WO2019080661A1 - 用户意图识别方法及装置 - Google Patents

用户意图识别方法及装置

Info

Publication number
WO2019080661A1
WO2019080661A1 PCT/CN2018/105192 CN2018105192W WO2019080661A1 WO 2019080661 A1 WO2019080661 A1 WO 2019080661A1 CN 2018105192 W CN2018105192 W CN 2018105192W WO 2019080661 A1 WO2019080661 A1 WO 2019080661A1
Authority
WO
WIPO (PCT)
Prior art keywords
question
user
predefined
predefined elements
elements
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2018/105192
Other languages
English (en)
French (fr)
Inventor
刘佳
崔恒斌
张家兴
吴丽娟
毛瑶瑶
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Alibaba Group Holding Ltd
Original Assignee
Alibaba Group Holding Ltd
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
Application filed by Alibaba Group Holding Ltd filed Critical Alibaba Group Holding Ltd
Publication of WO2019080661A1 publication Critical patent/WO2019080661A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3344—Query execution using natural language analysis

Definitions

  • One or more embodiments of the present disclosure relate to the field of computer technologies, and in particular, to a user intention identification method and apparatus.
  • the hotline system can receive various questions sent by users and analyze the received problems.
  • the user's intention is identified by analyzing the received problem. Therefore, there is a need to provide a quick solution for identifying user intent.
  • One or more embodiments of the present specification describe a user intent identification method and apparatus to quickly answer a user's question.
  • a method for user intent identification including:
  • the plurality of predefined elements are extracted from knowledge points of the knowledge base, and the plurality of predefined elements are respectively attributed to N categories, N Is a positive integer;
  • the user intent recognition result is determined.
  • a user intention identification device including:
  • An obtaining unit configured to acquire a dialogue text of the user
  • a determining unit configured to determine a relevance value of the dialog text acquired by the acquiring unit and a plurality of predefined elements, where the plurality of predefined elements are extracted from knowledge points of the knowledge base, and the plurality of The predefined elements are respectively assigned to N categories, and N is a positive integer;
  • a selecting unit configured to select a first element from the plurality of predefined elements according to the relevance value determined by the determining unit and a preset threshold
  • a matching unit configured to match corresponding knowledge points from the knowledge base according to the first element selected by the selecting unit
  • the determining unit is further configured to determine a user intention recognition result according to the knowledge point matched by the matching unit.
  • the user intention identification method and apparatus provided by one or more embodiments of the present specification first acquires a dialogue text of a user, and determines a relevance value of the conversation text and a plurality of predefined elements. Then, according to the correlation value determined above, the first element matching the dialog text is selected. Finally, according to the selected first element, the corresponding knowledge points are matched from the knowledge base. After matching to the knowledge point, the user intent recognition result is determined based on the knowledge point. Thereby, the user's intention can be quickly recognized.
  • FIG. 1 is a schematic diagram of an application scenario of a user intent identification method according to an embodiment of the present disclosure
  • FIG. 2 is a flowchart of a method for identifying a user intent provided by an embodiment of the present specification
  • FIG. 3 is a schematic diagram of a method for identifying a user intent according to another embodiment of the present disclosure
  • FIG. 4 is a schematic diagram of a user intention identification device provided by an embodiment of the present specification.
  • the hotline system can be any system that can provide "smart robot" service. Specifically, it can receive various questions sent by the user and analyze the received questions. The user's intention is identified by analyzing the received problem. It should be noted that in the process of identifying the user's intention, the hotline system can conduct multiple rounds of conversation with the user. After identifying the user's intent, answer the user's question.
  • a knowledge base may be stored in the hotline system, and the knowledge base is composed of one or more knowledge points. Among them, each knowledge point has a corresponding solution.
  • the knowledge points herein can be understood as a canonical representation of the problems of users collected in advance. Specifically, after receiving the user's question, the hotline system can match the problem with the knowledge points in the knowledge base. The solution corresponding to the matched knowledge point is sent to the user as the answer to the user's question.
  • FIG. 2 is a flowchart of a method for identifying a user intent provided by an embodiment of the present specification.
  • the execution body of the method may be a device having processing capabilities: a server or a system or device, such as the hotline system of FIG. As shown in FIG. 2, the method may specifically include:
  • Step 210 Acquire a dialogue text of the user.
  • the dialog text here can be any text that can express the question the user wants to ask. For example, it can be: “How can I do if the flower garden is not on the money?", "How to repay the flower basket?” and "How about the income of the balance treasure?”
  • Step 220 Determine a relevance value of the conversation text and the plurality of predefined elements.
  • the machine is not easy to recognize because the dialogue text of the directly acquired user is relatively colloquial. Therefore, after the user's conversation text is obtained, it is usually necessary to perform preprocessing, such as word segmentation processing.
  • preprocessing such as word segmentation processing here is a conventional conventional technique and will not be repeated here.
  • the predefined elements here can be extracted from the knowledge points. For example, for the knowledge point of “failure of flower repayment”, “flower”, “repayment” and “failure” can be used as predefined elements.
  • the above predefined elements may have corresponding categories, and one predefined element may be uniquely attributed to one category, that is, the predefined elements may not intersect each other.
  • the predefined elements of the above extraction may be attributed to the following three categories: "business type”, "framework verb", and “question type”. As the foregoing example, "Flower” can be attributed to "business type”, “repayment” can be attributed to "framework verb”, and “failure” can be attributed to "question type”.
  • the “business type” is a relatively broad category, and the predefined elements belonging to the category may also be: “account”, “password”, “yield treasure”, “flower”, “borrowing” , “regular financial management,” and so on.
  • the predefined elements attributed to "frame verbs” are usually the most central verbs or verb phrases in the knowledge points. For example, they can be: “login”, “forget”, “payment”, “repayment”, “borrowing”, etc.
