WO2021164199A1 - 基于多粒度融合模型的中文句子语义智能匹配方法及装置 - Google Patents
基于多粒度融合模型的中文句子语义智能匹配方法及装置 Download PDFInfo
- Publication number
- WO2021164199A1 WO2021164199A1 PCT/CN2020/104723 CN2020104723W WO2021164199A1 WO 2021164199 A1 WO2021164199 A1 WO 2021164199A1 CN 2020104723 W CN2020104723 W CN 2020104723W WO 2021164199 A1 WO2021164199 A1 WO 2021164199A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- sentence
- word
- character
- matching
- vector
- Prior art date
Links
- 230000004927 fusion Effects 0.000 title claims abstract description 72
- 238000000034 method Methods 0.000 title claims abstract description 46
- 238000012549 training Methods 0.000 claims abstract description 115
- 238000013507 mapping Methods 0.000 claims abstract description 40
- 238000010276 construction Methods 0.000 claims abstract description 29
- 238000006243 chemical reaction Methods 0.000 claims abstract description 28
- 239000013598 vector Substances 0.000 claims description 208
- 238000000605 extraction Methods 0.000 claims description 40
- 238000012545 processing Methods 0.000 claims description 24
- 239000011159 matrix material Substances 0.000 claims description 21
- 230000002452 interceptive effect Effects 0.000 claims description 17
- 230000011218 segmentation Effects 0.000 claims description 17
- 238000005457 optimization Methods 0.000 claims description 14
- 108010001267 Protein Subunits Proteins 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 11
- 238000007781 pre-processing Methods 0.000 claims description 8
- 230000001174 ascending effect Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 6
- 230000007246 mechanism Effects 0.000 claims description 5
- 238000013461 design Methods 0.000 claims description 4
- 239000000284 extract Substances 0.000 claims description 4
- 230000009193 crawling Effects 0.000 claims description 3
- 238000003058 natural language processing Methods 0.000 abstract description 3
- 238000004458 analytical method Methods 0.000 abstract description 2
- 238000013473 artificial intelligence Methods 0.000 abstract description 2
- 230000003993 interaction Effects 0.000 abstract 1
- 235000019580 granularity Nutrition 0.000 description 30
- 230000006870 function Effects 0.000 description 29
- 238000010586 diagram Methods 0.000 description 9
- 238000007476 Maximum Likelihood Methods 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- 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
- G06F16/3344—Query execution using natural language analysis
-
- 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
- G06F16/35—Clustering; Classification
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- 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
-
- 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
Definitions
- the invention relates to the field of artificial intelligence and natural language processing, in particular to a method and device for intelligently matching Chinese sentence semantics based on a multi-granularity fusion model.
- Sentence semantic matching plays a key role in many natural language processing tasks, such as question answering (QA), natural language inference (NLI), machine translation (MT), and so on.
- the key to sentence semantic matching is to calculate the degree of matching between the semantics of a given sentence pair.
- Sentences can be segmented at different granularities, such as characters, words, and phrases.
- the commonly used text segmentation granularity is words, especially in the Chinese field.
- the patent document with the patent number CN106569999A discloses a multi-granularity short text semantic similarity comparison method, which includes the following steps: S1, short text preprocessing; the preprocessing includes Chinese word segmentation and part-of-speech tagging; S2. Feature selection is performed on the preprocessed short text; S3, distance measurement is performed on the vector set after feature selection to determine the similarity of the short text.
- S1 short text preprocessing
- the preprocessing includes Chinese word segmentation and part-of-speech tagging
- S2 Feature selection is performed on the preprocessed short text
- S3 distance measurement is performed on the vector set after feature selection to determine the similarity of the short text.
- the technical task of the present invention is to provide a Chinese sentence semantic intelligent matching method and device based on a multi-granularity fusion model to solve the problems of incomplete semantic analysis of a single-granularity model and imprecise sentence matching.
- the intelligent matching method for Chinese sentence semantics based on a multi-granularity fusion model is specifically as follows:
- S303 Construct a multi-granularity embedding layer: perform vector mapping on words and characters in the sentence to obtain word-level sentence vectors and character-level sentence vectors;
- S304 Construct a multi-granularity fusion coding layer: perform coding processing on word-level sentence vectors and character-level sentence vectors to obtain sentence semantic feature vectors;
- S305 Construct an interactive matching layer: perform hierarchical comparison of sentence semantic feature vectors to obtain matching representation vectors of sentence pairs;
- the construction of the text matching knowledge base in step S1 is specifically as follows:
- Preprocess the original data preprocess the similar text in the original similar sentence knowledge base, and perform word segmentation and hyphenation processing on each sentence to obtain a text matching knowledge base; wherein, the word segmentation processing is based on each word in Chinese Unit, perform word segmentation operations on each piece of data; hyphenation processing is based on each character in Chinese as the basic unit, and each piece of data is hyphenated; each Chinese character and word is divided by a space, and each word is reserved. All contents including numbers, punctuation and special characters in the data;
- the training data set for constructing the text matching model in the step S2 is specifically as follows:
- Q1-char represents sentence 1 at character level granularity
- Q1-word represents sentence 1 at word level granularity
- Q2-char represents sentence 2 at character level granularity
- Q2-word represents sentence 2 at word level granularity
- 1 represents sentence 1 It matches the two texts of sentence 2, which is a positive example
- Q1-char represents sentence 1 at character level granularity
- Q1-word represents sentence 1 at word level granularity
- Q2-char represents sentence 2 at character level granularity
- Q2-word represents sentence 2 at word level granularity
- 0 represents sentence Q1 It does not match the two texts of sentence Q2, which is a negative example
- step S203 Construct a training data set: Combine all the positive samples and negative samples obtained after the operations of step S201 and step S202, and disrupt their order to construct a final training data set; among them, whether it is positive data
- Still negative data includes five dimensions, namely Q1-char, Q1-word, Q2-char, Q2-word, 0 or 1.
- step S301 the construction of the character word mapping conversion table in step S301 is specifically as follows:
- the character vocabulary table is constructed by the text matching knowledge base obtained after preprocessing;
- each character and word in the table is mapped to a unique digital identifier.
- the mapping rule is: start with the number 1, and then follow the order in which each character and word is entered into the character vocabulary table Sort in ascending order to form a character-word mapping conversion table;
- the construction of the input layer in the step S302 is specifically as follows:
- the input layer includes four inputs.
- the two sentences to be matched are preprocessed to obtain Q1-char, Q1-word, Q2-char, and Q2-word respectively, which are formalized as: (Q1-char, Q1- word,Q2-char,Q2-word);
- step S303 the construction of the multi-granularity embedding layer in step S303 is specifically as follows:
- the word-level sentence vector and character-level sentence vector Q1-word Emd, Q1-char Emd, Q2-word Emd, Q2-char Emd are obtained after multi-granular embedding layer processing; among them, text
- Each sentence in the matching knowledge base can transform text information into vector form through character word vector mapping;
- the construction of the multi-granularity fusion coding layer in step S304 takes the word-level sentence vector and character-level sentence vector output by the multi-granularity embedding layer in step S303 as input, and obtains text semantic features from two perspectives, namely, character-level semantic feature extraction and Word-level semantic feature extraction; then through the form of bitwise addition, the text semantic features from the two perspectives are integrated to obtain the final sentence semantic feature vector; for sentence Q1, the final sentence semantic feature vector is obtained as follows:
- i represents the relative position of the corresponding character vector in the sentence
- Q i is the corresponding vector representation of each character in the sentence Q1
- Q′ i is the corresponding vector representation of each character after the initial LSTM encoding
- Q′′ i is the corresponding vector representation of each character after the initial LSTM encoding.
