WO2020143844A1 - 意图分析方法、装置、显示终端及计算机可读存储介质 - Google Patents
意图分析方法、装置、显示终端及计算机可读存储介质 Download PDFInfo
- Publication number
- WO2020143844A1 WO2020143844A1 PCT/CN2020/074688 CN2020074688W WO2020143844A1 WO 2020143844 A1 WO2020143844 A1 WO 2020143844A1 CN 2020074688 W CN2020074688 W CN 2020074688W WO 2020143844 A1 WO2020143844 A1 WO 2020143844A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- sentence
- sample
- intent
- samples
- intention
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Definitions
- the present application relates to the technical field of deep learning algorithms, and in particular, to a deep learning-based intent analysis method, device, display terminal, and computer-readable storage medium.
- smart TVs have been widely popularized. Users can input voice to the smart TV to search for movies, music, and chat.
- the smart TV can identify the user's intention through the user's dialogue. If the user's intention matches the business of the smart TV Then, smart TV can carry out accurate product recommendation. For example, in an online advertising scenario, smart TV can select an advertising page based on user dialogue to make accurate advertising recommendations.
- Intent recognition generally uses short text matching, which plays a vital role in natural language processing tasks such as information retrieval, question answering, and dialogue systems.
- Conventional text matching methods rely on manually defined templates and rules.
- the adjusted intent analysis model lacks generalization and portability for different task requirements.
- the main purpose of the present application is to provide an intent recognition analysis method, device, display terminal and computer readable storage medium based on deep learning, aiming to solve the technical problem that the existing intent recognition module lacks generalization capability and portability.
- the intent analysis method based on deep learning includes:
- the sample text to be recognized and sample application sentences arbitrarily selected from each of the intention categories are input as comparison text into the deep network model;
- the present application also provides an intention analysis device.
- the intention analysis device includes:
- the sentence allocation module is configured to determine a preset intention category involved in the intention analysis for user sentences; assign a preset first value sentence sample to each preset intention category;
- the sample generation module is configured to select sentence pairs from the same sentence sample of the intention classification to generate a positive sample; select sentence pairs from sentence samples of different said intention classification to generate a negative sample; the positive sample and the negative sample are common Constitute a training set;
- the training module is configured to train a preset intent analysis model based on the training set to generate a deep network model, wherein the similarities obtained by inputting the positive samples and negative samples to the deep network model tend to 1 and 0, respectively;
- An intent recognition module configured to, when receiving the text to be recognized, input the text to be recognized and sample application sentences arbitrarily selected from each of the intent categories as comparison text into the deep network model; If the target application sentence sample with the similarity of the text to be recognized is greater than a preset similarity threshold, it is determined that the intent of the text to be recognized belongs to the target intent category to which the target application sentence sample with the highest similarity belongs.
- sample generation module is also set to:
- a pair of sentences equal to the number of positive samples is randomly selected from the primary negative samples to form a negative sample.
- the training module includes:
- the word segmentation unit is set to perform word segmentation on the positive and negative samples of the training set, based on the preset word vector algorithm, the words in the positive and negative samples after word segmentation are sequentially converted into corresponding word vectors to generate and Describe the positive sample sentence vector and the negative sample sentence vector corresponding to the positive sample and the negative sample, respectively;
- the training unit is configured to input each of the positive sample sentence vector and the negative sample sentence vector to a preset intention analysis model for training to generate a deep network model.
- the training unit is also set to:
- the node weights of the intention analysis model are adjusted one by one, so that the subsequent first test similarity tends to 1, and the subsequent second test similarity tends to At 0, until the traversal of the positive sample sentence vector and the negative sample sentence vector is completed;
- the intention analysis model after training all the positive sample sentence vectors and negative sample sentence vectors is taken as the deep network model.
- the training unit is also set to:
- the intent recognition module is also set to:
- the average value of the similarity of each unit in each of the intention categories is taken as the similarity between the corresponding intention category and the text to be recognized.
- the intent recognition module is also set to:
- the present application also provides a display terminal including a memory, a processor, and computer-readable instructions stored on the memory and operable on the processor, the computer-readable instructions are described by When the processor executes, it implements the steps of the deep learning-based intent analysis method as described above.
- the present application also provides a computer-readable storage medium on which computer-readable instructions are stored.
- a deep learning-based intent analysis method as described above is implemented step.
- This application constructs the original data by assigning sentence samples with the first numerical value to each preset intention classification.
- features can be automatically extracted from the original data to form sentence pairs, eliminating the overhead of manually designing features, and can be more Extracting features finely and accurately enhances the generalization ability of the intention analysis model and its portability for different task requirements.
- select sentence pairs based on sentence samples assigned by each intention to generate positive samples and negative samples to form a training set generation engine. Only a small amount of training data (that is, sentence samples) is needed to generate a large number of training sets ( That is positive samples and negative samples) for the intention analysis model training.
- a deep network model is generated, and after receiving the text to be recognized, the text to be recognized and sample application sentences are input into the deep network model in a sentence pairing manner to analyze and identify the waiting
- the intent classification of the recognized text improves the recognition accuracy of the intent classification.
- FIG. 1 is a schematic structural diagram of an embodiment of a display terminal of this application.
- FIG. 2 is a schematic flowchart of an embodiment of a deep learning-based intent analysis method of the present application
- Figure 3 is the frame structure diagram of the DSSM model
- FIG. 4 is a schematic structural diagram of an embodiment of an intention analysis model of the present application.
- generalization ability refers to the adaptability of machine learning algorithms to fresh samples.
- the purpose of learning is to learn the rules hidden behind the data.
- the trained network can also give a suitable output. This ability is called generalization ability.
- the main solutions of the embodiments of the present application are: determining the preset intent classification involved in the intention analysis for user sentences; assigning a preset first value sentence sample to each preset intent category; sentence samples from the same intent category Select sentence pairs in the middle to generate positive samples; select sentence pairs from the sentence samples of different intent classifications to generate negative samples; positive samples and negative samples together constitute the training set; train the preset intent analysis model based on the training set to generate depth Network model, wherein the similarity of the positive and negative samples input to the deep network model tends to 1 and 0, respectively; when the text to be recognized is received, the text to be recognized and the application sentence samples randomly selected from each intention classification Entered into the deep network model as the comparison text; if there is a target application sentence sample whose similarity to the text to be recognized is greater than the preset similarity threshold, it is determined that the intention of the text to be recognized belongs to the target application sentence sample with the largest similarity Target intention classification.
- features can be automatically extracted from sentence samples as raw data (forming sentence pairs), eliminating the overhead of manually designing features, and enabling more precise and accurate feature extraction.
- the deep learning model is like an engine, and the massive training data is like fuel. These two are also indispensable for the artificial intelligence of the machine.
- the training data is very critical in deep learning. This application establishes a deep network model based on the intention analysis model to improve the accuracy of intention recognition from the automatic generation of the training set.
- FIG. 1 is a schematic structural diagram of a portable terminal of a hardware operating environment involved in a solution of an embodiment of the present application.
- the deep learning-based intent analysis method, intent analysis device, display terminal, and computer-readable storage medium provided by the present application relate to the technical field of deep learning algorithms.
- the display terminal may be a terminal with a display function such as a smart TV, a tablet computer, and a smart phone. equipment.
- the display terminal may include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002.
- the communication bus 1002 is used to implement connection communication between these components.
- the user interface 1003 may include an input unit such as a keyboard, and the optional user interface 1003 may also include a standard wired interface and a wireless interface.
- the network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).
- the memory 1005 may be a high-speed RAM memory or a non-volatile memory (non-volatile memory), such as a disk memory.
- the memory 1005 may optionally be a storage device independent of the foregoing processor 1001.
- the structure of the display terminal shown in FIG. 1 does not constitute a limitation on the portable terminal, and may include more or less components than those illustrated, or combine certain components, or arrange different components.
- the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and computer-readable instructions for implementing an intent analysis method based on deep learning.
- the network interface 1004 is mainly configured to connect with a background server (such as a positioning server) to perform data communication with the background server;
- the user interface 1003 is mainly configured to connect with a client (such as an associated host terminal), and The client performs data communication;
- the processor 1001 may be set to call computer-readable instructions stored in the memory 1005 and perform the following operations:
- the sample text to be recognized and sample application sentences arbitrarily selected from each of the intention categories are input as comparison text into the deep network model;
- the intent analysis method based on deep learning includes:
- Step S10 Determine a preset intention category involved in the intention analysis of the user sentence; assign a preset first value sentence sample to each preset intention category;
- the intention classification of the user sentence intention analysis applicable to the present application can be set at the factory of the display terminal or set by the user on the display terminal.
- the set intention classifications together constitute a set of preset intention classifications. It is possible to grab real user questions from the Internet for each preset intent category, until a sentence sample with the first numerical value assigned to each preset intent category (that is, a real user question).
- the sentence samples may also include artificially designed question sentences other than real user questions.
- the preset intention classification includes 12 categories, and the preset first value may be 1000.
- Examples of intention classification and sentence samples are as follows in Table 1:
- the 12 intent categories correspond to different sentence sample examples.
- the sentence samples in Table 1 have completed word segmentation processing.
