CN104424290A - Voice based question-answering system and method for interactive voice system - Google Patents

Voice based question-answering system and method for interactive voice system Download PDF

Info

Publication number
CN104424290A
CN104424290A CN 201310390944 CN201310390944A CN104424290A CN 104424290 A CN104424290 A CN 104424290A CN 201310390944 CN201310390944 CN 201310390944 CN 201310390944 A CN201310390944 A CN 201310390944A CN 104424290 A CN104424290 A CN 104424290A
Authority
CN
Grant status
Application
Patent type
Prior art keywords
word
question
bag
step
unit
Prior art date
Application number
CN 201310390944
Other languages
Chinese (zh)
Inventor
左祥
金浩
Original Assignee
佳能株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/30Information retrieval; Database structures therefor ; File system structures therefor
    • G06F17/3061Information retrieval; Database structures therefor ; File system structures therefor of unstructured textual data
    • G06F17/30634Querying
    • G06F17/30637Query formulation
    • G06F17/30654Natural language query formulation or dialogue systems
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/30Information retrieval; Database structures therefor ; File system structures therefor
    • G06F17/3061Information retrieval; Database structures therefor ; File system structures therefor of unstructured textual data
    • G06F17/30634Querying
    • G06F17/30657Query processing
    • G06F17/30675Query execution
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/30Information retrieval; Database structures therefor ; File system structures therefor
    • G06F17/3061Information retrieval; Database structures therefor ; File system structures therefor of unstructured textual data
    • G06F17/30634Querying
    • G06F17/30657Query processing
    • G06F17/30675Query execution
    • G06F17/30684Query execution using natural language analysis
    • GPHYSICS
    • G10MUSICAL INSTRUMENTS; ACOUSTICS
    • G10LSPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00Speech recognition
    • G10L15/28Constructional details of speech recognition systems
    • G10L15/32Multiple recognisers used in sequence or in parallel; Score combination systems therefor, e.g. voting systems

Abstract

The invention relates to a voice based question-answering system and a method for an interactive voice system. The voice based question-answering system comprises a question-answering data storing unit, a voice recognizing unit, a semantic similarity calculating unit and a classifying unit, wherein the question-answering data storing unit is used for storing questions and the answers corresponding to the questions by an association manner; the voice recognizing unit is used for performing voice recognition for questions asked by a user through a language model; the semantic similarity calculating unit is used for calculating the semantic similarity of the question asked by the user and each question stored in the question-answering data storing unit according to the recognizing result performed by the voice recognizing unit for the questions asked by the user, wherein the semantic similarity is used for describing the similarity of the questions asked by the user and each question stored in the question-answering data storing unit based on the meanings; the classifying unit is used for classifying the questions asked by the user into the questions in the storing unit and the questions outside the storing unit according to the semantic similarity calculated by the semantic similarity calculating unit.

Description

基于语音的问答系统和用于交互式语音系统的方法 The method of speech-based question answering system for interactive voice

技术领域 FIELD

[0001] 本发明涉及基于语音的问答系统和用于交互式语音系统的方法。 [0001] The present invention relates to a method and a speech-based question answering system for interactive voice system.

背景技术 Background technique

[0002] 近年来,能够自动回答用户所问的问题的基于语音的问答系统得到研究。 Studies [0002] In recent years, a user can automatically answer the questions asked of speech-based Q & A system has been. 对于该类系统,基于实例的答复生成方法是一种有效的方法,所述基于实例的答复生成方法利用问答数据库,所述问答数据库中包括很多问题-答案对。 For such systems, an example of generating method based on the reply is an effective method, based on the example of using the method for generating answer quiz database, the database includes a number of quiz questions - answer pair. 当一个问答系统从用户接收到问题时,它首先通过语音识别器来识别用户所问的问题,然后从问答数据库中选择问答数据库中的与语音识别器所识别的结果最相似的问题,并将问答数据库中的与所选择的问题成一对(即,相对应)的答案提供给用户。 When the system receives a quiz question from the user, it first identifies the user questions asked by the speech recognizer, then the speech recognizer selects the identified result most similar database from quiz question Q database, and Q database with the selected question to answer one pair of (i.e., corresponding) provided to the user.

[0003] 但是,基于实例的答复生成方法的问题是开发人员无法事先把所有类型的问题-答案对都记录在问答数据库中。 [0003] However, based on the reply to question generation method is an example of developers can not advance to all types of questions - answers to the questions and answers are recorded in the database. 由于一个要问的问题的变型可以有很多种,对于开发人员而言,很难将它们都事先记录到问答数据库中;进一步地,将很多个问题的各种变型都事先记录到问答数据库中更是无法实现。 Because of variations to ask a question can have a variety, for developers, they are difficult to pre-recorded questions and answers database; further, various variations are a lot of questions in advance to the Q & A database records more It can not be achieved. 当问答数据库中没有记录与所问的问题类似的问题(和相应的答案)时,该问答系统会请用户问另外的问题,或者仅向用户提供不正确的答案。 When there is no record similar to the questions asked questions (and the corresponding answer) questions and answers in the database, the questions and answers will be prompted to ask another question, or provide incorrect answers only to the user.

[0004] 鉴于以上原因,需要提高基于语音的问答系统的问答数据库中记录的问题-答案对的数量。 The answer to the number of - [0004] For these reasons, problem-based Q & A Q & A database system voice recorded in the need to improve.

[0005] -种解决的方法例如是,开发人员可以先记录用户所问的问题。 [0005] - Method kinds of solutions are, for example, developers can record the user first questions asked. 然后,人为地听取所记录的问题并转录这些问题,并检测这些问题在问答数据库中是否有记录(问题检测步骤)。 Then, the problem artificially hear the recorded and reproduced these problems and these problems in the Q detecting whether the database record (detecting step problems). 然后,开发人员为每个未被记录的问题提供答案,并将所述未被记录的问题和提供的其答案添加到问答数据库中(答案生成步骤)。 Then, the developer provide answers to each question is not recorded and the unrecorded problem and provide the answer to the quiz added database (answer generating step).

[0006] 但是,该方法的问题检测步骤和答案生成步骤都需要人工操作,这使得该方法效率低下。 [0006] However, the method is problematic and answers generating step detecting step requires manual operation, which makes the process inefficient.

[0007] 另外,常规的基于语音的问答系统仅根据用户所问的问题的声学和语言特征来确定用户所问的问题是否在系统中有记录。 [0007] Further, the conventional voice-based question answering system according to the acoustic and linguistic features only questions asked of the user to determine whether the user has asked the question recorded in the system. 从而导致如果用户所问的问题与系统中的例如问答数据库中记录的问题在字面意思上不一致(即,单纯的文字上不一致)时,常规的系统会把用户所问的问题作为系统中未记录的问题对待,从而使得这样的问答系统的识别问题的精度较低,能够准确回答的问题也较少。 If the user resulting in the question asked questions and answers and problems such records in the database system does not match (i.e., inconsistency mere words) in the literal meaning, the conventional system will ask the user a question system unrecorded treatment of the problem of low, so that the recognition of the problem with such precision answering system can accurately answer the question less. 而如果想覆盖更多的问题,则必须增加系统中记录的问题-答案对的数量,这对于系统的存储容量提出更高的要求。 And if you want to cover more of a problem, you must increase the problems recorded in the system - the number of answers, which put forward higher requirements for the storage capacity of the system.

[0008] 例如,假设在问答数据库中记录了"给我操作手册看看",而用户所问的问题是"我想读一本手册",那么常规的基于语音的问答系统会将用户所问的该问题分类为未记录的问题。 [0008] For example, suppose the record "operating manual to see me" in the question and answer database, and the user asked the question "I want to read a manual," then conventional voice-based Q & A system will ask the user the problem is the problem of classification unrecorded. 但是,这两句话的意思差不多,也就是说,用户所问的该问题应当被认为等同于问答数据库中记录的相应问题而被分类为已记录的问题。 However, these two words mean the same, that is, the user is asking the question should be considered equivalent to the corresponding questions and answers questions recorded in the database and are classified as problem recorded. 即,语义对于基于语音的问答系统也是很重要的一个方面,但是在已有的方法中却仅考虑了用户所问的问题的声学和语言特征(即,字面意思),语义上的相似却未被考虑。 That is, the semantics of questions and answers for voice-based system is also very important aspect, but the existing methods but only considered the acoustic and linguistic characteristics of the questions asked of the user (ie, literally), but did not semantically similar be considered.

发明内容 SUMMARY

[0009] 综上可知,需要一种以下这样的基于语音的问答系统:其能够自动确定用户所问的问题是否在系统中(例如,系统的数据存储单元中)有记录。 [0009] To sum up, the following such a need for a speech answering system based on: the user is capable of automatically determining whether the questions asked in the system (e.g., the data storage unit system) record. 优选地,还需要该系统能够提高其中所存储的数据的有效性从而减轻对于存储容量的需求,需要使得该系统的识别问题的精度更高,需要使得该系统所覆盖的问题的范围更大。 Preferably, the system can also need to improve the effectiveness of data stored therein so as to reduce the need for storage capacity, so that the need for higher accuracy of the system identification problem, the problem that the system requires a larger covered.

[0010] 本发明旨在解决上面描述的至少一个问题。 [0010] The present invention aims to solve at least one problem described above. 本发明的一个目的是提供一种解决以上问题中的任何一个的基于语音的问答系统和用于交互式语音系统的方法。 An object of the present invention is to solve the above problems to provide a method of any one of the speech-based question answering system for interactive voice system.

[0011] 具体地,对于用户所问的问题,通过计算语音识别单元对于该所问问题的识别结果与系统中所存储的相关数据的在语义(所表达的含义)上的相似程度(简称语义相似度), 确定该所问问题在系统中是否有与其匹配的数据。 The degree of similarity (the semantics of the [0011] (meaning of the expression) In particular, the user is asked questions for the Semantics Recognition results with the system of the questions asked in the stored data by calculating a speech recognition unit similarity), determine the question asked whether there is data that matches it in the system.

[0012] 如果该所问问题在系统中有与其匹配的数据,则可以输出与该匹配数据对应的输出数据。 [0012] If the question asked its matching data in the system, it can be output to the output data corresponding to the matching data.

[0013] 如果该所问问题在系统中没有与其匹配的数据,则可以将该所问问题存储到系统中以扩展系统中所记录的数据。 [0013] If the data is not matched therewith ask questions in the system, the data system may be extended to systems in which questions are asked recorded is stored.

[0014] 根据本公开内容的一个方面,提供一种基于语音的问答系统,所述系统包括:问答数据存储单元,在该问答数据存储单元中相关联地存储问题以及对应于所述问题的答案; 语音识别单元,通过使用语言模型对于用户说出的问题进行语音识别;语义相似度计算单元,根据语音识别单元对于用户说出的问题的识别结果,计算用户说出的问题与所述问答数据存储单元中存储的每个问题之间的语义相似度,其中,所述语义相似度用于表示用户说出的问题与所述问答数据存储单元中存储的每个问题所表达的意思上的相似程度;以及分类单元,基于所述语义相似度计算单元所计算出的所述语义相似度,将用户说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 [0014] In accordance with one aspect of the present disclosure, there is provided a speech answering system, the system is based comprises: a data storage unit Q, Q in the data storage unit stored in association with the question and the corresponding answer to question ; voice recognition unit, by using a language model for speech recognition for the user issues spoken; semantic similarity calculating unit according to the result of the speech recognition unit question spoken by the user, the calculation of the Q data of the user's spoken semantic similarity between each question stored in the storage unit, wherein, the semantic similarity used to indicate similar meaning expressed on each question with the quiz question spoken by the user of the data storage unit stored degrees; and a classification unit, based on the semantic similarity of the semantic similarity calculation unit is calculated, the user issues spoken classified as a problem in the storage unit or an external storage unit in question.

[0015] 根据本公开内容的另一个方面,提供一种用于交互式语音系统的方法,该交互式语音系统包括语音识别单元和问答数据存储单元,在所述问答数据存储单元中相关联地存储了问题以及对应于所述问题的答案,所述方法包括:语音识别步骤,通过语音识别单元使用语言模型对于用户说出的问题进行语音识别;语义相似度计算步骤,根据语音识别单元对于用户说出的问题的识别结果,计算用户说出的问题与所述问答数据存储单元中存储的每个问题之间的语义相似度,其中,所述语义相似度用于表示用户说出的问题与所述问答数据存储单元中存储的每个问题所表达的意思上的相似程度;以及分类步骤,基于计算出的所述语义相似度,将用户说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 [0015] In accordance with another aspect of the disclosure, there is provided a method for interactive voice system, the interactive voice system comprises a voice recognition unit and Q data storing unit in association with said data storage unit Q is stored corresponding to the question and the answer, the method comprising: a speech recognition step of performing speech recognition for the user issues spoken by a voice recognition unit using a language model; semantic similarity calculation step, the speech recognition unit for the user recognition result spoken questions, each question semantic similarity between the question of the quiz data storing unit calculates spoken by the user is stored, wherein, the semantic similarity for indicating the user issues spoken the degree of similarity on the meaning expressed for each of the quiz question data storage unit; and a step of classification, based on the semantic similarity calculated, the user issues spoken classified as problems in the storage unit or problems outside the storage unit.

[0016] 由此,本发明能够自动确定用户所问的问题是否在系统中(例如,系统的数据存储单元中)有记录。 [0016] Accordingly, the present invention can automatically determine whether the user is asked questions in the system (e.g., the data storage unit system) record. 而且,本发明能够提高基于语音的问答系统中所存储的数据的有效性,从而减轻其对于存储容量的需求,并且,能够使得该系统的识别问题的精度更高,并且使得该系统所覆盖的问题的范围更大。 Further, the present invention can improve the effectiveness of Q-based voice data stored in the system, thus lessening the need for storage capacity, and it is possible to identify the problem that the accuracy of the system is higher, and so the system is covered larger scope of the problem.

[0017] 参照附图来阅读对于示例性实施例的以下描述,本发明的其他特性特征和优点将变得清晰。 [0017] For the following description to be read with reference to exemplary embodiments, other characteristic features and advantages of the present invention will become apparent from the accompanying drawings.

附图说明 BRIEF DESCRIPTION

[0018] 并入到说明书中并且构成说明书的一部分的附图示出了本发明的实施例,并且与描述一起用于解释本发明的原理。 [0018] a part of the specification are incorporated in and constitute a specification, illustrate an embodiment of the present invention, and together with the description serve to explain the principles of the invention. 在这些附图中,除非另外指明,否则,相同的附图标记用于表示具有相同功能的要素。 In these figures, unless otherwise indicated, the same reference numerals are used for elements having the same functions.