- i ' represents the relative position of the corresponding word in the sentence vector;
- Q i' expressed as Q1 respective sentence vectors of each word;
- Q 'i' expressed as a respective vector for each of words after the initial coding LSTM;
- Q " i′ is the corresponding vector representation of each word after the second LSTM encoding;
- step S30403 After step S30401 and step S30402, the feature vector of the corresponding character level is obtained And word-level feature vectors will with Add bitwise to get the final sentence semantic feature vector for text Q1
- the formula is as follows:
- step S30401 For sentence Q2, obtain the final sentence semantic feature vector
- step S305 the construction of the interactive matching layer in step S305 is specifically as follows:
- step S30501 After processing in step S304, the sentence semantic feature vectors of Q1 and Q2 are obtained with against with Perform three operations of subtraction, cross product, and dot product to get The formula is as follows:
- dot product also called the quantified product
- the result is the length of a vector projected in the direction of another vector, which is a scalar
- cross product also called the vector product
- the result is a vector that is perpendicular to the two existing vectors
- i represents the relative position of the corresponding semantic feature in the sentence;
- Q1 i is the text Q1 obtained by feature extraction in step S304
- Each respective feature vector is the semantic representation;
- I Q2 Q2 through the text feature extraction step S304 obtained The corresponding vector representation of each semantic feature in;
- For sentence semantic feature vector with Use Dense to further extract the feature vector obtained; Indicates that the coding dimension is 300;
- step S30504 will The result after a layer of fully connected layer encoding is the same as in step S30501 Sum, get the matching representation vector of the sentence pair
- the formula is as follows:
- step S306 The construction of the prediction layer in step S306 is specifically as follows:
- the prediction layer receives the matching representation vector output in step S305, and uses the Sigmoid function for calculation to obtain a matching degree representation y pred between [0, 1];
- the training of the multi-granularity fusion model in step S4 is specifically as follows:
- y true represents the true label, that is, the 0 and 1 flags that indicate match or not in each training example;
- y pred represents the prediction result;
- the use of balanced cross entropy can automatically balance the positive and negative samples and improve the accuracy of classification; it fuses the cross entropy with the mean square error, and the formula is as follows:
- a Chinese sentence semantic intelligent matching device based on a multi-granularity fusion model comprising:
- the text matching knowledge base building unit is used to use crawlers to crawl question sets on the Internet public question and answer platform, or use text matching data sets published on the Internet as the original similar sentence knowledge base, and then preprocess the original similar sentence knowledge base ,
- the main operation is to perform hyphenation and word segmentation on each sentence in the original similar sentence knowledge base, so as to construct a text matching knowledge base for model training;
- the training data set generation unit is used to construct training positive example data and training negative example data according to the sentences in the text matching knowledge base, and construct the final training data set based on the positive example data and the negative example data;
- the multi-granularity fusion model building unit is used to construct the character word mapping conversion table, and to construct the input layer, the multi-granularity embedding layer, the multi-granularity fusion coding layer, the interactive matching layer, and the prediction layer at the same time; among them, the multi-granularity fusion model building unit includes:
- the character word mapping conversion table constructs a subunit, which is used to segment each sentence in the text matching knowledge base according to characters and words, and store each character and word in a list in turn to obtain a character word table. Subsequently, starting with the number 1, the characters and words are entered in the character word table in ascending order, thereby forming the character word mapping conversion table required by the present invention; after the character word mapping conversion table is constructed, each character in the table Characters and words are mapped to unique digital identifiers; after that, use Word2Vec to train the character word vector model to obtain the weight of the character word vector matrix;
- the input layer constructs sub-units, which are used to convert each character and word in the input sentence into a corresponding digital identifier according to the character-word mapping conversion table, thereby completing the data input, specifically to obtain q1 and q2 respectively, and convert them Formalized as: (q1-char,q1-word,q2-char,q2-word);
- the multi-granular embedding layer constructs sub-units, which are used to load pre-trained character word vector weights, convert the character words in the input sentence into the character word vector form, and then form a complete sentence vector representation; this operation is based on the digital identification of the character words Completed by looking up the character word vector matrix;
- the multi-granularity fusion coding layer constructs sub-units, which are used to take the word-level sentence vector and character-level sentence vector output by the multi-granularity embedding layer as input; first obtain text semantic features from two perspectives, namely, character-level semantic feature extraction and word-level Semantic feature extraction; then through the form of bitwise addition, the semantic features of the text from the two perspectives are integrated to obtain the final sentence semantic feature vector;
- the interactive matching layer constructs sub-units, which are used to perform hierarchical matching calculations on the semantic feature vectors of the two input sentences to obtain the matching representation vectors of the sentence pairs;
- the prediction layer constructs a sub-unit for receiving the matching representation vector output by the interactive matching layer, using the Sigmoid function to calculate, and obtaining the matching degree between [0,1], and finally judging the sentence pair by comparing with the established threshold Matching degree;
- the multi-granularity fusion model training unit is used to construct the loss function needed in the model training process and complete the optimization training of the model.
- the text matching knowledge base building unit includes:
- the original data processing subunit is used to hyphenate and segment the sentences in the original similar sentence knowledge base, thereby constructing the text matching knowledge base for model training;
- the training data set generating unit includes:
- the training positive example data construction subunit is used to combine semantically matched sentences in the text matching knowledge base, and add matching label 1 to it to construct the training positive example data;
- the training negative example data constructs a subunit, which is used to first select a sentence q 1 from the text matching knowledge base, and then randomly select a sentence q 2 that does not match the sentence q 1 semantically from the text matching knowledge base, and compare q 1 with q 2 Combine and add a matching label 0 to it, and construct it as training negative example data;
- the training data set construction subunit is used to combine all the training positive example data and the training negative example data, and disrupt the order to construct the final training data set;
- the multi-granularity fusion model training unit includes:
- the loss function construction subunit is used to construct the loss function and calculate the error of the text matching degree between sentence 1 and sentence 2;
- the model optimization training subunit is used to train and adjust the parameters in the model training, thereby reducing the error between the predicted matching degree between sentence 1 and sentence 2 and the true matching degree in the model training process.
- a storage medium stores a plurality of instructions, and the instructions are loaded by a processor to execute the steps of the above-mentioned intelligent matching method for Chinese sentence semantics based on a multi-granularity fusion model.
- An electronic device which includes:
- the processor is configured to execute instructions in the storage medium.
- the present invention integrates word vectors and character vectors, and effectively extracts the semantic information of Chinese sentences from the two granularities of characters and words, thereby improving the accuracy of Chinese sentence coding;
- the present invention can accurately realize the task of matching Chinese sentences
- the present invention uses mean square error (MSE) as a balance factor to improve the cross-entropy loss function, thereby designing a balanced cross-entropy loss function; the loss function can solve the problem of overfitting, and the classification boundary is processed during the training process. Fuzzy processing; at the same time, it can alleviate the problem of category imbalance between positive and negative samples;
- MSE mean square error
- the multi-granularity fusion model uses different encoding methods to generate character-level sentence vectors and word-level sentence vectors; for word-level sentence vectors, two LSTM networks are used for sequential encoding, and then the attention mechanism is used for depth Feature extraction; for character-level sentence vectors, in addition to using the same processing method as word-level sentence vectors, a layer of LSTM network and attention mechanism are added for encoding; the encoding of word-level sentence vectors and character-level sentence vectors are finally superimposed Together, as a multi-granularity fusion coding representation of a sentence, it can make the coding representation of a sentence more accurate and comprehensive;
- the present invention uses mean square error (MSE) as a balance factor to improve the cross-entropy loss function.
- MSE mean square error
- LCQMC public data set
- the present invention realizes a multi-granularity fusion model, which considers both Chinese word-level granularity and character-level granularity, and integrates multi-granularity coding to better capture semantic features.