  • the “question type” can be: “what”, “when”, “where”, “whether", "failure”, “what” means.
  • the number of categories to which a predefined element of certain knowledge points in the knowledge base belongs may be allowed to be less than the total number of categories. For example, suppose that the total number of categories to which a predefined element in a knowledge base belongs is three, namely: "business type”, "framework verb", and "problem type”.
  • a certain knowledge point is “flower repayment”, that is, the categories to which the predefined elements constituting the knowledge point belong are “business type” and “frame verb”, that is, the number of categories is two.
  • the predefined element may further have a corresponding extended element, and the number of the extended element may be plural. Its role can be to better identify the predefined elements in the dialog text.
  • the extension element can be an alias for the corresponding predefined element, which is a name that is convenient for the user to understand. For example, when the predefined element is “cost modification”, the corresponding expansion element may be “Alipay” or the like.
  • the predefined elements of this specification can usually express meanings accurately and have generalization. It can be understood that preferred predefined elements usually appear in more than two knowledge points.
  • the knowledge point may further include an affiliate description portion that may be used when forming a rheme question in a course of multiple rounds of conversation with the user.
  • the categories of "business type", "framework verb” and "problem type” are as follows. The content of a complete knowledge point can be as shown in Table 1.
  • the predefined element 1, the predefined element 2 and the predefined element 3 can constitute a knowledge point.
  • the method of determining the relevance value of the dialog text and the plurality of predefined elements may include: a method of model recognition and/or a method of calculating a text matching degree value.
  • the method for model identification is specifically: inputting the pre-processed dialog text into the feature recognition model, and the feature recognition model may be a pre-trained classification model, and the probability that the dialogue text is classified into each predefined element can be directly obtained by the classification model. value.
  • the above probability value is taken as the above-mentioned relevance value.
  • the method for calculating the text matching degree value is specifically: calculating a similarity value of the dialogue text and each predefined element by using the similarity algorithm, and using the similarity value as the correlation value.
  • the results obtained by the two methods may be directly added or the results obtained by the two methods may be merged by other methods.
  • each of the predefined elements and the corresponding relevance value can be as shown in Table 2.
  • the predefined elements of each row in Table 1 can constitute a knowledge point, that is, these predefined elements are extracted from the corresponding knowledge points.
  • Table 2 is merely an exemplary description for facilitating understanding of the present embodiment, and is not intended to be a limitation of the embodiment.
  • other categories of predefined elements and the like may also be included in Table 2.
  • Step 230 Select a first element from a plurality of predefined elements according to the relevance value and the preset threshold.
  • the level information of each predefined element may be determined, and then the first element is selected from the plurality of predefined elements according to the level information.
  • the specific process may be: determining the level information corresponding to each predefined element according to the association value and the preset threshold. The first element is selected from a plurality of predefined elements based on the level information.
  • two thresholds can be set in advance: 0.8 and 0.6.
  • the level information of the predefined element may be determined to be a high level (eg, the level information of “Flower”, “Repayment”, and “Fail” in Table 2 For high grade).
  • the level information of the predefined element may be determined as the medium level (for example, the level information of "forgotten” in Table 2 is the medium level).
  • the level information of the predefined element may be determined to be a low level.
  • a high-level predefined element may be referred to as a trusted element
  • a medium-level predefined element may be referred to as a candidate element
  • a low-level predefined element may be referred to as an unknown element.
  • thresholds may be set for the predefined elements belonging to different categories, which are not limited in this specification.
  • the high-level predefined element may be selected as the first element, that is, the trusted element may be selected as the first element.
  • the trusted element may be selected as the first element.
  • “Flower”, "Repayment” and “Failure” in Table 2 can be selected as the first element.
  • Step 240 Match corresponding knowledge points from the knowledge base according to the first element.
  • the unique knowledge points can be matched when the predefined elements belonging to the same category are unique. Therefore, the present specification requires that the first element belonging to the same category is unique, that is, the trusted element in the same category in the present specification is unique, and the candidate element may have multiple.
  • the knowledge point can be matched as: “Flower Repayment Failure”.
  • Step 250 Determine a user intent recognition result according to the knowledge point.
  • the user's intent to identify the result can be determined.
  • trusted elements in all categories cannot be selected, and multiple knowledge points are matched at this time.
  • only the trusted elements may be selected: “Flower” and “Repayment”, that is, the trusted elements under “Question Type” are not selected. From Table 2, you can match two knowledge points: “Flower Repayment Failure” and “How to Repay the Flower”.
  • step A the second element is selected from a plurality of predefined elements according to the level information.
  • the medium-level predefined element may be selected as the second element, that is, the candidate element may be selected as the second element.
  • the candidate element may be selected as the second element.
  • "failure" in Table 2 can be selected as the second element.
  • Step B reading the matched question template from the query template library according to the first element, the category to which the second element belongs, and the level information.
  • the questioning template library here is used to record the correspondence between the level information of the elements belonging to different categories and the questioning template.
  • the query template library can be as shown in Table 3.
  • step C the corresponding question question is determined according to the first element, the second element, and the question template.