[0019] 图1是示出能够实施本发明的实施例的计算机系统的示例性的硬件配置的框图。 [0019] FIG. 1 is a block diagram of an exemplary hardware configuration of a computer system capable of implementing embodiments of the present invention.

[0020] 图2示例性地示出根据本发明的实施例的基于语音的问答系统的框图。 [0020] FIG. 2 shows a block diagram of an exemplary voice-based question answering system of the embodiment of the present invention.

[0021] 图3示例性地示出根据本发明的实施例的用于交互式语音系统的方法的流程图。 [0021] FIG. 3 illustrates an example flowchart of a method for interactive voice system embodiment of the present invention.

[0022] 图4示例性地示出根据本发明的实施例的作为例子的语义相似度计算单元的框图。 [0022] FIG. 4 shows a block diagram of an exemplary computing unit according to an embodiment of the semantic similarity of the present invention as an example.

[0023] 图5示例性地示出根据本发明的实施例的作为例子的用于计算语义相似度的方法的流程图。 [0023] FIG. 5 shows a flowchart of an exemplary method according to an embodiment of the present invention as an example for calculating the semantic similarity.

[0024] 图6示例性地示出根据本发明的实施例的另一基于语音的问答系统的框图。 [0024] FIG. 6 shows a block diagram of an exemplary voice-based question answering system according to another embodiment of the present invention.

[0025] 图7示例性地示出根据本发明的实施例的另一用于交互式语音系统的方法的流程图。 [0025] illustrates a method for interactive voice system according to another embodiment of the present invention. FIG. 7 is a flowchart exemplarily.

[0026] 图8示例性地示出文字相似度计算单元的框图。 [0026] FIG. 8 shows an example block diagram of a similarity calculation unit character.

[0027] 图9示例性地示出文字相似度计算方法的流程图。 [0027] FIG. 9 shows a flowchart of exemplary character similarity calculation method.

[0028] 图10示出本发明与常规技术的效果对比。 Contrast effect [0028] FIG. 10 illustrates the present invention and the conventional art.

具体实施方式 detailed description

[0029] 应当注意,以下的示例性实施例不意欲限制所附权利要求的范围,并且在示例性实施例中描述的特征的所有组合对于本发明的解决手段并不一定是必需的。 [0029] It should be noted that the following exemplary embodiments are not intended to limit the scope of the appended claims, and all combinations of features described in the exemplary embodiment of the present invention is solving means is not necessarily required. 以下描述的本发明的示例性实施例中的每一个都可单独地实施,或者在必要的情况下或在单个实施例中组合来自各个实施例的要素或特征是有益的情况下作为多个实施例或者它们的特征的组合来实施。 The following exemplary embodiments of the present invention are described each of which can be implemented separately, or if necessary, or from a combination of embodiments or features of the various elements of the embodiments is advantageous as a plurality of embodiments in the case where a single embodiment embodiment or a combination thereof characteristics thereof.

[0030] 由于在这些附图中类似的附图标记用于表示类似的要素,所以,在说明书中对于这些类似的要素将不会重复描述,而本领域普通技术人员应当明白这些类似的要素表示类似的含义。 [0030] Since numerals in these figures like reference refer to similar elements, therefore, will not be repeated description of these similar elements in the specification, but one of ordinary skill in the art should understand that these represent similar elements similar meaning.

[0031] 并且,在本公开中,本发明的基于语音的问答系统的各单元、部件和/或组件等等可以用硬件、固件、软件、或其任意组合来实施,并且如果这些单元、部件和/或组件要执行的操作与根据本发明的用于交互式语音系统的方法中的步骤是类似的,则为了简要起见, 可能将仅具体描述相应的步骤而省略对于操作的详细描述,但是本领域普通技术人员会明白,这些单元、部件和/或组件所要执行的操作的具体内容。 [0031] In the present disclosure, the present invention is based on the units Q speech system, components and / or components, etc. may be implemented in hardware, firmware, software, or any combination thereof, and if these elements, components, operations and / or components to perform the steps of a method for interactive voice system of the present invention are similar, then the sake of brevity, only the specifically described may be appropriate for the operation of the steps detailed description is omitted, but those of ordinary skill in the art will appreciate that the details of the operation of these elements, components, and / or components to be executed. 另外,根据本发明的用于交互式语音系统的方法也是可以用硬件、固件、软件、或其任意组合来实施的。 Further, a method for interactive voice system of the present invention are possible in hardware, firmware, software, or any combination thereof used to be implemented. 即,本发明的方法和系统不受实施方式的限制,其保护范围仅由所附权利要求来限定。 That is, the method and system of embodiments of the present invention are not by way of limitation, the scope of which is defined only by the appended claims.

[0032] 在本公开中,术语"第一"、"第二"等仅仅被用来在要素之间进行区分,而并不意图表示时间顺序、优先级或重要性等。 [0032] In the present disclosure, the terms "first," "second," and the like are only used to distinguish between the elements, and are not intended to represent a chronological order, the priority or importance like.

[0033] 而且,在本公开中,步骤的执行顺序不是必须要按照流程图所示出和实施例中所提到的那样,而是可以根据实际情况来灵活变通的,即,本发明不应该受到流程图所示出的步骤的执行顺序的限制。 [0033] Further, in the present disclosure, the order of the steps is not necessarily executed according to the flowchart shown in the embodiment and as mentioned embodiments, but may be flexible according to the actual situation, i.e., the present invention should not be limited by the order of execution of steps shown in the flowchart.

[0034] 以下,将参照附图来对本发明的示例性实施例进行详细描述。 [0034] Hereinafter, exemplary embodiments will be exemplary embodiments of the present invention are described in detail with reference to the accompanying drawings.

[0035] 图1是示出能够实施本发明的实施例的计算机系统1的硬件配置的框图。 [0035] FIG. 1 is a block diagram of a computer system according to embodiments of the present invention can be implemented in a hardware configuration.

[0036] 如图1中所示,计算机系统1包括计算机1110。 [0036] As shown in FIG, computer system 1 includes a computer 1110. 计算机1110包括经由系统总线1121连接的处理单元1120、系统存储器1130、固定非易失性存储器接口1140、可移动非易失性存储器接口1150、用户输入接口1160、网络接口1170、视频接口1190和输出外围接口1195。 The computer 1110 includes 1120, a system memory 1130, fixed non-volatile memory interface 1140, a removable nonvolatile memory interface 1150, a user input interface 1160, network interface 1170 via a processing unit connected to a system bus 1121, a video interface 1190 and an output peripheral interface 1195.

[0037] 系统存储器1130包括ROM (只读存储器)1131和RAM (随机存取存储器)1132。 [0037] The system memory 1130 includes a ROM (Read Only Memory) 1131 and a RAM (Random Access Memory) 1132. BIOS (基本输入输出系统)1133驻留在R0M1131中。 BIOS (Basic Input Output System) 1133 reside in R0M1131. 操作系统1134、应用程序1135、其他程序模块1136和某些程序数据1137驻留在RAMl 132中。 Operating system 1134, application programs 1135, other program modules 1136 and program data 1137 resides in some RAMl 132 in.

[0038] 诸如硬盘之类的固定非易失性存储器1141连接到固定非易失性存储器接口1140。 [0038] such as a hard disk fixed non-volatile memory 1141 is connected to the fixed non-volatile memory interface 1140. 固定非易失性存储器1141例如可以存储操作系统1144、应用程序1145、其他程序模块1146和某些程序数据1147。 A fixed non-volatile memory 1141 may store, for example, an operating system 1144, application programs 1145, other program modules 1146 and program data 1147 some.

[0039] 诸如软盘驱动器1151和⑶-ROM驱动器1155之类的可移动非易失性存储器连接到可移动非易失性存储器接口1150。 [0039] such as a floppy disk drive 1151 and the movable ⑶-ROM drive 1155 or the like connected to the non-volatile memory removable nonvolatile memory interface 1150. 例如,软盘1152可以被插入到软盘驱动器1151中,以及⑶(光盘)1156可以被插入到⑶-ROM驱动器1155中。 For example, a flexible disk 1152 may be inserted into the floppy disk drive 1151, and ⑶ (disc) 1156 may be inserted into ⑶-ROM drive 1155.

[0040] 诸如麦克风1161和键盘1162之类的输入设备被连接到用户输入接口1160。 [0040] Input devices such as a keyboard 1162 and a microphone 1161 or the like are connected to the user input interface 1160.

[0041] 计算机1110可以通过网络接口1170连接到远程计算机1180。 [0041] Computer 1110 can connect to remote computers through a network interface 1180 1170. 例如,网络接口1170可以经由局域网1171连接到远程计算机1180。 For example, network interface 1170 can be connected to the remote computer 1180 via LAN 1171. 或者,网络接口1170可以连接到调制解调器(调制器-解调器)1172,以及调制解调器1172经由广域网1173连接到远程计算机1180。 Alternatively, network interface 1170 can be connected to a modem (modulator - demodulator) 1172, and a modem 1172 connected to the remote computer 1180 via the WAN 1173.

[0042] 远程计算机1180可以包括诸如硬盘之类的存储器1181,其存储远程应用程序1185。 [0042] The remote computer 1180 can include a memory 1181 such as a hard disk, which stores remote application programs 1185.

[0043] 视频接口1190连接到监视器1191。 [0043] The video interface 1190 is connected to a monitor 1191.

[0044] 输出外围接口1195连接到打印机1196和扬声器1197。 [0044] an output peripheral interface 1195 connected to a printer 1196 and a speaker 1197.

[0045] 图1所示的计算机系统可以被实施于任何实施例,可作为独立计算机,或者也可作为设备中的处理系统,可以移除一个或更多个不必要的组件,也可以向其添加一个或更多个附加的组件。 The computer system shown [0045] FIG. 1 may be implemented in any embodiment, as a standalone computer, or may be used as a processing system apparatus, you may remove one or more unwanted components, to which can be add one or more additional components.

[0046] 用户可以采用任何方式使用图1所示的计算机系统,本发明对于用户使用计算机系统的方式不作限制。 [0046] The user can employ any manner using the computer system shown in FIG. 1, for the embodiment of the present invention using the computer system is not limited.

[0047] 显然,图1所示的计算机系统只是示例性的,并且决不意图限制本发明、本发明的应用或用途。 [0047] Obviously, the computer system shown in FIG 1 is merely exemplary in nature and is in no way intended to limit the present disclosure, application, or uses of the invention.

[0048] [第一实施例] [0048] [First Embodiment]

[0049] 以下,将参照图2和图3来详细描述本发明的第一实施例。 [0049] Hereinafter, a first embodiment of the present invention will be described in detail with reference to FIGS. 2 and 3.

[0050] 图2示例性地示出根据本发明的第一实施例的基于语音的问答系统100的框图。 [0050] FIG. 2 shows an example of the voice-based question answering system block diagram of a first embodiment of the present invention embodiment 100 according to.

[0051] 如图2所示,根据本发明的第一实施例的基于语音的问答系统100可以包括问答数据存储单元101、语音识别单元102、语义相似度计算单元103、分类单元104。 [0051] As shown, a speech-based system 100 may comprise a quiz question and answer data storage unit 101, a voice recognition unit 102, semantic similarity calculating a first embodiment of the present invention 103, the classification unit 104 2.

[0052] 其中,在问答数据存储单元101中相关联地存储有问题以及对应于所述问题的答案。 [0052] wherein Q in the data storage unit 101 are stored in association with a question and an answer corresponding to the question.

[0053] 并且,语音识别单元102被配置为通过例如使用语言模型对于用户说出的问题进行语音识别。 [0053] Further, the speech recognition unit 102 is configured by, for example, using speech recognition language model for the user issues spoken.

[0054] 并且,语义相似度计算单元103被配置为根据语音识别单元102对于用户说出的问题的识别结果,计算用户说出的问题与所述问答数据存储单元101中存储的每个问题之间的语义相似度。 [0054] Further, the semantic similarity calculation unit 103 is configured to spoken voice recognition unit 102 according to the recognition result of the user spoken questions, each question is calculated and the user questions quiz data stored in the storage unit 101 of the semantic similarity between. 在本文中,该语义相似度用于表示用户说出的问题与所述问答数据存储单元101中存储的每个问题所表达的意思上的相似程度。 Herein, the degree of semantic similarity used to indicate similar meaning expressed on the user issues spoken questions and each of said Q data in the storage unit 101.

[0055] 并且,分类单元104被配置为基于所述语义相似度计算单元103所计算出的语义相似度,将用户说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 [0055] Then, the classification unit 104 is configured based on the semantic similarity calculating unit 103 calculates the semantic similarity, the user issues spoken classified as a problem in the storage unit or an external storage unit in question.

[0056] 当然,本领域普通技术人员会明白,上面描述的基于语义相似度来对问题进行分类以区分其是在存储单元内还是在存储单元外只是一个例子,实际上,在本发明中,分类单元可以基于能够想到的任何其他方式来将用户所问的问题进行分类,只要能够将其区分开是在存储单元内还是在存储单元外即可。 [0056] Of course, those of ordinary skill in the art will appreciate that, to the problem based on semantic similarity of the above-described classification is to distinguish which is just one example out of a storage unit in the storage unit, in fact, in the present invention, user classification unit may be asked questions based classification can think of any other way, as long as it can separate area within the storage unit or in an external storage unit can be.

[0057] 另外,需要注意的是,图2所示的基于语音的问答系统仅是示例性的,本发明的基于语音的问答系统不限于此。 [0057] Further, it is noted that, as shown in FIG. 2 Q speech-based system is only exemplary, speech-based question answering system of the present invention is not limited thereto. 并且,图2中所示的这些部件对于解决本发明的技术问题而言并不一定都是必须的,并且,这些部件可以采取单个部件的形式或者可以采取任何方式进行组合。 Further, the components shown in FIG. 2 for solving the technical problem of this invention, are not necessarily required, and these members may take the form of a single component or may be combined in any manner.

[0058] 以下将参照图3来描述用于交互式语音系统的方法。 [0058] Here will be described with reference to a method for interactive voice system 3. FIG.

[0059] 图3示例性地示出根据本发明的实施例的用于交互式语音系统的方法的流程图。 [0059] FIG. 3 illustrates an example flowchart of a method for interactive voice system embodiment of the present invention.

[0060] 在本发明中,交互式语音系统包括上面提到的基于语音的问答系统,还可以包括能够想到的其他形式的可以和用户进行交互的基于语音的系统。 [0060] In the present invention, the interactive voice system comprises a voice-based question answering system mentioned above, but also may include other forms may be contemplated capable of voice-based systems and user interaction.