- Step S20 select sentence pairs from sentence samples of the same intention category to generate positive samples; select sentence pairs from sentence samples of different intention categories to generate negative samples; the positive samples and the negative samples together constitute the training set;
- Pairs of sentence samples can be paired to generate a training set. If two sentence samples are marked as 1 in the same intent classification and 0 in different intent classifications, refer to the following table 2:
- the generated training set In order to train an effective intent analysis model, the generated training set must be large enough, at least tens of millions of pairs of matching groups (that is, the number of pairs of sentence sample matching groups), through the above sentence sample matching method to generate the training set, Each intent type only needs a small number of sentence samples to generate a large amount of training data (ie sentence sample pairing group).
- the selected sentence sample pairing group needs to cover all intent classifications. For example, there are 12 preset intention classifications. If there are two pairs of intention classifications (including pairings of the same intention), 12 quadratic intention classification pairs can be generated, that is, 144 intention classification pairs, 12 of which It is a pair of sentence samples of the same intent classification (ie positive samples), 132 is a pair of sentence samples of different intent allocations (negative samples).
- the training set should contain 12 positive and negative samples of intent classification , Positive samples and negative samples together constitute the training set.
- step S20 sentence pairs are selected from sentence samples of different intention classifications.
- the steps of generating negative samples include:
- the training set is more suitable for the real scene of intent analysis, and the training set is targeted and accurate for the intent analysis model.
- N n (c-1)*P n
- the number of positive samples is 5.994*10 6 and the number of primary negative samples is 65.934*10 6.
- Statements with the same number of positive samples ie 5.994*10 6
- this application has designed a training set generation mechanism. With a small amount of training data (12 intent categories totaling 12,000 sentence samples), at least 12 million sentence pairs are generated to form a massive training set for intent analysis model training.
- Step S30 Train the preset intention analysis model based on the training set to generate a deep network model, wherein the similarity obtained by inputting the positive samples and negative samples to the deep network model tends to 1 and 0, respectively;
- the intention analysis model of this application can be improved based on DSSM (Deep Structured Similarity Model).
- DSSM Deep Structured Similarity Model
- the main purpose of DSSM is to calculate the similarity of the semantic space.
- the basic structure of DSSM is shown in Figure 3.
- DSSM is mainly used to model and calculate the similarity between each input sentence Q and multiple sentences D1...Dn, through the original text, after many Layer linear projection, each sentence input will obtain 128-dimensional hidden semantic vector for similarity calculation.
- step S30 may include:
- Step S31 segment the positive and negative samples of the training set, and convert the words in the positive and negative samples after word segmentation into corresponding word vectors based on the preset word vector algorithm, and generate corresponding to the positive and negative samples respectively Positive sample sentence vector and negative sample sentence vector;
- Step S32 input each positive sample sentence vector and negative sample sentence vector to a preset intention analysis model for training, and generate a deep network model.
- the positive and negative samples of the training set can be segmented based on the preset word segmentation algorithm.
- the preset word segmentation algorithm can include character matching method, semantic syntax based understanding method, statistical method, etc., so that the sentence samples in the positive sample and the negative sample are Split into multiple words.
- the words in the positive and negative samples after word segmentation can be sequentially converted into corresponding word vectors based on a preset word vector algorithm (such as word2vec word vector algorithm).
- word vectors such as word2vec word vector algorithm.
- the basic idea of word vectors is to express discrete words as vectors in a continuous space, and express the correlation (semantics, grammar, etc.) between words as the distance between vectors.
- the positive samples and the negative samples that is, the paired sentences (X, Y)
- Sim(X,Y) is directly proportional to the similarity between the two matching sentences.
- the word vectors in the same sentence sample together constitute the sentence vector set of the sentence sample, that is, the sentence vector of the sentence sample, thereby generating positive sample sentence vectors and negative sample sentence vectors corresponding to positive samples and negative samples, respectively.
- the paired sentences in the positive sample and the negative sample are all word vector sets (that is, sentence vectors).
- each positive sample sentence vector and negative sample sentence vector are input to a preset intention analysis model for model training, for example, a negative sample sentence vector includes sentence X(w1, w2..., wn) and sentence Y(w1 , W2,...wn), input X and Y into the DSSM model (that is, an embodiment of the intention analysis module) as shown in FIG. 4, and the blank boxes f(.) and g(.) can be replaced by any model , For example: fully connected FC, convolutional neural network CNN, recurrent neural network RNN, etc., the intention analysis model is used to measure the semantic distance between two sentences.
- a negative sample sentence vector includes sentence X(w1, w2..., wn) and sentence Y(w1 , W2,...wn), input X and Y into the DSSM model (that is, an embodiment of the intention analysis module) as shown in FIG. 4, and the blank boxes f(.) and g(.) can be replaced by any model , For example: fully connected
- this application selects CNN convolutional neural networks to convert high-dimensional feature vectors into continuous vectors in low-dimensional spaces, because CNN is more suitable for extracting the features of sentences, and the features are time-series.
- Feature dimensionality reduction is to reduce the number of features. This has two purposes. One is to speed up the calculation of the algorithm. In addition, it can reduce noise and effectively improve the accuracy of the intention classification analysis.
- the DSSM module can be used with different loss functions to achieve specific functions.
- the cosine similarity can be used to calculate the similarity and become a Regression model, where sentences are short for sample sentences.
- this application uses cosine similarity (cosine similarity) to measure the semantic relevance between the user’s search term (ie, the text to be recognized) and the candidate document (ie, the application sentence sample).
- cosine similarity cosine similarity
- the model task type Supports regression in the [0,1] range. To aid understanding, a specific example is shown, if the sentence X and the sentence Y corresponding to the sentence vector are:
- the semantic distance used in the training process of the intention analysis model is Relu(cos ⁇ ).
- step S32 includes:
- Step S321 Input each positive sample sentence vector and negative sample sentence vector to the preset intention analysis model one by one for training;
- Step S322 Based on the first test similarity of each positive sample vector and the second test similarity of negative sample vector one by one, adjust the node weight of the intention analysis model one by one, so that the subsequent first test similarity tends to 1, and the subsequent second test The similarity tends to 0 until the positive sample sentence vector and the negative sample sentence vector are traversed;
- Step S323 the intention analysis model after training all positive sample sentence vectors and negative sample sentence vectors as a deep network model.
- the first test similarity of the two sentences of each positive sample sentence vector and the second of the two sentences of each negative sample sentence vector Test similarity Since the positive sample sentence vectors have the same intent classification, the standard reference value of the first test similarity is 1; therefore, when training the intent analysis model based on the positive sample sentence vectors, the weight of each node in the preset intent analysis model is continuously adjusted, In order to gradually increase the first test similarity to its standard parameter value of 1.
- the standard reference value of the second test similarity is 0, so when training the intention analysis model based on the negative sample sentence vectors, the weights of each node in the preset intention analysis model are continuously adjusted, In order to gradually move the similarity of the second test towards its standard parameter value of 0. In this way, input positive samples and negative samples to the intent analysis model cyclically, and adjust the weight of each node in the intent analysis model synchronously until the traversal of the positive sample sentence vector and the negative sample sentence vector is completed. After completing all positive sample sentence vector and negative sample sentence vector pairs After training the intent analysis model, the intent analysis module after training is the deep network model.
- Step S40 when receiving the text to be recognized, input the sample of the application sentence to be recognized from the text to be recognized and each intention classification as the comparison text into the deep network model;
- the text to be recognized is obtained based on the semantic recognition technology, and one or more sentence samples are randomly selected from each preset intent classification as the waiting
- the application sentence samples of the recognition text for intent classification and comparison, and the word to be recognized and the application sentence samples are segmented into the intent analysis model as the comparison text, and the similarity between the text to be recognized and each application sentence sample is obtained.
- step S50 if there is a target application sentence sample whose similarity with the text to be recognized is greater than a preset similarity threshold, it is determined that the intent of the text to be recognized belongs to the target intention category to which the target application sentence sample with the highest similarity belongs.
- the display terminal After calculating the similarity between all application sentence samples and the text to be recognized, compare the calculated similarity with a preset similarity threshold, if there is a target application sentence sample with a similarity to the text to be recognized greater than the preset similarity threshold , It is determined that the intent of the text to be recognized belongs to the target intent category to which the target application sentence sample with the highest similarity belongs, and the display terminal can be controlled to perform precise advertisement push and page push based on the target intent category; if there is no similarity to the text to be recognized
- the target application sentence sample with the preset similarity threshold value determines that the intention of the text to be recognized does not belong to the preset intention classification.
- the preset similarity threshold is 0.6, and 12 application sentence samples corresponding to 12 intent categories are randomly selected to obtain the similarity between the text to be recognized and the 12 application sentence samples. If these 12 similarities are not If it is greater than 0.6, it is determined that the intent of the text to be recognized does not belong to the preset 12 intent categories. If there are two similarities of 0.65 and 0.7 with a similarity greater than 0.6, the sample sentence of the application sentence with a similarity of 0.7 is matched with the type of the text intent The target application sentence sample.
- the original data is constructed by assigning a sentence sample with a first numerical value to each preset intention classification.
