WO2018153217A1 - 一种句子相似度判断方法 - Google Patents

一种句子相似度判断方法 Download PDF

Info

Publication number
WO2018153217A1
WO2018153217A1 PCT/CN2018/074336 CN2018074336W WO2018153217A1 WO 2018153217 A1 WO2018153217 A1 WO 2018153217A1 CN 2018074336 W CN2018074336 W CN 2018074336W WO 2018153217 A1 WO2018153217 A1 WO 2018153217A1
Authority
WO
WIPO (PCT)
Prior art keywords
sentence
neural network
similarity
network model
vector
Prior art date
Application number
PCT/CN2018/074336
Other languages
English (en)
French (fr)
Inventor
沈磊
陈见耸
Original Assignee
芋头科技（杭州）有限公司
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 芋头科技（杭州）有限公司 filed Critical 芋头科技（杭州）有限公司
Priority to US16/488,572 priority Critical patent/US10949709B2/en
Publication of WO2018153217A1 publication Critical patent/WO2018153217A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/205—Parsing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06F16/3347—Query execution using vector based model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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

Definitions

  • the invention relates to the technical field of natural language processing, and in particular to a method for judging sentence similarity.
  • the word vector matrix of two sentences is first obtained and input into the deep neural network model.
  • the sentence vector is obtained by the processing of the deep neural network and spliced as the input of the classification neural network model.
  • the word sequence is formed by the word sequence mapping in the sentence, and the parameters are generally initialized by the word vector formed by the pre-training of the language model, so the parameter quality comparison depends on the pre-training.
  • the quality of the word vector Moreover, if there are words or words (ie, unregistered words) that are not in the word vector dictionary in the calculation, the map will be mapped into a random vector for calculation, thereby affecting the measurement effect of the model.
  • a technical solution for a sentence similarity judging method which aims to solve the problem that the calculated sentence similarity comparison in the prior art depends on the quality of the pre-trained word/word vector and the unregistered word. Problem, thereby improving the measure of calculating the similarity of sentences.
  • a sentence similarity judgment method wherein a sentence similarity judgment model is formed by pre-training, the sentence similarity judgment model includes a first neural network model for processing a sentence vector and a process for obtaining a representation a second neural network model of similarity measure of sentence similarity;
  • the sentence similarity judgment method further includes:
  • Step S1 acquiring, according to two externally input sentence samples, a word vector matrix in each of the sentence samples;
  • Step S2 respectively extracting overlapping features in each of the sentence samples to form an overlapping feature matrix, and combining the corresponding word vector matrix with the overlapping feature matrix as the first for each of the sentence samples Input data of a neural network model;
  • Step S3 respectively processing the sentence vector for each of the sentence samples according to the first neural network model and operating to form a sentence merge vector, and combining with the overlapping feature vector formed according to the overlapping feature Input data of the second neural network model;
  • Step S4 according to the second neural network model processing, obtaining a similarity measure associated with two of the sentence samples and outputting them as a basis for determining the similarity between the two sentence samples;
  • the sentence merging vector is formed by an operation mode in which the sentence vector is directly subtracted, or the sentence merging vector is formed by an operation manner of splicing the sentence vector.
  • the sentence similarity judging method wherein in the step S1, the word vector matrix of each of the sentence samples comprises:
  • the sentence sample is sliced into word sequences and the word sequences are mapped into the word vector matrix.
  • the sentence similarity judging method wherein in the step S2, the overlapping feature matrix is formed by:
  • Step S21 replacing the mutually overlapping words or words in the two sentence samples with a first character
  • Step S22 replacing words or words that do not overlap in the two sentence samples with a second character
  • Step S23 forming an overlapping feature sequence associated with each of the sentence samples according to the first character and the second character, respectively;
  • Step S24 mapping each of the overlapping feature sequences into the overlapping feature matrix
  • Step S25 each of the word vector matrix and the corresponding overlapping feature matrix are respectively combined as the input data of the first neural network model.
  • the sentence similarity judging method in the step S3, the processing obtains the similarity product of the two sentence vectors, and then performs subtraction operations on the two sentence vectors, and the similarity
  • the product and the overlapping feature vector are combined as the input data of the second neural network.
  • the sentence similarity judging method wherein the similarity product is obtained by calculating a dot product between two of the sentence vectors; or
  • the parameter matrix is simultaneously trained.
  • the sentence similarity judging method wherein the first neural network model is a deep neural network model.
  • the sentence similarity judging method wherein the first neural network model is a convolutional neural network model or a cyclic neural network model.
  • the sentence similarity judging method wherein the second neural network model is a classification neural network model.
  • the above technical solution has the beneficial effects of providing a sentence similarity judgment method, which can solve the problem that the calculated sentence similarity comparison depends on the quality of the pre-trained word/word vector and the unregistered word in the prior art, thereby improving the calculation sentence similarity.
  • Degree measure The above technical solution has the beneficial effects of providing a sentence similarity judgment method, which can solve the problem that the calculated sentence similarity comparison depends on the quality of the pre-trained word/word vector and the unregistered word in the prior art, thereby improving the calculation sentence similarity. Degree measure.
  • 2-3 is a schematic diagram showing the overall flow of a sentence similarity judging method in a preferred embodiment of the present invention.
  • FIG. 4 is a schematic flow chart showing the formation of an overlapping feature matrix in a preferred embodiment of the present invention.
  • FIG. 5 is a schematic structural diagram of a first neural network model in a preferred embodiment of the present invention.
  • Figure 6 is a block diagram showing the structure of a second neural network model in a preferred embodiment of the present invention.
  • a sentence similarity judgment model can be formed by pre-training, and the sentence similarity judgment model includes a sentence for processing A first neural network model of the vector and a second neural network model for processing a similarity measure representing the similarity of the sentences.
  • the first neural network model and the second neural network model are integrally formed by unified training, that is, the sentence similarity judgment model including the first neural network model and the second neural network model is first constructed (the first neural network is to be used) The output of the model is used as an input to the second neural network model, and then the entire sentence similarity judgment model is trained by inputting the training samples to the first neural network model.
  • the method is specifically as shown in FIG. 2, including:
  • Step S1 acquiring a word vector matrix in each sentence sample according to two externally input sentence samples
  • Step S2 respectively extracting overlapping features in each sentence sample to form an overlapping feature matrix, and combining corresponding word vector matrix and overlapping feature matrix as input data of the first neural network model for each sentence sample;
  • Step S3 respectively, processing a sentence vector for each sentence sample according to the first neural network model and operating to form a sentence merge vector, and combining with the overlapping feature vector formed according to the overlapping feature as input data of the second neural network model;
  • step S4 the similarity measure associated with the two sentence samples is processed according to the second neural network model and outputted as a basis for determining the similarity of the two sentence samples.
  • the sentence merging vector is formed by the operation method in which the sentence vector is directly subtracted, or the sentence merging vector is formed by the operation method of splicing the sentence vector.
  • the word vector matrix in each sentence sample is first obtained separately.
  • the so-called word vector matrix refers to a matrix formed by mapping the word vectors in a sentence.
  • the overlapping features in each sentence sample are acquired to form an overlapping feature matrix
  • the overlapping feature is an overlapping word feature extracted according to mutually overlapping words/words in the two sentence samples