[0061] 根据本发明的实施例的交互式语音系统可以包括语音识别单元和问答数据存储单元。 [0061] The voice recognition unit and a data storage unit Q according to an embodiment of the interactive voice system of the present invention may comprise. 其中,语音识别单元可以通过例如语言模型来识别输入的语音,例如用户所问的问题。 Wherein the voice recognition unit may recognize the voice input through the language model, for example, such as the user questions asked. 并且,在所述问答数据存储单元中可以相关联地存储有多种数据,例如,其中可以存储有问题以及对应于所述问题的答案。 Further, there may be associated a plurality of data stored in said data storage unit Q, for example, which may be stored in question and the answer corresponding to the question.

[0062] 在根据本发明的实施例的用于交互式语音系统的方法中,在语音识别步骤S101, 对于语音输入(例如,用户说出的问题),通过语音识别单元利用例如语言模型等进行语音识别,从而可以得到语音识别的结果。 [0062] In the method for interactive voice system according to embodiments of the present invention, the speech recognition step S101, the speech input for (e.g., the user issues spoken), using, for example by a voice recognition unit according to the language models voice recognition, the voice recognition result can be obtained in. 这里,语音识别的结果由词构成,例如,其可以包括构成所问问题的词。 Here, the speech recognition result of the word constituted by, for example, it may include words constituting the questions asked.

[0063] 这里,上面提到的语言模型还可以通过例如更新单元被更新(后面将提到),以使得语音识别结果更精确。 [0063] Here, the above-mentioned language model can also be updated by the updating unit, for example (which will be mentioned later), so that a more accurate speech recognition result.

[0064] 这里,可用的识别语音的方法有很多,例如,基于HMM (隐式马尔科夫模型)的方法等,在此对其不作任何限制,只要其能够识别出输入的语音即可。 [0064] Here, there is a method available for recognizing a voice of many, for example, the HMM (Hidden Markov Model) method based on this thereof without any restrictions, as long as it can recognize the speech input can be. 另外,语言模型例如可以是通过领域内文本数据训练的领域特定的N元语言模型。 Further, for example, the language model may be trained in the art of the text data within a domain-specific N-gram language model.

[0065] 另外,从所述语音识别单元还可以输出语音识别结果(更具体地,词)的声学得分和语言得分、以及所问问题的帧大小等。 [0065] Further, the speech recognition result may be output from the voice recognition unit (more specifically, the words) of the acoustic score and the linguistic score, and the questions asked frame size.

[0066] 在语义相似度计算步骤S102,根据语音识别单元对于用户说出的问题的语音识别结果,计算用户说出的问题与所述问答数据存储单元中所存储的每个问题之间的语义相似度。 [0066] In the semantic similarity calculation step S102, the speech recognition unit the speech recognition result to the user issues spoken calculates the semantic question spoken by the user between each question and the question and answer data stored in the storage unit similarity. 这里,语义相似度用于表示用户说出的问题与所述问答数据存储单元中存储的每个问题所表达的意思上的相似程度。 Here, the degree of semantic similarity used to indicate similar meaning expressed on the user issues spoken questions and each of said Q data in the storage unit. 由此可以看出,本发明不是仅考虑文字(即单纯的字面意思)上的相似程度。 It can be seen that the present invention not only considers the text (i.e. purely literally) on the degree of similarity.

[0067] 关于计算语义相似度的方法将在后面作为例子进行描述。 [0067] on semantic similarity calculation method will be described later by way of example. 但是,需要注意的是,本发明不限于后面将描述的计算语义相似度的方法,本领域普通技术人员能够想到的任何计算语义相似度的方法都应包含在本发明的保护范围内。 Note, however, that the present invention is not limited to the method of calculating the semantic similarity will be described later, those of ordinary skill in the art that any conceivable method of calculating semantic similarity to be included within the scope of the present invention.

[0068] 在分类步骤S103,基于语义相似度计算步骤S102所计算出的语义相似度,将用户说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 [0068] In the classification step S103, the semantic similarity calculating semantic similarity calculated in step S102 based on the user issues spoken classified as a problem in the storage unit or an external storage unit in question.

[0069] 这里,上述的分类可以通过例如支持向量机(SVM)来实现,但是这仅是一种分类的实现方式的例子,本发明对于分类的实现方式并不进行任何限制,即,只要能够实现本发明所需的分类,任何分类方式都是可以的。 [0069] Here, the classification can be achieved by, for example, support vector machines (the SVM), this is merely an example of an implementation of a classification of the present invention for implementation of the classification is not any limitation, i.e., as long as the present invention to achieve the desired classification, any classification is possible.

[0070] 另外,需要注意的是,参照图3所描述的用于交互式语音系统的方法仅是示例性的,本发明的用于交互式语音系统的方法不限于此。 [0070] Further, it is noted that a method for interactive voice system described with reference to FIG. 3 are only exemplary, and a method for interactive voice system of the present invention is not limited thereto. 并且,图3中所示的这些步骤对于解决本发明的技术问题而言并不一定都是必须的,并且,上面所描述的这些步骤之间的顺序也仅仅是示例性的,本发明并不限制于这样的步骤顺序。 Then, the steps shown in FIG. 3 for solving the technical problem of this invention, are not necessarily required, and the sequence between the steps described above is only exemplary, and the present invention is not limited to such a sequence of steps.

[0071] 通过本发明,具体地,借助于计算所问的问题与所存储的问题之间的语义相似度, 能够自动确定用户所问的问题在系统中(例如,系统的数据存储单元中)是否有记录,并且能够提高系统中所存储的数据的有效性,从而减轻了对于系统的存储容量的要求,并且还使得系统的识别精度更高,从而使得该系统所覆盖的问题的范围更大。 [0071] By the present invention, in particular, by means of a semantic similarity between the calculated and the question asked question stored, a user can automatically determine the questions asked in the system (e.g. data storage unit, system) if greater recording, and to improve the validity of the data stored in the system, thereby reducing the storage capacity required for the system, and also makes the system a higher recognition accuracy, so that the problem that the range covered by the system .

[0072] [第二实施例] [0072] [Second Embodiment]

[0073] 以下,将参照图4、图5来描述根据本发明的第二实施例。 [0073] Hereinafter, with reference to FIG. 4, FIG. 5 will be described a second embodiment of the present invention.

[0074] 在本发明中,可以进行多种语义相似度的计算,具体地,可以计算用户所说出的问题与问答数据存储单元中记录的问题之间的多种语义相似度。 [0074] In the present invention, the plurality of semantic similarity may be calculated, in particular, can be calculated more semantic similarity between the user and the spoken question quiz question recorded in the data storage unit. 也就是说,各种语义相似度的计算方式都在本发明的保护范围之内。 That is, the various semantic similarity calculation are within the scope of the present invention. 但是,为了简要起见,下面将在本实施例中仅参照图4、图5来描述语义相似度的计算的例子。 However, for brevity, will be 4, FIG. 5 describes an example of semantic similarity is calculated only with reference to the embodiments in the present embodiment FIG.

[0075] 图4示例性地示出根据本发明的实施例的作为例子的语义相似度计算单元103的框图。 [0075] FIG. 4 shows a block diagram of an exemplary calculation unit 103 according to the embodiment of semantic similarity embodiment of the present invention as an example.

[0076] 如图4所示,该语义相似度计算单元103可以包括第一词袋产生部件1031、第一词选择部件1032、第一检索部件1033、第二词袋产生部件1034、第二检索部件1035、以及语义相似度计算部件1036。 [0076] As shown in FIG. 4, the semantic similarity calculation unit 103 may include a first bag of words generating member 1031, a first word selection means 1032, a first retrieval part 1033, a second bag of words generating member 1034, second search member 1035, and the semantic similarity calculation section 1036.

[0077] 其中,第一词袋产生部件1031被配置为根据语音识别单元对于用户所说出的问题的识别结果,产生第一词袋,其中所述第一词袋包括该识别结果中所含的词。 [0077] wherein the first bag of words generating member 1031 is configured as a result of the speech recognition unit for the user issues spoken generates a first bag of words, wherein the bag comprises a first word included in the recognition result word.

[0078] 第一词选择部件1032被配置为根据语音识别单元对于用户所说出的问题的识别结果,从第一词袋选择其概率大于第一阈值的词。 [0078] The first member 1032 is configured to select the word speech recognition unit according to the result of the user issues spoken selects the probability is greater than a first threshold value words from the first word to the bag. 其中,词的概率例如可以由通过N元语言模型得到的声学得分来确定。 Wherein, the probability of the word may be determined by, for example, through the acoustic score obtained N-gram language model. 当然,也可以通过其他方式来确定词的概率。 Of course, you can also determine the probability of the word by other means.

[0079] 第一检索部件1033被配置为根据所述第一词选择部件所选择的词,从数据源检索文档。 [0079] The first member 1033 is configured to retrieve member selected word based on the first word is selected from the data source to retrieve the document.

[0080] 第二词袋产生部件1034被配置为针对所述问答数据存储单元中存储的每个问题产生相应的第二词袋,其中每个第二词袋包括所述问答数据存储单元中所存储的相应的问题所含的词。 [0080] The second bag of words generating member 1034 is configured to generate a respective second bag of words for each question the question and answer data stored in the storage unit, wherein each of said Q second word bag comprises as data storage unit word corresponding problems contained in storage.

[0081] 第二检索部件1035被配置为基于每个第二词袋中的词,从数据源检索文档。 [0081] The second member 1035 is configured to retrieve each word of the second word based on the bag, from a data source to retrieve the document.

[0082] 语义相似度计算部件1036被配置为基于第一检索部件所检索到的文档和第二检索部件所检索到的文档,计算用户所说出的问题与所述问答数据存储单元中存储的每个问题之间的语义相似度。 [0082] The semantic similarity calculating section 1036 is configured to retrieve a first part of the retrieved documents based on the retrieval means and the second retrieved documents, calculating a user issues spoken by the Q of the data storage unit semantic similarity between each question.

[0083] 下面,将更具体地结合图5来描述作为例子的用于计算语义相似度的方法。 [0083] Next, more specifically 5 incorporated to the method described for calculating the semantic similarity as an example in FIG.

[0084] 图5示例性地示出根据本发明的实施例的用于计算语义相似度的方法的流程图。 [0084] FIG. 5 shows a flowchart of exemplary semantic similarity calculation method according to an embodiment of the present invention.

[0085] 如图5所示,在第一词袋产生步骤S1021,基于语音识别单元对于用户所说出的问题的识别结果,产生第一词袋。 [0085] As shown in FIG 5, a first bag of words is generated in step S1021, the speech recognition unit based on the recognition result for the user issues spoken generates a first bag of words. 该第一词袋可以包括该识别结果中所含的所有的词。 The bag may comprise a first word all the words contained in the recognition result.

[0086] 例如,对于用户所说出的一个问题的语音识别结果中有以下字母所代表的词:A、 B、C、D、E、F、G、H、I、J、K、L、M,则该第一词袋可以为由这些词所组成的集合{A,B,C,D,E, F,G,H,I,J,K,L,M}。 [0086] For example, a question for the speech recognition result of the user has spoken words represented by the following letters: A, B, C, D, E, F, G, H, I, J, K, L, M, then the first word by the pouch may be a collection of these words consisting of {a, B, C, D, E, F, G, H, I, J, K, L, M}.

[0087] 当然,该第一词袋也可以根据情况而仅包含该语音识别结果的一部分,也就是说, 并不一定必须要包含该识别结果中所含的所有的词。 [0087] Of course, the first bag of words may comprise only a portion of the speech recognition result according to the situation, that is to say, it does not necessarily have to contain all the words included in the recognition result. 或者说,第一词袋可以包括该识别结果中所含的部分或全部的词。 Or, the bag may comprise a first word portion or all of the words contained in the recognition result.

[0088] 然后,在第一词选择步骤S1022,将第一词袋中的其概率小于等于第一阈值(例如, 0. 3)的词去除,换而言之,选择第一词袋中其概率大于第一阈值的词。 [0088] Then, the first word in the selection step S1022, the first word bag probability less than or equal a first threshold value (e.g., 0.3) removing the word, in other words, it selects the first word bag the probability is greater than the threshold value of the first term. 这里,词的概率可以使用已知的方法来计算,例如可以通过模糊网络来计算。 Here, the word probabilities can be calculated using known methods, for example, may be calculated by the fuzzy network. 这里,词的概率例如可以通过由例如N元语言模型所得到的声学得分而得到。 Here, for example, the probability of the word may be obtained by the acoustic score from the N-gram language model, for example, obtained. 当然,也可以采用其他方式来选择词,本发明对此不作任何限制。 Of course, other means may be used to select a word, the present invention does not set any limit.

[0089] 请注意,这里第一阈值可以为范围在0〜1的值,并且,在本发明中,第一阈值的具体数值不受限制。 [0089] Note that there may be a first threshold value in the range of 0~1, and, in the present invention, the specific value of the first threshold value is not limited.

[0090] 然后,在第一检索步骤S1023,基于在第一词选择步骤S1022中所选择的第一词袋中的词,从数据源(例如,从外部的大型数据库或者从网络等)检索文档,从而可以得到一组文档。 [0090] Then, in the first search step S1023, based on the selection of the first word in step S1022 bag word selected in the first word from a data source (e.g., from an external database or from the large-scale network, etc.) to retrieve the document so you can get a set of documents. 这里,可以使用任何已知的方法或者本领域普通技术人员能够想到的方法来检索文档。 Here, the method may be used, or any method known to those of ordinary skill in the art can think to retrieve the document.

[0091] 另一方面,在第二词袋产生步骤S1024,针对问答数据存储单元中所存储的每个问题,产生相应的第二词袋。 [0091] On the other hand, in step S1024 generates a second bag of words, for each question quiz data stored in the storage unit, generating a second word corresponding to the bag. 每个相应的第二词袋可以包括相应问题中所含的所有的词(即, 构成该问题的全部的词),当然,第二词袋也可以根据情况而仅包含相应问题中所含的一部分词。 Each respective second bag of words may include all of the words contained in the respective problems (i.e., all of the words constituting the problem), of course, the second bag of words may only comprise respective question contained in some cases part of the word. 也就是说,如第一词袋那样,第二词袋并不一定必须要包含相应问题中所含的所有的。 That is, as the first word as bags, bags second term does not necessarily have to include the appropriate questions contained in all.