  • the rhetorical question when the name of the first element or the second element is difficult for a particular user to understand, it can be replaced with the corresponding subsidiary description to achieve the purpose of changing the expression.
  • the generated rhetorical question if it is not fluent enough and needs to be interacted again, it can be implemented by configuring a specific rhetorical question.
  • step D a question question is sent to the user.
  • step E the user intention identification result is determined according to the answer of the received rhetorical question.
  • the user's intention to identify the result can be finally determined.
  • the first element and the second element are combined with the inverse question template to adapt the personalized rhetorical question, thereby reducing the difficulty of the interaction as much as possible.
  • the first element is not selected, and the first element may be selected, but the knowledge point cannot be matched from the knowledge base.
  • the user sends a preset question question, for example, "What problem have you encountered?".
  • a preset question question for example, "What problem have you encountered?".
  • there may be a specific question which can be done through configuration. For example, when the first element is the same as the element in the preset set, a preset question question corresponding to the element is obtained. The user's intention to identify the result is determined based on the received answer to the preset question question.
  • the user intention identification method proposed by the above embodiment of the present specification can increase the user intention recognition accuracy rate and the recall rate by 5%.
  • the user is more willing to engage in multiple rounds of dialogue with the robot.
  • FIG. 3 is a schematic diagram of a method for identifying a user intent provided by another embodiment of the present specification.
  • three elements can be extracted from the knowledge points in the knowledge base in advance, that is, elements belonging to the “business type”, “frame verb” and “question type” are extracted from the knowledge points of the knowledge base.
  • the elements in the dialog text are identified by calculating the recognition probability values of the dialog text and the previously extracted elements. Then, based on the calculated recognition probability and the threshold, the identified elements are divided into trusted elements, candidate elements, and unknown elements. Finally, the corresponding knowledge points are matched from the knowledge base through trusted elements. If the corresponding knowledge point is matched, the user intent is directly output.
  • the corresponding challenge template is matched from the question template library according to the trusted element and the candidate element. Then, according to the matching question template, the question question is output to the user, and the user's intention is identified according to the answer to the question.
  • an embodiment of the present specification further provides a user intention identification device.
  • the device includes:
  • the obtaining unit 401 is configured to acquire the dialogue text of the user.
  • the determining unit 402 is configured to determine a relevance value of the dialog text acquired by the obtaining unit 401 and the plurality of predefined elements, where the plurality of predefined elements are extracted from the knowledge points of the knowledge base, and the plurality of predefined elements are respectively attributed to N categories, N is a positive integer.
  • Categories here can include: business type, framework verbs, and question types.
  • the determining unit 402 is specifically configured to:
  • the dialog text is input into the feature recognition model, and the probability values corresponding to the plurality of predefined elements are outputted, wherein the feature recognition model is a predefined machine learning model for calculating the probability value of the input text and the plurality of predefined elements.
  • a relationship value of the dialog text and the plurality of predefined elements is determined according to the probability value and/or the matching value.
  • the selecting unit 403 is configured to select the first element from the plurality of predefined elements according to the relevance value determined by the determining unit 402 and the preset threshold.
  • the selecting unit 403 is specifically configured to:
  • the level information corresponding to the plurality of predefined elements is determined according to the association value and the preset threshold.
  • the first element is selected from a plurality of predefined elements based on the level information.
  • the matching unit 404 is configured to match the corresponding knowledge point from the knowledge base according to the first element selected by the selecting unit 403.
  • the determining unit 402 is further configured to determine a user intent recognition result according to the knowledge points matched by the matching unit 404.
  • the determining unit 402 is further configured to:
  • the second element is selected from a plurality of predefined elements based on the level information.
  • the matching question template is read from the query template library, and the template library is used to record the level information between the elements belonging to different categories and the question template. Correspondence relationship.
  • Corresponding question questions are determined according to the first element, the second element, and the question template.
  • the user's intention to identify the result is determined based on the answer to the received rhetorical question.
  • the device may further include:
  • the sending unit 405 is configured to send a preset question question to the user.
  • the determining unit 402 is further configured to determine a user intention recognition result according to the received answer to the preset question question.
  • the obtaining unit 401 is further configured to: when the first element is the same as the element in the preset set, obtain a preset question question corresponding to the element.
  • the determining unit 402 is further configured to determine a user intention recognition result according to the received answer to the preset question question.
  • the acquisition unit 401 acquires the conversation text of the user.
  • the determining unit 402 determines the relevance value of the acquired dialog text and the plurality of predefined elements.
  • the selecting unit 403 selects the first element from the plurality of predefined elements according to the relevance value and the preset threshold.
  • the matching unit 404 matches the corresponding knowledge points from the knowledge base according to the selected first element.
  • the determining unit 402 determines the user intent recognition result based on the matched knowledge points. Thereby, the user's intention can be quickly recognized.
  • the functions described herein can be implemented in hardware, software, firmware, or any combination thereof.
  • the functions may be stored in a computer readable medium or transmitted as one or more instructions or code on a computer readable medium.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Machine Translation (AREA)
  • User Interface Of Digital Computer (AREA)