  • the matrix is a matrix formed by the same mapping method in which the overlapping features form the word vector matrix according to the above mapping.
  • the two matrices associated with the same sentence sample are combined as input data of the first neural network model, and then the first neural network model is processed.
  • the sentence vectors of the two sentence samples are subjected to a subtraction operation, the specific method of which is detailed below. And, an overlapping feature vector is formed for the overlapping features obtained above, and combined with the formed sentence merging vector as input data of the second neural network model.
  • the sentence vectors of the two sentence samples are spliced, and the specific method of the splicing operation is the same as in the prior art. And, an overlapping feature vector is formed for the overlapping features obtained above, and combined with the formed sentence merging vector as input data of the second neural network model.
  • the similarity measure of the two sentence samples is finally obtained by the second neural network model processing, as a basis for judging the similarity of the two sentence samples.
  • the improved portion of the technical solution of the present invention is shown in FIG. .
  • the way of splicing the sentence vectors in the prior art is changed to either splicing or subtracting.
  • the above method improves the model for calculating the similarity of sentences, and finally improves the measurement method for calculating the similarity of sentences.
  • the word vector matrix of each sentence sample includes:
  • the above word vector matrix includes a word/word vector matrix of each sentence sample.
  • step S1 in the above step S1:
  • the sentence sample is divided into word sequences and the word sequence is mapped into a word vector matrix.
  • the overlapping feature matrix is formed by the following method as shown in FIG. 4:
  • Step S21 replacing words or words overlapping each other in the two sentence samples with a first character
  • Step S22 replacing non-overlapping words or words in the two sentence samples with a second character
  • Step S23 forming an overlapping feature sequence associated with each sentence sample according to the first character and the second character, respectively;
  • Step S24 mapping each overlapping feature sequence into an overlapping feature matrix
  • step S25 each word vector matrix and the corresponding overlapping feature matrix are respectively combined as input data of the first neural network model.
  • the first character in order to facilitate the processing by the computer, the first character may be 1 and the second character may be 0, and a binary overlapping feature vector associated with each sentence sample may be formed.
  • the overlapping parts ie overlapping features
  • the overlapping feature sequence is 1001
  • the overlapping feature sequence for "Give me a song” is 01001
  • the above two overlapping feature sequences 1001 and 01001 are respectively mapped to form overlapping features according to the same method in which the word vector is mapped into a word vector matrix.
  • the matrix that is, the character 0 is mapped into a one-dimensional vector, the character 1 is mapped into a one-dimensional vector, and then a matrix is formed, and the word vector matrix and the overlapping feature matrix of each sentence sample are combined as input data of the first neural network model.
  • first character and the second character may also be selected in other forms suitable for processing, and details are not described herein again.
  • the manner in which the overlapping feature vectors are formed may include the following:
  • s1 represents one sentence sample
  • s2 represents another sentence sample
  • IDF_overlap is used to represent the sum of IDF (Inverse Document Frequency) of mutually overlapping words in two sentence samples
  • length is used to represent each The sentence length of the sentence sample
  • the IDF number of a particular word/word can be obtained by dividing the total number of files by the number of files containing the word/word and then taking the obtained quotient logarithm. This will not be repeated below.
  • one sentence sample is represented by s1, s2 represents another sentence sample, and IDF_overlap is used to represent the sum of IDFs of mutually overlapping words in two sentence samples, and IDF_sum is used to indicate IDF of all words in each sentence sample.
  • s1 represents one sentence sample
  • s2 represents another sentence sample
  • length is used to represent the sentence length of each sentence sample
  • word_overlap is used to represent the word overlap number in two sentence samples
  • the above three methods can process the overlapping feature vectors and directly splicing the overlapping feature vectors into the input data of the second neural network model.
  • Stop Words mainly include English characters, numbers, mathematical characters, punctuation marks, and single Chinese characters with extremely high frequency. If a stop word is encountered during text processing, stop processing immediately and throw it away. .
  • performing the subtraction operation on the two sentence vectors in the above step S3 can better find the difference between the two sentence vectors (as shown in FIG. 3). Further, the subtraction operation of two sentence vectors can be implemented in the following ways:
  • the first neural network model may be a convolutional neural network model, and the convolutional neural network is divided into a convolution layer and a sampling layer (as shown in FIG. 5), and the above two types may be directly applied after the convolution layer processing.
  • One of the ways subtracts the two vectors and then samples them at the sampling layer to get the result.
  • step S3 while processing two sentence vectors in a subtractive manner, the similarity product of the two sentence vectors is obtained, and the similarity product, sentence is obtained.
  • the result of the vector subtraction and the overlapping feature vector are combined as input data of the second neural network (as shown in FIG. 3).
  • the above similarity product can be processed in the following ways:
  • the similarity product can be expressed as x*M*y.
  • the parameter matrix M can be trained together when training forms a sentence similarity judgment model (ie, when unified training forms the first neural network model and the second neural network model).
  • the sentence vector in the foregoing step S3, may be subjected to a subtraction operation, and the sentence vector mosaic method similar to the prior art may be used to splicing the two sentence vectors, and formed according to the overlapping features.
  • the overlapping feature vector is combined with the input data as the second neural network model (as shown in FIG. 3, in FIG. 3, it may be selected to be processed by sentence vector stitching or sentence vector subtraction).
  • the first neural network model may be a deep neural network model, and may further be a Convolutional Neural Network (CNN) or a Recurrent Neural Network (Recurrent Neural Network). , RNN), even a variant of the cyclic neural network model, such as the Long Short Term Memory (LSTM) or the Gated Recurrent Unit (GRU).
  • CNN Convolutional Neural Network
  • Recurrent Neural Network Recurrent Neural Network
  • RNN Recurrent Neural Network
  • RNN Recurrent Neural Network
  • RNN Recurrent Neural Network
  • LSTM Long Short Term Memory
  • GRU Gated Recurrent Unit
  • the second neural network model may be a classification neural network model, and the general structure of the second neural network model is shown in FIG. 6, and the second neural network model may be divided into an input layer.
  • the hidden layer and the output layer, the output layer is also the classification layer, and the above hidden layer can also be removed, that is, only the input layer and the output layer (classification layer) exist.
  • the technical solution of the present invention provides a sentence similarity judgment method, which introduces overlapping features of sentence vectors and respectively serves as input data of a deep neural network model and a classification neural network model, and splicing sentence vectors in the process of processing.
  • the process is changed to the process of subtracting the sentence vector, so it can solve the problem that the calculated sentence similarity comparison depends on the quality of the pre-trained word/word vector and the unregistered word in the prior art, thereby improving the measure of calculating the similarity of the sentence. method.
  • the sentence similarity judgment method provided in the technical solution of the present invention can be applied to a scenario of "chat" between a user and a smart device.
  • the process in which the smart device gives a response through background processing is generally: obtaining a preliminary candidate sentence set by using an alternative database in the background of the smart device, and then adopting the technical solution of the present invention.
  • the provided sentence similarity judging method obtains a similar sentence associated with the user's spoken words from the candidate sentence set, and then feeds back the answer corresponding to the similar sentence to the user.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Databases & Information Systems (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Medical Informatics (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Algebra (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Machine Translation (AREA)