- Figure 1 is a flow chart of a Chinese sentence semantic intelligent matching method based on a multi-granularity fusion model
- Figure 2 is a block diagram of the process of constructing a text matching knowledge base
- Figure 3 is a block diagram of the process of constructing the training data set of the text matching model
- Figure 4 is a block diagram of the process of constructing a multi-granularity fusion model
- Figure 5 is a block diagram of the process of training a multi-granularity fusion model
- Figure 6 is a schematic diagram of a multi-granularity fusion model
- Figure 7 is a schematic diagram of a multi-granularity embedding layer
- Fig. 8 is a schematic diagram of a multi-granularity fusion coding layer
- Fig. 9 is a schematic diagram of an interactive matching layer
- Fig. 10 is a block diagram of a device for intelligent semantic matching of Chinese sentences based on a multi-granularity fusion model.
- the intelligent matching method for Chinese sentence semantics based on the multi-granularity fusion model of the present invention is specifically as follows:
- the text matching data set publicly available on the Internet as the original knowledge base.
- the LCQMC data set [Liu,X.,Chen,Q.,Deng,C.,Zeng,H.,Chen,J.,Li,D.,Tang,B.:LCQMC:A large-scale Chinese question matching corpus .In:Proceedings of the 27th International Conference on Computational Linguistics.pp.1952-1962(2018)]
- this data set has a total of 260,068 pairs of annotation results, divided into three parts: 238766 training set, 8802 verification set and 12500 test set. It is a Chinese data set specially used for text matching tasks.
- Preprocess the original data preprocess the similar texts in the original similar sentence knowledge base, and perform word segmentation and hyphenation processing on each sentence to obtain a text matching knowledge base;
- step S101 The similar text obtained in step S101 is preprocessed to obtain a text matching knowledge base.
- step S102 in order to avoid the loss of semantic information, the present invention reserves all stop words in the sentence.
- the word segmentation processing takes each word in Chinese as the basic unit, and performs word segmentation operations on each piece of data; for example, take the sentence 2 "Can you apply for a one-day extension of repayment?" shown in step S101 as an example. After the word segmentation is processed, you get "Can you apply for a one-day extension of repayment?".
- the present invention records the sentence after word segmentation processing as a sentence with word level granularity.
- Hyphenation processing takes each Chinese character as the basic unit, and performs hyphenation operations on each piece of data; each Chinese character is divided by spaces, and the numbers, punctuation and special characters included in each piece of data are kept in For example, take the sentence 2 "Can you apply for a one-day repayment extension?" shown in step S101 as an example, after hyphenating it, you get "Can you apply for a one-day extension of repayment?".
- the present invention records the hyphenated sentence as a sentence with character level granularity.
- Q1-char represents sentence 1 at character level granularity
- Q1-word represents sentence 1 at word level granularity
- Q2-char represents sentence 2 at character level granularity
- Q2-word represents sentence 2 at word level granularity
- 1 represents sentence 1 It matches the two texts of sentence 2, which is a positive example
- Q1-char represents sentence 1 at character level granularity
- Q1-word represents sentence 1 at word level granularity
- Q2-char represents sentence 2 at character level granularity
- Q2-word represents sentence 2 at word level granularity
- 0 represents sentence Q1 It does not match the two texts of sentence Q2, which is a negative example
- step S203 Construct a training data set: Combine all the positive samples and negative samples obtained after the operations of step S201 and step S202, and disrupt their order to construct a final training data set; among them, whether it is positive data
- Still negative data includes five dimensions, namely Q1-char, Q1-word, Q2-char, Q2-word, 0 or 1.
- the core of the present invention is a multi-granularity fusion model, which can be divided into four parts: multi-granularity embedding layer, multi-granularity fusion coding layer, interactive matching layer, prediction layer ;
- multi-granularity embedding layer to perform vector mapping on words and characters in the sentence to obtain word-level sentence vectors and character-level sentence vectors;
- build a multi-granularity fusion coding layer to encode word-level sentence vectors and character-level sentence vectors
- the semantic feature vector of the sentence is obtained; then the interactive matching layer is constructed, and the semantic feature vector of the sentence is hierarchically compared to obtain the matching representation vector of the sentence pair; finally, the Sigmoid function of the prediction layer is processed to determine the semantic matching degree of the sentence pair.
- the details are as follows:
- the character vocabulary table is constructed by the text matching knowledge base obtained after preprocessing;
- each character and word in the table is mapped to a unique digital identifier.
- the mapping rule is: start with the number 1, and then follow the order in which each character or word is entered into the character vocabulary table Sort in ascending order to form a character-word mapping conversion table;
- embedding_matrix numpy.zeros([len(tokenizer.word_index)+1,embedding_dim])
- w2v_corpus is the training corpus, that is, all data in the text matching knowledge base; embedding_dim is the dimension of the character word vector, embedding_dim is set to 300 in the present invention, and word_set is the word list.
- the input layer includes four inputs.
- the two sentences to be matched are preprocessed to obtain Q1-char, Q1-word, Q2-char, and Q2-word respectively, which are formalized as: (Q1-char, Q1- word,Q2-char,Q2-word);
- the present invention uses the positive example text displayed in step S201 as an example to form a piece of input data.
- the result is as follows:
- mapping in the character vocabulary table the above input data is converted into a numerical representation (assuming that the mappings of characters and words that appear in sentence 2 but not in sentence 1 are "Yes”: 18, “No”: 19, “Apply”: 20, “Please”: 21, “Whether”: 22, “Apply”: 23, “Extension”: 24), the results are as follows:
- the word-level sentence vector and character-level sentence vector Q1-word Emd, Q1-char Emd, Q2-word Emd, Q2-char Emd are obtained after multi-granular embedding layer processing; among them, text
- Each sentence in the matching knowledge base can transform text information into a vector form by means of character word vector mapping; embedding_dim is set to 300 in the present invention.
- embedding_matrix is the weight of the character word vector matrix trained in step S301
- embedding_matrix.shape[0] is the size of the word table (dictionary) of the character word vector matrix
- embedding_dim is the dimension of the output character word vector
- input_length is the input The length of the sequence.
- the corresponding texts Q1 and Q2 are processed by the multi-granular embedding layer to obtain word-level sentence vectors and character-level sentence vectors Q1-word Emd, Q1-char Emd, Q2-word Emd, Q2-char Emd.
- step S304 Construct a multi-granularity fusion coding layer: as shown in FIG. 8, the word-level sentence vector and character-level sentence vector are coded to obtain the sentence semantic feature vector; the construction of the multi-granularity fusion coding layer in step S304 is the step S303
- the word-level sentence vector and character-level sentence vector output by the multi-granular embedding layer are used as input to obtain text semantic features from two perspectives, namely, character-level semantic feature extraction and word-level semantic feature extraction;
- the text semantic features from the two perspectives are integrated to obtain the final sentence semantic feature vector; for the sentence Q1, the final sentence semantic feature vector is obtained as follows:
- i represents the relative position of the corresponding character vector in the sentence
- Q i is the corresponding vector representation of each character in the sentence Q1
- Q′ i is the corresponding vector representation of each character after the initial LSTM encoding
- Q′′ i is the corresponding vector representation of each character after the initial LSTM encoding.