- features can be automatically extracted from the original data to form sentence pairs, eliminating the overhead of manually designing features , And can extract features more finely and accurately, enhancing the generalization ability of the intention analysis model and its portability for different task requirements.
- a deep network model is generated, and after receiving the text to be recognized, the text to be recognized and sample application sentences are input into the deep network model in a sentence-pairing manner for analysis and recognition
- the intention classification of the text to be recognized improves the recognition accuracy of the intention classification.
- the intent analysis method based on deep learning of the present application further includes:
- Step S61 Obtain the false acceptance rate of the deep network model that judges sentences of different intention classifications as the same intention classification under different test similarity thresholds;
- Step S62 Receive an externally input intention analysis requirement, and use the test similarity threshold corresponding to the false acceptance rate adapted to the intention analysis requirement as the final preset similarity threshold of the deep network model.
- Different test similarity thresholds can be set to test the deep network model of the application based on the pre-selected test set to obtain errors under different test similarity thresholds that the deep network model judges different intention classification sentences in the test set as the same intent classification Acceptance rate.
- false acceptance refers to: the sentence with different intentions is processed by the deep network model and the test similarity threshold, and the judgment result is the same intention. That is, the false acceptance situation means that the similarity of sentences with different intent classifications analyzed by the deep network model is greater than the test similarity threshold selected at this time.
- the false acceptance rate is equal to the ratio of the number of false acceptances to the total logarithm of sentences with different intention classification.
- the test similarity thresholds include 0.6, 0.7, and 0.8. Enter 100 pairs of sentences with different intention classifications into the depth Network model, corresponding to 100 similarities corresponding to 100 pairs of sentences with different intention classification. Among these 100 similarities, there are 5 values greater than 0.6, 3 values greater than 0.7, and 1 greater than 0.8. Then, the false acceptance rate of the test similarity threshold 0.6 is 5%, the false acceptance rate of the test similarity threshold 0.7 is 3%, and the false acceptance rate of the test similarity threshold 0.8 is 1%.
- the intent analysis needs are customer service transfers with high professional requirements and high real-time answers.
- the customer service with the wrong professional direction will be assigned to dock the user consultation, resulting in customer service and user
- the waste of time for example, when the user inquires about the product upgrade service to the display terminal, and the matching intention is classified as shopping service at this time, it is easy to cause the customer service of the shopping service to fail to solve the doubts of the user's product upgrade service, resulting in a waste of time. Therefore, the customer service transfer has a low tolerance for the false acceptance rate, and the test similarity threshold should be selected to a higher value.
- the test similarity threshold 0.8 is selected as the default similarity threshold for the customer service transfer. Therefore, in this embodiment, the preset similarity threshold of the deep network model can be flexibly adjusted according to the intention analysis requirements, which extends the application scenario of the deep network model of the present application.
- the step of inputting the sample text to be recognized and the application sentence samples arbitrarily selected from each intent classification as the comparison text into the deep network model includes:
- the average value of the similarity of each unit in each intent category is taken as the similarity of the corresponding intent category and the text to be recognized.
- the intent recognition task can be abstracted into a classification problem.
- the text to be recognized is Tx.
- the preset intent classification includes 12 categories. Determine which one belongs to the 12 intent categories.
- the to-be-recognized text is compared with the application sentence samples (that is, ⁇ T i c ⁇ ) of each intent classification in the intent list.
- the deep network model calculates the to-be-recognized text Tx and the intent test text list ⁇ T i c ⁇
- the matching degree that is, unit similarity of all texts (ie sentence samples) sim(T x ,T i c ) of each type of intention classification in ⁇ ,sim(T x ,T i c ) ⁇ (0,1),
- the closer sim(T x ,T i c ) is to 0, the lower the similarity of two sentences is.
- a second number of application sentence samples are randomly selected in each intent category, that is, m samples in each intent category are compared with the input to-be-recognized text Tx to calculate