[0092] 然后,在第二检索步骤S1025,对于每个第二词袋,基于该第二词袋中的词,从数据源(例如,从系统中已有的一个数据库,或者从外部的大型数据库,或者从网络,等等)检索文档,从而同样可以得到一组文档。 [0092] Then, in a second search step S1025, bags for each second word, the second word based on the word pocket, from a data source (e.g., an existing system from a database, or from the outside of the large database, or from the network, etc.) to retrieve a document, so that the same can be obtained a set of documents. 同样,这里,可以使用任何已知的方法或者本领域普通技术人员能够想到的方法来检索文档。 Also, here, a method may be used, or any method known to those of ordinary skill in the art can think to retrieve the document.

[0093] 这里,第一检索步骤S1023和第二检索步骤S1025中所用的数据源可以是同一个数据源,也可以不是同一个。 [0093] Here, the first search and a second search step S1023 in step S1025 the data source can be used in the same data source, may not be the same. 本发明对此并没有任何限制或者特别的要求。 The present invention does not have any special requirements or restrictions.

[0094] 然后,在语义相似度计算子步骤S1026,基于在第一检索步骤S1023所检索到的一组文档和在第二检索步骤S1025所检索到的一组文档,计算用户所说出的问题与问答数据存储单元中所存储的每个问题之间的语义相似度。 [0094] Then, in sub-step S1026 semantic similarity calculation, based on the first search step S1023 is a set of documents retrieved in the second retrieving step S1025 and the retrieved document set, calculation of user spoken and semantic similarity between each quiz question data stored in the storage unit of.

[0095] 计算语义相似度的方法很多,下面将例举两个根据本发明的实施例的语义相似度的计算例子。 [0095] Many semantic similarity calculation method, will include two semantic similarity is calculated according to an embodiment of the present invention is an example.

[0096] 首先描述计算例子1,在该例子中,通过向量空间模型(VSM),基于在第一检索步骤S1023和在第二检索步骤S1025检索得到的两组文档中的每个词来计算语义相似度。 [0096] First, a calculation example is described, in this example, by a vector space model (the VSM), based on the first search step S1023 and sets for each word in a second search documents retrieved in step S1025 to calculate the semantic similarity. 为了简便起见,将利用计算例子1得到的语义相似度表示为SS1。 For simplicity, the example of calculation obtained using semantic similarity is expressed as 1 SS1.

[0097] 其中,SSl可由下式(1)得到。 [0097] wherein SSl obtained by the following formula (1).

Figure CN104424290AD00131

[0099] 这里,SS1U q)表示用户所说出的问题s与问答数据存储单元中存储的问题q之间的语义相似度,ys和y,分别表示经过上述第一词选择步骤S1022中的词选择之后得到的第一词袋和从上述第二词袋产生步骤得到的第二词袋,Ays和分别表示从第一检索步骤S1023得到的一组文档和从第二检索步骤S1025得到的一组文档,和,分5 JS J(i 别表示该组文档和该组文档中的每个词的词权重。 Semantic similarity between the [0099] Here, SS1U q) s indicate a problem with the data storage unit Q stored in the user issues spoken q, ys, and y, respectively, after the first word in the word selection step S1022 after a first bag of words obtained from said selecting and generating a second bag of words obtained in the second step bag of words, Ays and S1023 represent a set of documents obtained from the first step and retrieving from the second search step S1025 to obtain a set of documents, and divided 5 JS J (i denote the set of documents the right word and the set of documents for each word weight.

[0100]上述的词权重可以通过例如TFIDF (term frequency - inverse document frequency)算法(参见http://zh. wikipedia. org/wiki/TF-IDF)来计算。 [0100] The weight of the words, for example, by TFIDF (term frequency - inverse document frequency) algorithm (see http: // zh wikipedia org / wiki / TFIDF..) Is calculated. 例如,对于一组文档(称为一个文集)中的一个文档d中的词t的词权重,可以通过下式(2)来计算。 For example, for a set of document term weight t word (called a corpus) of a document d a weight may be calculated by the following formula (2).

Figure CN104424290AD00132

[0102] 这里,tfi;d是该文档d中的该词ti的频率,M是该文集中的文档的总个数,Clf i是该组文集中包含h的文档的总个数。 [0102] Here, tfi; d is the frequency of term ti of the document d, M is the total number of documents in the corpus, Clf i is the total number of the set of corpus of documents containing h.

[0103] 然后,再来描述计算例子2,在该例子中,基于在第一检索步骤S1023和在第二检索步骤S1025检索得到的两组文档的重合来计算语义相似度。 [0103] Then, again described calculation example 2, in this example, based on the first search step S1023 and sets the coincidence document retrieved second search step S1025 to calculate the semantic similarity. 为了简便起见,将利用计算例子2得到的语义相似度表示为SS2。 For simplicity, the example 2 will be calculated using semantic similarity is expressed as SS2. 可通过下式(3)来得到语义相似度SS2。 SS2 semantic similarity may be obtained by the following formula (3).

Figure CN104424290AD00133

[0105] 这里,表示从第一检索步骤S1023和第二检索步骤S1025得到的文档组dv和d 中的重合文档(即,重复的文档或者也可以认为是其中的某些内容相同的}s ycI 文档)的总个数,LWy表示文档组中的文档的总个数(即文档组\,5 中的文档的个数与文档组中的文档的个数之和)。 [0105] Here, dv and represents d S1025 document groups obtained from the first and the second retrieving step S1023 coincides document retrieval step (i.e., duplicate documents may be considered, or some of the content of the same} s ycI document) on the total number, LWY represents the total number of documents in the document group (i.e. document group \, document number, and the number of group 5 of the documents in the document).

[0106] 这里,需要注意的是,尽管上面给出了用于计算语义相似度的两个例子,但是,在本发明中,可以仅使用得到的一个语义相似度来进行后续的分类操作,也可以使用两个或更多个语义相似度来进行后续的分类操作。 [0106] Here, it should be noted that, although the two examples given above for calculating the semantic similarity, however, in the present invention may use only a semantic similarity obtained for subsequent classification operations, also two or more can be used for subsequent semantic similarity sort operation. 所使用的语义相似度可以是通过上面的计算例子给出的方式得到的,也可以是通过其他方式得到的。 Semantic similarity obtained may be used by way of example of calculation given above, it may be obtained by other means. 本发明对于这些并不进行任何限制, 而是只要能够得到所需的语义相似度并能够进行后续的分类即可。 The present invention is not any limitation to these, but can be obtained as long as the desired semantic similarity and to enable subsequent classification. 这里再顺便提及,本发明对于后续的分类操作所使用的分类方法也没有任何限制。 Here again Incidentally, the present invention is a method for classification of the subsequent sorting operation can be used without any limitation.

[0107] 另外,优选地,如图4所示,语义相似度计算单元103还可以包括:第一过滤部件1037,被配置为根据词性,从所述第一词选择部件1032自所述第一词袋所选择的词中过滤掉一部分词,和/或根据词性,从所述第二词袋中过滤掉一部分词。 [0107] Further, preferably, as shown in FIG. 4, the semantic similarity calculation unit 103 may further comprises: a first filter member 1037 configured according to part of speech, from the first word is selected 1032 from the first member bag of words of the selected word to filter out a portion of the words, and / or speech in accordance with, the second word from the filter bag away a portion of the word. 例如,如果第一词袋为集合{A,B,C,D,E,F,G,H,I,J,K,L,M},其中,A、K 为助词,D、F、I 为名词,E、M 为动词,B、 C、G、H、J、L为形容词,则可以根据词性,将第一词袋中的对于后续的检索操作用处可能不大的助词去除。 For example, if the first word of the bag set {A, B, C, D, E, F, G, H, I, J, K, L, M}, where, A, K is a particle, D, F, I is a noun, E, M is a verb, B, C, G, H, J, L is an adjective, according to the speech, the retrieval operation for the subsequent use of the particle may not be removed first word bag.

[0108] 这里,第一检索部件1033可以针对经过第一过滤部件1037过滤后的第一词袋进行检索,第二检索部件1035也可以针对经过第一过滤部件1037过滤后的第二词袋进行检索,以使得到的检索文档更准确。 [0108] Here, the first retrieval part 1033 may be retrieved through a first filter element for a first word after the bag filter 1037, a second retrieval part 1035 may be performed through a first filter for the second member 1037 after the filter bag of words search, retrieve documents in order to get the more accurate.

[0109] 再优选地,如图4所示,所述语义相似度计算单元103还可以包括:领域词添加部件1038,被配置为向过滤后的第一词袋和/或过滤后的第二词袋添加与领域有关的词。 [0109] Still more preferably, as shown in FIG. 4, the semantic similarity calculating unit 103 may further comprise: adding the word field member 1038, and is configured to the second or the first word after the bag filter / filter bag of words added to the areas related words. [0110] 例如,如果该基于语音的问答系统是领域特定的,例如,针对的是"照相机"领域, 那么可以添加的领域词可以是"照相机"或者相关的词。 [0110] For example, if the voice answering system is based on domain-specific, for example, for a "camera" area, then you can add fields word could be "camera" or related words.

[0111] 由此,所述第一检索部件1033可以针对进行上述添加后的第一词袋进行检索,第二检索部件1035也可以针对进行上述添加后的第二词袋进行检索。 [0111] Accordingly, the first retrieval member 1033 may be retrieved for the first word after the bag was added the above, the second retrieval part 1035 may also retrieve the bag for a second term after the above addition.

[0112] 这样,可以提高检索部件1033和/或1035所执行的检索操作的准确度,S卩,使得检索的命中率更高。 [0112] Thus, the search means can improve the accuracy of the retrieval operation 1033 and / or 1035 is executed, S Jie, so that retrieval hit rate higher.

[0113] 这里,需要注意的是,上述的过滤和添加不是必须的,并且,过滤和添加操作之一或者两者可以仅针对第一词袋或第二词袋进行,也可以针对第一词袋和第二词袋两者都进行。 [0113] Here, it should be noted that the above-described filter is not necessary and is added, and the filtered one or both of the add operation may be performed for the first word and the bag or bags a second word only, may for the first word both the bag and a second bag of words are carried out.

[0114] 另一方面,在用于计算语义相似度的方法中,还可以相应地优选包括第一过滤步骤S1027和/或领域词添加步骤S1028,如图5所示。 [0114] On the other hand, in the method for calculating the semantic similarity, it may accordingly also preferably comprises the step of adding a first filtration step S1027 and / or S1028 art words, as shown in FIG.

[0115] 在第一过滤步骤S1027,根据词性,从在第一词选择步骤S1022所选择的第一词袋的词中过滤掉一部分词,和/或根据词性,从所述第二词袋中过滤掉一部分词。 [0115] In a first filtration step S1027, according to part of speech, filtering the first word from the word selecting step S1022 the selected first word of the bag away a portion of the word, and / or speech in accordance with, the second word from the bag filter out part of the word.

[0116] 可以在第一过滤步骤S1027之后进行第一检索步骤S1023, S卩,可以针对经过第一过滤步骤S1027过滤后的第一词袋进行检索,也可以针对经过第一过滤步骤S1027过滤后的第二词袋进行检索,以使得到的检索文档更准确。 [0116] can be the first search step S1023, S Jie after the first filtration step S1027, can be retrieved for the first word after the bag through a first filtration step S1027 filtering may be passed through the first filter for the filtration step S1027 the second word to retrieve the bag, so to get a more accurate document retrieval.

[0117] 在领域词添加步骤S1028,向过滤后的第一词袋和/或过滤后的第二词袋添加与领域有关的词。 [0117] In the field of word addition step S1028, the art related words added to the bag and the second word or the first word after the bag filters / filtration.

[0118] 由此,可以在上述的领域词添加步骤S1028之后进行各检索步骤S1023和S1025。 [0118] Accordingly, steps may be added in the field of each search word is the above-described steps S1023 and S1025 after S1028. 艮P,可以在第一检索步骤S1023针对在第一过滤步骤S1027过滤了的第一词袋进行检索,也可以在第二检索步骤S1025针对在第一过滤步骤S1027过滤了的第二词袋进行检索,以使得到的检索文档更准确。 Gen P, can be retrieved in the first retrieving step S1023 in a first filter for filtering the first step S1027 bag of words, may be carried out in a first filtration step S1027 for the second word filter bags second search step S1025 search, retrieve documents in order to get the more accurate.

[0119] 另外,在本发明中,也可以不经过领域词添加步骤S1028而从第一过滤步骤S1027 直接进行到相应的检索步骤。 [0119] Further, in the present invention, without the art may be added in step S1028 and S1027 word from a first filtration step directly to the corresponding search step.

[0120] 请注意,虽然在上面的例子中描述的是ys和y,分别表示经过上述第一词选择步骤S1022中的词选择之后得到的第一词袋和从上述第二词袋产生步骤得到的第二词袋,但是,^和&应当是经过相应的处理步骤之后最终获得的第一词袋和第二词袋。 [0120] Note that although in the example described above is ys and y, respectively, a first bag of words obtained through said selected first word after word selection step S1022 is generated and the bag of words from the second step to give a second bag of words, however, and ^ & bag of words should be the first and second bag of words after a respective processing step of finally obtained. 也就是说, 参照图5,如果用于计算语义相似度的方法中不包含第一过滤步骤S1027和领域词添加步骤S1028,则ys和y,分别为第一词选择步骤输出的第一词袋和第二词选择步骤输出的第二词袋。 That is, referring to FIG 5, if the method for calculating the semantic similarity does not contain the step of adding a first filtering step of S1027 and S1028 art words, the ys and y, respectively, step of selecting the first bag of words as a first output word word selection step and a second output from the second bag of words. 如果用于计算语义相似度的方法中包含第一过滤步骤S1027,则ys和y,为经由该过滤步骤后输出的第一词袋和第二词袋(即,经过过滤后的第一词袋和第二词袋)。 If the method for calculating the semantic similarity comprises a first filtration step S1027, the ys and y, is the first word and the second word the bag after the bag via the filter output step (i.e., the first word after the filter bags and second bag of words). 另外,如果用于计算语义相似度的方法中进一步包含领域词添加步骤S1028,则ys和y,为领域词添加步骤输出的第一词袋和第二词袋(即,经过词添加之后的第一词袋和第二词袋)。 Further, if the method for calculating the semantic similarity further comprises the step of adding the word field S1028, the ys and y, adding the first output word and the second step of the bag the bag for the word art word (i.e., the first word after adding The word bags and bags second term).

[0121] 需要注意的是,在本发明中,实施例中所描述的步骤顺序仅是示例性的,实际上本发明并不局限于所描述的步骤顺序,而是本领域技术人员能够想到的任何能够实现本发明的效果的步骤的顺序,都是可以的。 [0121] Note that, in the present invention, the steps described in the embodiments are merely exemplary of the sequence, in fact, the present invention is not limited to the steps described order, but those skilled in the art can occur any order of the steps to achieve the effect of the present invention, are possible.