- i ' represents the relative position of the corresponding word in the sentence vector;
- Q i' expressed as Q1 respective sentence vectors of each word;
- Q 'i' expressed as a respective vector for each of words after the initial coding LSTM;
- Q " i′ is the corresponding vector representation of each word after the second LSTM encoding;
- step S30403 After step S30401 and step S30402, the feature vector of the corresponding character level is obtained And word-level feature vectors
- the coding dimension of the present invention is uniformly set to 300, and the present invention sets with Add bitwise to get the final sentence semantic feature vector for text Q1
- the formula is as follows:
- step S30401 For sentence Q2, obtain the final sentence semantic feature vector
- step S30501 After processing in step S304, the sentence semantic feature vectors of Q1 and Q2 are obtained with against with Perform three operations of subtraction, cross product, and dot product to get The formula is as follows:
- dot product also called the quantified product
- the result is the length of a vector projected in the direction of another vector, which is a scalar
- cross product also called the vector product
- the result is a vector that is perpendicular to the two existing vectors
- i represents the relative position of the corresponding semantic feature in the sentence;
- Q1 i is the text Q1 obtained by feature extraction in step S304
- Each respective feature vector is the semantic representation;
- I Q2 Q2 through the text feature extraction step S304 obtained The corresponding vector representation of each semantic feature in;
- For sentence semantic feature vector with Use Dense to further extract the feature vector obtained; Indicates that the coding dimension is 300;
- step S30504 will The result after a layer of fully connected layer encoding is the same as in step S30501 Sum, get the matching representation vector of the sentence pair
- the formula is as follows:
- the prediction layer receives the matching representation vector output in step S305, and uses the Sigmoid function for calculation to obtain a matching degree representation y pred between [0, 1];
- y true represents the true label, that is, the 0 and 1 flags that indicate match or not in each training example;
- y pred represents the prediction result;
- the use of balanced cross entropy can automatically balance the positive and negative samples and improve the accuracy of classification; it fuses the cross entropy with the mean square error, and the formula is as follows:
- the present invention designs a cross-entropy loss function to prevent over-fitting problems.
- cross entropy is a common loss function for training models.
- the method based on maximum likelihood estimation will generate input noise. This method may divide the training sample into 0 or 1, leading to the problem of overfitting.
- the present invention proposes to use mean square error (MSE) as a balance parameter to balance positive samples and negative samples, thereby greatly improving the performance of the model.
- MSE mean square error
- model keras.models.Model([Q1-char,Q1-word,Q2-char,Q2-word],[y pred ])
- the loss function loss selects the custom Loss in this step S401; the optimization algorithm optimizer selects the previously defined optim; Q1-char, Q1-word, Q2-char, and Q2-word are the model inputs, and y pred is the model output; evaluation Indicator metrics, the present invention selects accuracy, precision, recall, F 1 -score calculated based on recall and precision.
- the model of the present invention has achieved better results than the current model on the LCQMC public data set.
- the comparison of the experimental results is shown in the following table:
- the first fourteen lines are the experimental results of the prior art model [Liu, X., Chen, Q., Deng, C., Zeng, H., Chen, J., Li, D., Tang, B. ,2018.Lcqmc:A large-scale chinese question matching corpus,in:Proceedings of the 27th International Conference on Computational Linguistics,pp.1952–1962]. Comparing the model of the present invention with the existing model, it can be seen that the method of the present invention has the best performance compared with other methods.
- the intelligent matching device for Chinese sentence semantics based on the multi-granularity fusion model of the present invention includes:
- the text matching knowledge base building unit is used to use crawlers to crawl question sets on the Internet public question and answer platform, or use the text matching data set published on the Internet as the original similar sentence knowledge base, and then preprocess the original similar sentence knowledge base ,
- the main operation is to perform hyphenation and word segmentation on each sentence in the original similar sentence knowledge base, thereby constructing a text matching knowledge base for model training;
- the text matching knowledge base building unit includes:
- the original data processing subunit is used to hyphenate and segment the sentences in the original similar sentence knowledge base, thereby constructing the text matching knowledge base for model training;
- the training data set generation unit is used to construct training positive example data and training negative example data according to the sentences in the text matching knowledge base, and build the final training data set based on the positive and negative example data; training data set generation unit include,
- the training positive example data construction subunit is used to combine semantically matched sentences in the text matching knowledge base, and add matching label 1 to it to construct the training positive example data;
- the training negative example data constructs a subunit, which is used to first select a sentence q 1 from the text matching knowledge base, and then randomly select a sentence q 2 that does not match the sentence q 1 semantically from the text matching knowledge base, and compare q 1 with q 2 Combine and add a matching label 0 to it, and construct it as training negative example data;
- the training data set construction subunit is used to combine all the training positive example data and the training negative example data, and disrupt the order to construct the final training data set;
- the multi-granularity fusion model building unit is used to construct the character word mapping conversion table, and to construct the input layer, the multi-granularity embedding layer, the multi-granularity fusion coding layer, the interactive matching layer, and the prediction layer at the same time; among them, the multi-granularity fusion model building unit includes:
- the character word mapping conversion table constructs a subunit, which is used to segment each sentence in the text matching knowledge base according to characters and words, and store each character and word in a list in turn to obtain a character word table. Subsequently, starting with the number 1, the characters and words are entered in the character word table in ascending order, thereby forming the character word mapping conversion table required by the present invention; after the character word mapping conversion table is constructed, each character in the table Characters and words are mapped to unique digital identifiers; after that, use Word2Vec to train the character word vector model to obtain the weight of the character word vector matrix;
- the input layer constructs sub-units, which are used to convert each character and word in the input sentence into a corresponding digital identifier according to the character-word mapping conversion table, thereby completing the data input, specifically to obtain q1 and q2 respectively, and convert them Formalized as: (q1-char,q1-word,q2-char,q2-word);
- the multi-granular embedding layer constructs sub-units, which are used to load pre-trained character word vector weights, convert the character words in the input sentence into the character word vector form, and then form a complete sentence vector representation; this operation is based on the digital identification of the character words Completed by looking up the character word vector matrix;
- the multi-granularity fusion coding layer constructs sub-units, which are used to take the word-level sentence vector and character-level sentence vector output by the multi-granularity embedding layer as input; first obtain text semantic features from two perspectives, namely, character-level semantic feature extraction and word-level Semantic feature extraction; then through the form of bitwise addition, the semantic features of the text from the two perspectives are integrated to obtain the final sentence semantic feature vector;
- the interactive matching layer constructs sub-units, which are used to perform hierarchical matching calculations on the semantic feature vectors of the two input sentences to obtain the matching representation vectors of the sentence pairs;
- the prediction layer constructs a sub-unit for receiving the matching representation vector output by the interactive matching layer, using the Sigmoid function to calculate, and obtaining the matching degree between [0,1], and finally judging the sentence pair by comparing with the established threshold Matching degree;
- the multi-granular fusion model training unit is used to construct the loss function needed in the model training process and complete the optimization training of the model; the multi-granular fusion model training unit includes:
- the loss function construction subunit is used to construct the loss function and calculate the error of the text matching degree between sentence 1 and sentence 2;
- the model optimization training subunit is used to train and adjust the parameters in the model training, thereby reducing the error between the predicted matching degree between sentence 1 and sentence 2 and the true matching degree in the model training process.
- the device for intelligent semantic matching of Chinese sentences based on the multi-granularity fusion model shown in FIG. 10 can be integrated and deployed in various hardware devices, such as personal computers, workstations, and smart mobile devices.
- a plurality of instructions are stored therein, and the instructions are loaded by the processor to execute the steps of the method for intelligent semantic matching of Chinese sentences based on the multi-granularity fusion model of Embodiment 1.
- the electronic device includes:
- the processor is used to execute instructions in the storage medium.