Abstract

本发明公开了一种句子相似度判断方法，属于自然语言处理技术领域；方法包括：根据两个外部输入的句子样本，获取句子样本中的字词向量矩阵；提取句子样本中的重叠特征以形成重叠特征矩阵，并将字词向量矩阵与重叠特征矩阵结合作为第一神经网络模型的输入数据；根据第一神经网络模型处理得到针对句子样本的句子向量并进行操作形成一句子合并向量，并与根据重叠特征形成的重叠特征向量结合作为第二神经网络模型的输入数据；根据第二神经网络模型处理得到相似性度量并输出，以作为判断两个句子样本的相似度的依据。上述技术方案的有益效果是：解决现有技术中计算句子相似度比较依赖预训练的字/词向量的质量以及未登录词的问题。

Description

一种句子相似度判断方法 技术领域
本发明涉及自然语言处理技术领域，尤其涉及一种句子相似度判断方法。
背景技术
在自然语言处理的技术领域中，对于两个句子之间判断相似度的应用非常广泛。现有技术中通常会采用如图1所示的以下方法来计算两个句子之间的相似度：
对于句子1和句子2，首先分别获取两个句子的字词向量矩阵并输入到深度神经网络模型中，通过深度神经网络的处理得到句子向量并进行拼接以作为分类神经网络模型的输入，最后得到两个句子的相似性度量。
上述处理方法在计算句子相似度时，由句子中的字词序列映射形成字词向量矩阵，其参数一般都会使用由语言模型预训练形成的字词向量进行初始化，因此参数质量比较依赖预训练的字词向量的质量。并且，若在进行计算时，句子中存在字词向量词典中没有的字或词(即未登录词)，则会将其映射成随机向量进行计算，从而影响模型的度量效果。
发明内容
根据现有技术中存在的上述问题，现提供一种句子相似度判断方法的技术方案，旨在解决现有技术中计算句子相似度比较依赖预训练的字/词向量的质量和未登录词的问题，从而改进计算句子相似度的度量方法。
上述技术方案具体包括：
一种句子相似度判断方法，其中，通过预先训练形成一句子相似度判断模型，所述句子相似度判断模型中包括一用于处理得到句子向量的第一神经网络模型以及一用于处理得到表示句子相似度的相似性度量的第二神经网络模型；
所述句子相似度判断方法还包括：
步骤S1，根据两个外部输入的句子样本，分别获取每个所述句子样本中的字词向量矩阵；
步骤S2,分别提取每个所述句子样本中的重叠特征以形成重叠特征矩阵，并针对每个所述句子样本将对应的所述字词向量矩阵与所述重叠特征矩阵结合作为所述第一神经网络模型的输入数据；
步骤S3，根据所述第一神经网络模型分别处理得到针对每个所述句子样本的所述句子向量并进行操作形成一句子合并向量，并与根据所述重叠特征形成的重叠特征向量结合作为所述第二神经网络模型的输入数据；
步骤S4，根据所述第二神经网络模型处理得到关联于两个所述句子样本的相似性度量并输出，以作为判断两个所述句子样本的相似度的依据；
所述步骤S3中，采用所述句子向量直接相减的操作方式形成所述句子合并向量，或者采用拼接所述句子向量的操作方式形成所述句子合并向量。
优选的，该句子相似度判断方法，其中，所述步骤S1中，每个所述句子样本的字词向量矩阵包括：
每个所述句子样本的字向量矩阵；或者
每个所述句子样本的词向量矩阵；
则所述步骤S1中：
将所述句子样本切分成字序列，并将所述字序列映射成所述字向量矩阵；或者
将所述句子样本切分成词序列，并将所述词序列映射成所述词向量矩阵。
优选的，该句子相似度判断方法，其中，所述步骤S2中，采用下述方式处理形成所述重叠特征矩阵：
步骤S21，将所述两个所述句子样本中相互重叠的字或词分别替换成一第一字符；
步骤S22，将所述两个句子样本中不相重叠的字或词分别替换成一第二字符；
步骤S23，根据所述第一字符和所述第二字符分别形成关联于每个所述句子样本的重叠特征序列；
步骤S24，将每个所述重叠特征序列映射成所述重叠特征矩阵；
步骤S25，每个所述字词向量矩阵和对应的所述重叠特征矩阵分别结合 作为所述第一神经网络模型的所述输入数据。
优选的，该句子相似度判断方法，其中，所述步骤S3中，处理得到两个所述句子向量的相似度乘积，随后对两个所述句子向量做相减操作，并与所述相似度乘积以及所述重叠特征向量结合作为所述第二神经网络的所述输入数据。
优选的，该句子相似度判断方法，其中，通过计算两个所述句子向量之间的点积得到所述相似度乘积；或者
根据一参数矩阵处理得到所述相似度乘积；
在预先对所述句子相似度判断模型进行训练的过程中，同时训练得到所述参数矩阵。
优选的，该句子相似度判断方法，其中，所述第一神经网络模型为深度神经网络模型。
优选的，该句子相似度判断方法，其中，所述第一神经网络模型为卷积神经网络模型或者循环神经网络模型。
优选的，该句子相似度判断方法，其中，所述第二神经网络模型为分类神经网络模型。
上述技术方案的有益效果是：提供一种句子相似度判断方法，能够解决现有技术中计算句子相似度比较依赖预训练的字/词向量的质量和未登录词的问题，从而改进计算句子相似度的度量方法。
附图说明
图1是现有技术中，处理得到句子相似度的流程示意图；
图2-3是本发明的较佳的实施例中，一种句子相似度判断方法的总体流程示意图；
图4是本发明的较佳的实施例中，形成重叠特征矩阵的具体流程示意图；
图5是本发明的一个较佳的实施例中，第一神经网络模型的结构示意图；
图6是本发明的一个较佳的实施例中，第二神经网络模型的结构示意图。
具体实施方式
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行 清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动的前提下所获得的所有其他实施例，都属于本发明保护的范围。