- the similarity between the text and each intention classification improves the accuracy of the similarity between the text to be recognized and each intention classification.
- r(x, c) can be based on the following formula
- the methods in the above embodiments can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware, but in many cases the former is better Implementation.
- the technical solution of the present application can be embodied in the form of a software product in essence or part that contributes to the prior art, and the computer software product is stored in a storage medium (such as ROM/RAM) as described above , Disk, CD), including several instructions to make a terminal device (which can be a mobile phone, computer, server, air conditioner, or network equipment, etc.) to perform the method described in each embodiment of the present application.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Machine Translation (AREA)
Abstract
一种基于深度学习的意图分析方法、装置、显示终端和计算机可读存储介质,该方法包括:为预设的各意图分类分配预设第一数值的语句样品;从同一意图分类的语句样品中挑选语句配对,生成正样品;从不同意图分类的语句样品中挑选语句配对,生成负样品;基于训练集对预设的意图分析模型进行训练,生成深度网络模型,其中正样品和负样品输入至深度网络模型所得相似度分别趋向1和0(S30);当接收到待识别文本时,将待识别文本和各意图分类中的应用语句样品作为比对文本输入至意图分析模型中;若存在目标应用语句样品,则确定待识别文本的意图属于目标意图分类。该方法增强了意图分析模型的泛化能力及其面向不同任务需求的可移植性。
Description
本申请要求于2019年1月10日提交中国专利局、申请号为201910029832.9、发明名称为“意图分析方法、装置、显示终端及计算机可读存储介质”的中国专利申请的优先权,其全部内容通过引用结合在申请中。
本申请涉及深度学习算法的技术领域,尤其涉及一种基于深度学习的意图分析方法、装置、显示终端及计算机可读存储介质。
随着科技技术的发展,智能电视得到广泛普及,用户对智能电视输入语音来搜索电影、音乐和聊天,智能电视可通过用户的对话来识别用户的意图,如果用户的意图和智能电视的业务吻合,那么智能电视就可以开展精准的产品推荐,例如在线广告场景中,智能电视可以基于用户对话选出广告页面,以做精准的广告推荐。
意图识别一般会应用短文本匹配,它在信息检索、问答、对话系统等自然语言处理任务中起着至关重要的作用。常规的文本匹配方法都依赖于手动定义的模板和规则,调整的意图分析模型缺乏泛化能力及其面向不同任务需求的可移植性。
上述内容仅用于辅助理解本申请的技术方案,并不代表承认上述内容是现有技术。
发明内容
本申请的主要目的在于提供一种基于深度学习的意图识别分析方法、装置、显示终端及计算机可读存储介质,旨在解决现有的意图识别模块缺乏泛化能力和可移植性的技术问题。
为实现上述目的,本申请提供一种基于深度学习的意图分析方法,所述基于深度学习的意图分析方法包括:
确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;
从同一所述意图分类的语句样品中挑选语句配对,生成正样品;从不同所述意图分类的语句样品中挑选语句配对,生成负样品;所述正样品和负样品共同构成训练集;
基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至所述深度网络模型所得相似度分别趋向1和0;
当接收到待识别文本时,将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中;
若存在与所述待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定所述待识别文本的意图属于相似度最大的所述目标应用语句样品所属的目标意图分类。
本申请还提供一种意图分析装置,所述意图分析装置包括:
语句分配模块,设置为确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;
样品生成模块,设置为从同一所述意图分类的语句样品中挑选语句配对,生成正样品;从不同所述意图分类的语句样品中挑选语句配对,生成负样品;所述正样品和负样品共同构成训练集;
训练模块,设置为基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至所述深度网络模型所得相似度分别趋向1和0;
意图识别模块,设置为当接收到待识别文本时,将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中;若存在与所述待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定所述待识别文本的意图属于相似度最大的所述目标应用语句样品所属的目标意图分类。
可选地,所述样品生成模块还设置为:
从不同所述意图分类的语句样品中挑选语句配对,生成初级负样品;
从所述初级负样品中随机选取与所述正样品个数相等的语句配对,构成负样品。
可选地,所述训练模块包括:
分词单元,设置为对所述训练集的正样品和负样品进行分词,基于预设词向量算法将分词后的所述正样品和负样品中的词语依次转换为对应的词向量,生成与所述正样品和负样品分别对应的正样品句向量和负样品句向量;
训练单元,设置为将各所述正样品句向量和负样品句向量输入至预设的意图分析模型进行训练,生成深度网络模型。
可选地,所述训练单元还设置为:
将各所述正样品句向量和负样品句向量逐个输入至预设的意图分析模型进行训练;
逐个基于各正样品向量的第一测试相似度和负样品向量的第二测试相似度,逐次调整意图分析模型的节点权重,使后续第一测试相似度趋向于1,后续第二测试相似度趋向于0,直至所述正样品句向量和负样品句向量遍历完毕;
将所有所述正样品句向量和负样品句向量训练后的意图分析模型,作为深度网络模型。
可选地,所述训练单元还设置为:
在开始接收待识别文本之前,获取在不同测试相似度阈值下,所述深度网络模型对不同意图分类句子判定为相同意图分类的错误接受率;
接收外部输入的意图分析需求,将所述意图分析需求适配的错误接受率所对应的测试相似度阈值,作为所述深度网络模型最终的预设相似度阈值。
可选地,所述意图识别模块还设置为:
为每个所述意图分类随机选取预设第二个数的应用语句样品;
基于所述深度网络模型,获取所述待识别文本分别与各所述意图分类的应用语句样品的单元相似度;
将各所述意图分类中各单元相似度的平均值作为对应所述意图分类与待识别文本的相似度。
可选地,所述意图识别模块还设置为:
若不存在与所述待识别文本的相似度大于所述相似度阈值的目标应用语 句样品,则确定所述待识别文本的意图不属于预设的意图分类。
本申请还提供一种显示终端,所述显示终端包括存储器、处理器、及存储在所述存储器上并可在所述处理器上运行的计算机可读指令,所述计算机可读指令被所述处理器执行时实现如上所述的基于深度学习的意图分析方法的步骤。
本申请还提供一种计算机可读存储介质,所述计算机可读存储介质上存储有计算机可读指令,所述计算机可读指令被处理器执行时实现如上述的基于深度学习的意图分析方法的步骤。
本申请通过为预设的各意图分类分配第一数值的语句样品构建原始数据,基于深度学习思想,可自动从原始数据中抽取特征以形成语句配对,免去人工设计特征的开销,并且能够更精细和准确地抽取特征,增强了意图分析模型的泛化能力及其面向不同任务需求的可移植性。再者,基于各意图分配的语句样品挑选语句配对,生成正样品和负样品,形成一种训练集产生机擎,只需要少量的训练数据(即语句样品),就可以产生大量的训练集(即正样品和负样品)以供意图分析模型训练。再者,在深度学习和训练集产生机擎的支持下,生成深度网络模型,在接收待识别文本后,以句子配对方式将待识别文本和应用语句样品输入深度网络模型,以分析和识别待识别文本的意图分类,提高了意图分类的识别准确性。
图1为本申请显示终端一实施例的结构示意图;
图2为本申请基于深度学习的意图分析方法一实施例的流程示意图;
图3为DSSM模型的框架结构图;
图4为本申请意图分析模型一实施例的框架结构示意图。
本申请目的的实现、功能特点及优点将结合实施例,参照附图做进一步说明。
应当理解,此处所描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
在意图识别或文本识别技术领域,过去对于研究文本匹配主要集中在人工定义特征的关系学习,这些方法都依赖于手动定义的模板和规则,因此限制了调整良好的模型的泛化能力及其面向不同人物需求的可移植性。其中,泛化能力(generalization ability)是指机器学习算法对新鲜样本的适应能力。学习的目的是学到隐含在数据背后的规律,对具有同一规律的学习集以外的数据,经过训练的网络也能给出合适的输出,该能力称为泛化能力。
本申请实施例的主要解决方案是:确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;从同一意图分类的语句样品中挑选语句配对,生成正样品;从不同意图分类的语句样品中挑选语句配对,生成负样品;正样品和负样品共同构成训练集;基于训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至所述深度网络模型所得相似度分别趋向1和0;当接收到待识别文本时,将待识别文本和各意图分类中任意选取的应用语句样品作为比对文本输入至深度网络模型中;若存在与待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定待识别文本的意图属于相似度最大的目标应用语句样品所属的目标意图分类。
在本申请中,基于深度学习可以自动从作为原始数据的语句样品中抽取特征(形成语句配对),免去人工设计特征的开销,并且能够更精细和准确地抽取特征。此外,在深度学习算法中,深度学习模型就好像引擎,海量的训练数据就好像燃料,这两者对于机器的人工智能而言同样缺一不可,训练数据在深度学习中是非常关键的。本申请建立一个基于意图分析模型生成深度网络模型,从训练集自动生成方面提升意图识别的准确率。
如图1所示,图1是本申请实施例方案涉及的硬件运行环境的随身终端结构示意图。
本申请提供的基于深度学习的意图分析方法、意图分析装置、显示终端及计算机可读存储介质涉及深度学习算法的技术领域,显示终端可以为智能电视、平板电脑、智能手机等具有显示功能的终端设备。
如图1所示,该显示终端可以包括:处理器1001,例如CPU,网络接口1004,用户接口1003,存储器1005,通信总线1002。其中,通信总线1002用于实现这些组件之间的连接通信。用户接口1003可以包括输入单元比如微键盘(Keyboard),可选用户接口1003还可以包括标准的有线接口、无线接口。网络接口1004可选的可以包括标准的有线接口、无线接口(如WI-FI接口)。存储器1005可以是高速RAM存储器,也可以是稳定的存储器(non-volatile memory),例如磁盘存储器。存储器1005可选的还可以是独立于前述处理器1001的存储装置。