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)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Machine Translation (AREA)
Abstract
Description
句子1 | 还款期限可以延后一天吗? |
句子2 | 是否可以申请延期一天还款? |
Claims (10)
- 基于多粒度融合模型的中文句子语义智能匹配方法,其特征在于,该方法具体如下:S1、构建文本匹配知识库;S2、构建文本匹配模型的训练数据集;S3、构建多粒度融合模型;具体如下:S301、构建字符词语映射转换表;S302、构建输入层;S303、构建多粒度嵌入层:对句子中的词语和字符进行向量映射,得到词语级句子向量和字符级句子向量;S304、构建多粒度融合编码层:对词语级句子向量和字符级句子向量进行编码处理,得到句子语义特征向量;S305、构建交互匹配层:对句子语义特征向量进行分层比较,得到句子对的匹配表征向量;S306、构建预测层:经预测层的Sigmoid函数处理,判断句子对的语义匹配程度;S4、训练多粒度融合模型。
- 根据权利要求1所述的基于多粒度融合模型的中文句子语义智能匹配方法,其特征在于,所述步骤S1中构建文本匹配知识库具体如下:S101、使用爬虫获取原始数据:在互联网公共问答平台爬取问题集,得到原始相似句子知识库;或者使用网上公开的句子匹配数据集,作为原始相似句子知识库;S102、预处理原始数据:预处理原始相似句子知识库中的相似文本,对每 个句子进行分词和断字处理,得到文本匹配知识库;其中,分词处理是以中文里的每个词语作为基本单位,对每条数据进行分词操作;断字处理是以中文里的每个字作为基本单位,对每条数据进行断字操作;每个汉字或词语之间用空格进行切分,并保留每条数据中包括的数字、标点以及特殊字符在内的所有内容;所述步骤S2中构建文本匹配模型的训练数据集具体如下:S201、构建训练正例:将句子与其对应的语义匹配的句子进行组合,构建训练正例,形式化为:(Q1-char,Q1-word,Q2-char,Q2-word,1);其中,Q1-char表示字符级粒度的句子1;Q1-word表示词语级粒度的句子1;Q2-char表示字符级粒度的句子2;Q2-word表示词语级粒度的句子2;1表示句子1和句子2这两个文本相匹配,是正例;S202、构建训练负例:选中一个句子Q1,再从文本匹配知识库中随机选择一个与句子Q1不匹配的句子Q2,将Q1与Q2进行组合,构建负例,形式化为:(Q1-char,Q1-word,Q2-char,Q2-word,0);其中,Q1-char表示字符级粒度的句子1;Q1-word表示词语级粒度的句子1;Q2-char表示字符级粒度的句子2;Q2-word表示词语级粒度的句子2;0表示句子Q1和句子Q2这两个文本不匹配,是负例;S203、构建训练数据集:将经过步骤S201和步骤S202操作后所获得的全部的正例样本和负例样本进行组合,并打乱其顺序,构建最终的训练数据集;其中,无论是正例数据还是负例数据均包含五个维度,即Q1-char、Q1-word、Q2-char、Q2-word、0或1。
- 根据权利要求1或2所述的基于多粒度融合模型的中文句子语义智能匹配方法,其特征在于,所述步骤S301中构建字符词语映射转换表具体如下:S30101、字符词语表通过预处理后得到的文本匹配知识库来构建;S30102、字符词语表构建完成后,表中每个字符、词语均被映射为唯一的 数字标识,映射规则为:以数字1为起始,随后按照每个字符、词语被录入字符词语表的顺序依次递增排序,从而形成字符词语映射转换表;S30103、使用Word2Vec训练字符词语向量模型,得到字符词语向量矩阵权重embedding_matrix;所述步骤S302中构建输入层具体如下:S30201、输入层包括四个输入,对两个待匹配的句子进行预处理分别获取Q1-char、Q1-word、Q2-char、Q2-word,将其形式化为:(Q1-char,Q1-word,Q2-char,Q2-word);S30202、对于输入句子中的每个字符和词语均按照步骤S301中构建完成的字符词语映射转换表将其转化为相应的数字标识。
- 根据权利要求3所述的基于多粒度融合模型的中文句子语义智能匹配方法,其特征在于,所述步骤S303中构建多粒度嵌入层具体如下:S30301、通过加载步骤S301中训练所得的字符词语向量矩阵权重来初始化当前层的权重参数;S30302、针对输入句子Q1和Q2,经过多粒度嵌入层处理后得到其词语级句子向量和字符级句子向量Q1-word Emd、Q1-char Emd、Q2-word Emd、Q2-char Emd;其中,文本匹配知识库中每一个句子均能通过字符词语向量映射的方式,将文本信息转化为向量形式;所述步骤S304中构建多粒度融合编码层是将步骤S303中多粒度嵌入层输出的词语级句子向量和字符级句子向量作为输入,从两个角度获取文本语义特征,即字符级别语义特征提取和词语级别语义特征提取;再通过按位相加的形式,对两个角度的文本语义特征进行整合,得到最终的句子语义特征向量;对于句子Q1求取最终的句子语义特征向量具体如下:S30401、针对字符级别语义特征提取,具体如下:其中,i表示相应字符向量在句子中的相对位置,Q i为句子Q1中每个字符的相应向量表示;Q′ i为经过初次LSTM编码后每个字符的相应向量表示;Q″ i为经过第二次LSTM编码后每个字符的相应向量表示;S30402、针对词语级别语义特征提取,具体如下:其中,i′表示相应词语向量在句子中的相对位置;Q i′为句子Q1中每个词语的相应向量表示;Q′ i′为经过初次LSTM编码后每个词语的相应向量表示;Q″ i′为经过第二次LSTM编码后每个词语的相应向量表示;
- 根据权利要求4所述的基于多粒度融合模型的中文句子语义智能匹配方法,其特征在于,所述步骤S305构建交互匹配层具体如下:其中,i表示相应语义特征在句子中的相对位置;Q1 i为文本Q1经过步骤S304特征提取得到的 中每个语义特征的相应向量表示;Q2 i为文本Q2经过步骤S304特征提取得到的 中每个语义特征的相应向量表示; 为针对句子语义特征向量 和 使用Dense进一步提取,得到的特征向量; 表示编码维度为300;所述步骤S306中构建预测层具体如下:S30601、预测层接收步骤S305输出的匹配表征向量,使用Sigmoid函数进行计算,得到处于[0,1]之间的匹配度表示y pred;S30602、将y pred与设立的阈值进行比较来判别句子对的匹配程度,具体如下:①、当y pred≥0.5时,表示句子Q1以及句子Q2相匹配;②、当y pred<0.5时,表示句子Q1以及句子Q2不匹配。
- 一种基于多粒度融合模型的中文句子语义智能匹配装置,其特征在于,该装置包括,文本匹配知识库构建单元,用于使用爬虫程序,在互联网公共问答平台爬取问题集,或者使用网上公开的文本匹配数据集,作为原始相似句子知识库,再对原始相似句子知识库进行预处理,主要操作为对原始相似句子知识库中的每个句子进行断字处理和分词处理,从而构建用于模型训练的文本匹配知识库;训练数据集生成单元,用于根据文本匹配知识库中的句子来构建训练正例数据和训练负例数据,并且基于正例数据与负例数据来构建最终的训练数据集;多粒度融合模型构建单元,用于构建字符词语映射转换表,并同时构建输入层、多粒度嵌入层、多粒度融合编码层、交互匹配层、预测层;其中,多粒度融合模型构建单元包括,字符词语映射转换表构建子单元,用于对文本匹配知识库中的每个句子按字 符和词语进行切分,并将每个字符和词语依次存入一个列表中,从而得到一个字符词语表,随后以数字1为起始,按照每个字符和词语被录入字符词语表的顺序依次递增排序,从而形成本发明所需的字符词语映射转换表;字符词语映射转换表构建完成后,表中每个字符和词语均被映射为唯一的数字标识;其后,使用Word2Vec训练字符词语向量模型,得到字符词语向量矩阵权重;输入层构建子单元,用于根据字符词语映射转换表,将输入句子中的每个字符和词语转化为相应的数字标识,从而完成数据的输入,具体来说就是分别获取q1与q2,将其形式化为:(q1-char,q1-word,q2-char,q2-word);多粒度嵌入层构建子单元,用于加载预训练好的字符词语向量权重,将输入句子中的字符词语转换为字符词语向量形式,进而构成完整的句子向量表示;该操作根据字符词语的数字标识查找字符词语向量矩阵而完成;多粒度融合编码层构建子单元,用于将多粒度嵌入层输出的词语级句子向量和字符级句子向量作为输入;先从两个角度来获取文本语义特征,即字符级别语义特征提取和词语级别语义特征提取;再通过按位相加的形式,对两个角度的文本语义特征进行整合,得到最终的句子语义特征向量;交互匹配层构建子单元,用于将输入的两个句子语义特征向量,经过分层匹配计算,得到句子对的匹配表征向量;预测层构建子单元,用于接收交互匹配层输出的匹配表征向量,使用Sigmoid函数进行计算,得到处于[0,1]之间的匹配度,最终通过与设立的阈值进行比较来判别句子对的匹配程度;多粒度融合模型训练单元,用于构建模型训练过程中所需要的损失函数,并完成模型的优化训练。