Abstract

一种用户意图识别方法及装置，在用户意图识别方法中，首先获取用户的对话文本（S210），并确定该对话文本与多个预定义要素的关联度值（S220）。之后根据关联度值以及预设阈值，从多个预定义要素中选取第一要素（S230）。最后根据选取的第一要素，从知识库中匹配对应的知识点（S240）。在匹配到知识点之后，根据该知识点确定用户意图识别结果（S250）。

Description

用户意图识别方法及装置 技术领域
本说明书一个或多个实施例涉及计算机技术领域，尤其涉及一种用户意图识别方法及装置。
背景技术
在在线或者热线服务中，热线系统可以接收用户发送的各种问题，并对接收的问题进行分析。通过对接收的问题进行分析，识别出用户的意图。因此，需要提供一种快速的识别用户意图的方案。
发明内容
本说明书一个或多个实施例描述了一种用户意图识别方法及装置，以快速地对用户的问题进行解答。
第一方面，提供了一种用户意图识别方法，包括：
获取用户的对话文本；
确定所述对话文本与多个预定义要素的关联度值，所述多个预定义要素是从知识库的知识点中抽取的，且所述多个预定义要素分别归属于N个类别，N为正整数；
根据所述关联度值以及预设阈值，从所述多个预定义要素中选取第一要素；
根据所述第一要素，从所述知识库中匹配对应的知识点；
根据所述知识点，确定用户意图识别结果。
第二方面，提供了一种用户意图识别装置，包括：
获取单元，用于获取用户的对话文本；
确定单元，用于确定所述获取单元获取的所述对话文本与多个预定义 要素的关联度值，所述多个预定义要素是从知识库的知识点中抽取的，且所述多个预定义要素分别归属于N个类别，N为正整数；
选取单元，用于根据所述确定单元确定的所述关联度值以及预设阈值，从所述多个预定义要素中选取第一要素；
匹配单元，用于根据所述选取单元选取的所述第一要素，从所述知识库中匹配对应的知识点；
所述确定单元，还用于根据所述匹配单元匹配的所述知识点，确定用户意图识别结果。
本说明书一个或多个实施例提供的用户意图识别方法及装置，首先获取用户的对话文本，并确定该对话文本与多个预定义要素的关联度值。之后根据上述确定的关联度值，选取与对话文本相匹配的第一要素。最后根据选取的第一要素，从知识库中匹配对应的知识点。在匹配到知识点之后，根据该知识点确定用户意图识别结果。由此，可以快速地识别用户意图。
附图说明
为了更清楚地说明本发明实施例的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其它的附图。
图1为本说明书一个实施例提供的用户意图识别方法的应用场景示意图；
图2为本说明书一个实施例提供的用户意图识别方法流程图；
图3为本说明另一个实施例提供的用户意图识别方法示意图；
图4为本说明书一个实施例提供的用户意图识别装置示意图。
具体实施方式
下面结合附图，对本说明书提供的方案进行描述。
本说明书一个实施例提供的用户意图识别方法可以应用于如图1所示的场景中，图1中，热线系统可以为任一能提供“智能机器人”服务的系统。具体地，其可以接收用户发送的各种问题，并对接收的问题进行分析。通过对接收的问题进行分析，识别出用户的意图。需要说明的是，在识别用户意图的过程中，热线系统可以与用户进行多轮会话。在识别出用户的意图之后，对用户的问题进行解答。
图1中，热线系统中可以存储有知识库，该知识库由一个或多个知识点构成。其中，每个知识点具有对应的解答方案。此处的知识点可以理解为对预先收集的用户的问题的规范表述。具体地，热线系统在接收到用户的问题之后，可以将问题与知识库中的知识点进行匹配。将相匹配的知识点对应的解答方案作为用户的问题的答案发送给用户。
图2为本说明书一个实施例提供的用户意图识别方法流程图。所述方法的执行主体可以为具有处理能力的设备：服务器或者系统或者装置，如，图1中的热线系统。如图2所示，所述方法具体可以包括：
步骤210，获取用户的对话文本。
此处的对话文本可以为任一能表述用户想要问的问题的文字。如，可以为：“花呗还不上钱，怎么办？”、“花呗如何还款？”以及“余额宝收益怎么样？”等。
步骤220，确定对话文本与多个预定义要素的关联度值。
可以理解的是，由于直接获取的用户的对话文本比较口语化，机器不容易识别。因此，在获取到用户的对话文本之后，通常先要进行预处理，如，分词处理等。此处的分词处理等预处理过程为传统常规技术，在此不复赘述。
此处的预定义要素可以是从知识点中抽取的。如，对于“花呗还款失败”这个知识点，“花呗”、“还款”和“失败”就可以作为预定义要素。上述预定义要素可以具有对应的类别，一个预定义要素可以唯一地归属于一个类别，也即预定义要素之间可以不交叉存在。在一个例子中，上述抽 取的预定义要素可以归属于如下三个类别：“业务类型”、“框架动词”以及“问题类型”。如前述例子，“花呗”可以归属于“业务类型”，“还款”可以归属于“框架动词”，“失败”可以归属于“问题类型”。需要说明的是，“业务类型”是一个比较宽泛的类别，归属于该类别的预定义要素还可以为：“账户”、“密码”、“余额宝”、“花呗”、“借呗”、“定期理财”等。归属于“框架动词”的预定义要素通常是知识点中最核心的动词或者动词短语，如，可以为：“登录”、“忘记”、“支付”、“还款”、“借款”等。归属于“问题类型”可以为：“什么”、“何时”、“何地”、“是否”、“失败”、“什么意思”等。
在本说明书的一个例子中，可以允许知识库中的某些知识点的预定义要素所归属的类别数目小于总的类别数目。举例来说，假设某知识库中的预定义要素所归属的总的类别数目为3个，分别为：“业务类型”、“框架动词”以及“问题类型”。某个知识点为“花呗还款”，即构成该知识点的预定义要素所归属的类别分别为“业务类型”和“框架动词”，也即类别的数目为2个。