本领域技术人员可以理解,图1中示出的显示终端结构并不构成对随身终端的限定,可以包括比图示更多或更少的部件,或者组合某些部件,或者不同的部件布置。
如图1所示,作为一种计算机存储介质的存储器1005中可以包括操作系统、网络通信模块、用户接口模块以及实现基于深度学习的意图分析方法的计算机可读指令。
在图1所示的显示终端中,网络接口1004主要设置为连接后台服务器(如定位服务器),与后台服务器进行数据通信;用户接口1003主要设置为连接客户端(如关联的主机终端),与客户端进行数据通信;而处理器1001可以设置为调用存储器1005中存储的计算机可读指令,并执行以下操作:
确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;
从同一所述意图分类的语句样品中挑选语句配对,生成正样品;从不同所述意图分类的语句样品中挑选语句配对,生成负样品;所述正样品和负样品共同构成训练集;
基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至所述深度网络模型所得相似度分别趋向1和0;
当接收到待识别文本时,将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中;
若存在与所述待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定所述待识别文本的意图属于相似度最大的所述目标应用语句样 品所属的目标意图分类。
基于上述显示终端,提供一种基于深度学习的意图分析方法,参照图2,基于深度学习的意图分析方法包括:
步骤S10,确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;
本申请适用的用户语句意图分析的意图分类可在显示终端出厂设置或用户对显示终端进行自行设置,设置的意图分类共同构成预设意图分类的集合。可为每个预设意图分类从网上抓取真实的用户问句,直至为预设的各意图分类分配第一数值的语句样品(即真实的用户问句)。当然,语句样品也可包括除真实用户问句之外的人工设计问句。
例如,预设的意图分类包括12类,预设第一数值可为1000,意图分类和语句样品的示例如下表1:
表1
在表1中,12个意图分类分别对应不同语句样品示例,表1中的语句样品已经完成了分词处理。
步骤S20,从同一意图分类的语句样品中挑选语句配对,生成正样品;从不同意图分类的语句样品中挑选语句配对,生成负样品;正样品和负样品共同构成训练集;
可对语句样品进行两两配对以生成训练集,如果两个语句样品在同一个意图分类中标记为1、在不同意图分类中标记为0,可参照下表2:
表2
为了训练有效的意图分析模型,生成的训练集的集合要足够大,至少要千万级别的配对组数(即语句样本配对组的组数),通过以上的语句样本配对生成训练集的方式,每个意图类型只需要少量的语句样本就可以生成大量的训练数据(即语句样本配对组)。
为了有效训练意图分析模型产生训练集时,选取的语句样本配对组需涵盖所有意图分类。比如,预设的意图分类有12钟,若意图分类之间两两配对(包括同一意图的配对),则可以产生12的二次方个意图分类配对,即144个意图分类配对,其中12个是同一意图分类的语句样品配对(即正样品), 132个是不同意图分配的语句样品配对(负样品),为了有效训练意图分析模型,训练集应当包含12个意图分类的正样品和负样品,正样品和负样品共同构成训练集。
可选地,构成训练集的正样品和负样品的数量相对平衡,步骤S20中从不同意图分类的语句样品中挑选语句配对,生成负样品的步骤包括:
从不同意图分类的语句样品中挑选语句配对,生成初级负样品;从初级负样品中随机选取与正样品个数相等的语句配对,构成负样品。使训练集更加贴合意图分析的真实场景,提供训练集对意图分析模型的训练针对性和准确有效性。
接上一段例子,12个意图分类(即c=12),每个意图分类有1000个语句样品(N
s=1000),从而可以产生正样品量Pn为:
N
n=(c-1)*P
n
N
n=(12-1)*5.994*10
6=65.934*10
6
因此,从而可以产生正样品数量为5.994*10
6和初级负样品数量为65.934*10
6,可从初级负样品中随机挑选与正样品个数相同(即5.994*10
6)的语句配对以构成负样品,从而正样品和负样品共接近1200万个句子对(即语句样本配对)。
从而本申请设计了一个训练集产生机制,以少量训练数据(12个意图分类共计1.2万个语句样品),产生了至少1200万个句子对以构成海量的训练集以供意图分析模型训练。
步骤S30,基于训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至深度网络模型所得相似度分别趋向1和0;
本申请的意图分析模型可基于DSSM(Deep Structured Similarity Model,深度语义相似度模型)改进得出。DSSM主要用途在于计算语义空间的相似度。DSSM基本结构如图3所示,在DSSM的相关论文中,DSSM主要用来建模计算每一个输入句子Q与多个句子D1...Dn之间的相似度,通过原始的文本,经过多层线性投影,每个句子输入都会获得128维的隐语义向量,以进行相似度计算。
在本申请中,步骤S30可包括:
步骤S31,对训练集的正样品和负样品进行分词,基于预设词向量算法将分词后的正样品和负样品中的词语依次转换为对应的词向量,生成与正样品和负样品分别对应的正样品句向量和负样品句向量;
步骤S32,将各正样品句向量和负样品句向量输入至预设的意图分析模型进行训练,生成深度网络模型。
可基于预设分词算法对训练集的正样品和负样品进行分词,预设分词算法可包括字符匹配法、基于语义句法的理解法、统计法等,从而正样品和负样品中的语句样品被拆分为多个词语。然后可基于预设词向量算法(如word2vec词向量算法)将分词后的正样品和负样品中的词语依次转换为对应的词向量。词向量的基本思想是将离散的词表示成连续空间中的向量,将词间的相关性(语义、语法等)表示为向量间的距离。从而将正样品和负样品,即两两配对好的句子(X,Y),意图分类相同的句子对,希望sim(X,Y)=1,意图分类不同的句子对,希望sim(X,Y)=0。sim(X,Y)与两配的句子的相似度成正比。
同一语句样品中的词向量一起构成该语句样品的词向量集合,即该语句样品的句向量,从而生成与正样品和负样品分别对应的正样品句向量和负样品句向量。正样品和负样品中的两两配对好的句子都是词向量集合(即句向量)。
参照图4,将各正样品句向量和负样品句向量输入至预设意图分析模型进行模型训练,例如一负样品句向量包括句子X(w1,w2...,wn)和句子Y(w1,w2,...wn),将X和Y输入至如图4的DSSM模型(即意图分析模块的一实施例)中,空白框的f(.)和g(.)可以用任何模型替代,例如:全连接FC,卷积神经网络CNN,循环神经网络RNN等,该意图分析模型用于衡量两句子间的语义距离。在生成低维度向量的模型结构上,本申请选CNN卷积神经网络将高维特征向量转化为低维空间的连续向量,因为CNN更适合提取句子的特征,且特征带有时序性。特征降维就是减少特征的数量,这有两个目的,一个是可以加快算法计算的速度,另外可以减少噪音,有效提高意图分类分析的准确率。
在实际任务中,DSSM模块可以搭配不同的损失函数来实现具体的功能,在本申请中,因为需要对两句子X和Y计算语义距离时,可以使用余弦相似 度来计算相似度,变成一个回归模型,其中句子为语句样品的简称。可选地,在语义距离计算方面,本申请使用余弦相似性(cosine imilarity)来衡量用户搜索词(即待识别文本)与候选文档(即应用语句样品)间的语义相关性,在模型任务类型上支持[0,1]值域内的回归。为辅助理解,以一具体例子说明,假如句子X和句子Y对应句向量分别是:
x1,x2,...,xn和y1,y2,...,yn
则,句子X和句子Y的语义距离可以用两者之间夹角的余弦值来表示:
当两句子向量夹角余弦值等于1时,这两句子完全重复;当夹角余弦值接近于1时,两句子相似;夹角的余弦值越小,两句子越不相关。在本实施例中,n=128。从而,在意图分析模型的训练过程中所用的语义距离是Relu(cosθ)。
可选地,步骤S32包括:
步骤S321,将各正样品句向量和负样品句向量逐个输入至预设的意图分析模型进行训练;
步骤S322,逐个基于各正样品向量的第一测试相似度和负样品向量的第二测试相似度,逐次调整意图分析模型的节点权重,使后续第一测试相似度趋向于1,后续第二测试相似度趋向于0,直至所述正样品句向量和负样品句向量遍历完毕;
步骤S323,将所有正样品句向量和负样品句向量训练后的意图分析模型,作为深度网络模型。
在将各正样品句向量和负样品句向量逐个输入至预设意图分析模型进行模型训练,各正样品句向量的两句子的第一测试相似度以及各负样品句向量的两句子的第二测试相似度。由于正样品句向量意图分类相同,则第一测试相似度的标准参考值为1;因此在基于正样品句向量进行意图分析模型训练时,对预设意图分析模型中各节点的权重不断调整,以将第一测试相似度逐渐趋向其标准参数值1。由于负样品句向量意图分类不同,则第二测试相似度的标准参考值为0,因此在基于负样品句向量进行意图分析模型训练时,对预设意图分析模型中各节点的权重不断调整,以将第二测试相似度逐渐趋向其标准参数值0。如此循环输入正样品和负样品至意图分析模型,并同步调整意 图分析模型中各节点权重,直至正样品句向量和负样品句向量遍历完毕,在完成所有正样品句向量和负样品句向量对意图分析模型的训练后,训练完成的意图分析模块即为深度网络模型。
步骤S40,当接收到待识别文本时,将待识别文本和各意图分类中任意选取的应用语句样品作为比对文本输入至深度网络模型中;
在意图分析模型训练完成并生成深度网络模型之后,用户对显示终端输入对话时,基于语义识别技术得到待识别文本,从每个预设的意图分类中任意选取一个或多个语句样品作为与待识别文本进行意图分类对比的应用语句样品,对待识别文本和应用语句样品进行分词后分别作为比对文本输入至意图分析模型中,得出待识别文本与各应用语句样品的相似度。
步骤S50,若存在与待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定待识别文本的意图属于相似度最大的目标应用语句样品所属的目标意图分类。
在计算出所有应用语句样品与待识别文本的相似度之后,将计算出的相似度与预设相似度阈值比较,若存在与待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定待识别文本的意图属于相似度最大的目标应用语句样品所属的目标意图分类,可控制显示终端基于目标意图分类进行精准广告推送和页面推送;若不存在与待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定待识别文本的意图不属于预设的意图分类。
例如,预设相似度阈值为0.6,任意选取了与12个意图分类一一对应的12个应用语句样品,得到待识别文本与12个应用语句样品的相似度,若这12个相似度均不大于0.6,则确定待识别文本的意图不属于预设的12个意图分类,若存在0.65、0.7两个相似度大于0.6,则将相似度为0.7的应用语句样品作为与待识别文本意图类型匹配的目标应用语句样品。
在本实施例中,通过为预设的各意图分类分配第一数值的语句样品构建原始数据,基于深度学习思想,可自动从原始数据中抽取特征以形成语句配对,免去人工设计特征的开销,并且能够更精细和准确地抽取特征,增强了意图分析模型的泛化能力及其面向不同任务需求的可移植性。再者,基于各意图分配的语句样品挑选语句配对,生成正样品和负样品,形成一种训练集产生机擎,只需要少量的训练数据(即语句样品),就可以产生大量的训练集 (即正样品和负样品)以供意图分析模型训练。再者,在深度学习和训练集产生机擎的支持下,生成深度网络模型,在接收待识别文本后,以句子配对方式将待识别文本和应用语句样品输入深度网络模型中,以分析和识别待识别文本的意图分类,提高了意图分类的识别准确性。
进一步地,在开始接收待识别文本之前,本申请基于深度学习的意图分析方法还包括:
步骤S61,获取在不同测试相似度阈值下,深度网络模型对不同意图分类句子判定为相同意图分类的错误接受率;
步骤S62,接收外部输入的意图分析需求,将意图分析需求适配的错误接受率所对应的测试相似度阈值,作为深度网络模型最终的预设相似度阈值。