需要说明的是，在不冲突的情况下，本发明中的实施例及实施例中的特征可以相互组合。
下面结合附图和具体实施例对本发明作进一步说明，但不作为本发明的限定。
根据现有技术中存在的上述问题，现提供一种句子相似度判断方法，该判断方法中，可以通过预先训练形成一句子相似度判断模型，句子相似度判断模型中包括一用于处理得到句子向量的第一神经网络模型以及一用于处理得到表示句子相似度的相似性度量的第二神经网络模型。
具体地，上述第一神经网络模型和第二神经网络模型是通过统一训练一体形成的，即首先搭建包括第一神经网络模型和第二神经网络模型的句子相似度判断模型(将第一神经网络模型的输出作为第二神经网络模型的输入)，随后通过向第一神经网络模型输入训练样本的方式训练形成整个句子相似度判断模型。
则该方法具体如图2所示，包括：
步骤S1，根据两个外部输入的句子样本，分别获取每个句子样本中的字词向量矩阵；
步骤S2,分别提取每个句子样本中的重叠特征以形成重叠特征矩阵，并针对每个句子样本将对应的字词向量矩阵与重叠特征矩阵结合作为第一神经网络模型的输入数据；
步骤S3，根据第一神经网络模型分别处理得到针对每个句子样本的句子向量并进行操作形成一句子合并向量，并与根据重叠特征形成的重叠特征向量结合作为第二神经网络模型的输入数据；
步骤S4，根据第二神经网络模型处理得到关联于两个句子样本的相似性度量并输出，以作为判断两个句子样本的相似度的依据。
上述步骤S3中，采用句子向量直接相减的操作方式形成句子合并向量，或者采用拼接句子向量的操作方式形成句子合并向量。
具体地，本实施例中，对于两个给定的句子样本，首先分别获取每个句 子样本中的字词向量矩阵。所谓字词向量矩阵，是指由句子中的字词向量映射形成的矩阵。
随后，本实施例中，获取每个句子样本中的重叠特征以形成重叠特征矩阵，该重叠特征为根据两个句子样本中相互重叠的字/词提取到的重叠的字词特征，该重叠特征矩阵为重叠特征按照上述映射形成字词向量矩阵相同的映射方法形成的矩阵。
本实施例中，获取到上述字词向量矩阵和重叠特征矩阵后，将关联于同一个句子样本的这两个矩阵结合作为第一神经网络模型的输入数据，随后第一神经网络模型通过处理得到针对每个句子样本的句子向量。
本发明的一个较佳的实施例中，将两个句子样本的句子向量进行相减操作，该相减操作的具体方法在下文中详述。并且，针对上文中得到的重叠特征形成一重叠特征向量，与形成的句子合并向量一起结合作为第二神经网络模型的输入数据。
本发明的另一个较佳的实施例中，将两个句子样本的句子向量进行拼接操作，该拼接操作的具体方法与现有技术中相同。并且，针对上文中得到的重叠特征形成一重叠特征向量，与形成的句子合并向量一起结合作为第二神经网络模型的输入数据。
本实施例中，最后通过第二神经网络模型处理得到两个句子样本的相似性度量，以作为判断两个句子样本的相似度的依据。
本发明的一个较佳的实施例中，相对于现有技术中对句子相似度的判断方法(如图1中所示)，在图3中示出了本发明技术方案中做出改进的部分。主要在于引入了两个句子的重叠特征，并且将该重叠特征进行处理以分别作为第一神经网络模型的输入数据(重叠特征矩阵)以及作为第二神经网络模型的输入数据(重叠特征向量)，因此使得神经网络模型较少依赖预训练的字词向量的质量，并且解决了未登录词的问题，同时，将现有技术中对句子向量进行拼接的方式更改为既可以拼接也可以相减。上述方法改进了计算句子相似度的模型，最终改进了计算句子相似性的度量方法。
本发明的较佳的实施例中，上述步骤S1中，每个句子样本的字词向量矩阵包括：
每个句子样本的字向量矩阵；或者
每个句子样本的词向量矩阵。
即上述字词向量矩阵包括每个句子样本的字/词向量矩阵。
则本实施例中，在上述步骤S1中：
将句子样本切分成字序列，并将字序列映射成字向量矩阵；或者
将句子样本切分成词序列，并将词序列映射成词向量矩阵。
本发明的较佳的实施例中，上述步骤S2中，采用如图4所示的下述方式处理形成重叠特征矩阵：
步骤S21，将两个句子样本中相互重叠的字或词分别替换成一第一字符；
步骤S22，将两个句子样本中不相重叠的字或词分别替换成一第二字符；
步骤S23，根据第一字符和第二字符分别形成关联于每个句子样本的重叠特征序列；
步骤S24，将每个重叠特征序列映射成重叠特征矩阵；
步骤S25，每个字词向量矩阵和对应的重叠特征矩阵分别结合作为第一神经网络模型的输入数据。
具体地，本实施例中，上述步骤中，为了方便计算机进行处理，上述第一字符可以为1，第二字符可以为0，则可以形成关联于每个句子样本的二进制的重叠特征向量。例如，对于两个句子样本“我要听歌”和“给我放首歌”，相互重叠的部分(即重叠特征)分别为“我”和“歌”，则针对“我要听歌”的重叠特征序列为1001，针对“给我放首歌”的重叠特征序列为01001，随后根据字词向量映射成字词向量矩阵的相同方法将上述两个重叠特征序列1001和01001分别映射形成重叠特征矩阵，即字符0映射成一维向量，字符1映射成一维向量，随后形成矩阵，并将每个句子样本的字词向量矩阵和重叠特征矩阵结合作为第一神经网络模型的输入数据。
本发明的其他实施例中，上述第一字符和第二字符也可以选择其他适于处理的形式，在此不再赘述。
本发明的较佳的实施例中，上述步骤S3中，重叠特征向量的形成方式可以包括如下几种：
1)以s1表示其中一个句子样本，s2表示另一个句子样本，并采用IDF_overlap表示两个句子样本中相互重叠的字词的IDF(Inverse Document Frequency，逆向文档频率)之和，采用length表示每个句子样本的句子长度， 则重叠特征向量feat可以被表示为feat＝IDF_overlap/(length(s1)+length(s2))。