- 根据权利要求7所述的基于多粒度融合模型的中文句子语义智能匹配装置,其特征在于,所述文本匹配知识库构建单元包括,爬取原始数据子单元,用于在互联网公共问答平台爬取问题集,或者使用网上公开的文本匹配数据集,构建原始相似句子知识库;原始数据处理子单元,用于将原始相似句子知识库中的句子进行断字处理和分词处理,从而构建用于模型训练的文本匹配知识库;所述训练数据集生成单元包括,训练正例数据构建子单元,用于将文本匹配知识库中语义匹配的句子进行组合,并对其添加匹配标签1,构建为训练正例数据;训练负例数据构建子单元,用于先从文本匹配知识库中选取一个句子q 1,再从文本匹配知识库中随机选择一个与句子q 1语义不匹配的句子q 2,将q 1与q 2进行组合,并对其添加匹配标签0,构建为训练负例数据;训练数据集构建子单元,用于将所有的训练正例数据与训练负例数据组合在一起,并打乱其顺序,从而构建最终的训练数据集;所述多粒度融合模型训练单元包括,损失函数构建子单元,用于构建损失函数,计算句子1和句子2间文本匹配度的误差;模型优化训练子单元,用于训练并调整模型训练中的参数,从而减小模型训练过程中预测的句子1与句子2间匹配度与真实匹配度之间的误差。
- 一种存储介质,其中存储有多条指令,其特征在于,所述指令由处理器加载,执行权利要求1-6中所述的基于多粒度融合模型的中文句子语义智能匹配方法的步骤。
- 一种电子设备,其特征在于,所述电子设备包括:权利要求9所述的存储介质;以及处理器,用于执行所述存储介质中的指令。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010103529.1 | 2020-02-20 | ||
CN202010103529.1A CN111310438B (zh) | 2020-02-20 | 2020-02-20 | 基于多粒度融合模型的中文句子语义智能匹配方法及装置 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021164199A1 true WO2021164199A1 (zh) | 2021-08-26 |
Family
ID=71151080
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2020/104723 WO2021164199A1 (zh) | 2020-02-20 | 2020-07-27 | 基于多粒度融合模型的中文句子语义智能匹配方法及装置 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN111310438B (zh) |
WO (1) | WO2021164199A1 (zh) |
Cited By (41)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113705197A (zh) * | 2021-08-30 | 2021-11-26 | 北京工业大学 | 一种基于位置增强的细粒度情感分析方法 |
CN114153839A (zh) * | 2021-10-29 | 2022-03-08 | 杭州未名信科科技有限公司 | 多源异构数据的集成方法、装置、设备及存储介质 |
CN114218380A (zh) * | 2021-12-03 | 2022-03-22 | 淮阴工学院 | 基于多模态的冷链配载用户画像标签抽取方法及装置 |
CN114238563A (zh) * | 2021-12-08 | 2022-03-25 | 齐鲁工业大学 | 基于多角度交互的中文句子对语义智能匹配方法和装置 |
CN114239566A (zh) * | 2021-12-14 | 2022-03-25 | 公安部第三研究所 | 基于信息增强实现两步中文事件精准检测的方法、装置、处理器及其计算机可读存储介质 |
CN114281987A (zh) * | 2021-11-26 | 2022-04-05 | 重庆邮电大学 | 一种用于智能语音助手的对话短文本语句匹配方法 |
CN114297390A (zh) * | 2021-12-30 | 2022-04-08 | 江南大学 | 一种长尾分布场景下的方面类别识别方法及*** |
CN114357121A (zh) * | 2022-03-10 | 2022-04-15 | 四川大学 | 一种基于数据驱动的创新方案设计方法和*** |
CN114357158A (zh) * | 2021-12-09 | 2022-04-15 | 南京中孚信息技术有限公司 | 基于句粒度语义和相对位置编码的长文本分类技术 |
CN114416930A (zh) * | 2022-02-09 | 2022-04-29 | 上海携旅信息技术有限公司 | 搜索场景下的文本匹配方法、***、设备及存储介质 |
CN114461806A (zh) * | 2022-02-28 | 2022-05-10 | 同盾科技有限公司 | 广告识别模型的训练方法及装置、广告屏蔽方法 |
CN114492451A (zh) * | 2021-12-22 | 2022-05-13 | 马上消费金融股份有限公司 | 文本匹配方法、装置、电子设备及计算机可读存储介质 |
CN114547256A (zh) * | 2022-04-01 | 2022-05-27 | 齐鲁工业大学 | 面向消防安全知识智能问答的文本语义匹配方法和装置 |
CN114595306A (zh) * | 2022-01-26 | 2022-06-07 | 西北大学 | 基于距离感知自注意力机制和多角度建模的文本相似度计算***及方法 |
CN114742016A (zh) * | 2022-04-01 | 2022-07-12 | 山西大学 | 一种基于多粒度实体异构图的篇章级事件抽取方法及装置 |
CN115048944A (zh) * | 2022-08-16 | 2022-09-13 | 之江实验室 | 一种基于主题增强的开放域对话回复方法及*** |
CN115238684A (zh) * | 2022-09-19 | 2022-10-25 | 北京探境科技有限公司 | 一种文本采集方法、装置、计算机设备及可读存储介质 |
CN115438674A (zh) * | 2022-11-08 | 2022-12-06 | 腾讯科技(深圳)有限公司 | 实体数据处理、实体链接方法、装置和计算机设备 |
CN115600945A (zh) * | 2022-09-07 | 2023-01-13 | 淮阴工学院(Cn) | 基于多粒度的冷链配载用户画像构建方法及装置 |
CN115910345A (zh) * | 2022-12-22 | 2023-04-04 | 广东数业智能科技有限公司 | 一种心理健康测评智能预警方法及存储介质 |
CN115936014A (zh) * | 2022-11-08 | 2023-04-07 | 上海栈略数据技术有限公司 | 一种医学实体对码方法、***、计算机设备、存储介质 |
CN116071759A (zh) * | 2023-03-06 | 2023-05-05 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | 一种融合gpt2预训练大模型的光学字符识别方法 |
CN116204642A (zh) * | 2023-03-06 | 2023-06-02 | 上海阅文信息技术有限公司 | 数字阅读中角色隐式属性智能识别分析方法、***和应用 |
CN116306558A (zh) * | 2022-11-23 | 2023-06-23 | 北京语言大学 | 一种计算机辅助中文文本改编的方法及装置 |
CN116304745A (zh) * | 2023-03-27 | 2023-06-23 | 济南大学 | 基于深层次语义信息的文本话题匹配方法及*** |
CN116629275A (zh) * | 2023-07-21 | 2023-08-22 | 北京无极慧通科技有限公司 | 一种基于大数据的智能决策支持***及方法 |
CN116680590A (zh) * | 2023-07-28 | 2023-09-01 | 中国人民解放军国防科技大学 | 基于工作说明书解析的岗位画像标签提取方法及装置 |
CN116822495A (zh) * | 2023-08-31 | 2023-09-29 | 小语智能信息科技(云南)有限公司 | 基于对比学***行句对抽取方法及装置 |
CN117271438A (zh) * | 2023-07-17 | 2023-12-22 | 乾元云硕科技(深圳)有限公司 | 用于大数据的智能存储***及其方法 |
CN117390141A (zh) * | 2023-12-11 | 2024-01-12 | 江西农业大学 | 一种农业社会化服务质量用户评价数据分析方法 |
CN117556027A (zh) * | 2024-01-12 | 2024-02-13 | 一站发展(北京)云计算科技有限公司 | 基于数字人技术的智能交互***及方法 |
CN117590944A (zh) * | 2023-11-28 | 2024-02-23 | 上海源庐加佳信息科技有限公司 | 实体人对象和数字虚拟人对象的绑定*** |
CN117633518A (zh) * | 2024-01-25 | 2024-03-01 | 北京大学 | 一种产业链构建方法及*** |