可设置不同的测试相似度阈值,基于事先选取好的测试集对本申请深度网络模型进行测试,获取在不同测试相似度阈值下,深度网络模型对测试集中不同意图分类句子判定为相同意图分类的错误接受率。其中错误接受是指:不同意图分类句子经深度网络模型和测试相似度阈值处理后,判定结果为相同意图。即错误接受情形指:不同意图分类句子经深度网络模型分析后所得相似度大于此时选定的测试相似度阈值。错误接受率等于错误接受个数与不同意图分类句子的总对数之比。
为辅助理解上述错误接受率的概念,以一具体例子辅助说明:假如测试集有100对意图分类不同的句子,测试相似度阈值包括0.6、0.7和0.8,将100对意图分类不同的句子输入深度网络模型,对应得到与100对意图分类不同的句子一一对应的100个相似度,这100个相似度中大于0.6的值有5个、大于0.7的值有3个、大于0.8的值有1个,则测试相似度阈值0.6的错误接受率为5%,测试相似度阈值0.7的错误接受率为3%,测试相似度阈值0.8的错误接受率为1%。
在获取在不同测试相似度阈值下深度网络模型的错误接受率之后,可接收外部输入的意图分析需求,将意图分析需求适配的错误接受率所对应的测试相似度阈值,作为深度网络模型最终的预设相似度阈值。接上一段的例子,例如意图分析需求是专业要求高、解答实时性高的客服转接,若对用户的意图归类错误,则会分配错误专业方向的客服去对接用户咨询,造成客服和用 户时间的浪费,例如用户对显示终端询问产品升级服务,此时匹配意图分类为购物服务,则容易造成购物服务的客服无法解决用户的产品升级服务的疑惑,造成时间的浪费。因此,客服转接对于错误接受率容忍度很低,测试相似度阈值应当选取较高值,如选取测试相似度阈值0.8作为客服转接的预设相似度阈值。从而本实施例可根据意图分析需求灵活调整深度网络模型的预设相似度阈值,扩展了本申请深度网络模型的适用场景。
可选地,将待识别文本和各意图分类中任意选取的应用语句样品作为比对文本输入至深度网络模型中的步骤包括:
为每个意图分类随机选取预设第二个数的应用语句样品;
基于深度网络模型,获取待识别文本分别与各意图分类的应用语句样品的单元相似度;
将各意图分类中各单元相似度的平均值作为对应意图分类与待识别文本的相似度。
在接收到待识别文本时,表明一个实际的意图识别任务开始了,意图识别任务可以抽象成一个分类问题,待识别文本为Tx,预设的意图分类包括12类,需要对待识别分析进行分析以确定属于12意图分类的哪一个。首先建立一个意图测试文本列表{T
i
c},其中C是意图分类,如C=12,i是每个意图分类中随机选取语句样品,i=0,...9。基于深度网络模型,将待识别文本与意图列表中各意图分类的应用语句样品(即{T
i
c})进行比对,深度网络模型会计算待识别文本Tx和意图测试文本列表{T
i
c}中每一类意图分类的所有文本(即语句样品)sim(T
x,T
i
c)的匹配度(即单元相似度),sim(T
x,T
i
c)∈(0,1),sim(T
x,T
i
c)越接近1,标识两句子的相似度越高,相反,如果sim(T
x,T
i
c)越接近0,表示两句子的相似度越低。
在待识别文本的意图分析过程中,在每个意图分类中随机各取第二个数的应用语句样品,即在每个意图分类中取m个样品和输入的待识别文本Tx做对比,计算各单元相似度sim(T
x,T
i
c)的平均值r(x,c),每个意图分类对应一个相似度平均值r(x,c),将r(x,c)作为待识别文本与各意图分类的相似度,提高了待识别文本与各意图分类相似度的准确性。r(x,c)可基于以下公式
最后,比较C个意图分类相似度r(x,c),其中最大的r(x,c)就是待 识别文本匹配的意图分类。
需要说明的是,在本文中,术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、方法、物品或者系统不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、方法、物品或者系统所固有的要素。在没有更多限制的情况下,由语句“包括一个……”限定的要素,并不排除在包括该要素的过程、方法、物品或者系统中还存在另外的相同要素。
上述本申请实施例序号仅仅为了描述,不代表实施例的优劣。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。基于这样的理解,本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来,该计算机软件产品存储在如上所述的一个存储介质(如ROM/RAM、磁碟、光盘)中,包括若干指令用以使得一台终端设备(可以是手机,计算机,服务器,空调器,或者网络设备等)执行本申请各个实施例所述的方法。
以上仅为本申请的优选实施例,并非因此限制本申请的专利范围,凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换,或直接或间接运用在其他相关的技术领域,均同理包括在本申请的专利保护范围内。
Claims (20)
- 一种基于深度学习的意图分析方法,其中,所述基于深度学习的意图分析方法包括:确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;从同一所述意图分类的语句样品中挑选语句配对,生成正样品;从不同所述意图分类的语句样品中挑选语句配对,生成负样品;所述正样品和负样品共同构成训练集;基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至所述深度网络模型所得相似度分别趋向1和0;当接收到待识别文本时,将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中;若存在与所述待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定所述待识别文本的意图属于相似度最大的所述目标应用语句样品所属的目标意图分类。
- 如权利要求1所述的基于深度学习的意图分析方法,其中,所述从不同所述意图分类的语句样品中挑选语句配对,生成负样品的步骤包括:从不同所述意图分类的语句样品中挑选语句配对,生成初级负样品;从所述初级负样品中随机选取与所述正样品个数相等的语句配对,构成负样品。
- 如权利要求2所述的基于深度学习的意图分析方法,其中,所述基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型的步骤包括:对所述训练集的正样品和负样品进行分词,基于预设词向量算法将分词后的所述正样品和负样品中的词语依次转换为对应的词向量,生成与所述正样品和负样品分别对应的正样品句向量和负样品句向量;将各所述正样品句向量和负样品句向量输入至预设的意图分析模型进行训练,生成深度网络模型。
- 如权利要求3所述的基于深度学习的意图分析方法,其中,所述将各 所述正样品句向量和负样品句向量输入至预设的意图分析模型进行训练,生成深度网络模型的步骤包括:将各所述正样品句向量和负样品句向量逐个输入至预设的意图分析模型进行训练;逐个基于各正样品向量的第一测试相似度和负样品向量的第二测试相似度,逐次调整意图分析模型的节点权重,使后续第一测试相似度趋向于1,后续第二测试相似度趋向于0,直至所述正样品句向量和负样品句向量遍历完毕;将所有所述正样品句向量和负样品句向量训练后的意图分析模型,作为深度网络模型。
- 如权利要求1所述的基于深度学习的意图分析方法,其中,在开始接收待识别文本之前,所述基于深度学习的意图分析方法还包括:获取在不同测试相似度阈值下,所述深度网络模型对不同意图分类句子判定为相同意图分类的错误接受率;接收外部输入的意图分析需求,将所述意图分析需求适配的错误接受率所对应的测试相似度阈值,作为所述深度网络模型最终的预设相似度阈值。
- 如权利要求1所述的基于深度学习的意图分析方法,其中,所述将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中的步骤包括:为每个所述意图分类随机选取预设第二个数的应用语句样品;基于所述深度网络模型,获取所述待识别文本分别与各所述意图分类的应用语句样品的单元相似度;将各所述意图分类中各单元相似度的平均值作为对应所述意图分类与待识别文本的相似度。
- 如权利要求1所述的基于深度学习的意图分析方法,其中,在所述将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中的步骤之后包括:若不存在与所述待识别文本的相似度大于所述相似度阈值的目标应用语句样品,则确定所述待识别文本的意图不属于预设的意图分类。
- 一种意图分析装置,其中,所述意图分析装置包括:语句分配模块,设置为确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;样品生成模块,设置为从同一所述意图分类的语句样品中挑选语句配对,生成正样品;从不同所述意图分类的语句样品中挑选语句配对,生成负样品;所述正样品和负样品共同构成训练集;训练模块,设置为基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至所述深度网络模型所得相似度分别趋向1和0;意图识别模块,设置为当接收到待识别文本时,将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中;若存在与所述待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定所述待识别文本的意图属于相似度最大的所述目标应用语句样品所属的目标意图分类。
- 一种显示终端,其中,所述显示终端包括存储器、处理器、及存储在所述存储器上并可在所述处理器上运行的计算机可读指令,所述计算机可读指令被所述处理器执行时实现如下步骤:确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;从同一所述意图分类的语句样品中挑选语句配对,生成正样品;从不同所述意图分类的语句样品中挑选语句配对,生成负样品;所述正样品和负样品共同构成训练集;基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至所述深度网络模型所得相似度分别趋向1和0;当接收到待识别文本时,将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中;若存在与所述待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定所述待识别文本的意图属于相似度最大的所述目标应用语句样品所属的目标意图分类。
- 如权利要求9所述的显示终端,其中,所述从不同所述意图分类的 语句样品中挑选语句配对,生成负样品的步骤包括:从不同所述意图分类的语句样品中挑选语句配对,生成初级负样品;从所述初级负样品中随机选取与所述正样品个数相等的语句配对,构成负样品。
- 如权利要求10所述的显示终端,其中,所述基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型的步骤包括:对所述训练集的正样品和负样品进行分词,基于预设词向量算法将分词后的所述正样品和负样品中的词语依次转换为对应的词向量,生成与所述正样品和负样品分别对应的正样品句向量和负样品句向量;将各所述正样品句向量和负样品句向量输入至预设的意图分析模型进行训练,生成深度网络模型。
- 如权利要求11所述的显示终端,其中,所述将各所述正样品句向量和负样品句向量输入至预设的意图分析模型进行训练,生成深度网络模型的步骤包括:将各所述正样品句向量和负样品句向量逐个输入至预设的意图分析模型进行训练;逐个基于各正样品向量的第一测试相似度和负样品向量的第二测试相似度,逐次调整意图分析模型的节点权重,使后续第一测试相似度趋向于1,后续第二测试相似度趋向于0,直至所述正样品句向量和负样品句向量遍历完毕;将所有所述正样品句向量和负样品句向量训练后的意图分析模型,作为深度网络模型。
- 如权利要求9所述的显示终端,其中,在开始接收待识别文本之前,所述计算机可读指令被所述处理器执行时实现如下步骤:获取在不同测试相似度阈值下,所述深度网络模型对不同意图分类句子判定为相同意图分类的错误接受率;接收外部输入的意图分析需求,将所述意图分析需求适配的错误接受率所对应的测试相似度阈值,作为所述深度网络模型最终的预设相似度阈值。