在本说明书的另一个例子中，上述预定义要素还可以具有对应的扩展要素，该扩展要素的个数可以为多个。其作用可以是为了更好的识别出对话文本中的预定义要素。在一个例子中，该扩展要素可以为对应的预定义要素的别名，该别名是方便用户理解的名称。如，在预定义要素为“费用改造”时，对应的扩展要素可以为“支付宝”等。
此外，本说明书的预定义要素通常可以准确表达含义，同时具有泛化性。可以理解的是，较优的预定义要素通常出现在两个以上的知识点中。在本说明书的再一个例子中，知识点还可以包括附属描述部分，该部分可以在与用户进行多轮会话的过程中形成反问问句时使用。以类别分别为：“业务类型”、“框架动词”以及“问题类型”为例来说，一个完整的知识点所包含的内容部分可以如表1所示。
表1
Figure PCTCN2018105192-appb-000001
可以理解的是，表1中，预定义要素1、预定义要素2以及预定义要素3可以构成一个知识点。
当然，在实际应用中，表1中的知识点还可以包括其它部分内容，如，解答方案等，本说明书对此不作限定。
步骤220中，确定对话文本与多个预定义要素的关联度值的方法可以包括：模型识别的方法和/或计算文本匹配度值的方法。模型识别的方法具体为：将预处理后的对话文本输入要素识别模型，该要素识别模型可以是预先训练好的分类模型，通过该分类模型可以直接得到将对话文本分类到各个预定义要素的概率值。将上述概率值作为上述关联度值。计算文本匹配度值的方法具体为：通过相似度算法，计算对话文本与各个预定义要素的相似度值，将该相似度值作为上述关联度值。
当通过两种方法来确定上述关联度值时，可以直接将两种方法得到的结果相加或者也可以通过别的方法将两种方法得到的结果进行融合。
在一个例子中，各个预定义要素以及对应的关联度值可以如表2所示。
表2
Figure PCTCN2018105192-appb-000002
可以理解的是，表1中每行的预定义要素可以构成一个知识点，也即这些预定义要素是从对应的知识点中抽取的。需要说明的是，表2只是为了便于理解本实施例而给出的示例性说明，并不作为本实施例的限制。如，表2中还可以包括其它类别的预定义要素等。
步骤230，根据关联度值以及预设阈值，从多个预定义要素中选取第一要素。
在一种实现方式中，在执行步骤230之前，可以先确定各个预定义要素的等级信息，之后根据等级信息，从多个预定义要素中选取第一要素。其具体过程可以为：根据关联度值以及预设阈值，确定各个预定义要素对应的等级信息。根据等级信息，从多个预定义要素中选取第一要素。
以表2为例来说，可以预先设定两个阈值：0.8和0.6。具体地，当预定义要素的关联度值大于等于0.8时，可以将预定义要素的等级信息确定为高等级(如，表2中“花呗”、“还款”和“失败”的等级信息为高等级)。当预定义要素的关联度值在0.6与0.8之间时，可以将预定义要素的等级信息确定为中等级(如，表2中“忘记”的等级信息为中等级)。当预定义要素的关联度值小于0.6时，可以将预定义要素的等级信息确定为低等级。还需要说明的是，在本说明书中，高等级的预定义要素可以称为可信要素，中等级的预定义要素可以称为候选要素，低等级的预定义要 素可以称为未知要素。
需要说明的是，上述只是一种设定阈值的方式。在其它实现方式中，对归属于不同类别的预定义要素，可以设定不同的阈值，本说明书对此不作限定。
在确定出各个预定义要素的等级信息之后，可以将高等级的预定义要素选取为第一要素，也即可以将可信要素选取为第一要素。如，可以将表2中的“花呗”、“还款”和“失败”选取为第一要素。
步骤240，根据第一要素，从知识库中匹配对应的知识点。
从表1中可以看出，当归属于同一类别的预定义要素唯一时，才可以匹配到唯一的知识点。因此，本说明书要求归属于同一类别的第一要素唯一，也即本说明书中同一类别下的可信要素是唯一的，而候选要素则可以有多个。
如前述例子，可以匹配到知识点为：“花呗还款失败”。
步骤250，根据知识点，确定用户意图识别结果。
在匹配到唯一的知识点之后，就可以确定用户意图识别结果。
当然，在实际应用中，有可能会出现不能选取出所有类别下的可信要素的情况，此时会匹配到多个知识点。如前述例子，可能只选取出可信要素：“花呗”和“还款”，也即未选取出“问题类型”下的可信要素。则从表2中可以匹配到两个知识点：“花呗还款失败”和“花呗如何还款”。
当匹配到多个知识点时，需要与用户进行多轮交互，以确定用户意图。在与用户进行多轮交互的过程中，如何确定反问问句，以快速地确定用户意图成为关键。在本说明书中，提出了如下确定反问问句的方法：
步骤A，根据等级信息，从多个预定义要素中选取第二要素。
在一种实现方式中，可以将中等级的预定义要素选取为第二要素，也即可以将候选要素选取为第二要素。如，可以将表2中的“失败”选取为第二要素。
步骤B，根据第一要素、第二要素所归属的类别以及等级信息，从反问 模板库中读取相匹配的反问模板。
此处的反问模板库用于记录归属于不同类别的要素的等级信息与反问模板之间的对应关系。在一个例子中，反问模板库可以如表3所示。
表3
Figure PCTCN2018105192-appb-000003
表3中，
Figure PCTCN2018105192-appb-000004
表示高等级，“√”表示中等级，“×”表示低等级。举例来说，假设选取的第一要素(即可信要素)为：“花呗”和“还款”，选取的第二要素(即候选要素)为：“失败”。因为“花呗”所归属的类别为：“业务类型”，“还款”所归属的类别为：“框架动词”，“失败”所归属的类别为“问题类型”。此外，“花呗”和“还款”所对应的等级信息为：“高等级”，“失败”所对应的等级信息为：“中等级”。所以可以匹配到表3中的第5行，也即读取的反问模板为：“请问您在#bus iness##frame#时，#type#了还是怎么了？”。