上文中，某一个特定字/词的IDF数，可以由总文件的数目除以包含该字/词的文件数目，再将得到的商取对数得到。下文中不再赘述。
2)同样以s1表示其中一个句子样本，s2表示另一个句子样本，并采用IDF_overlap表示两个句子样本中相互重叠的字词的IDF之和，采用IDF_sum表示每个句子样本中所有字词的IDF之和，则重叠特征向量feat可以被表示为feat＝IDF_overlap/(IDF_sum(s1)+IDF_sum(s2))。
3)同样以s1表示其中一个句子样本，s2表示另一个句子样本，并采用length表示每个句子样本的句子长度，采用word_overlap表示两个句子样本中的字重叠数，则上述重叠特征向量feat可以被表示为feat＝word_overlap/(length(s1)+length(s2))。
上述三种方法都能处理得到重叠特征向量，并直接将重叠特征向量拼接到第二神经网络模型的输入数据中。
本发明的一个较佳的实施例中，计算上述重叠特征向量的时候，也可以先将句子中的停止词去掉，再计算重叠特征向量。所谓停止词(Stop Words)，主要包括英文字符、数字、数学字符、标点符号及使用频率特高的单汉字等，在文本处理过程中如果遇到停止词，则立即停止处理，将其扔掉。
本发明的较佳的实施例中，上述步骤S3中，对两个句子向量执行相减操作能够更好地找到两个句子向量之间的差异(如图3所示)。进一步地，可以采用下述几种方式实现两个句子向量的相减操作：
1)直接将两个句子向量相减得到结果；
2)将两个句子向量相减，再取绝对值得到结果；
3)上述第一神经网络模型可以为一卷积神经网络模型，卷积神经网络分为卷积层和采样层(如图5所示)，则可以在卷积层处理之后直接应用上述两种方式中的一种对两个向量进行相减，随后再在采样层进行采样，最终得到结果。
本发明的较佳的实施例中，在上述步骤S3中，在采用相减的方式对两个句子向量进行处理的同时，处理得到两个句子向量的相似度乘积，并将相似度乘积、句子向量相减的结果以及重叠特征向量结合作为第二神经网络的输入数据(如图3所示)。
具体地，上述相似度乘积可以采用下述几种方式处理得到：
1)计算两个句子向量的点积，以作为上述相似度乘积；
2)引入一参数矩阵M，并以x和y分别表示两个句子向量，则上述相似度乘积可以被表示为x*M*y。本发明的较佳的实施例中，上述参数矩阵M可以在训练形成句子相似度判断模型(即统一训练形成第一神经网络模型和第二神经网络模型时)时一起训练形成。
本发明的其他实施例中，上述步骤S3中，可以不对句子向量进行相减操作，而采用与现有技术中类似的句子向量拼接方式对两个句子向量进行拼接处理，并与根据重叠特征形成的重叠特征向量结合作为第二神经网络模型的输入数据(如图3中所示，在图3中，可以选择采用句子向量拼接或者句子向量相减的方式进行处理)。
本发明的较佳的实施例中，上述第一神经网络模型可以为深度神经网络模型，进一步地可以为卷积神经网络模型(Convolutional Neural Network，CNN)，或者为循环神经网络模型(Recurrent Neural Network，RNN)，甚至可以为循环神经网络模型的变体，例如长短期记忆神经网络模型(Long Short Term Memory，LSTM)或者门限循环神经网络模型(Gated Recurrent Unit，GRU)。
本发明的较佳的实施例中，上述第二神经网络模型可以为分类神经网络模型，如图6所示为第二神经网络模型的一般结构，该第二神经网络模型可以被划分为输入层、隐层和输出层，输出层也就是分类层，上述隐层也可以去除，即只存在输入层和输出层(分类层)。
本发明技术方案中提供了一种句子相似度判断方法，该方法引入了句子向量的重叠特征并分别作为深度神经网络模型和分类神经网络模型的输入数据，并且在处理过程中将句子向量的拼接过程更改为对句子向量做相减操作的过程，因此能够解决现有技术中计算句子相似度比较依赖预训练的字/词向量的质量以及未登录词的问题，从而改进计算句子相似度的度量方法。
值得注意的是，在不考虑相似性度量的质量比较依赖预训练的字/词以及未登录词等问题的前提下，本发明技术方案中的一些技术特征都可以被替代或者被移除，而依然可以作为一个完整的句子相似度判断方法进行应用。例如：
1)对句子向量进行相减操作的过程可以修改为传统流程中对句子向量进行拼接的过程，不影响整体判断流程的进行；
2)在第一神经网络模型的输入数据中去除由重叠特征形成的重叠特征矩阵，而只将重叠特征形成的重叠特征向量作为第二神经网络模型的输入数据，同样不影响整体判断流程的进行；
3)在第二神经网络模型的输入数据中去除由重叠特征形成的重叠特征向量，而只将重叠特征形成的重叠特征矩阵作为第一神经网络模型的输入数据，同样不影响整体判断流程的进行；
4)去除重叠特征，只将现有技术中的句子向量拼接的操作修改为句子向量相减，同样不影响整体判断流程的进行。
本发明技术方案中提供的句子相似度判断方法，能够适用于使用者与智能设备之间进行“聊天”的场景。例如：当使用者向智能设备说一句话时，智能设备通过后台处理给出应答的过程通常为：通过智能设备后台的备选资料库检索得到初步的候选句子集，随后采用本发明技术方案中提供的句子相似度判断方法从候选句子集中得到关联于使用者说的话的相似句子，随后将该相似句子对应的回答反馈给使用者。
以上所述仅为本发明较佳的实施例，并非因此限制本发明的实施方式及保护范围，对于本领域技术人员而言，应当能够意识到凡运用本发明说明书及图示内容所作出的等同替换和显而易见的变化所得到的方案，均应当包含在本发明的保护范围内。