CN117669593A (zh) * | 2024-01-31 | 2024-03-08 | 山东省计算中心(国家超级计算济南中心) | 基于等价语义的零样本关系抽取方法、***、设备及介质 |
CN117744787A (zh) * | 2024-02-20 | 2024-03-22 | 中国电子科技集团公司第十研究所 | 一阶研判规则知识合理性的智能度量方法 |
CN117874209A (zh) * | 2024-03-12 | 2024-04-12 | 深圳市诚立业科技发展有限公司 | 基于nlp的诈骗短信监控告警*** |
CN117910460A (zh) * | 2024-03-18 | 2024-04-19 | 国网江苏省电力有限公司南通供电分公司 | 一种基于bge模型的电力科研知识关联性构建方法及*** |
CN118093791A (zh) * | 2024-04-24 | 2024-05-28 | 北京中关村科金技术有限公司 | 结合云计算的ai知识库生成方法及*** |
CN118132683A (zh) * | 2024-05-07 | 2024-06-04 | 杭州海康威视数字技术股份有限公司 | 文本抽取模型的训练方法、文本抽取方法和设备 |
CN118153553A (zh) * | 2024-05-09 | 2024-06-07 | 江西科技师范大学 | 基于多任务的社交网络用户心理危机原因抽取方法与*** |
CN118193743A (zh) * | 2024-05-20 | 2024-06-14 | 山东齐鲁壹点传媒有限公司 | 一种基于预训练模型的多层级文本分类模型 |
Families Citing this family (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111310438B (zh) * | 2020-02-20 | 2021-06-08 | 齐鲁工业大学 | 基于多粒度融合模型的中文句子语义智能匹配方法及装置 |
CN111753524A (zh) * | 2020-07-01 | 2020-10-09 | 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***、电子设备及存储介质 |
CN111914551B (zh) * | 2020-07-29 | 2022-05-20 | 北京字节跳动网络技术有限公司 | 自然语言处理方法、装置、电子设备及存储介质 |
CN112149410A (zh) * | 2020-08-10 | 2020-12-29 | 招联消费金融有限公司 | 语义识别方法、装置、计算机设备和存储介质 |
CN112000772B (zh) * | 2020-08-24 | 2022-09-06 | 齐鲁工业大学 | 面向智能问答基于语义特征立方体的句子对语义匹配方法 |
CN112101030B (zh) * | 2020-08-24 | 2024-01-26 | 沈阳东软智能医疗科技研究院有限公司 | 建立术语映射模型、实现标准词映射的方法、装置及设备 |
CN112328890B (zh) * | 2020-11-23 | 2024-04-12 | 北京百度网讯科技有限公司 | 搜索地理位置点的方法、装置、设备及存储介质 |
CN112256841B (zh) * | 2020-11-26 | 2024-05-07 | 支付宝(杭州)信息技术有限公司 | 文本匹配和对抗文本识别方法、装置及设备 |
CN112463924B (zh) * | 2020-11-27 | 2022-07-05 | 齐鲁工业大学 | 面向智能问答基于内部相关性编码的文本意图匹配方法 |
CN112560502B (zh) * | 2020-12-28 | 2022-05-13 | 桂林电子科技大学 | 一种语义相似度匹配方法、装置及存储介质 |
CN112613282A (zh) * | 2020-12-31 | 2021-04-06 | 桂林电子科技大学 | 一种文本生成方法、装置及存储介质 |
CN112966524B (zh) * | 2021-03-26 | 2024-01-26 | 湖北工业大学 | 基于多粒度孪生网络的中文句子语义匹配方法及*** |
CN113065358B (zh) * | 2021-04-07 | 2022-05-24 | 齐鲁工业大学 | 面向银行咨询服务基于多粒度对齐的文本对语义匹配方法 |
CN113593709B (zh) * | 2021-07-30 | 2022-09-30 | 江先汉 | 一种疾病编码方法、***、可读存储介质及装置 |
CN113569014B (zh) * | 2021-08-11 | 2024-03-19 | 国家电网有限公司 | 基于多粒度文本语义信息的运维项目管理方法 |
CN113780006B (zh) * | 2021-09-27 | 2024-04-09 | 广州金域医学检验中心有限公司 | 医学语义匹配模型的训练方法、医学知识匹配方法及装置 |
CN114090747A (zh) * | 2021-10-14 | 2022-02-25 | 特斯联科技集团有限公司 | 基于多重语义匹配的自动问答方法、装置、设备及介质 |
CN114049884B (zh) * | 2022-01-11 | 2022-05-13 | 广州小鹏汽车科技有限公司 | 语音交互方法、车辆、计算机可读存储介质 |
CN115422362B (zh) * | 2022-10-09 | 2023-10-31 | 郑州数智技术研究院有限公司 | 一种基于人工智能的文本匹配方法 |
CN115688796B (zh) * | 2022-10-21 | 2023-12-05 | 北京百度网讯科技有限公司 | 用于自然语言处理领域中预训练模型的训练方法及其装置 |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180137404A1 (en) * | 2016-11-15 | 2018-05-17 | International Business Machines Corporation | Joint learning of local and global features for entity linking via neural networks |
CN108984532A (zh) * | 2018-07-27 | 2018-12-11 | 福州大学 | 基于层次嵌入的方面抽取方法 |
CN109299262A (zh) * | 2018-10-09 | 2019-02-01 | 中山大学 | 一种融合多粒度信息的文本蕴含关系识别方法 |
CN110083692A (zh) * | 2019-04-22 | 2019-08-02 | 齐鲁工业大学 | 一种金融知识问答的文本交互匹配方法及装置 |
CN110321419A (zh) * | 2019-06-28 | 2019-10-11 | 神思电子技术股份有限公司 | 一种融合深度表示与交互模型的问答匹配方法 |
CN110502627A (zh) * | 2019-08-28 | 2019-11-26 | 上海海事大学 | 一种基于多层Transformer聚合编码器的答案生成方法 |
CN111310438A (zh) * | 2020-02-20 | 2020-06-19 | 齐鲁工业大学 | 基于多粒度融合模型的中文句子语义智能匹配方法及装置 |
Family Cites Families (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104408153B (zh) * | 2014-12-03 | 2018-07-31 | 中国科学院自动化研究所 | 一种基于多粒度主题模型的短文本哈希学习方法 |
CN107315772B (zh) * | 2017-05-24 | 2019-08-16 | 北京邮电大学 | 基于深度学习的问题匹配方法以及装置 |
CN108268643A (zh) * | 2018-01-22 | 2018-07-10 | 北京邮电大学 | 一种基于多粒度lstm网络的深层语义匹配实体链接方法 |
CN109408627B (zh) * | 2018-11-15 | 2021-03-02 | 众安信息技术服务有限公司 | 一种融合卷积神经网络和循环神经网络的问答方法及*** |
CN110032639B (zh) * | 2018-12-27 | 2023-10-31 | ***股份有限公司 | 将语义文本数据与标签匹配的方法、装置及存储介质 |
CN110032635B (zh) * | 2019-04-22 | 2023-01-20 | 齐鲁工业大学 | 一种基于深度特征融合神经网络的问题对匹配方法和装置 |
CN110334184A (zh) * | 2019-07-04 | 2019-10-15 | 河海大学常州校区 | 基于机器阅读理解的智能问答*** |
-
2020
- 2020-02-20 CN CN202010103529.1A patent/CN111310438B/zh active Active
