inputted to the dialogue information extracting device 100 (S206).

  FIG. 6 is an explanatory diagram showing an example of the listening text data 170. Here, the primary operator summarizes the requirements from users as “customer requirements”. At this time, in addition to the text information indicating the user's requirements, information such as the name of the corresponding primary operator and the corresponding date and time may be described in the listening text data 170.

  Finally, the difference sentence output unit 124 performs morphological analysis on the input listening text data 170 and outputs sentences other than sentences in which important words in each sentence of the dialogue sentence group are all in the sentences of the listening text data 170. (S208). Thus, important sentences that are not described in the listening text data 170 and that are important are extracted from the dialogue text data 152. Hereinafter, the operation of extracting such a differential sentence group will be described in detail.

  FIG. 7 is an explanatory diagram in which the attribute of each word of the listening text data 170 shown in FIG. 6 is analyzed based on the attribute dictionary 160 shown in FIG. Here, the difference sentence output unit 124 determines whether all the important words in each sentence of the conversation sentence group are present in the sentence of the listening text data 170. Hereinafter, the sentences of identifiers 3, 4, 6, 7, 8, 9, and 10 extracted as a dialogue sentence group in FIG.

  First, the important words “MP200”, “paper”, and “jamming” in the sentence of identifier 3 are all present in the listening text data 170, and their attributes also correspond. Therefore, it can be determined that the sentence with the identifier 3 is already described in the listening text data 170. Similarly, since the sentence with identifier 7 has the word “last month” in the listening text data 170, it can be determined that such sentence has already been described.

  On the other hand, the sentences with identifiers 4, 6, 8, 9, and 10 are not included in the listening text data 170 because the important words included in each of the sentences are not included in the listening text data 170. , Difference sentence group. Therefore, the difference sentence group is a sentence with identifiers 4, 6, 8, 9, and 10.

  Here, the difference sentence output unit 124 selects a sentence corresponding to the sentence of the listening text data 170 from the dialogue sentence group (identifiers 3, 4, 6, 7, 8, 9, 10) and a question-to-answer answer. Related sentences may be excluded.

  In this embodiment, among the sentences corresponding to the sentence of the listening text data (sentences of identifiers 3 and 7) in the dialogue sentence group, the sentence of identifier 7 is an answer to the question of identifier 6 I understand. In this way, when only a question or only an answer is written, it can be considered that the corresponding answer or question is also included in the listening text data 170, so the sentence of the identifier 6 corresponding to the question May be removed by the difference text output unit 124.

  Here, when the utterance type of the sentence described in the listening text data 170 corresponds to “answer”, it is considered that “question” is also described, but conversely, an undescribed sentence (difference sentence) is “answer” "Question" can be a difference sentence. Similarly, when an undescribed sentence (difference sentence) is a “question” and an “answer” for the “question” is not extracted as a dialogue sentence group, the “answer” is also regarded as a difference sentence. It may be forcibly extracted.

  As a result of the dialogue information extraction method according to the present embodiment, the difference sentence group is a sentence having identifiers 4, 8, 9, and 10. The difference sentence output unit 124 may output the difference voice data 162 obtained by converting the difference sentence group into speech and combining it, or may output the difference text data 164 combined with the difference sentence group. , Both data may be output at once.

  The dialog text group extraction and the differential text group extraction in the dialog text extraction unit 118 and the difference text output unit 124 described above may be repeatedly performed one by one in the order of the identifiers.

  FIG. 8 is a flowchart illustrating in detail the dialogue information extraction method for processing each sentence. Here, first, the identifier number is set to the initial value 1 (S300), and the sentence with the identifier number 1 is extracted from the dialogue text data 152 (S302).

  Then, the dialogue sentence extraction unit 118 determines whether or not the extracted sentence includes an important word or an important utterance type based on the attribute dictionary 160 (S304). Here, if an important word or an important utterance type is included, whether or not the sentence further includes a stop word (a word such as “name” or “address”) is determined (S310). It is extracted as an important sentence (dialogue sentence group) (S312).

  Next, the difference sentence output unit 124 determines whether or not the extracted dialogue sentence group is included in the listening text data 170 (S320). If not included, the difference sentence output unit 124 corresponds to the sentence of the listening text data 170. It is determined whether there is a question-answer relationship (S322). If there is no question-answer relationship, it is extracted as a difference sentence group (S324). In addition, if the above judgment (S304, S310, S320, S322) is omitted, the sentence does not become a difference sentence group.

  When the processing of the sentence with the identifier number is completed, it is determined whether or not the sentence still remains in the dialog text data 152 (S330). If it remains, the identifier number is incremented (S332), and the sentence extraction step (S302). Repeat the process from. When all the sentences in the dialog text data 152 are processed, the extracted difference sentence group is output (S334), and the dialog information extraction method is terminated.