[0122] 通过本发明,具体地,借助于计算所问的问题与所存储的问题之间的语义相似度, 能够自动确定用户所问的问题在系统中(例如,系统的数据存储单元中)是否有记录,并且能够提高系统中所存储的数据的有效性,从而减轻了对于系统的存储容量的要求,并且还使得系统的识别精度更高,从而使得该系统所覆盖的问题的范围更大。 [0122] By the present invention, in particular, by means of a semantic similarity between the calculated and the question asked question stored, a user can automatically determine the questions asked in the system (e.g. data storage unit, system) if greater recording, and to improve the validity of the data stored in the system, thereby reducing the storage capacity required for the system, and also makes the system a higher recognition accuracy, so that the problem that the range covered by the system .

[0123] [第三实施例] [0123] [Third Embodiment]

[0124] 以下,将参照图6、图7、图8、图9来描述根据本发明的第三实施例。 [0124] Hereinafter, with reference to FIG. 6, FIG. 7, FIG. 8, FIG. 9 will be described a third embodiment of the present invention. 在用于描述本实施例的这些附图中,除非另外指明,否则与第一和第二实施例中的附图标记相同的附图标记用于表示具有相同功能的部件。 In the drawings for describing the embodiment of the present embodiment, unless otherwise specified, the same first and second embodiment, reference numerals used to denote components having the same function.

[0125] 首先,图6示例性地示出根据本发明的实施例的另一基于语音的问答系统200的框图。 [0125] First, FIG. 6 shows a block diagram of an exemplary voice-based question answering system 200 according to another embodiment of the present invention.

[0126] 优选地,基于语音的问答系统200还可以包括置信度计算单元205。 [0126] Preferably, speech-based question answering system 200 may further comprise a confidence calculation unit 205. 该置信度计算单元205被配置为基于语音识别单元102输出的有关用户所说出的问题的识别结果,为用户所说出的问题计算置信度。 The confidence calculation unit 205 is configured to identify issues related to the user based on a result output from the speech recognition unit 102 spoken calculates a confidence of the user issues spoken.

[0127] 例如,可以通过将用户所说出的问题经由N元语言模型进行识别而得到的声学得分除以用户所说出的该问题经由音素网络进行识别而得到的声学得分来确定用户所说出的该问题的置信度。 Acoustic [0127] For example, the user can issue spoken identification via N-gram language model score obtained by dividing the problem of identifying the user spoken phoneme network via the acoustic score obtained by a user of said determined confidence out of the question.

[0128] 更具体而言,可以使用下式(4)来计算用户所说出的问题s的置信度CMS。 [0128] More specifically, using the (4) to calculate the user issues spoken formula s confidence CMS.

Figure CN104424290AD00151

[0130] 这里,n(s)表示问题s的巾贞大小,PG I IF; G_v)表示通过N元语言模型Gn得到的语音识别结果澤F的声学得分,u表示音素序列,L(Gp)表示音素网络Gp接受的一组可能的音素序列,P(s|u)表示音素序列U的声学得分。 [0130] Here, n (s) represented towel Zhen size of the problem s., PG I IF; G_v) represents a speech recognition result by the N-gram language model Gn obtained Ze F acoustic score, u represents a phoneme sequence, L (Gp) a set of possible phoneme sequence representing the phoneme network Gp received, P (s | u) represents the phoneme sequence U acoustic score. 这里,澤^尤选地是语音识别结果中的前N个最好识别结果中的第一个最好识别结果,当然,其也可以是通过N元语言模型得到的语音识别结果中的其他识别结果。 Here, in particular optionally Ze ^ is the best recognition result first top N best recognition result of the speech recognition result is, of course, other identification which may be the speech recognition result obtained by the N-gram language model in result.

[0131] 其中,帧大小n(s)、语音识别结果矿、语音识别结果p的声学得分/3Oi IG y )、以及音素序列U的声学得分p (s Iu)例如可以由上述的语音识别单元通过上面描述的语音识别步骤而得到。 [0131] wherein, frame size n (s), the speech recognition result ore, the speech recognition result p is the acoustic score / 3Oi IG y), and a phoneme sequence U acoustic score p (s Iu), for example by the speech recognition unit obtained by the above described voice recognition step.

[0132] 由此,分类单元104还可以基于所述置信度计算单元205计算出的用户所说出的问题的置信度而将该问题分类为在存储单元内的问题或者在存储单元外的问题。 And the problem is classified as a problem in the storage unit [0132] Accordingly, the classification unit 104 may also be based on the confidence calculation unit 205 calculates the user's spoken confidence or the outside of the memory cell in question .

[0133] 例如,如果用户所问的一个问题的置信度小于等于置信度阈值(例如,0. 3 ),则可以直接将该问题分类到在存储单元外的问题,即,可以不必执行例如后续的语义相似度的计算操作。 [0133] For example, if the confidence of a problem the user asked less confidence threshold (e.g., 0.3), can be directly question classification The problem in the external memory unit, i.e., may not necessarily be performed, for example, subsequent semantic similarity computing operations.

[0134] 这里,置信度阈值的取值范围可以为0〜1。 [0134] Here, the range of the confidence threshold can be 0~1.

[0135] 另外,优选地,根据本实施例的基于语音的问答系统200还可以包括输出单元206。 [0135] In addition, preferably, a speech-based question answering system of the present embodiment may further include an output 200 in accordance with section 206. 如果用户所说出的问题被分类为在存储单元内的问题,则该输出单元被配置为能够输出与用户所说出的问题相应的在所述问答数据存储单元中所存储的问题对应的答案。 If the user issues spoken is classified as a problem in the storage unit, the output unit can be configured to output a user issues spoken quiz questions corresponding to the data stored in the storage unit corresponding to answers .

[0136] 另外,优选地,根据本实施例的基于语音的问答系统200还可以包括更新单元207。 [0136] In addition, preferably, a speech-based system of the present embodiment according to Q 200 may further include an update unit 207. 如果用户所说出的问题被分类为在存储单元外的问题,则该更新单元207被配置为基于所述在存储单元外的问题而更新所述问答数据存储单元中存储的问题和/或更新所述语目模型。 If the user issues spoken question is classified as an outer memory unit, the updating unit 207 is configured based on the problem in an external storage unit updating the quiz question in the data storage unit and / or update the language model mesh.

[0137] S卩,在本发明中,对于在存储单元内的问题(S卩,问答数据存储单元中已经记录了的问题),可以通过输出单元206向用户提供所需的输出,例如该问题的相应答案。 [0137] S Jie, in the present invention, the problems (S Jie, Q data storing means has been recorded problems) in the storage unit, can provide the desired output by the output unit 206 to the user, for example, the problem the corresponding answers.

[0138] 对于不在存储单元内的问题,即问答数据存储单元中未记录的问题,可以将该问题添加到系统中,以对系统中存储的数据进行扩展,从而使得能够更多地满足用户需求。 [0138] For the problem is not in the storage unit, i.e. the data storage unit quiz question is not recorded, the problem can be added to the system to data stored in the system be extended, thereby enabling to meet user needs more .

[0139] 具体而言,例如,如果用户说出的问题被分类为在存储单元外的问题,则可以将用户说出的问题通过更新单元而添加到问答数据存储单元101中,以便扩展问答数据存储单元101中所存储的数据。 [0139] Specifically, for example, if the user speaks the question is classified as a problem in an external storage unit, the user may be added to the spoken quiz question data storage unit 101 through the updating means so as to expand the data Q a data storage unit 101 is stored. 当然,新添加的问题的答案也可以相应地添加到问答数据存储单元101中,以使得问答数据存储单元101中所存储的数据更有效。 Of course, the answer to add new questions may be added to the corresponding data storage unit 101 Q, Q data so that the data stored in the storage unit 101 is more effective.

[0140] 另外,上面提到的语言模型的更新可以通过本领域普通技术人员能够想到的任何方法来实现,本发明对此不作任何限制。 Any method [0140] Furthermore, update the language model mentioned above by those of ordinary skill in the art can think of is achieved, according to the present invention does not set any limit. 也就是说,任何能够实现语言模型的更新的方法都可以应用于本发明。 That is, any method can be implemented to update the language model can be applied to the present invention.

[0141] 优选地,根据本实施例的基于语音的问答系统200还可以包括文字相似度计算单元208。 [0141] Preferably, speech-based question answering system of the present embodiment may further include text 200 according to the similarity calculating unit 208. 文字相似度计算单元208被配置为,根据语音单元102对于用户说出的问题的识别结果,计算用户所说出的问题与所述问答数据存储单元中的每个问题之间的文字相似度。 Text similarity calculation unit 208 is configured to, according to the degree of similarity between the character recognition result for the speech unit 102 issues spoken by the user, the user is calculated with each of the spoken question to the quiz question data storage unit. 这里,所述文字相似度用于表示用户说出的问题与所述问答数据存储单元中存储的每个问题在文字上的相似程度。 Here, the similarity of the text spoken by the user for indicating the degree of similarity with the question of the quiz question data each storage unit in the text.

[0142] 由此,所述分类单元还能够基于所述文字相似度计算单元208所计算出的所述文字相似度,将用户所说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 [0142] Accordingly, the classification unit can also calculate the similarity of text similarity unit 208 calculated based on the text, the user issues spoken classified as a problem in the storage unit or a storage unit outside the problem.

[0143] 例如,对于置信度、文字相似度、以及语义相似度,可以按照以下顺序来执行判断: 首先,可以判断用户所说出的问题的置信度是否大于所述置信度阈值。 [0143] For example, the confidence, text similarity, and the semantic similarity determination may be performed in the following order: First, the user can estimate the confidence level of the spoken question is greater than the confidence threshold. 如果不是,则可以直接将该问题分类为在存储单元外的问题;如果是,则继续判断用户所说出的问题的文字相似度是否大于文字相似度阈值。 If not, then the problem can be directly classified as a problem in an external storage unit; if yes, issue the user continues to judge the spoken text character similarity is greater than a threshold similarity. 如果是,则可以直接将该问题分类为在存储单元内的问题; 如果用户所说出的问题的文字相似度不大于所述文字相似度阈值,则可以根据语义相似度来对用户所说出的问题进行分类。 If so, the problem may be directly classified as a problem in the storage unit; text if the user issues spoken text similarity is not greater than the similarity threshold, then the user can be spoken in accordance with the semantic similarity the problem classification.

[0144] 根据语义相似度来进行分类的方式例如可以如下:判断用户所问问题的语义相似度是否大于语义相似度阈值,如果用户所说出的问题的语义相似度不大于所述语义相似度阈值,则将该问题分类为在存储单元外的问题,否则,将该问题分类为在存储单元内的问题。 [0144] The approach to classifying semantic similarity example, as follows: determining whether the user is asked questions semantic similarity semantic similarity is greater than a threshold value, if the semantic similarity of user issues spoken semantic similarity is not larger than the threshold, then the problem in question is classified as an external storage unit, otherwise, the problem is classified as a problem in the storage unit.

[0145] 这里的语义相似度可以是通过一种方法得到的一种语义相似度,也可以是通过两种或多种方法得到的两种或多种语义相似度的组合(可以称其为组合的语义相似度)。 [0145] A semantic similarity semantic similarity herein may be obtained by a process, it may be obtained by two or more methods in combination of two or more semantic similarity (which may be referred to as a combination of semantic similarity). 至少一种语义相似度组合的方式例如可以为求平均、加权、加权平均、平方根、以及能够想到的任何其他的形式。 Any other form of embodiment at least one semantic similarity may be combined, for example, averaging, weighted, the weighted average, square root, and can be contemplated.

[0146] 为了更详细地描述本发明,图7示例性地示出根据本发明的实施例的另一用于交互式语音系统的方法的流程图。 [0146] For the present invention will be described in more detail, FIG. 7 shows an example flowchart of a method for interactive voice system according to another embodiment of the present invention.

[0147] 如图7所示,在步骤S201,对于用户所说出的问题进行语音识别。 [0147] As shown in FIG. 7, in step S201, the problem for the user's spoken voice recognition. 该步骤与图3中的步骤SlOl类似,在此省略对其的详细描述。 This step is the step in FIG. 3 SlOl Similarly, in detail description thereof is omitted.

[0148] 在步骤S202,基于上述语音识别步骤S201得到的结果,计算用户所说出的问题的置信度。 [0148] In step S202, based on the speech recognition result obtained in step S201, calculating the spoken user confidence problem. 这里,该置信度的计算方法可以采用上面所描述的示例性的方法。 Here, the calculation method may confidence exemplary method described above employed.

[0149] 在步骤S203,判断所计算的置信度是否大于置信度阈值。 [0149] In step S203, it determines whether the calculated confidence level is greater than the confidence threshold. 如果不是,则在分类步骤S208,将用户的该问题分类为在存储单元外的问题(S2082)。 If not, then the classification step S208, the user of the problem in question is classified as an external storage unit (S2082). 如果是,则进行到步骤S204。 If so, then proceeds to step S204.

[0150] 在步骤S204,对于用户所问的问题和问答数据存储单元内存储的问题进行文字相似度的计算。 [0150] In step S204, the user issues to the questions asked and Q data storing means for storing the text similarity calculation. 为了便于理解,后面将给出文字相似度的计算方法的例子。 For ease of understanding, the example will be given later in the text similarity calculation method. 然后进行到步骤S205。 Then proceeds to step S205.

[0151] 在步骤S205,判断计算出的文字相似度是否大于文字相似度阈值。 [0151] In step S205, it determines whether the calculated similarity is larger than the literal text similarity threshold. 如果计算出的文字相似度大于文字相似度阈值,则在分类步骤S208将用户所问的问题分类为在存储单元内的问题(S2081)。 If the calculated similarity is larger than the text character similarity threshold value, then at step S208 the classification questions asked by the user is classified as a problem in the storage unit (S2081). 否则,进行到步骤S206。 Otherwise, proceed to step S206.

[0152] 在步骤S206,对于用户所问的问题和问答数据存储单元内存储的问题进行语义相似度的计算。 [0152] In step S206, the user issues and questions asked in the quiz data stored in the storage unit to calculate semantic similarity. 这里,语义相似度的计算方法可以采用上面提到的方法。 Here, the semantic similarity calculation method mentioned above can be employed. 然后进行到步骤S207。 Then proceeds to step S207.