Claims (8)

  1. 一种句子相似度判断方法，其特征在于，通过预先训练形成一句子相似度判断模型，所述句子相似度判断模型中包括一用于处理得到句子向量的第一神经网络模型以及一用于处理得到表示句子相似度的相似性度量的第二神经网络模型；
    所述句子相似度判断方法还包括：
    步骤S1，根据两个外部输入的句子样本，分别获取每个所述句子样本中的字词向量矩阵；
    步骤S2,分别提取每个所述句子样本中的重叠特征以形成重叠特征矩阵，并针对每个所述句子样本将对应的所述字词向量矩阵与所述重叠特征矩阵结合作为所述第一神经网络模型的输入数据；
    步骤S3，根据所述第一神经网络模型分别处理得到针对每个所述句子样本的所述句子向量并进行操作形成一句子合并向量，并与根据所述重叠特征形成的重叠特征向量结合作为所述第二神经网络模型的输入数据；
    步骤S4，根据所述第二神经网络模型处理得到关联于两个所述句子样本的相似性度量并输出，以作为判断两个所述句子样本的相似度的依据；
    所述步骤S3中，采用所述句子向量直接相减的操作方式形成所述句子合并向量，或者采用拼接所述句子向量的操作方式形成所述句子合并向量。
  2. 如权利要求1所述的句子相似度判断方法，其特征在于，所述步骤S1中，每个所述句子样本的字词向量矩阵包括：
    每个所述句子样本的字向量矩阵；或者
    每个所述句子样本的词向量矩阵；
    则所述步骤S1中：
    将所述句子样本切分成字序列，并将所述字序列映射成所述字向量矩阵；或者
    将所述句子样本切分成词序列，并将所述词序列映射成所述词向量矩阵。
  3. 如权利要求1所述的句子相似度判断方法，其特征在于，所述步骤S2中，采用下述方式处理形成所述重叠特征矩阵：
    步骤S21，将所述两个所述句子样本中相互重叠的字或词分别替换成一 第一字符；
    步骤S22，将所述两个句子样本中不相重叠的字或词分别替换成一第二字符；
    步骤S23，根据所述第一字符和所述第二字符分别形成关联于每个所述句子样本的重叠特征序列；
    步骤S24，将每个所述重叠特征序列映射成所述重叠特征矩阵；
    步骤S25，每个所述字词向量矩阵和对应的所述重叠特征矩阵分别结合作为所述第一神经网络模型的所述输入数据。
  4. 如权利要求1所述的句子相似度判断方法，其特征在于，所述步骤S3中，处理得到两个所述句子向量的相似度乘积，随后对两个所述句子向量做相减操作，并与所述相似度乘积以及所述重叠特征向量结合作为所述第二神经网络的所述输入数据。
  5. 如权利要求4所述的句子相似度判断方法，其特征在于，通过计算两个所述句子向量之间的点积得到所述相似度乘积；或者
    根据一参数矩阵处理得到所述相似度乘积；
    在预先对所述句子相似度判断模型进行训练的过程中，同时训练得到所述参数矩阵。
  6. 如权利要求1所述的句子相似度判断方法，其特征在于，所述第一神经网络模型为深度神经网络模型。
  7. 如权利要求7所述的句子相似度判断方法，其特征在于，所述第一神经网络模型为卷积神经网络模型或者循环神经网络模型。
  8. 如权利要求1所述的句子相似度判断方法，其特征在于，所述第二神经网络模型为分类神经网络模型。
PCT/CN2018/074336 2017-02-27 2018-01-26 一种句子相似度判断方法 WO2018153217A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/488,572 US10949709B2 (en) 2017-02-27 2018-01-26 Method for determining sentence similarity