步骤C，根据第一要素、第二要素以及反问模板，确定对应的反问问句。
具体地，可以将反问模板中不同类别的信息替换为该类别下的第一要素或者第二要素。如前述例子，可以假设“#bus iness#”为“业务类型”的信息，“#frame#”为“框架动词”的信息，“#type#”为“问题类型”的信息。则确定的对应的反问问句为：“请问您在花呗还款时，失败了还是怎么了？”。
在确定反问问句的时候，当第一要素或者第二要素的名称比较特殊用户不易懂的，则可以替换为对应的附属描述，以达到换一种表达方式的目的。此外，如果生成的反问问句不够通顺，且需要再次进行交互，则可以通过配置特定的反问问句来实现。
步骤D，向用户发送反问问句。
步骤E，根据接收的反问问句的答案，确定用户意图识别结果。
此处，在接收到用户对反问问句的答案，就可以最终确定用户意图识别结果了。
综上，本说明书上述实施例中借助第一要素和第二要素，结合反问模版，适配出个性化的反问问句，尽可能地降低了交互的难度。
需要说明的是，在执行上述步骤210-步骤250的过程中，有可能选取不到第一要素，也有可能可以选取到第一要素，但是从知识库中匹配不到知识点，此时可以向用户发送预设的反问问句，如，“请问您遇到了什么 问题？”。还需要说明的是，在本说明书中，对于某些特殊的对话文本，可以有特定的问法，其可以通过配置完成。如，在第一要素与预设集合中的要素相同时，获取与要素相对应的预设的反问问句。根据接收的预设的反问问句的答案，确定用户意图识别结果。
经实验表明，通过本说明书上述实施例提出的用户意图识别方法，可以将用户意图识别准确率和召回率都提升5％。此外，通过上述针对性的反问问句的交互，用户更愿意与机器人进行多轮对话。
图3为本说明书另一个实施例提供的用户意图识别方法示意图。图3中，可以预先从知识库中的知识点抽取出三要素，即从知识库的知识点中抽取分别归属于“业务类型”、“框架动词”以及“问题类型”的要素。当接收到用户的对话文本时，通过计算对话文本与预先抽取的要素的识别概率值，来识别对话文本中的要素。之后，根据计算的识别概率以及阈值，将识别的要素划分为可信要素、候选要素以及未知要素。最后，通过可信要素从知识库中匹配对应的知识点。若匹配到对应的知识点，则直接输出用户意图。若匹配到多个知识点，则根据可信要素和候选要素从反问模板库中匹配对应的反问模板。之后根据匹配的反问模板来向用户输出反问问句，并根据反问问句的答案，识别用户意图。
与上述用户意图识别方法对应地，本说明书一个实施例还提供的一种用户意图识别装置，如图4所示，该装置包括：
获取单元401，用于获取用户的对话文本。
确定单元402，用于确定获取单元401获取的对话文本与多个预定义要素的关联度值，多个预定义要素是从知识库的知识点中抽取的，且多个预定义要素分别归属于N个类别，N为正整数。
此处的类别可以包括：业务类型、框架动词以及问题类型等。
可选地，确定单元402具体可以用于：
将对话文本输入要素识别模型，输出多个预定义要素对应的概率值，其中要素识别模型是预定义的用于计算输入的文本与多个预定义要素相匹 配的概率值的机器学习模型。
和/或，
计算对话文本与多个预定义要素的匹配度值。
根据概率值和/或匹配度值，确定对话文本与多个预定义要素的关联度值。
选取单元403，用于根据确定单元402确定的关联度值以及预设阈值，从多个预定义要素中选取第一要素。
可选地，选取单元403具体可以用于：
根据关联度值以及预设阈值，确定多个预定义要素对应的等级信息。
根据等级信息，从多个预定义要素中选取第一要素。
匹配单元404，用于根据选取单元403选取的第一要素，从知识库中匹配对应的知识点。
确定单元402，还用于根据匹配单元404匹配的知识点，确定用户意图识别结果。
可选地，确定单元402还可以用于：
根据等级信息，从多个预定义要素中选取第二要素。
根据第一要素、第二要素所归属的类别以及等级信息，从反问模板库中读取相匹配的反问模板，反问模板库用于记录归属于不同类别的要素的等级信息与反问模板之间的对应关系。
根据第一要素、所述第二要素以及反问模板，确定对应的反问问句。
向用户发送反问问句。
根据接收的反问问句的答案，确定用户意图识别结果。
可选地，该装置还可以包括：
发送单元405，用于向用户发送预设的反问问句。
确定单元402，还用于根据接收的预设的反问问句的答案，确定用户意图识别结果。
可选地，获取单元401，还用于当第一要素与预设集合中的要素相同时， 获取与要素相对应的预设的反问问句。
确定单元402，还用于根据接收的预设的反问问句的答案，确定用户意图识别结果。
本说明书上述实施例装置的各功能模块的功能，可以通过上述方法实施例的各步骤来实现，因此，本说明书一个实施例提供的装置的具体工作过程，在此不复赘述。
本说明书一个实施例提供的用户意图识别装置，获取单元401获取用户的对话文本。确定单元402确定获取的对话文本与多个预定义要素的关联度值。选取单元403根据关联度值以及预设阈值，从多个预定义要素中选取第一要素。匹配单元404根据选取的第一要素，从知识库中匹配对应的知识点。确定单元402根据匹配的知识点，确定用户意图识别结果。由此，可以快速地识别用户意图。
本领域技术人员应该可以意识到，在上述一个或多个示例中，本发明所描述的功能可以用硬件、软件、固件或它们的任意组合来实现。当使用软件实现时，可以将这些功能存储在计算机可读介质中或者作为计算机可读介质上的一个或多个指令或代码进行传输。
以上所述的具体实施方式，对本发明的目的、技术方案和有益效果进行了进一步详细说明，所应理解的是，以上所述仅为本发明的具体实施方式而已，并不用于限定本发明的保护范围，凡在本发明的技术方案的基础之上，所做的任何修改、等同替换、改进等，均应包括在本发明的保护范围之内。