[0153] 在步骤S207,判断计算出的语义相似度是否大于语义相似度阈值。 [0153] In step S207, it determines whether the calculated similarity is greater than a semantic semantic similarity threshold. 如果计算出的语义相似度大于语义相似度阈值,则在分类步骤S208将用户所问的问题分类为在存储单元内的问题(S2081)。 If the calculated similarity is greater than semantic semantic similarity threshold value, then at step S208 the classification questions asked by the user is classified as a problem in the storage unit (S2081). 否则,在分类步骤S208将用户所问的问题分类为在存储单元外的问题(S2082)。 Otherwise, in step S208 the user classification questions asked in question is classified as an external storage unit (S2082).

[0154] 在步骤S2081将用户所问的问题分类为在存储单元内的问题之后,进行到步骤S209。 [0154] After the step S2081 questions asked by the user is classified as a problem in the storage unit, proceeds to step S209.

[0155] 在步骤S209,向用户输出用户所问的问题(属于在存储单元内的问题)在存储单元内存储的相应的答案。 [0155] In step S209, the user output to a user the question asked (question belongs in the storage unit) corresponding answer stored in the storage unit.

[0156] 另一方面,在步骤S2082将用户所问的问题分类为在存储单元外的问题之后,进行到步骤S210。 [0156] On the other hand, in step S2082 after the questions asked by the user is classified as a problem in an external storage unit proceeds to step S210.

[0157] 在步骤S210,针对所述在存储单元外的问题,对于问答数据存储单元内存储的数据进行扩展,即,可以将所述在存储单元外的问题添加到问答数据存储单元内,并且,也可以相应地将所述在存储单元外的问题的答案也添加到问答数据存储单元内,以扩展该交互式语音系统的应对能力。 [0157] In step S210, for the problems in the external memory unit, the data storage unit stores the data Q to be extended, i.e., the problem may be added in the external memory unit into the data storage unit Q, and may accordingly the problem outside of the memory cell is also added to the quiz answer data storage unit, to extend the ability to respond to the interactive voice system.

[0158] 下面将参照图8和图9来描述文字相似度的计算的例子。 [0158] FIG. 9 will be described with examples and text similarity is calculated with reference to FIG 8.

[0159] 图8示例性地示出文字相似度计算单元208的框图。 [0159] Figure 8 illustrates a block diagram schematically showing a character similarity calculation unit 208.

[0160] 文字相似度计算单元208包括:第三词袋产生部件2081,被配置为,根据语音识别单元102对于用户所说出的问题的识别结果,产生第三词袋,其中所述第三词袋包括所述语音识别单元对于用户所说出的问题的识别结果中所含的词;第四词袋产生部件2083,被配置为针对所述问答数据存储单元中存储的每个问题产生相应的第四词袋,其中每个第四词袋包括所述问答数据存储单元中所存储的相应的问题所含的词;第二词选择部件2082, 被配置为根据语音识别单元对于用户所说出的问题的识别结果,从所述第三词袋选择其概率大于第二阈值的词,其中所述概率能够由通过N元语言模型得到的声学得分来确定;以及文字相似度计算部件2084,被配置为根据所述第二词选择部件所选择的第三词袋中的词和所述第四词袋中的词,计算用户所说出的问题与所述问答数据 [0160] The text similarity calculating unit 208 includes: a third bag of words generating means 2081 configured to, according to the recognition result for the speech recognition unit 102 issues a user spoken generates a third bag of words, wherein the third bag of words comprising the word recognition result to the voice recognition unit user issues spoken contained; the fourth bag of words generating member 2083, configured to generate, for each respective question of the quiz data stored in the storage unit the fourth bag of words, wherein each word comprises a fourth bag of words corresponding to the quiz question data stored in the storage unit contained; second word selection member 2082 configured according to the user of said voice recognition unit a recognition result of the problems, a third bag of words selected from the words whose probability is greater than a second threshold, wherein the probability score can be determined acoustically from the N-gram language model obtained; and text similarity calculating section 2084, words words words is arranged a third member of the selected word and the fourth bag pouch is selected in accordance with the second word, the spoken user of the computing problems of the Q data 储单元中存储的每个问题之间的文字相似度。 Text between each question stored in the storage unit similarity.

[0161] 优选地,文字相似度计算单元208还包括:第二过滤部件2085,被配置为根据词性,从所述第二词选择部件2082自所述第三词袋所选择的词中过滤掉一部分词,和/或根据词性,从所述第二词袋中过滤掉一部分词。 [0161] Preferably, text similarity calculation unit 208 further comprises: a second filter member 2085 configured according to part of speech, from the third selection means 2082 from the bag the second word is a word in the selected word filtered out word part, and / or speech in accordance with, the second word from the filter bag away a portion of the word. 其中,所述文字相似度计算部件2084能够在第二过滤部件2085进行上述过滤之后再进行文字相似度的计算。 Wherein, the text can be re-similarity calculating section 2084 calculates the similarity of the text after the second filter 2085 for the filtering member.

[0162] 其中,对于图8中的部件,第三词袋产生部件2081与图4中的第一词袋产生部件1031、第二词选择部件2082与图4中的第一词选择部件1032、第四词袋产生部件2083与图4中的第二词袋产生部件1034、第二过滤部件2085与图4中的第一过滤部件1037具有类似的功能。 [0162] wherein, for the member in FIG. 8, a third word generating section 2081 generates the bag member 1031 in FIG. 4 of the first bag of words, selecting a second word of the first word in the selection member 2082 and member 1032 in FIG. 4, the fourth bag of words generating member 2083 and a second bag of words in FIG. 4 generating member 1034, a second filter member 2085 and the first filter member of FIG. 41 037 have similar functions.

[0163] 这里,文字相似度计算部件2084进行文字相似度的计算可以利用向量空间模型(VSM)。 [0163] Here, text similarity calculating section 2084 calculates the similarity may be utilized for text vector space model (VSM). 例如,可以采用与上面在描述式1时提到的向量空间模型相同的向量空间模型。 For example, in the above Formula 1 the vector space model described referring to the same vector space model may be employed. 当然,也可以使用不同的向量空间模型。 Of course, you can also use a different vector space model.

[0164] 下面将结合图9来简要描述文字相似度的计算方法。 [0164] The following text in conjunction with a similarity calculation method in FIG. 9 be briefly described.

[0165] 图9示例性地示出了文字相似度的计算方法的流程图。 [0165] FIG. 9 exemplarily shows a flowchart of a method of calculating the similarity of text.

[0166] 在第三词袋产生步骤S2081,根据语音识别步骤对于用户所说出的问题的识别结果,产生第三词袋。 [0166] Step S2081 is generated in the third bag of words, the speech recognition result to the recognition step the spoken question to the user, generating a third bag of words. 其中所述第三词袋包括在语音识别时得到的对于用户所说出的问题的识别结果中所含的词。 Wherein said bag comprises a third word for word recognition result of the user issues spoken obtained when contained in the speech recognition. 该步骤与图5中的步骤S1021类似。 This step and step S1021 is similar to FIG. 5. 然后,步骤S2081进行到步骤S2082。 Then, step S2081 proceeds to step S2082.

[0167] 另一方面,在第四词袋产生步骤S2083,为所述问答数据存储单元中存储的每个问题产生相应的第四词袋,其中每个第四词袋包括所述问答数据存储单元中所存储的相应的问题所含的词。 [0167] On the other hand, is generated in step S2083 in the fourth bag of words, to produce a fourth bag of words corresponding to the question and answer each question stored in the data storage unit, wherein each said bag comprises fourth word data storage Q words corresponding unit stored in question contained. 该步骤与图5中的步骤S1024类似。 Step 5 is similar to the step S1024. 然后,步骤S2083进行到步骤S2085。 Then, step S2083 proceeds to step S2085.

[0168] 在第二词选择步骤S2082,根据从所述语音识别步骤得到的对于用户所说出的问题的识别结果,从所述第三词袋选择其概率大于第三阈值的词,其中所述概率能够由通过N 元语言模型得到的声学得分来确定。 [0168] In a second word selection step S2082, based on the identification results for the user issues spoken obtained from said speech recognition step of selecting the probability is greater than the third threshold value words from the third bag of words, wherein said probability score can be determined by the N-gram language model acoustically obtained. 该步骤与图5中的步骤S1022类似。 This step and step S1022 is similar to FIG. 5. 然后,步骤S2082 也进行到步骤S2085。 Then, step S2082 also proceeds to step S2085.

[0169] 在文字相似度计算子步骤S2085,根据在所述第二词选择步骤S2082中所选择的第三词袋中的词和在第四词袋产生步骤S2083得到的所述第四词袋中的词,计算用户所说出的问题与所述问答数据存储单元中存储的每个问题之间的文字相似度。 [0169] The text similarity calculating sub-step S2085, according to the word in the third word of the second word selecting step S2082 the selected bag and the bag is generated in step S2083 to obtain the fourth word of the fourth word in the bag in words, the words spoken between the user of the computing problems and issues each of said data storage unit Q similarity.

[0170] 优选地,所述文字相似度计算步骤还包括:第二过滤步骤S2084,其根据词性,从所述第二词选择步骤S2082中所选择的第三词袋的词中过滤掉一部分词,和/或根据词性, 从所述第四词袋中过滤掉一部分词。 [0170] Preferably, the character similarity calculating step further comprises: a second filtering step of S2084, based on part of speech, a third bag of words selected in step S2082 in the selected word from said second filter out a portion of the word word and / or in accordance with speech, word from the fourth bag filter away a portion of the word. 该步骤S2084与图5中的步骤S1027类似。 This step S2084 in FIG. 5 is similar to S1027.

[0171] 由此,根据本发明,能够在进行了第二过滤步骤S2084中的过滤之后,进行所述文字相似度计算子步骤S2085中的文字相似度的计算,以便简化计算。 After [0171] Thus, according to the present invention, performing the filtration in the second filtration step S2084, the character count for the sub-step S2085 text similarity similarity calculation, in order to simplify calculations.

[0172] 需要注意的是,在本发明中,实施例中所描述的步骤顺序仅是示例性的,实际上本发明并不局限于所描述的步骤顺序,而是本领域技术人员能够想到的任何能够实现本发明的效果的步骤的顺序,都是可以的。 [0172] Note that, in the present invention, the steps described in the embodiments are merely exemplary of the sequence, in fact, the present invention is not limited to the steps described order, but those skilled in the art can occur any order of the steps to achieve the effect of the present invention, are possible.

[0173] 通过将置信度、文字相似度和语义相似度相结合,根据本实施例的基于语音的问答系统和用于交互式语音系统的方法与第一和第二实施例描述的系统和方法相比能够获得更有利的效果,具体地,除了识别精度高、覆盖范围广之外,根据本实施例的系统和方法能够更高效地进行处理,并且系统的应对能力更强。 [0173] system and method by confidence, text similarity, and the combination of semantic similarity, according to speech-based question answering system for interactive voice system and method of the present embodiment described in the first embodiment and the second embodiment compared more advantageous effect can be obtained, in particular, in addition to the identification of high precision, covering a wide range, can be treated more efficiently in accordance with the method and system of the present embodiment, and greater capacity of the system to respond. 而且,由于能够方便地输出数据和更新数据,所以其也更加灵活。 Further, since it is possible to easily update the data and output data, so the more flexible.

[0174] 这里,需要指出的是,置信度、文字相似度和语义相似度结合的方式可以是任意的,即,它们可以采用两两结合的方式(即,置信度与语义相似度相结合,置信度与文字相似度相结合,语义相似度与文字相似度相结合),也可以采用将这三个都结合起来的方式(即, 置信度、文字相似度和语义相似度这三个相结合)。 [0174] Here, it should be noted that the confidence level, text similarity, and a combination of semantic similarity may be arbitrary, i.e., they may be paired together in the manner employed (i.e., the confidence and semantic similarity combination, confidence and text similarity combined semantic similarity and the similarity of text in combination) can also be used to combine all these three ways (i.e., the confidence, text similarity, and the combination of these three semantic similarity ). 另外,它们结合的顺序也不受任何限制。 Further, the order thereof is not subject to any bound. 总之,在本发明的一些实施例中,用户所问问题的置信度、文字相似度和语义相似度可以任意组合,以判断该问题是否在存储单元内等等。 In summary, in some embodiments of the present invention, a user confidence to ask questions, and text similarity semantic similarity can be any combination, to determine whether the problem is in the storage unit and the like.

[0175] [本发明与现有技术的效果对比] [0175] [Comparative effects of the present invention and the prior art]

[0176] 以下,为了示例本发明的技术效果,将参照图10来描述本发明和现有技术的比较结果。 [0176] Here, for exemplary technical effect of the present invention will be described with reference to 10 to compare the results of the present invention and the prior art of FIG.

[0177] 在此将要进行比较的现有技术的方法为单纯地采用文字相似度来进行判断的方法,例如由S. Takeuchi, T. Cincarek, H. Kawanami, H. Saruwatari 和K. Shikano 发表的论文"Question and Answer Database Optimization Using Speech Recognition Results,' (proceeding of:INTERSPEECH2008,9th Annual Conference of the International Speech Communication Association, Brisbane, Australia, September22_26, 2008)。艮P, 根据用户所说出的问题与存储单元中所存储的问题之间的文字相似度,来将该问题分类为在存储单元内的问题还是在存储单元外的问题,并由此来决定是否存在与该问题相对应的答案和/或是否能够输出该相对应的答案。 Method [0177] In this prior art will be compared to simply using character similarity determination method is performed, for example, published by S. Takeuchi, T. Cincarek, H. Kawanami, H. Saruwatari and K. Shikano of papers "question and Answer Database Optimization Using Speech Recognition Results, '(proceeding of: INTERSPEECH2008,9th Annual Conference of the International Speech Communication Association, Brisbane, Australia, September22_26, 2008) Gen P, the user issues spoken and stored. problems between the text stored in the unit similarity to this problem is classified as a problem in the storage unit or an external storage unit in question, and thus to determine whether there is an answer corresponding to the question and / or whether the output of the corresponding answer.

[0178] 与该现有技术相比较的本发明的方法是采用第一实施例所示的仅利用语义相似度来进行判断的方法,没有组合置信度和/或文字相似度。 [0178] The method of the present invention compared with the prior art method is performed using only the semantic similarity determining a first embodiment illustrated embodiment, there is no combination of confidence and / or text similarity. 当然,如果组合了置信度和/或文字相似度,效果应该会更好。 Of course, if the combination of confidence and / or text similarity, the effect should be better.