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710109324.2A CN108509408B (zh) 2017-02-27 2017-02-27 一种句子相似度判断方法
CN201710109324.2 2017-02-27

Publications (1)

Publication Number Publication Date
WO2018153217A1 true WO2018153217A1 (zh) 2018-08-30

Family

ID=63252399

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/074336 WO2018153217A1 (zh) 2017-02-27 2018-01-26 一种句子相似度判断方法

Country Status (4)

Country Link
US (1) US10949709B2 (zh)
CN (1) CN108509408B (zh)
TW (1) TWI657342B (zh)
WO (1) WO2018153217A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110287312A (zh) * 2019-05-10 2019-09-27 平安科技（深圳）有限公司 文本相似度的计算方法、装置、计算机设备及计算机存储介质
CN110750977A (zh) * 2019-10-23 2020-02-04 支付宝(杭州)信息技术有限公司 一种文本相似度计算方法及系统
CN111400462A (zh) * 2019-01-02 2020-07-10 珠海格力电器股份有限公司 问答机器人训练方法、系统及问答机器人
CN111489754A (zh) * 2019-01-28 2020-08-04 国家电网有限公司客户服务中心 一种基于智能语音技术的话务数据分析方法
CN111666482A (zh) * 2019-03-06 2020-09-15 珠海格力电器股份有限公司 查询方法及装置、存储介质和处理器

Families Citing this family (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108509408B (zh) * 2017-02-27 2019-11-22 芋头科技（杭州）有限公司 一种句子相似度判断方法
CN108287858B (zh) * 2017-03-02 2021-08-10 腾讯科技（深圳）有限公司 自然语言的语义提取方法及装置
CN109522921A (zh) * 2018-09-18 2019-03-26 义语智能科技（上海）有限公司 语句相似度判别方法及设备
CN109614485B (zh) * 2018-11-19 2023-03-14 中山大学 一种基于语法结构的分层Attention的句子匹配方法及装置
CN109740728B (zh) * 2018-12-10 2019-11-01 杭州世平信息科技有限公司 一种基于多种神经网络组合的量刑计算方法
CN109766547B (zh) * 2018-12-26 2022-10-18 重庆邮电大学 一种句子相似度计算方法
CN109740126B (zh) * 2019-01-04 2023-11-21 平安科技（深圳）有限公司 文本匹配方法、装置及存储介质、计算机设备
CN109992659B (zh) * 2019-02-12 2023-02-17 创新先进技术有限公司 用于文本排序的方法和装置
WO2020198855A1 (en) * 2019-03-29 2020-10-08 Knowtions Research Inc. Method and system for mapping text phrases to a taxonomy
CN110298035B (zh) * 2019-06-04 2023-12-01 平安科技（深圳）有限公司 基于人工智能的字向量定义方法、装置、设备及存储介质
CN110362681B (zh) * 2019-06-19 2023-09-22 平安科技（深圳）有限公司 问答系统重复问题识别方法、装置及存储介质
CN110347776A (zh) * 2019-07-17 2019-10-18 北京百度网讯科技有限公司 兴趣点名称匹配方法、装置、设备及存储介质
TWI745724B (zh) * 2019-07-25 2021-11-11 國泰人壽保險股份有限公司 行動文件辨識系統
KR102098734B1 (ko) * 2019-08-06 2020-04-08 전자부품연구원 대화 상대의 외형을 반영한 수어 영상 제공 방법, 장치 및 단말
CN110689023B (zh) * 2019-08-15 2024-01-16 平安科技（深圳）有限公司 可靠组合特征提取方法、装置、计算机设备及存储介质
CN111666755A (zh) * 2020-06-24 2020-09-15 深圳前海微众银行股份有限公司 一种复述句识别的方法及装置
CN111737991B (zh) * 2020-07-01 2023-12-12 携程计算机技术（上海）有限公司 文本断句位置的识别方法及系统、电子设备及存储介质
CN111814717B (zh) * 2020-07-17 2022-09-27 腾讯科技（深圳）有限公司 人脸识别方法、装置及电子设备
CN111783458B (zh) * 2020-08-20 2024-05-03 支付宝(杭州)信息技术有限公司 叠字错误检测方法及装置
CN113468872B (zh) * 2021-06-09 2024-04-16 大连理工大学 基于句子级别图卷积的生物医学关系抽取方法及系统
CN113378970B (zh) * 2021-06-28 2023-08-22 山东浪潮成方数字服务有限公司 语句相似性检测方法、装置、电子设备及存储介质
CN113688571B (zh) * 2021-08-25 2023-06-30 哈尔滨工程大学 一种基于卷积相似度的水声环境参数分类方法
CN115017915B (zh) * 2022-05-30 2023-05-30 北京三快在线科技有限公司 一种模型训练、任务执行的方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6810376B1 (en) * 2000-07-11 2004-10-26 Nusuara Technologies Sdn Bhd System and methods for determining semantic similarity of sentences
CN101566998A (zh) * 2009-05-26 2009-10-28 华中师范大学 一种基于神经网络的中文问答系统
CN104834735A (zh) * 2015-05-18 2015-08-12 大连理工大学 一种基于词向量的文档摘要自动提取方法
CN105183714A (zh) * 2015-08-27 2015-12-23 北京时代焦点国际教育咨询有限责任公司 句子相似度计算方法及装置