Claims (14)

  1. 一种用户意图识别方法，其特征在于，包括：
    获取用户的对话文本；
    确定所述对话文本与多个预定义要素的关联度值，所述多个预定义要素是从知识库的知识点中抽取的，且所述多个预定义要素分别归属于N个类别，N为正整数；
    根据所述关联度值以及预设阈值，从所述多个预定义要素中选取第一要素；
    根据所述第一要素，从所述知识库中匹配对应的知识点；
    根据所述知识点，确定用户意图识别结果。
  2. 根据权利要求1所述的方法，其特征在于，所述确定所述对话文本与多个预定义要素的关联度值，包括：
    将所述对话文本输入要素识别模型，输出所述多个预定义要素对应的概率值，其中要素识别模型是预定义的用于计算输入的文本与多个预定义要素相匹配的概率值的机器学习模型；
    和/或，
    计算所述对话文本与所述多个预定义要素的匹配度值；
    根据所述概率值和/或所述匹配度值，确定所述对话文本与多个预定义要素的关联度值。
  3. 根据权利要求1或2所述的方法，其特征在于，所述根据所述关联度值以及预设阈值，从所述多个预定义要素中选取第一要素，包括：
    根据所述关联度值以及所述预设阈值，确定所述多个预定义要素对应的等级信息；
    根据所述等级信息，从所述多个预定义要素中选取第一要素。
  4. 根据权利要求3所述的方法，其特征在于，当匹配的知识点为多个时，所述根据所述知识点，确定用户意图识别结果，包括：
    根据所述等级信息，从所述多个预定义要素中选取第二要素；
    根据所述第一要素、所述第二要素所归属的类别以及等级信息，从反问模板库中读取相匹配的反问模板，所述反问模板库用于记录归属于不同类别的要素的等级信息与反问模板之间的对应关系；
    根据所述第一要素、所述第二要素以及所述反问模板，确定对应的反问问句；
    向用户发送所述反问问句；
    根据接收的所述反问问句的答案，确定用户意图识别结果。
  5. 根据权利要求3所述的方法，其特征在于，当所述等级信息为低等级时或者当未匹配到知识点时，还包括：
    向用户发送预设的反问问句；
    根据接收的所述预设的反问问句的答案，确定用户意图识别结果。
  6. 根据权利要求1所述的方法，其特征在于，还包括：
    当所述第一要素与预设集合中的要素相同时，获取与所述要素相对应的预设的反问问句；
    根据接收的所述预设的反问问句的答案，确定用户意图识别结果。
  7. 根据权利要求1-6任一项所述的方法，其特征在于，所述类别包括：业务类型、框架动词以及问题类型。
  8. 一种用户意图识别装置，其特征在于，包括：
    获取单元，用于获取用户的对话文本；
    确定单元，用于确定所述获取单元获取的所述对话文本与多个预定义要素的关联度值，所述多个预定义要素是从知识库的知识点中抽取的，且所述多个预定义要素分别归属于N个类别，N为正整数；
    选取单元，用于根据所述确定单元确定的所述关联度值以及预设阈值，从所述多个预定义要素中选取第一要素；
    匹配单元，用于根据所述选取单元选取的所述第一要素，从所述知识库中匹配对应的知识点；
    所述确定单元，还用于根据所述匹配单元匹配的所述知识点，确定用 户意图识别结果。
  9. 根据权利要求8所述的装置，其特征在于，所述确定单元具体用于：
    将所述对话文本输入要素识别模型，输出所述多个预定义要素对应的概率值，其中要素识别模型是预定义的用于计算输入的文本与多个预定义要素相匹配的概率值的机器学习模型；
    和/或，
    计算所述对话文本与所述多个预定义要素的匹配度值；
    根据所述概率值和/或所述匹配度值，确定所述对话文本与多个预定义要素的关联度值。
  10. 根据权利要求8或9所述的装置，其特征在于，所述选取单元具体用于：
    根据所述关联度值以及所述预设阈值，确定所述多个预定义要素对应的等级信息；
    根据所述等级信息，从所述多个预定义要素中选取第一要素。
  11. 根据权利要求10所述的装置，其特征在于，所述确定单元还用于：
    根据所述等级信息，从所述多个预定义要素中选取第二要素；
    根据所述第一要素、所述第二要素所归属的类别以及等级信息，从反问模板库中读取相匹配的反问模板，所述反问模板库用于记录归属于不同类别的要素的等级信息与反问模板之间的对应关系；
    根据所述第一要素、所述第二要素以及所述反问模板，确定对应的反问问句；
    向用户发送所述反问问句；
    根据接收的所述反问问句的答案，确定用户意图识别结果。
  12. 根据权利要求10所述的装置，其特征在于，还包括：
    发送单元，用于向用户发送预设的反问问句；
    所述确定单元，还用于根据接收的所述预设的反问问句的答案，确定用户意图识别结果。
  13. 根据权利要求8所述的装置，其特征在于，
    所述获取单元，还用于当所述第一要素与预设集合中的要素相同时，获取与所述要素相对应的预设的反问问句；
    所述确定单元，还用于根据接收的所述预设的反问问句的答案，确定用户意图识别结果。
  14. 根据权利要求8-13任一项所述的装置，其特征在于，所述类别包括：业务类型、框架动词以及问题类型。
PCT/CN2018/105192 2017-10-25 2018-09-12 用户意图识别方法及装置 Ceased WO2019080661A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711005661.3 2017-10-25
CN201711005661.3A CN107862005A (zh) 2017-10-25 2017-10-25 用户意图识别方法及装置

Publications (1)

Publication Number Publication Date
WO2019080661A1 true WO2019080661A1 (zh) 2019-05-02

Family

ID=61696342

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/105192 Ceased WO2019080661A1 (zh) 2017-10-25 2018-09-12 用户意图识别方法及装置

Country Status (3)

Country Link
CN (1) CN107862005A (zh)
TW (1) TWI700632B (zh)
WO (1) WO2019080661A1 (zh)