[0179] 由图10可以看出,与现有技术相比,本发明明显具有更高的识别精度和更高的召回率(recalI rate)。 [0179] As can be seen from Figure 10, compared to the prior art, the present invention has a significantly higher and higher recognition accuracy of recall (recalI rate). 这里,召回率为能够被正确地分类为在存储单元内的问题的用户所问问题的数目与真正应该属于在存储单元内的问题的用户所问问题的总数的比值。 Here, the number recall rate can be correctly classified as a problem in the storage unit of the user to ask questions of the ratio of the total number of the user in question should belong to the real memory cell questions asked.

[0180] 在本文中,为了简要起见,仅对比了识别精度和召回率这两个参数,本领域普通技术人员会明白,实际上,本发明的有利技术效果不是仅由这两个参数体现的。 [0180] Herein, for brevity, only compare the recognition precision and recall these two parameters, one of ordinary skill in the art will understand that, in fact, advantageous technical effects of the present invention are not embodied only by these two parameters .

[0181] 尽管已经参照附图描述了本发明,但是本发明的实施例不限于上面描述的那些, 本领域普通技术人员能够理解,其它的实施方式也可适用于本发明。 [0181] While the present invention has been described with reference to the drawings, embodiments of the present invention is not limited to those, of ordinary skill in the art to understand the above description, other embodiments are also applicable to the present invention. 另外,请注意,在实施例中示出的部件仅是例子,它们不限于那样的形式,而可以是单独的部件或者可以相互组合。 Also, note that, in the embodiment illustrated are only examples member, they are not limited to such a form, but may be separate components or may be combined with each other.

[0182] 此外,请注意,可以以许多方式实施本发明的基于语音的问答系统和用于交互式语音系统的方法。 [0182] In addition, note that embodiments may be voice-based question answering system for interactive voice system and method of the present invention in many ways. 例如,可以通过软件、硬件、固件、或其任何组合来实施本发明的基于语音的问答系统和用于交互式语音系统的方法。 For example, speech may be implemented based question answering system for interactive voice system and method of the present invention by software, hardware, firmware, or any combination thereof. 上述的方法步骤的次序仅是示例性的,本发明的方法步骤不限于以上具体描述的次序,除非以其他方式明确说明。 Sequence of method steps described above are merely exemplary, and the method steps of the present invention is not limited to the specific order described above, unless expressly described otherwise. 此外,在一些实施例中,本发明还可以被实施为记录在记录介质中的程序,其包括用于实现根据本发明的方法的机器可读指令。 Further, in some embodiments, the present invention also may be implemented as a program recorded in a recording medium, comprising machine readable instructions for implementing the method according to the present invention. 因而,本发明还覆盖存储用于实现根据本发明的方法的程序的记录介质。 Accordingly, the present invention also covers a recording medium storing a program according to implement the method of the present invention.

[0183] 虽然已通过示例详细展示了本发明的一些具体实施例,但是本领域普通技术人员应当理解,上述例子仅意图是示例性的而非限制本发明的范围。 [0183] Although by way of example of the present invention are shown in detail some specific embodiments, those skilled in the art will appreciate, the above examples are only intended to be exemplary and not limiting the scope of the present invention. 本领域普通技术人员应当理解,上述实施例可以被修改而不脱离本发明的范围和实质。 One of ordinary skill in the art should appreciate that the above embodiments may be modified without departing from the scope and spirit of the invention. 本发明的范围是通过所附的权利要求来限定的。 The scope of the present invention is achieved by the appended claims is defined.

Claims (24)

  1. 1. 一种基于语音的问答系统,包括: 问答数据存储单元,在该问答数据存储单元中相关联地存储问题以及对应于所述问题的答案; 语音识别单元,通过使用语言模型对于用户说出的问题进行语音识别; 语义相似度计算单元,根据语音识别单元对于用户说出的问题的识别结果,计算用户说出的问题与所述问答数据存储单元中存储的每个问题之间的语义相似度,其中,所述语义相似度用于表示用户说出的问题与所述问答数据存储单元中存储的每个问题所表达的意思上的相似程度;以及分类单元,基于所述语义相似度计算单元所计算出的所述语义相似度,将用户说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 A voice-based question answering system, comprising: a data storage unit Q, the problems associated with the stored question and answer storage means and data corresponding to the answer to the question; voice recognition unit, by using the language model for the user utters speech recognition problem; semantic similarity calculating unit according to the result of the speech recognition unit recognizes the user issues spoken calculated between the semantic question spoken by the user with the question and answer each question stored in the data storage unit is similar to degrees, wherein the degree of semantic similarity used to indicate similar meaning expressed on the user issues spoken questions and each of said data storage unit Q; and a classification unit, based on the semantic similarity calculation means the calculated semantic similarity, the user issues spoken classified as a problem in the storage unit or an external storage unit in question.
  2. 2. 根据权利要求1的问答系统,还包括: 置信度计算单元,基于语音识别单元输出的有关用户所说出的问题的识别结果,为用户所说出的问题计算置信度; 其中,所述分类单元还能够基于所述置信度计算单元计算出的置信度,将用户所说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 2. The question answering system of claim 1, further comprising: confidence calculation unit, based on the recognition result of the speech recognition unit outputs a user about the problem spoken, calculates a confidence of the user issues spoken; wherein said cell classification unit further calculates the degree of confidence can be calculated based on the confidence level, the user issues spoken classified as a problem in the storage unit or an external storage unit in question.
  3. 3. 根据权利要求1的问答系统,还包括: 输出单元,如果用户所说出的问题被分类为在存储单元内的问题,则该输出单元能够输出与用户所说出的问题相应的在所述问答数据存储单元中所存储的问题对应的答案。 Answering system according to claim 1, further comprising: an output unit, if the user issues spoken is classified as a problem in the storage unit, the output unit can output the user issues spoken in the respective said quiz question data stored in the storage unit corresponding to the answer.
  4. 4. 根据权利要求1的问答系统,还包括: 更新单元,如果用户所说出的问题被分类为在存储单元外的问题,则该更新单元基于所述在存储单元外的问题更新所述问答数据存储单元中存储的问题和/或更新所述语言模型。 Answering system according to claim 1, further comprising: an updating unit, if the user issues spoken question is classified as an outer memory unit, the update unit based on the problem of updating the external memory unit Q issue data storage unit and / or update the language model.
  5. 5. 根据权利要求2的问答系统,其中所述置信度是通过将用户所说出的问题经由N元语言模型进行识别而得到的声学得分除以用户所说出的该问题经由音素网络进行识别而得到的声学得分来确定的。 The phoneme network identification via answering system according to claim 2, wherein the confidence is spoken by the user via the problems identified N-gram language model and the acoustic score obtained by dividing the user issues spoken the obtained acoustic score determined.
  6. 6. 根据权利要求1的问答系统,其中,所述语义相似度计算单元包括: 第一词袋产生部件,根据语音识别单元对于用户所说出的问题的识别结果,产生第一词袋,其中所述第一词袋包括该识别结果中所含的词; 第一词选择部件,根据语音识别单元对于用户所说出的问题的识别结果,从所述第一词袋选择其概率大于第一阈值的词,其中所述概率能够由通过N元语言模型得到的声学得分来确定; 第一检索部件,根据所述第一词选择部件所选择的词,从数据源检索文档; 第二词袋产生部件,为所述问答数据存储单元中存储的每个问题产生相应的第二词袋,其中每个第二词袋包括所述问答数据存储单元中所存储的相应的问题所含的词; 第二检索部件,基于每个第二词袋中的词,从数据源检索文档;以及语义相似度计算部件,基于第一检索部件所检索到的文 6. The question answering system of claim 1, wherein said semantic similarity calculating unit comprises: a first member generating bag of words, the speech recognition result to the recognition unit as the user issues spoken generates a first bag of words, wherein the first word comprises a bag of words contained in the recognition result; a first word selection means, voice recognition unit according to the recognition result for the user issues spoken, the probability is greater than a first selection from the first bag of words word threshold value, wherein the probability can be determined by the acoustic score obtained by the N-gram language model; first retrieval means, the word selected components selected from the first word to retrieve the document from the data source; a second bag of words generation means generates a second bag of words corresponding to each of the quiz question data storage unit, wherein each word comprises a second bag of words corresponding issue data stored in the storage unit included in the Q & a; second search means, each second word based on the word bag, retrieved from the data source document; and semantic similarity calculating section based on the first retrieval member retrieved text 档和第二检索部件所检索到的文档,计算用户所说出的问题与所述问答数据存储单元中存储的每个问题之间的语义相似度。 A second retrieval member gear and the retrieved documents, calculating semantic similarity between the user and the spoken questions to the question and answer each question stored in the data storage unit.
  7. 7. 根据权利要求6的问答系统,其中,所述语义相似度计算单元还包括: 第一过滤部件,根据词性,从所述第一词选择部件自所述第一词袋所选择的词中过滤掉一部分词,和/或根据词性,从所述第二词袋中过滤掉一部分词; 其中,所述第一检索部件针对过滤后的第一词袋进行检索,和/或第二检索部件针对过滤后的第二词袋进行检索。 A first filter member, according to part of speech, the word selecting means from the first bag of words selected from said first word: 7. The question answering system of claim 6, wherein, the semantic similarity calculating unit further comprises word part was filtered off, and / or speech in accordance with, the second word from the filter bag away a portion of the word; wherein said first search means to search, and / or the second member for retrieving a first word after the bag filters to search for a second term after the bag filter.
  8. 8. 根据权利要求7的问答系统,其中,所述语义相似度计算单元还包括: 领域词添加部件,向过滤后的第一词袋和/或过滤后的第二词袋添加与领域有关的词; 其中,所述第一检索部件针对进行上述添加后的第一词袋进行检索,和/或第二检索部件针对进行上述添加后的第二词袋进行检索。 Answering system according to claim 7, wherein the semantic similarity calculating unit further comprises: adding the word field member, a second bag of words to the first word after the bag filter after filtration and / or addition of the art related to word; wherein said first search means to search, and / or the second member for retrieving a first word of the bag after adding the above-mentioned retrieval bags for a second term after the above addition.
  9. 9. 根据权利要求6所述的问答系统,其中,所述语义相似度计算单元基于以下来进行语义相似度的计算: 第一检索部件所检索到的文档和第二检索部件所检索到的文档中的每个词;和/或第一检索部件所检索到的文档和第二检索部件所检索到的文档之间的重合。 9. The question answering system of claim 6, wherein the means to the semantic similarity is calculated based on semantic similarity calculation: a first retrieval means and the documents retrieved by the second retrieval member retrieved documents each word; coincidence between and / or the first retrieval means and the documents retrieved by the second retrieving means to retrieve the document.
  10. 10. 根据权利要求1〜9中的任一项的问答系统,还包括: 文字相似度计算单元,根据语音识别单元对于用户说出的问题的识别结果,计算用户所说出的问题与所述问答数据存储单元中的每个问题之间的文字相似度,其中,所述文字相似度用于表示用户说出的问题与所述问答数据存储单元中存储的每个问题在文字上的相似程度; 其中,所述分类单元还能够基于所述文字相似度计算单元所计算出的所述文字相似度,将用户所说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 10. The question answering system of any one of claims 1~9, further comprising: a text similarity calculation unit, according to the recognition result to the voice recognition unit issues a user speaks, the user of the computing problems of the spoken text between each quiz question data storage unit similarity, wherein the similarity of the text spoken by the user for indicating the degree of similarity with the question of the quiz question data each storage unit in the text ; wherein the character, the similarity calculation can be further classification unit based on the calculated word unit similarity, the user issues spoken classified as a problem in the storage unit or an external storage unit in question.
  11. 11. 根据权利要求10的问答系统,其中,所述文字相似度计算单元包括: 第三词袋产生部件,根据语音识别单元对于用户所说出的问题的识别结果,产生第三词袋,其中所述第三词袋包括所述识别结果中所含的词; 第四词袋产生部件,为所述问答数据存储单元中存储的每个问题产生相应的第四词袋,其中每个第四词袋包括所述问答数据存储单元中所存储的相应的问题所含的词; 第二词选择部件,根据语音识别单元对于用户所说出的问题的识别结果,从所述第三词袋选择其概率大于第二阈值的词,其中所述概率能够由通过N元语言模型得到的声学得分来确定;以及文字相似度计算部件,根据所述第二词选择部件所选择的第三词袋中的词和所述第四词袋中的词,计算用户所说出的问题与所述问答数据存储单元中存储的每个问题之间的文字相似度。 Q 11. The system of claim 10, wherein said character similarity calculating unit comprises: means generating a third bag of words, the speech recognition unit recognition results for the user issues spoken of, generating a third bag of words, wherein the third word comprising a bag of words contained in the recognition result; fourth bag of words generating means generates a fourth bag of words corresponding to the question and answer each question stored in the data storage unit, wherein each of the fourth bag of words including words corresponding issue data stored in the storage unit contained in the question and answer; second word selection means, voice recognition unit according to the recognition result for the user issues spoken of, the third bag of words selected from word is greater than a second threshold probability value, wherein the probability score can be determined acoustically from the N-gram language model obtained; and text similarity calculating means selects the third word bag member according to the second selected word word and the fourth word of the bag, the word uttered by the user between computing problems and issues each of said data storage unit Q similarity.
  12. 12. 根据权利要求11的问答系统,其中,所述文字相似度计算单元还包括: 第二过滤部件,根据词性,从所述第二词选择部件自所述第三词袋所选择的词中过滤掉一部分词,和/或根据词性,从所述第四词袋中过滤掉一部分词; 其中,在第二过滤部件进行上述过滤后,所述文字相似度计算部件进行文字相似度的计算。 12. The question answering system of claim 11, wherein said character similarity calculating unit further comprises: a second filter element words, according to part of speech, selecting from the third member selected from the bag of words in the second word word part was filtered off, and / or in accordance with speech, word from the fourth bag filter away a portion of the word; wherein, after the second filter means for the filtering, the character text similarity calculating section calculates the degree of similarity.
  13. 13. -种用于交互式语音系统的方法,该交互式语音系统包括语音识别单元和问答数据存储单元,在所述问答数据存储单元中相关联地存储了问题以及对应于所述问题的答案,所述方法包括: 语音识别步骤,通过语音识别单元使用语言模型对于用户说出的问题进行语音识别; 语义相似度计算步骤,根据语音识别单元对于用户说出的问题的识别结果,计算用户说出的问题与所述问答数据存储单元中存储的每个问题之间的语义相似度,其中,所述语义相似度用于表示用户说出的问题与所述问答数据存储单元中存储的每个问题所表达的意思上的相似程度;以及分类步骤,基于计算出的所述语义相似度,将用户说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 13. The - method for an interactive voice systems, the interactive voice system comprising a voice recognition unit and a data storage unit Q, Q in the data storage unit stored in association with the question and the answer corresponding to the , said method comprising: a speech recognition step of performing speech recognition for the user issues spoken by a voice recognition unit using a language model; semantic similarity calculation step, the speech recognition result to the recognition unit question spoken by the user, the said user is calculated a semantic similarity between each question and the question of the quiz data stored in the storage unit, wherein, the semantic similarity is used to indicate problems with the storage unit of the user data Q stored in each spoken the degree of similarity on the meaning expressed problems; and a classification step of, based on the semantic similarity calculated, the user issues spoken classified as a problem in the storage unit or an external storage unit in question.
  14. 14. 根据权利要求13的方法,还包括: 置信度计算步骤,基于所述语音识别步骤中得到的有关用户所说出的问题的识别结果,为用户所说出的问题计算置信度; 其中,在所述分类步骤中,还能够基于计算出的所述置信度,将用户所说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 14. The method of claim 13, further comprising: confidence level calculation step, based on the speech recognition result obtained in the recognition step the spoken user about the problem, the problem of calculating the confidence spoken user; wherein, in the classification step, it is also based on the confidence level can be calculated, the user issues spoken classified as a problem in the storage unit or an external storage unit in question.
  15. 15. 根据权利要求13的方法,还包括: 输出步骤,如果用户所说出的问题被分类为在存储单元内的问题,则输出与用户所说出的问题相应的在所述问答数据存储单元中所存储的问题对应的答案。 15. The method of claim 13, further comprising: an output step, if the user issues spoken is classified as a problem in the storage unit is output by the user issues spoken in the respective data storing unit Q problems in stored corresponding answers.
  16. 16. 根据权利要求13的方法,还包括: 更新步骤,如果用户所说出的问题被分类为在存储单元外的问题,则基于所述在存储单元外的问题更新所述问答数据存储单元中存储的问题和/或更新所述语言模型。 16. The method of claim 13, further comprising: updating step, if the user issues spoken question is classified as an outer memory unit, based on the problems in the external memory unit updating the data storage unit Q storage problems and / or update the language model.
  17. 17. 根据权利要求14的方法,其中所述置信度是通过将用户所说出的问题经由N元语言模型进行识别而得到的声学得分除以用户所说出的该问题经由音素网络进行识别而得到的声学得分来确定的。 17. The method according to claim 14, wherein the confidence that the problem of the acoustic score of the spoken by a user via the problems identified N-gram language model obtained by dividing a user identified via the spoken phoneme network the obtained acoustic score determined.
  18. 18. 根据权利要求13的方法,其中,所述语义相似度计算步骤包括: 第一词袋产生步骤,根据所述语音识别步骤中得到的对于用户所说出的问题的识别结果,产生第一词袋,其中所述第一词袋包括该识别结果中所含的词; 第一词选择步骤,根据从所述语音识别步骤得到的对于用户所说出的问题的识别结果,从所述第一词袋选择其概率大于第三阈值的词,其中所述概率能够由通过N元语言模型得到的声学得分来确定; 第一检索步骤,根据所述第一词选择步骤中所选择的词,从数据源检索文档; 第二词袋产生步骤,为所述问答数据存储单元中存储的每个问题产生相应的第二词袋,其中每个第二词袋包括所述问答数据存储单元中所存储的相应的问题所含的词; 第二检索步骤,基于每个第二词袋中的词,从数据源检索文档;以及语义相似度计算子步骤,基 18. The method according to claim 13, wherein, the semantic similarity calculation step comprising: a step of generating a first bag of words, according to the recognition result to the user issues spoken by the voice recognition obtained in the step of generating a first bag of words, wherein the first bag of words comprises words contained in the recognition result; a first step of selecting words, according to the recognition result to the user issues spoken obtained from said speech recognition step, from the first the term probability greater than a selected word bags third threshold value, wherein the probability score can be determined acoustically from the N-gram language model obtained; a first retrieving step of selecting word step is selected in accordance with the first word, retrieve the document from a data source; bag of words a second generation step of generating a second bag of words corresponding to the question and answer each question stored in the data storage unit, wherein each of said Q second word bag comprises as data storage unit words corresponding problems contained in the storage; a second step of retrieving, based on the second word of each word bag, retrieved from the data source document; and semantic similarity calculation sub-step group 第一检索部件所检索到的文档和第二检索部件所检索到的文档,计算用户所说出的问题与所述问答数据存储单元中存储的每个问题之间的语义相似度。 First search means and the documents retrieved by the second retrieval member retrieved documents, calculating semantic similarity between the user and the spoken questions to the question and answer each question stored in the data storage unit.
  19. 19. 根据权利要求18的方法,其中,所述语义相似度计算步骤还包括: 第一过滤步骤,根据词性,从在所述第一词选择步骤中所选择的第一词袋的词中过滤掉一部分词,和/或根据词性,从所述第二词袋中过滤掉一部分词; 其中,在所述第一检索步骤中针对过滤后的第一词袋进行检索,和/或在第二检索步骤中针对过滤后的第二词袋进行检索。 19. The method according to claim 18, wherein, the semantic similarity calculation step further comprises: a first filtering step, according to part of speech, the filter bag of words from the first word in the selection step the first word in the selected away a portion of the word, and / or speech in accordance with, the second word from the filter bag away a portion of the word; wherein, performed for the first word after the bag filter in said first retrieval step, retrieval, and / or the second search step to search for a second term after the bag filter.
  20. 20. 根据权利要求19的方法,其中,所述语义相似度计算步骤还包括: 领域词添加步骤,向过滤后的第一词袋和/或过滤后的第二词袋添加与领域有关的词; 其中,在所述第一检索步骤中,针对进行上述添加后的第一词袋进行检索,和/或在所述第二检索步骤中,针对进行上述添加后的第二词袋进行检索。 20. The method according to claim 19, wherein, the semantic similarity calculating step further comprising: a field word addition step, adding the art related to the word or the second word of the first bag and the bag of words after filtration / filter ; wherein, in the first retrieving step, retrieving for the first word of the bag after the above additions, and / or to search for a second term after the bag is added to the above-described second search step.
  21. 21. 根据权利要求18所述的方法,其中,在所述语义相似度计算步骤中,基于以下来进行语义相似度的计算: 在第一检索步骤中所检索到的文档和在第二检索步骤中所检索到的文档中的每个词; 和/或在第一检索步骤中所检索到的文档和在第二检索步骤中所检索到的文档之间的重合。 21. The method according to claim 18, wherein, in the semantic similarity calculation step, calculation is performed based on the semantic similarity: in a first search step and the retrieved document in a second search step the retrieved documents in each word; and / or between the first step is to retrieve the documents retrieved in the second retrieving step and the retrieved documents coincide.
  22. 22. 根据权利要求13〜21中的任一项的方法,还包括: 文字相似度计算步骤,根据从所述语音识别步骤得到的对于用户说出的问题的识别结果,计算用户所说出的问题与所述问答数据存储单元中的每个问题之间的文字相似度,其中,所述文字相似度用于表示用户说出的问题与所述问答数据存储单元中存储的每个问题在文字上的相似程度; 其中,在所述分类步骤中,还能够基于所计算出的文字相似度,将用户所说出的问题分类为在存储单元内的问题或者在存储单元外的问题。 22. A method according to any one of claims 13~21, further comprising: a similarity calculating step of writing, in accordance with the recognition result to the user issues spoken obtained from said speech recognition step of calculating the user's spoken the degree of similarity between the text of each question of the quiz question data storage unit, wherein the text spoken by the user for indicating the degree of similarity with each question of the quiz question data storage unit in the text the degree of similarity; wherein, in the classification step, it is possible character based on the calculated degree of similarity, the user issues spoken classified as a problem in the storage unit or an external storage unit in question.
  23. 23. 根据权利要求22的方法,其中,所述文字相似度计算步骤包括: 第三词袋产生步骤,根据从所述语音识别步骤得到的对于用户所说出的问题的识别结果,产生第三词袋,其中所述第三词袋包括该识别结果中所含的词; 第四词袋产生步骤,为所述问答数据存储单元中存储的每个问题产生相应的第四词袋,其中每个第四词袋包括所述问答数据存储单元中所存储的相应的问题所含的词; 第二词选择步骤,根据从所述语音识别步骤得到的对于用户所说出的问题的识别结果,从所述第三词袋选择其概率大于第四阈值的词,其中所述概率能够由通过N元语言模型得到的声学得分来确定;以及文字相似度计算子步骤,根据在所述第二词选择步骤中所选择的第三词袋中的词和所述第四词袋中的词,计算用户所说出的问题与所述问答数据存储单元中存储的每个问题 23. The method according to claim 22, wherein said text similarity calculation step comprises: step of generating a third bag of words, according to the recognition result to the user issues spoken obtained from said speech recognition step of generating a third bag of words, wherein the third bag of words comprises words contained in the recognition result; fourth bag of words generating step of generating a fourth bag of words corresponding to the question and answer each question stored in the data storage unit, wherein each of the fourth bag of words comprising a word corresponding issue data stored in the storage unit contained in the question and answer; second word selection step, according to the recognition result to the user issues spoken obtained from said speech recognition step, a third bag of words selected from the words whose probability is greater than a fourth threshold, wherein the probability score can be determined acoustically from the N-gram language model obtained; and text similarity calculation sub-step, in accordance with the second word the third word word selection step, the selected bag and the bag fourth word word spoken by the user is calculated with each question to the quiz question data stored in the storage unit 之间的文字相似度。 Text similarity between.
  24. 24. 根据权利要求23的方法,其中,所述文字相似度计算步骤还包括: 第二过滤步骤,根据词性,从所述第二词选择步骤中所选择的第三词袋的词中过滤掉一部分词,和/或根据词性,从所述第四词袋中过滤掉一部分词; 其中,在进行了第二过滤步骤中的过滤之后,进行所述文字相似度计算子步骤中的文字相似度的计算。 24. The method according to claim 23, wherein said similarity calculating step further comprises the text: a second filtering step, according to part of speech, words from said second selecting step is selected word third word bag filter out word part, and / or in accordance with speech, word from the fourth bag filter away a portion of the word; wherein, after performing the filtration in the second filtration step, the substep of writing the text similarity calculation similarity calculations.
CN 201310390944 2013-09-02 2013-09-02 Voice based question-answering system and method for interactive voice system CN104424290A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 201310390944 CN104424290A (en) 2013-09-02 2013-09-02 Voice based question-answering system and method for interactive voice system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 201310390944 CN104424290A (en) 2013-09-02 2013-09-02 Voice based question-answering system and method for interactive voice system