- 如权利要求9所述的显示终端,其中,所述将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络 模型中的步骤包括:为每个所述意图分类随机选取预设第二个数的应用语句样品;基于所述深度网络模型,获取所述待识别文本分别与各所述意图分类的应用语句样品的单元相似度;将各所述意图分类中各单元相似度的平均值作为对应所述意图分类与待识别文本的相似度。
- 一种计算机可读存储介质,其中,所述计算机可读存储介质上存储有计算机可读指令,所述计算机可读指令被处理器执行时实现如下的步骤:确定用于用户语句的意图分析所涉及的预设意图分类;为预设的各意图分类分配预设第一数值的语句样品;从同一所述意图分类的语句样品中挑选语句配对,生成正样品;从不同所述意图分类的语句样品中挑选语句配对,生成负样品;所述正样品和负样品共同构成训练集;基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型,其中所述正样品和负样品输入至所述深度网络模型所得相似度分别趋向1和0;当接收到待识别文本时,将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中;若存在与所述待识别文本的相似度大于预设相似度阈值的目标应用语句样品,则确定所述待识别文本的意图属于相似度最大的所述目标应用语句样品所属的目标意图分类。
- 如权利要求15所述的计算机可读存储介质,其中,所述从不同所述意图分类的语句样品中挑选语句配对,生成负样品的步骤包括:从不同所述意图分类的语句样品中挑选语句配对,生成初级负样品;从所述初级负样品中随机选取与所述正样品个数相等的语句配对,构成负样品。
- 如权利要求16所述的计算机可读存储介质,其中,所述基于所述训练集对预设的意图分析模型进行训练,生成深度网络模型的步骤包括:对所述训练集的正样品和负样品进行分词,基于预设词向量算法将分词后的所述正样品和负样品中的词语依次转换为对应的词向量,生成与所述正 样品和负样品分别对应的正样品句向量和负样品句向量;将各所述正样品句向量和负样品句向量输入至预设的意图分析模型进行训练,生成深度网络模型。
- 如权利要求17所述的计算机可读存储介质,其中,所述将各所述正样品句向量和负样品句向量输入至预设的意图分析模型进行训练,生成深度网络模型的步骤包括:将各所述正样品句向量和负样品句向量逐个输入至预设的意图分析模型进行训练;逐个基于各正样品向量的第一测试相似度和负样品向量的第二测试相似度,逐次调整意图分析模型的节点权重,使后续第一测试相似度趋向于1,后续第二测试相似度趋向于0,直至所述正样品句向量和负样品句向量遍历完毕;将所有所述正样品句向量和负样品句向量训练后的意图分析模型,作为深度网络模型。
- 如权利要求15所述的计算机可读存储介质,其中,在开始接收待识别文本之前,所述计算机可读指令被所述处理器执行时实现如下步骤:获取在不同测试相似度阈值下,所述深度网络模型对不同意图分类句子判定为相同意图分类的错误接受率;接收外部输入的意图分析需求,将所述意图分析需求适配的错误接受率所对应的测试相似度阈值,作为所述深度网络模型最终的预设相似度阈值。
- 如权利要求15所述的计算机可读存储介质,其中,所述将所述待识别文本和各所述意图分类中任意选取的应用语句样品作为比对文本输入至所述深度网络模型中的步骤包括:为每个所述意图分类随机选取预设第二个数的应用语句样品;基于所述深度网络模型,获取所述待识别文本分别与各所述意图分类的应用语句样品的单元相似度;将各所述意图分类中各单元相似度的平均值作为对应所述意图分类与待识别文本的相似度。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910029832.9A CN109800306B (zh) | 2019-01-10 | 2019-01-10 | 意图分析方法、装置、显示终端及计算机可读存储介质 |
| CN201910029832.9 | 2019-01-10 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2020143844A1 true WO2020143844A1 (zh) | 2020-07-16 |
Family
ID=66558770
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2020/074688 Ceased WO2020143844A1 (zh) | 2019-01-10 | 2020-02-11 | 意图分析方法、装置、显示终端及计算机可读存储介质 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109800306B (zh) |
| WO (1) | WO2020143844A1 (zh) |
Cited By (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112329836A (zh) * | 2020-11-02 | 2021-02-05 | 成都网安科技发展有限公司 | 基于深度学习的文本分类方法、装置、服务器及存储介质 |
| CN112465555A (zh) * | 2020-12-04 | 2021-03-09 | 北京搜狗科技发展有限公司 | 一种广告信息推荐的方法及相关装置 |
| CN112800239A (zh) * | 2021-01-22 | 2021-05-14 | 中信银行股份有限公司 | 意图识别模型训练方法、意图识别方法及装置 |
| CN113127602A (zh) * | 2021-04-30 | 2021-07-16 | 竹间智能科技(上海)有限公司 | 一种意图识别方法及装置 |
| CN113158680A (zh) * | 2021-03-23 | 2021-07-23 | 北京新方通信技术有限公司 | 一种语料处理及意图识别的方法和装置 |
| CN113240438A (zh) * | 2021-05-11 | 2021-08-10 | 京东数字科技控股股份有限公司 | 意图识别方法、设备、存储介质及程序产品 |
| CN113420128A (zh) * | 2021-08-23 | 2021-09-21 | 腾讯科技(深圳)有限公司 | 文本匹配方法、装置、存储介质及计算机设备 |
| CN113792540A (zh) * | 2021-09-18 | 2021-12-14 | 平安科技(深圳)有限公司 | 意图识别模型更新方法及相关设备 |
| CN115115002A (zh) * | 2022-07-22 | 2022-09-27 | 宁波牛信网络科技有限公司 | 文本相似度计算模型生成方法、装置、设备及存储介质 |
| CN115238066A (zh) * | 2021-04-23 | 2022-10-25 | 北京三快在线科技有限公司 | 用户意图识别方法、装置、对话系统、设备以及存储介质 |
| CN117556022A (zh) * | 2023-12-18 | 2024-02-13 | 北京中关村科金技术有限公司 | 一种智能客服意图识别方法、装置及系统 |
| CN117973402A (zh) * | 2024-04-01 | 2024-05-03 | 腾讯科技(深圳)有限公司 | 文本转换的预处理方法和装置、存储介质及电子设备 |
| CN118012982A (zh) * | 2024-02-05 | 2024-05-10 | 山东浪潮科学研究院有限公司 | 一种可控的长文本生成方法及系统 |
Families Citing this family (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109800306B (zh) * | 2019-01-10 | 2023-10-17 | 深圳Tcl新技术有限公司 | 意图分析方法、装置、显示终端及计算机可读存储介质 |
| CN110287285B (zh) * | 2019-05-31 | 2023-06-16 | 平安科技(深圳)有限公司 | 一种问题意图识别方法、装置、计算机设备及存储介质 |
| CN110347792B (zh) * | 2019-06-25 | 2022-12-20 | 腾讯科技(深圳)有限公司 | 对话生成方法及装置、存储介质、电子设备 |
| CN110516056A (zh) * | 2019-08-20 | 2019-11-29 | 深圳市奥拓电子股份有限公司 | 人机对话的自主学习方法、自主学习系统及存储介质 |
| CN110853649A (zh) * | 2019-11-05 | 2020-02-28 | 集奥聚合(北京)人工智能科技有限公司 | 基于智能语音技术的标签提取方法、系统、设备及介质 |
| CN111046667B (zh) * | 2019-11-14 | 2024-02-06 | 深圳市优必选科技股份有限公司 | 一种语句识别方法、语句识别装置及智能设备 |
| CN113362096A (zh) * | 2020-03-04 | 2021-09-07 | 驰众信息技术(上海)有限公司 | 一种基于深度学习的框架广告图像匹配方法 |
| CN111126606B (zh) * | 2020-03-30 | 2020-07-28 | 同盾控股有限公司 | 基于平台的深度学习模型自主训练方法、装置、设备及介质 |
| CN111611383A (zh) * | 2020-05-26 | 2020-09-01 | 深圳壹账通智能科技有限公司 | 用户意图的识别方法、装置、计算机设备及存储介质 |
| CN113535945B (zh) * | 2020-06-15 | 2023-09-15 | 腾讯科技(深圳)有限公司 | 文本类别的识别方法、装置、设备及计算机可读存储介质 |
| CN113742480B (zh) * | 2020-06-18 | 2025-01-14 | 北京汇钧科技有限公司 | 客服应答方法和装置 |
| CN111931512B (zh) * | 2020-07-01 | 2024-07-26 | 联想(北京)有限公司 | 语句意图的确定方法及装置、存储介质 |
| CN111832305B (zh) * | 2020-07-03 | 2023-08-25 | 北京小鹏汽车有限公司 | 一种用户意图识别方法、装置、服务器和介质 |
| CN111833076A (zh) * | 2020-07-07 | 2020-10-27 | 信雅达系统工程股份有限公司 | 智能催收系统及方法 |
| CN112446209A (zh) * | 2020-11-27 | 2021-03-05 | 平安科技(深圳)有限公司 | 一种意图标签的设置方法、设备、装置及存储介质 |
| CN114580426B (zh) * | 2020-12-01 | 2025-02-14 | 阿里巴巴集团控股有限公司 | 用户意图识别方法、交互方法、装置、系统、设备和介质 |
| CN113064631B (zh) * | 2021-04-07 | 2025-10-31 | 泰豪软件股份有限公司 | 机器人版本回归管理方法、装置、机器人及存储介质 |
| CN113515594B (zh) * | 2021-04-28 | 2024-09-20 | 京东科技控股股份有限公司 | 意图识别方法、意图识别模型训练方法、装置及设备 |
| CN114492452B (zh) * | 2021-12-24 | 2025-04-01 | 深圳云天励飞技术股份有限公司 | 一种预训练语言模型的训练、诉求转接方法、装置和设备 |
| CN114662601A (zh) * | 2022-03-25 | 2022-06-24 | 鼎富智能科技有限公司 | 基于正负样本的意图分类模型训练方法及装置 |
| CN114861064A (zh) * | 2022-05-27 | 2022-08-05 | 深圳集智数字科技有限公司 | 基于双塔模型的对象推荐方法及装置 |
| CN116089604A (zh) * | 2022-08-10 | 2023-05-09 | 腾讯科技(深圳)有限公司 | 信息识别方法、装置、计算机设备、存储介质及程序产品 |