  Further, the dictionary storage unit 120 described above has a plurality of attribute dictionaries 160, and the dialog text extraction unit 118 selects according to another user (secondary operator) who uses the output result of the differential text output unit 124. The attribute dictionary 160 can be used.

  FIG. 9 is an explanatory diagram showing another example of the attribute dictionary 160 included in the dictionary storage unit 120. Here, as compared with FIG. 4, the attribute dictionary 160 is formed from only four attributes. The attribute dictionary 160 in FIG. 9 includes words suitable for the field of sales, for example, “XX company” which is a competitor.

  Here, the secondary operator or the primary operator selects the ID of the secondary operator that receives the differential text group output from the dialogue information extraction apparatus 100, and the attribute dictionary selection unit 180 shown in FIG. Judgment is made and associated with the attribute dictionary 160. For example, in the example of FIG. 9, “2” that is the ID of the secondary operator is selected, and the attribute dictionary 160 associated with the ID “2” is selected.

  Thus, by setting the attribute dictionary 160 individually for the secondary operator, it is possible to define important words for each user, and the dialogue sentence extraction unit 118 can improve the expertise for each user. Accordingly, it is possible to extract an appropriate sentence. In the above, the attribute dictionary 160 is distinguished from technology and sales, but is not limited thereto, and various distinction methods such as a technical field, a target product, a region, an age, an inquiry category, and the like can be considered.

  According to the first embodiment described above, it is possible to extract sentences that are missing from the listening text data written by the primary operator and important sentences from the listening text data 170 and the dialogue text data 152. The next operator can efficiently grasp the inquiry desired by the user without omission.

(Second Embodiment: Dialog Information Extraction Device)
In the first embodiment, the difference sentence output unit 124 compares the conversation sentence group and the listening text data 170 to determine sentences having the same word and attribute. However, the primary operator sometimes expresses the user's inquiry contents in different words so that the secondary operator can easily understand. In the second embodiment, it is possible to more accurately determine a duplicate sentence between the dialogue sentence group and the listening text data 170 by analyzing the sentence including the syntax analysis process.

  The dialogue information extraction apparatus according to the second embodiment has the same configuration as that of the dialogue information extraction apparatus 100 according to the first embodiment. Accordingly, as in the first embodiment, the second embodiment will be described using the dialogue information extraction apparatus 100 of FIG. The constituent elements already described as the constituent elements in the first embodiment have substantially the same functions, and therefore redundant description is omitted. Here, the difference sentence extraction unit 124 and the dictionary storage unit 120 having different functions are omitted. Is mainly explained.

  The differential sentence output unit 124 is dependent on the listening text data 170 instead of, or in addition to, one or more sentences in which the important words in the sentence are all in the listening text data 170 from the dialogue text group. It is possible to extract a relationship and output a sentence other than the sentence having the dependency relationship from the dialogue sentence group.

  Since the listening text data 170 is data input by the primary operator, it usually has a subject and a predicate and is a relatively well-organized text. Therefore, the listening text data 170 can be analyzed by a syntax analysis technique. The difference sentence output unit 124 performs a syntax analysis process on the listening text data 170 and the conversation sentence group, and outputs the difference sentence group using the result.

  FIG. 10 is an explanatory diagram in which the dependency of the listening text data 170 shown in FIG. 7 is extracted. In the explanatory diagram of FIG. 10, the relations “use”, “clog”, and “purchase” correspond to the dependency sources “MP200”, “paper”, and “last month” as dependency.

  The difference sentence output unit 124 receives the dependency relationship extracted in this way, and first determines whether there is a sentence having the dependency relationship from the dialogue sentence group. If there is no corresponding dependency relationship in the sentence, the dependency relationship is searched including a predetermined number of sentences before and after. The dependency-related search range may be one sentence before and after, and may include any number of sentences.

  For example, the dependency function shown in FIG. 10 is made to correspond to the dialog text group (sentences of identifiers 3, 4, 6, 7, 8, 9, 10 in FIG. 5) extracted by the dialog text extraction unit 118 in the first embodiment. As can be seen, the dependency of “MP200” and “use” is in the sentence of identifier 3, and the dependency of “paper” and “clog” is also in the sentence of identifier 3.

  The dictionary storage unit 120 may store a synonym dictionary 190 for adapting words in the text of the dialog text data 152 to the dependency relationship together with the attribute dictionary 160.

  The primary operator may generate the listening text data 170 by replacing the word being spoken with an easy-to-understand word. At this time, although the listening text data 170 and the dialogue text data 152 include a pair of words having the same dependency relationship, a situation occurs in which the words do not match. The differential sentence output unit 124 can refer to the synonym dictionary 190 to replace the word with a synonym and match the dependency relationship.