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107862005A (zh) * 2017-10-25 2018-03-30 阿里巴巴集团控股有限公司 用户意图识别方法及装置
CN110517672B (zh) * 2018-05-03 2023-08-15 阿里巴巴集团控股有限公司 用户意图识别方法、用户指令执行方法、系统及设备
CN108597503B (zh) * 2018-05-09 2021-04-30 科大讯飞股份有限公司 测试语料生成方法、装置、设备及可读写存储介质
CN110489635B (zh) * 2018-05-15 2023-05-05 阿里巴巴集团控股有限公司 数据对象搜索控制方法、装置及系统
CN109635271A (zh) * 2018-10-22 2019-04-16 阿里巴巴集团控股有限公司 一种用户意图识别方法、客服服务系统、装置及电子设备
CN110008308B (zh) * 2019-01-24 2023-06-02 创新先进技术有限公司 针对用户问句补充信息的方法和装置
CN113348502A (zh) * 2019-04-17 2021-09-03 深圳市欢太科技有限公司 语音识别方法、装置、存储介质及电子设备
CN110231985A (zh) * 2019-06-17 2019-09-13 三角兽（北京）科技有限公司 业务流程数据处理方法、装置、电子设备及存储介质
CN110765759B (zh) * 2019-10-21 2023-05-19 普信恒业科技发展（北京）有限公司 意图识别方法及装置
CN110727783B (zh) * 2019-10-23 2021-03-02 支付宝(杭州)信息技术有限公司 一种基于对话系统对用户问句提出反问的方法和装置
CN111368045B (zh) * 2020-02-21 2024-05-07 平安科技（深圳）有限公司 用户意图识别方法、装置、设备及计算机可读存储介质
CN111553162B (zh) * 2020-04-28 2023-09-22 腾讯科技（深圳）有限公司 一种意图识别的方法以及相关装置
CN111930854B (zh) * 2020-10-10 2021-01-08 北京福佑多多信息技术有限公司 意图预测的方法及装置
CN112328763A (zh) * 2020-11-04 2021-02-05 北京京东尚科信息技术有限公司 意图识别的方法、装置、对话方法和系统

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103049433A (zh) * 2012-12-11 2013-04-17 微梦创科网络科技（中国）有限公司 自动问答方法、自动问答系统及构建问答实例库的方法
CN103049548A (zh) * 2012-12-27 2013-04-17 安徽科大讯飞信息科技股份有限公司 电子渠道应用上的faq识别系统及方法
CN104050256A (zh) * 2014-06-13 2014-09-17 西安蒜泥电子科技有限责任公司 基于主动学习的问答方法及采用该方法的问答系统
US20170061004A1 (en) * 2015-08-31 2017-03-02 International Business Machines Corporation Managing question and answer sessions during a web conference
CN107862005A (zh) * 2017-10-25 2018-03-30 阿里巴巴集团控股有限公司 用户意图识别方法及装置

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8903924B2 (en) * 2011-12-09 2014-12-02 International Business Machines Corporation Aggregating data in electronic communications
CN108829682B (zh) * 2016-04-08 2022-10-28 上海智臻智能网络科技股份有限公司 计算机可读存储介质、智能问答方法及智能问答装置
CN106528759B (zh) * 2016-11-04 2019-04-02 上海智臻智能网络科技股份有限公司 智能问答系统的信息处理方法及装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103049433A (zh) * 2012-12-11 2013-04-17 微梦创科网络科技（中国）有限公司 自动问答方法、自动问答系统及构建问答实例库的方法
CN103049548A (zh) * 2012-12-27 2013-04-17 安徽科大讯飞信息科技股份有限公司 电子渠道应用上的faq识别系统及方法
CN104050256A (zh) * 2014-06-13 2014-09-17 西安蒜泥电子科技有限责任公司 基于主动学习的问答方法及采用该方法的问答系统
US20170061004A1 (en) * 2015-08-31 2017-03-02 International Business Machines Corporation Managing question and answer sessions during a web conference
CN107862005A (zh) * 2017-10-25 2018-03-30 阿里巴巴集团控股有限公司 用户意图识别方法及装置

Also Published As

Publication number Publication date
CN107862005A (zh) 2018-03-30
TW201917601A (zh) 2019-05-01
TWI700632B (zh) 2020-08-01

Similar Documents

Publication Publication Date Title
WO2019080661A1 (zh) 用户意图识别方法及装置
CN111666346B (zh) 信息归并方法、交易查询方法、装置、计算机及存储介质
CN112613917B (zh) 基于用户画像的信息推送方法、装置、设备及存储介质
CN110929125B (zh) 搜索召回方法、装置、设备及其存储介质
US9195910B2 (en) System and method for classification with effective use of manual data input and crowdsourcing
US9390378B2 (en) System and method for high accuracy product classification with limited supervision
US11176549B2 (en) Systems and methods for matching transactional data
CN107193974B (zh) 基于人工智能的地域性信息确定方法和装置
CN107015983A (zh) 一种用于智能问答中提供知识信息的方法与设备
CN112651236B (zh) 提取文本信息的方法、装置、计算机设备和存储介质
CN115618415A (zh) 敏感数据识别方法、装置、电子设备和存储介质
WO2017044409A1 (en) System and method of annotating utterances based on tags assigned by unmanaged crowds
US20240394629A1 (en) Recommendation system and operation method thereof
CN110321437A (zh) 一种语料数据处理方法、装置、电子设备及介质
CN108269122A (zh) 广告的相似度处理方法和装置
WO2019179030A1 (zh) 产品购买预测方法、服务器及存储介质
US11899770B2 (en) Verification method and apparatus, and computer readable storage medium
WO2016188334A1 (zh) 一种用于处理应用访问数据的方法与设备
CN114942979A (zh) 一种短文本问题匹配方法和装置
WO2017000341A1 (zh) 一种信息处理方法、装置以及终端
CN110020040A (zh) 查询数据的方法、装置和系统
CN111400440A (zh) 一种意图识别的方法及装置
CN119089237B (zh) 基于人工智能的精细化数据处理方法
CN111597453B (zh) 用户画像方法、装置、计算机设备及计算机可读存储介质
CN111782776B (zh) 一种通过填槽实现意图识别的方法和装置

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18869979

Country of ref document: EP

Kind code of ref document: A1