| CN115438153A (zh) * | 2022-09-14 | 2022-12-06 | 太保科技有限公司 | 一种意图匹配度分析模型的训练方法及装置 |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170358295A1 (en) * | 2016-06-10 | 2017-12-14 | Conduent Business Services, Llc | Natural language generation, a hybrid sequence-to-sequence approach |
| CN108228820A (zh) * | 2017-12-30 | 2018-06-29 | 厦门太迪智能科技有限公司 | 用户查询意图理解方法、系统及计算机终端 |
| US20180329990A1 (en) * | 2017-05-12 | 2018-11-15 | Sentient Technologies (Barbados) Limited | Intelligently driven visual interface on mobile devices and tablets based on implicit and explicit user actions |
| CN109800306A (zh) * | 2019-01-10 | 2019-05-24 | 深圳Tcl新技术有限公司 | 意图分析方法、装置、显示终端及计算机可读存储介质 |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107122346B (zh) * | 2016-12-28 | 2018-02-27 | 平安科技(深圳)有限公司 | 一种输入语句的纠错方法及装置 |
| CN106897263A (zh) * | 2016-12-29 | 2017-06-27 | 北京光年无限科技有限公司 | 基于深度学习的机器人对话交互方法及装置 |
| CN107329949B (zh) * | 2017-05-24 | 2021-01-01 | 北京捷通华声科技股份有限公司 | 一种语义匹配方法和系统 |
-
2019
- 2019-01-10 CN CN201910029832.9A patent/CN109800306B/zh active Active
-
2020
- 2020-02-11 WO PCT/CN2020/074688 patent/WO2020143844A1/zh not_active Ceased
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170358295A1 (en) * | 2016-06-10 | 2017-12-14 | Conduent Business Services, Llc | Natural language generation, a hybrid sequence-to-sequence approach |
| US20180329990A1 (en) * | 2017-05-12 | 2018-11-15 | Sentient Technologies (Barbados) Limited | Intelligently driven visual interface on mobile devices and tablets based on implicit and explicit user actions |
| CN108228820A (zh) * | 2017-12-30 | 2018-06-29 | 厦门太迪智能科技有限公司 | 用户查询意图理解方法、系统及计算机终端 |
| CN109800306A (zh) * | 2019-01-10 | 2019-05-24 | 深圳Tcl新技术有限公司 | 意图分析方法、装置、显示终端及计算机可读存储介质 |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112329836A (zh) * | 2020-11-02 | 2021-02-05 | 成都网安科技发展有限公司 | 基于深度学习的文本分类方法、装置、服务器及存储介质 |
| CN112465555A (zh) * | 2020-12-04 | 2021-03-09 | 北京搜狗科技发展有限公司 | 一种广告信息推荐的方法及相关装置 |
| CN112465555B (zh) * | 2020-12-04 | 2024-05-14 | 北京搜狗科技发展有限公司 | 一种广告信息推荐的方法及相关装置 |
| CN112800239A (zh) * | 2021-01-22 | 2021-05-14 | 中信银行股份有限公司 | 意图识别模型训练方法、意图识别方法及装置 |
| CN112800239B (zh) * | 2021-01-22 | 2024-04-12 | 中信银行股份有限公司 | 意图识别模型训练方法、意图识别方法及装置 |
| CN113158680A (zh) * | 2021-03-23 | 2021-07-23 | 北京新方通信技术有限公司 | 一种语料处理及意图识别的方法和装置 |
| CN113158680B (zh) * | 2021-03-23 | 2024-05-07 | 北京新方通信技术有限公司 | 一种语料处理及意图识别的方法和装置 |
| CN115238066A (zh) * | 2021-04-23 | 2022-10-25 | 北京三快在线科技有限公司 | 用户意图识别方法、装置、对话系统、设备以及存储介质 |
| CN113127602A (zh) * | 2021-04-30 | 2021-07-16 | 竹间智能科技(上海)有限公司 | 一种意图识别方法及装置 |
| CN113127602B (zh) * | 2021-04-30 | 2023-05-26 | 竹间智能科技(上海)有限公司 | 一种意图识别方法及装置 |
| CN113240438A (zh) * | 2021-05-11 | 2021-08-10 | 京东数字科技控股股份有限公司 | 意图识别方法、设备、存储介质及程序产品 |
| CN113420128B (zh) * | 2021-08-23 | 2021-11-19 | 腾讯科技(深圳)有限公司 | 文本匹配方法、装置、存储介质及计算机设备 |
| CN113420128A (zh) * | 2021-08-23 | 2021-09-21 | 腾讯科技(深圳)有限公司 | 文本匹配方法、装置、存储介质及计算机设备 |
| CN113792540B (zh) * | 2021-09-18 | 2024-03-22 | 平安科技(深圳)有限公司 | 意图识别模型更新方法及相关设备 |
| CN113792540A (zh) * | 2021-09-18 | 2021-12-14 | 平安科技(深圳)有限公司 | 意图识别模型更新方法及相关设备 |
| CN115115002A (zh) * | 2022-07-22 | 2022-09-27 | 宁波牛信网络科技有限公司 | 文本相似度计算模型生成方法、装置、设备及存储介质 |
| CN117556022A (zh) * | 2023-12-18 | 2024-02-13 | 北京中关村科金技术有限公司 | 一种智能客服意图识别方法、装置及系统 |
| CN118012982A (zh) * | 2024-02-05 | 2024-05-10 | 山东浪潮科学研究院有限公司 | 一种可控的长文本生成方法及系统 |
| CN117973402A (zh) * | 2024-04-01 | 2024-05-03 | 腾讯科技(深圳)有限公司 | 文本转换的预处理方法和装置、存储介质及电子设备 |
| CN117973402B (zh) * | 2024-04-01 | 2024-06-11 | 腾讯科技(深圳)有限公司 | 文本转换的预处理方法和装置、存储介质及电子设备 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109800306B (zh) | 2023-10-17 |
| CN109800306A (zh) | 2019-05-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2020143844A1 (zh) | 意图分析方法、装置、显示终端及计算机可读存储介质 | |
| CN109871446B (zh) | 意图识别中的拒识方法、电子装置及存储介质 | |
| CN109960725B (zh) | 基于情感的文本分类处理方法、装置和计算机设备 | |
| US11269665B1 (en) | Method and system for user experience personalization in data management systems using machine learning | |
| CN117312641B (zh) | 智能获取信息的方法、装置、设备及存储介质 | |
| WO2021143018A1 (zh) | 意图识别方法、装置、设备及计算机可读存储介质 | |
| WO2020140612A1 (zh) | 基于卷积神经网络的意图识别方法、装置、设备及介质 | |
| WO2021114841A1 (zh) | 一种用户报告的生成方法及终端设备 | |
| CN110059923A (zh) | 岗位画像和简历信息的匹配方法、装置、设备及存储介质 | |
| WO2018006727A1 (zh) | 机器人客服转人工客服的方法和装置 | |
| US12299392B2 (en) | Text processing method and apparatus for error correction of a plurality of types of non-words and real words | |
| CN110795913A (zh) | 一种文本编码方法、装置、存储介质及终端 | |
| CN110717021B (zh) | 人工智能面试中获取输入文本和相关装置 | |
| CN116150306A (zh) | 问答机器人的训练方法、问答方法及装置 | |
| CN111161726B (zh) | 一种智能语音交互方法、设备、介质及系统 | |
| CN111611383A (zh) | 用户意图的识别方法、装置、计算机设备及存储介质 | |
| CN117972160B (zh) | 一种多模态信息处理方法及装置 | |
| CN116628163A (zh) | 客服服务处理方法、装置、设备及存储介质 | |
| CN113362026A (zh) | 文本处理方法及装置 | |
| CN120045750A (zh) | 一种基于大语言模型的检索增强生成方法和系统 | |
| CN111126084A (zh) | 数据处理方法、装置、电子设备和存储介质 | |
| US12547835B2 (en) | Automatic extraction of semantically similar question topics | |
| JP2021081713A (ja) | 音声信号を処理するための方法、装置、機器、および媒体 | |
| WO2021077333A1 (zh) | 同声传译方法及装置、存储介质 | |
| CN119181102B (zh) | 短文本生成图像模型训练方法、系统、短文本到图像的生成方法、电子设备及存储介质 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 20738872 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 20738872 Country of ref document: EP Kind code of ref document: A1 |