Family Cites Families (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6137911A (en) * 1997-06-16 2000-10-24 The Dialog Corporation Plc Test classification system and method
US6026398A (en) * 1997-10-16 2000-02-15 Imarket, Incorporated System and methods for searching and matching databases
JP3099797B2 (ja) * 1998-03-19 2000-10-16 日本電気株式会社 文字認識装置
FR2825814B1 (fr) * 2001-06-07 2003-09-19 Commissariat Energie Atomique Procede de creation automatique d'une base de donnees images interrogeable par son contenu semantique
EP1668541A1 (en) * 2003-09-30 2006-06-14 British Telecommunications Public Limited Company Information retrieval
US8176054B2 (en) * 2007-07-12 2012-05-08 Ricoh Co. Ltd Retrieving electronic documents by converting them to synthetic text
US7646940B2 (en) * 2006-04-04 2010-01-12 Microsoft Corporation Robust indexing and retrieval of electronic ink
TWI396184B (zh) * 2009-09-17 2013-05-11 Tze Fen Li 一種語音辨認所有語言及用語音輸入單字的方法
CN103578464B (zh) * 2013-10-18 2017-01-11 威盛电子股份有限公司 语言模型的建立方法、语音辨识方法及电子装置
US9607245B2 (en) * 2014-12-02 2017-03-28 Xerox Corporation Adapted vocabularies for matching image signatures with fisher vectors
CN106445920A (zh) * 2016-09-29 2017-02-22 北京理工大学 利用句义结构特征的句子相似度计算方法
CN108509408B (zh) * 2017-02-27 2019-11-22 芋头科技（杭州）有限公司 一种句子相似度判断方法
US10268676B1 (en) * 2017-06-09 2019-04-23 Act, Inc. Automated detection of linguistic uniqueness
CN107491547B (zh) * 2017-08-28 2020-11-10 北京百度网讯科技有限公司 基于人工智能的搜索方法和装置
US10437936B2 (en) * 2018-02-01 2019-10-08 Jungle Disk, L.L.C. Generative text using a personality model

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6810376B1 (en) * 2000-07-11 2004-10-26 Nusuara Technologies Sdn Bhd System and methods for determining semantic similarity of sentences
CN101566998A (zh) * 2009-05-26 2009-10-28 华中师范大学 一种基于神经网络的中文问答系统
CN104834735A (zh) * 2015-05-18 2015-08-12 大连理工大学 一种基于词向量的文档摘要自动提取方法
CN105183714A (zh) * 2015-08-27 2015-12-23 北京时代焦点国际教育咨询有限责任公司 句子相似度计算方法及装置

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111400462A (zh) * 2019-01-02 2020-07-10 珠海格力电器股份有限公司 问答机器人训练方法、系统及问答机器人
CN111489754A (zh) * 2019-01-28 2020-08-04 国家电网有限公司客户服务中心 一种基于智能语音技术的话务数据分析方法
CN111666482A (zh) * 2019-03-06 2020-09-15 珠海格力电器股份有限公司 查询方法及装置、存储介质和处理器
CN111666482B (zh) * 2019-03-06 2022-08-02 珠海格力电器股份有限公司 查询方法及装置、存储介质和处理器
CN110287312A (zh) * 2019-05-10 2019-09-27 平安科技（深圳）有限公司 文本相似度的计算方法、装置、计算机设备及计算机存储介质
CN110287312B (zh) * 2019-05-10 2023-08-25 平安科技（深圳）有限公司 文本相似度的计算方法、装置、计算机设备及计算机存储介质
CN110750977A (zh) * 2019-10-23 2020-02-04 支付宝(杭州)信息技术有限公司 一种文本相似度计算方法及系统
CN110750977B (zh) * 2019-10-23 2023-06-02 支付宝(杭州)信息技术有限公司 一种文本相似度计算方法及系统

Also Published As

Publication number Publication date
US20200193217A1 (en) 2020-06-18
TWI657342B (zh) 2019-04-21
CN108509408A (zh) 2018-09-07
TW201839630A (zh) 2018-11-01
CN108509408B (zh) 2019-11-22
US10949709B2 (en) 2021-03-16

Similar Documents

Publication Publication Date Title
WO2018153217A1 (zh) 一种句子相似度判断方法
CN113962315B (zh) 模型预训练方法、装置、设备、存储介质以及程序产品
JP6718828B2 (ja) 情報入力方法および装置
CN111144131B (zh) 一种基于预训练语言模型的网络谣言检测方法
CN107301170B (zh) 基于人工智能的切分语句的方法和装置
WO2016197767A2 (zh) 一种表情输入方法、装置、终端和计算机可读存储介质
CN104598644B (zh) 喜好标签挖掘方法和装置
CN111241294A (zh) 基于依赖解析和关键词的图卷积网络的关系抽取方法
TW201837746A (zh) 特徵向量的產生、搜索方法、裝置及電子設備
WO2020052069A1 (zh) 用于分词的方法和装置
WO2021204014A1 (zh) 一种模型训练的方法及相关装置
CN107273348B (zh) 一种文本的话题和情感联合检测方法及装置
WO2018166115A1 (zh) 客服问答数据的处理方法、电子装置及计算机可读存储介质
WO2019154411A1 (zh) 词向量更新方法和装置
CN107515856A (zh) 一种基于局部信息表示的细粒度情感元素抽取方法
JP2018025874A (ja) テキスト解析装置及びプログラム
WO2022257452A1 (zh) 表情回复方法、装置、设备及存储介质
CN111813923A (zh) 文本摘要方法、电子设备及存储介质
CN111368066B (zh) 获取对话摘要的方法、装置和计算机可读存储介质
CN106776557B (zh) 情感机器人的情感状态记忆识别方法及装置
CN111832302A (zh) 一种命名实体识别方法和装置
CN111401038B (zh) 文本处理方法、装置、电子设备及存储介质
JP2019082860A (ja) 生成プログラム、生成方法及び生成装置
Bauersfeld et al. Cracking double-blind review: Authorship attribution with deep learning
CN111401069A (zh) 会话文本的意图识别方法、意图识别装置及终端

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: 18756542

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: 18756542

Country of ref document: EP

Kind code of ref document: A1