  FIG. 11 is an explanatory diagram for explaining the synonym dictionary 190. In the synonym dictionary 190, words that are treated as synonyms are listed together with synonym IDs. Here, in order to facilitate understanding, two word groups are described as synonyms, but naturally, synonyms are not limited to such word groups, and various word groups can be cited. Referring to the synonym dictionary 190, it can be understood that “purchase” and “purchase” are synonyms. Therefore, the dependency relationship between “last month” and “purchase” and the dependency relationship between “last month” and “purchase” are equal.

  Specifically, the above synonyms may be, for example, the same word expressed in kanji, katakana, kana, romaji, or the presence or absence of a long sound. That is, “computer” and “computer”, “watch” and “clock” are also synonymous.

  By using the synonym dictionary 190, the difference sentence output unit 124, from the conversation sentence group, the sentence of the identifier 6 having the word “purchase” and the same dependency as “last month” and “purchase”, and “last month” And the sentence of the identifier 7 having the word “can be extracted as a duplicate sentence. Accordingly, the overlapping sentences between the dialogue sentence group and the listening text data 170 are identification numbers 3, 6, and 7, and the difference sentence group is a sentence having identifiers 4, 8, 9, and 10, which are differences from the dialogue sentence group. . This result is the same as the result of the first embodiment.

  Thus, in the listening text data 170, when an arbitrary word pair is regarded as a dependency relationship, it is considered that a sentence or a sentence pair including a word having an equivalent dependency relationship is already included in the listening text data. Can do. Therefore, it is not necessary to leave the sentences with the identification numbers 3, 6, and 7 having such a relationship.

  The dialogue information extraction method to which the second embodiment is applied is processed according to the flowcharts of FIGS. 2 and 8 in the first embodiment, and includes a difference sentence group output step (S208) and a listening text data inclusion determination step. In (S320), the dependency relationship is applied.

  As described above, in the second embodiment, by using the parsing process, the listening text data and the dialogue text data can be linked more strictly, and a differential sentence group can be extracted with high accuracy. .

  By the dialogue information extraction apparatus or method in the first and second embodiments described above, the secondary operator can obtain necessary information without omission (accurate) and easy (rapid) without having to listen to all the contents of the dialogue of the primary operator. Therefore, the burden on the secondary operator can be reduced, and the efficiency and cost can be reduced. In addition, since it is not necessary for the primary operator to create listening text data again after the dialogue is completed, it contributes to reducing the burden on the primary operator.

  Furthermore, by responding efficiently to the user, opportunities to interact with the user increase and the amount of information that can be acquired increases, which is also effective in marketing and countermeasures against defects.

  As mentioned above, although preferred embodiment of this invention was described referring an accompanying drawing, it cannot be overemphasized that this invention is not limited to the example which concerns. It will be apparent to those skilled in the art that various changes and modifications can be made within the scope of the claims, and these are naturally within the technical scope of the present invention. Understood.

  For example, in the above-described embodiment, the use target person of the dialog information extraction apparatus is a user, a primary operator, and a secondary operator in a contact center. However, the present invention is not limited to this, and dialog voice data and listening text data are used. It can be applied to various situations that can be used.

  For example, when communicating with a doctor about the interaction between a patient and a nurse in a medical setting, a message from a parent company to a subsidiary in a company organization, a message from a school to a teacher or student in e-education, or a public office It can also be applied to matters to be communicated from reception to each person in charge.

  In the dialog text extraction unit and difference text output unit in the above embodiment, words are extracted by morphological analysis processing, but patterns using regular expressions can also be used.

  In addition, important words by the dialog sentence extraction unit and the difference sentence output unit are determined based on the attribute dictionary, but this is not the only case. For example, the volume at the time of utterance of the word and the emotion from the voice information It may be determined based on the recognition result, or may be used in combination with the attribute dictionary.

  In the embodiment described above, the dialogue sentence extraction unit once extracts an important sentence (dialog sentence group), and then the difference sentence output unit compares the dialogue sentence group and the listening text data. However, the present invention is not limited to this, and a configuration in which all dialogue text data and listening text data are first compared and important sentences are extracted from the difference sentences to form a difference sentence group is naturally within the technical scope of the present invention. It is understood that it belongs.

  Moreover, the dictionary storage part of the said dialog information extraction apparatus is provided in a separate server, and an attribute dictionary or a synonym dictionary may be referred to via a communication network. The communication network refers to, for example, a bidirectional communication network such as the Internet, an intranet, and a LAN (Local Area Network).

  In addition, the target conversation information is not limited to Japanese, and the present invention can deal with various languages such as alphabet and Korean.