Publications (1)

Publication Number Publication Date
CN104424290A true true CN104424290A (en) 2015-03-18

Family

ID=52973269

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 201310390944 CN104424290A (en) 2013-09-02 2013-09-02 Voice based question-answering system and method for interactive voice system

Country Status (1)

Country Link
CN (1) CN104424290A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105161098A (en) * 2015-07-31 2015-12-16 北京奇虎科技有限公司 Speech recognition method and speech recognition device for interaction system
CN105244024A (en) * 2015-09-02 2016-01-13 百度在线网络技术(北京)有限公司 Voice recognition method and device

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020065845A1 (en) * 2000-05-17 2002-05-30 Eiichi Naito Information retrieval system
CN101086843A (en) * 2006-06-07 2007-12-12 中国科学院自动化研究所 A sentence similarity recognition method for voice answer system
CN101373532A (en) * 2008-07-10 2009-02-25 昆明理工大学 FAQ Chinese request-answering system implementing method in tourism field
CN102013253A (en) * 2009-09-07 2011-04-13 株式会社东芝 Speech recognition method based on speed difference of voice unit and system thereof

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020065845A1 (en) * 2000-05-17 2002-05-30 Eiichi Naito Information retrieval system
CN101086843A (en) * 2006-06-07 2007-12-12 中国科学院自动化研究所 A sentence similarity recognition method for voice answer system
CN101373532A (en) * 2008-07-10 2009-02-25 昆明理工大学 FAQ Chinese request-answering system implementing method in tourism field
CN102013253A (en) * 2009-09-07 2011-04-13 株式会社东芝 Speech recognition method based on speed difference of voice unit and system thereof

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105161098A (en) * 2015-07-31 2015-12-16 北京奇虎科技有限公司 Speech recognition method and speech recognition device for interaction system
CN105244024A (en) * 2015-09-02 2016-01-13 百度在线网络技术(北京)有限公司 Voice recognition method and device

Similar Documents

Publication Publication Date Title
Jansen et al. Towards spoken term discovery at scale with zero resources
US20060129397A1 (en) System and method for identifying semantic intent from acoustic information
US20120253799A1 (en) System and method for rapid customization of speech recognition models
US20120271631A1 (en) Speech recognition using multiple language models
US9292487B1 (en) Discriminative language model pruning
CN102280106A (en) Voice network search method and apparatus for a mobile communication terminal,
US20100198577A1 (en) State mapping for cross-language speaker adaptation
US20100211391A1 (en) Automatic computation streaming partition for voice recognition on multiple processors with limited memory
Garnier-Rizet et al. CallSurf: Automatic Transcription, Indexing and Structuration of Call Center Conversational Speech for Knowledge Extraction and Query by Content.
Zazo et al. Language identification in short utterances using long short-term memory (LSTM) recurrent neural networks
Das et al. Recognition of isolated words using features based on LPC, MFCC, ZCR and STE, with neural network classifiers
US20150019221A1 (en) Speech recognition system and method
US20120158399A1 (en) Sample clustering to reduce manual transcriptions in speech recognition system
Schmitt et al. OpenXBOW: introducing the passau open-source crossmodal bag-of-words toolkit
JP2008216341A (en) Error-trend learning speech recognition device and computer program
US7966171B2 (en) System and method for increasing accuracy of searches based on communities of interest
US20150066502A1 (en) System and Method of Automated Model Adaptation
JP2013134430A (en) Device, method, and program for processing command
Yu et al. Context adaptive training with factorized decision trees for HMM-based statistical parametric speech synthesis
Mitra et al. Articulatory features from deep neural networks and their role in speech recognition
US20150066503A1 (en) System and Method of Automated Language Model Adaptation
Mariooryad et al. Compensating for speaker or lexical variabilities in speech for emotion recognition
US9600231B1 (en) Model shrinking for embedded keyword spotting
US9058811B2 (en) Speech synthesis with fuzzy heteronym prediction using decision trees
US20160078859A1 (en) Text-to-speech with emotional content

Legal Events

Date Code Title Description
C06 Publication
C10 Entry into substantive examination