  Note that the steps in the dialog information extraction method of the present specification do not necessarily have to be processed in time series in the order described in the flowchart, but are executed in parallel or individually (for example, parallel processing or object processing). May be included.

It is the functional block diagram which showed the schematic structure of the dialogue information extraction apparatus in 1st Embodiment. It is the flowchart which showed the flow of the process of the dialogue information extraction method in 1st Embodiment. It is explanatory drawing which showed an example of dialog text data. It is explanatory drawing which showed an example of the attribute dictionary. It is explanatory drawing which analyzed the attribute of each word of dialog text data based on the attribute dictionary. It is explanatory drawing which showed an example of listening text data. It is explanatory drawing which analyzed the attribute of each word of listening text data based on the attribute dictionary. It is the flowchart explaining in detail the dialog information extraction method which processes said sentence one by one. It is explanatory drawing which showed the other example of the attribute dictionary which a dictionary memory | storage part has. It is explanatory drawing which extracted the dependency of listening text data. It is explanatory drawing for demonstrating a synonym dictionary.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Dialogue information extraction apparatus 112 Dialogue voice generation part 114 Dialogue voice input part 116 Speech text conversion part 118 Dialogue sentence extraction part 120 Dictionary storage part 122 Listen text input part 124 Difference sentence output part 150 Dialogue voice data 152 Dialogue text data 160 Attribute dictionary 162 Difference voice data 164 Difference text data 170 Listening text data 180 Attribute dictionary selection unit 190 Synonym dictionary

Claims (13)

A dialog voice input unit for inputting dialog voice data;
A speech text conversion unit that performs speech recognition processing on the dialog speech data and converts the dialog speech data into dialog text data;
A dialog text extraction unit that extracts a dialog text group that is an important text group in the dialog text data;
A listening text input unit for inputting listening text data in which the listener has written the contents of the dialogue;
A differential sentence output unit that outputs a differential sentence group not corresponding to the sentence of the listening text data from the dialogue sentence group;
An apparatus for extracting dialogue information, comprising: An attribute dictionary storage unit for storing an attribute dictionary in which a word is associated with an attribute of the word;
2. The dialogue information extracting apparatus according to claim 1, wherein the dialogue sentence extracting unit extracts one or more sentences including words judged to be important based on attributes of the attribute dictionary. The dictionary storage unit has a plurality of attribute dictionaries,
The dialog information extracting apparatus according to claim 2, wherein the dialog text extracting unit uses an attribute dictionary selected in accordance with a use of an output result of the difference text output unit.   The dialogue information extracting apparatus according to claim 1, wherein the dialogue sentence extracting unit extracts at least any sentence indicating a question, an answer, and a wish.   The difference text output unit outputs a sentence other than a sentence in which all important words in the sentence are included in the listening text data from the dialogue sentence group. Interactive information extraction device.   6. The differential sentence output unit extracts a dependency relationship from the listening text data and outputs a sentence other than a sentence having the dependency relationship from the dialogue sentence group. The interactive information extraction device described in Crab.   The synonym dictionary memory | storage part which memorize | stores the synonym dictionary for adapting the word in the sentence of the said dialog text data to the said listening text data is further provided, It is characterized by the above-mentioned. Dialog information extraction device.   The difference text output unit excludes a sentence corresponding to a sentence corresponding to the sentence of the listening text data and a sentence in a question-answer relationship from the dialogue sentence group. The dialog information extraction device described in 1.   9. The dialogue information extracting apparatus according to claim 1, wherein the differential text output unit outputs differential voice data obtained by converting the differential text group into voice and combining the voices.   The dialogue information extraction device according to claim 1, wherein the differential text output unit outputs differential text data obtained by combining the differential text groups.   The dialogue information extracting apparatus according to claim 1, further comprising a dialogue voice generation unit that records the dialogue contents and generates dialogue voice data. Computer
A dialog voice input unit for inputting dialog voice data;
A speech text conversion unit that performs speech recognition processing on the dialog speech data and converts the dialog speech data into dialog text data;
A dialog text extraction unit that extracts a dialog text group that is an important text group in the dialog text data;
A listening text input unit for inputting listening text data in which the listener has written the contents of the dialogue;
A differential sentence output unit that outputs a differential sentence group not corresponding to the sentence of the listening text data from the dialogue sentence group;
A program characterized by functioning as A dialog voice input step for inputting dialog voice data;
A voice text conversion step of performing voice recognition processing on the dialog voice data and converting the dialog voice data into dialog text data;
A dialogue sentence extracting step of extracting a dialogue sentence group which is an important sentence group in the dialogue text data;
A listening text input step for inputting listening text data in which the listener has written the contents of the dialogue;
A differential sentence output step of outputting a differential sentence group not corresponding to the sentence of the listening text data from the dialogue sentence group;
Dialog information extraction method characterized by including

Images