- 2020-07-27 WO PCT/CN2020/104723 patent/WO2021164199A1/zh active Application Filing
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180137404A1 (en) * | 2016-11-15 | 2018-05-17 | International Business Machines Corporation | Joint learning of local and global features for entity linking via neural networks |
CN108984532A (zh) * | 2018-07-27 | 2018-12-11 | 福州大学 | 基于层次嵌入的方面抽取方法 |
CN109299262A (zh) * | 2018-10-09 | 2019-02-01 | 中山大学 | 一种融合多粒度信息的文本蕴含关系识别方法 |
CN110083692A (zh) * | 2019-04-22 | 2019-08-02 | 齐鲁工业大学 | 一种金融知识问答的文本交互匹配方法及装置 |
CN110321419A (zh) * | 2019-06-28 | 2019-10-11 | 神思电子技术股份有限公司 | 一种融合深度表示与交互模型的问答匹配方法 |
CN110502627A (zh) * | 2019-08-28 | 2019-11-26 | 上海海事大学 | 一种基于多层Transformer聚合编码器的答案生成方法 |
CN111310438A (zh) * | 2020-02-20 | 2020-06-19 | 齐鲁工业大学 | 基于多粒度融合模型的中文句子语义智能匹配方法及装置 |
Non-Patent Citations (1)
Title |
---|
NAIXIN ZHANG ET AL.: "MIFM: Multi-Granularity Information Fusion Model for Chinese Named Entity Recognition", IEEE ACCESS, vol. 2019, no. 7, 13 December 2019 (2019-12-13), ISSN: 2169-3536 * |
Cited By (66)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113705197A (zh) * | 2021-08-30 | 2021-11-26 | 北京工业大学 | 一种基于位置增强的细粒度情感分析方法 |
CN113705197B (zh) * | 2021-08-30 | 2024-04-02 | 北京工业大学 | 一种基于位置增强的细粒度情感分析方法 |
CN114153839A (zh) * | 2021-10-29 | 2022-03-08 | 杭州未名信科科技有限公司 | 多源异构数据的集成方法、装置、设备及存储介质 |
CN114281987A (zh) * | 2021-11-26 | 2022-04-05 | 重庆邮电大学 | 一种用于智能语音助手的对话短文本语句匹配方法 |
CN114218380A (zh) * | 2021-12-03 | 2022-03-22 | 淮阴工学院 | 基于多模态的冷链配载用户画像标签抽取方法及装置 |
CN114218380B (zh) * | 2021-12-03 | 2022-07-29 | 淮阴工学院 | 基于多模态的冷链配载用户画像标签抽取方法及装置 |
CN114238563A (zh) * | 2021-12-08 | 2022-03-25 | 齐鲁工业大学 | 基于多角度交互的中文句子对语义智能匹配方法和装置 |
CN114357158A (zh) * | 2021-12-09 | 2022-04-15 | 南京中孚信息技术有限公司 | 基于句粒度语义和相对位置编码的长文本分类技术 |
CN114357158B (zh) * | 2021-12-09 | 2024-04-09 | 南京中孚信息技术有限公司 | 基于句粒度语义和相对位置编码的长文本分类技术 |
CN114239566B (zh) * | 2021-12-14 | 2024-04-23 | 公安部第三研究所 | 基于信息增强实现两步中文事件精准检测的方法、装置、处理器及其计算机可读存储介质 |
CN114239566A (zh) * | 2021-12-14 | 2022-03-25 | 公安部第三研究所 | 基于信息增强实现两步中文事件精准检测的方法、装置、处理器及其计算机可读存储介质 |
CN114492451B (zh) * | 2021-12-22 | 2023-10-24 | 马上消费金融股份有限公司 | 文本匹配方法、装置、电子设备及计算机可读存储介质 |
CN114492451A (zh) * | 2021-12-22 | 2022-05-13 | 马上消费金融股份有限公司 | 文本匹配方法、装置、电子设备及计算机可读存储介质 |
CN114297390A (zh) * | 2021-12-30 | 2022-04-08 | 江南大学 | 一种长尾分布场景下的方面类别识别方法及*** |
CN114297390B (zh) * | 2021-12-30 | 2024-04-02 | 江南大学 | 一种长尾分布场景下的方面类别识别方法及*** |
CN114595306A (zh) * | 2022-01-26 | 2022-06-07 | 西北大学 | 基于距离感知自注意力机制和多角度建模的文本相似度计算***及方法 |
CN114595306B (zh) * | 2022-01-26 | 2024-04-12 | 西北大学 | 基于距离感知自注意力机制和多角度建模的文本相似度计算***及方法 |
CN114416930A (zh) * | 2022-02-09 | 2022-04-29 | 上海携旅信息技术有限公司 | 搜索场景下的文本匹配方法、***、设备及存储介质 |
CN114461806A (zh) * | 2022-02-28 | 2022-05-10 | 同盾科技有限公司 | 广告识别模型的训练方法及装置、广告屏蔽方法 |
CN114357121A (zh) * | 2022-03-10 | 2022-04-15 | 四川大学 | 一种基于数据驱动的创新方案设计方法和*** |
CN114742016A (zh) * | 2022-04-01 | 2022-07-12 | 山西大学 | 一种基于多粒度实体异构图的篇章级事件抽取方法及装置 |
CN114547256A (zh) * | 2022-04-01 | 2022-05-27 | 齐鲁工业大学 | 面向消防安全知识智能问答的文本语义匹配方法和装置 |
CN114547256B (zh) * | 2022-04-01 | 2024-03-15 | 齐鲁工业大学 | 面向消防安全知识智能问答的文本语义匹配方法和装置 |
CN115048944A (zh) * | 2022-08-16 | 2022-09-13 | 之江实验室 | 一种基于主题增强的开放域对话回复方法及*** |
CN115600945A (zh) * | 2022-09-07 | 2023-01-13 | 淮阴工学院(Cn) | 基于多粒度的冷链配载用户画像构建方法及装置 |
CN115238684B (zh) * | 2022-09-19 | 2023-03-03 | 北京探境科技有限公司 | 一种文本采集方法、装置、计算机设备及可读存储介质 |
CN115238684A (zh) * | 2022-09-19 | 2022-10-25 | 北京探境科技有限公司 | 一种文本采集方法、装置、计算机设备及可读存储介质 |
CN115936014A (zh) * | 2022-11-08 | 2023-04-07 | 上海栈略数据技术有限公司 | 一种医学实体对码方法、***、计算机设备、存储介质 |
CN115438674A (zh) * | 2022-11-08 | 2022-12-06 | 腾讯科技(深圳)有限公司 | 实体数据处理、实体链接方法、装置和计算机设备 |
CN115936014B (zh) * | 2022-11-08 | 2023-07-25 | 上海栈略数据技术有限公司 | 一种医学实体对码方法、***、计算机设备、存储介质 |
CN115438674B (zh) * | 2022-11-08 | 2023-03-24 | 腾讯科技(深圳)有限公司 | 实体数据处理、实体链接方法、装置和计算机设备 |
CN116306558B (zh) * | 2022-11-23 | 2023-11-10 | 北京语言大学 | 一种计算机辅助中文文本改编的方法及装置 |
CN116306558A (zh) * | 2022-11-23 | 2023-06-23 | 北京语言大学 | 一种计算机辅助中文文本改编的方法及装置 |
CN115910345A (zh) * | 2022-12-22 | 2023-04-04 | 广东数业智能科技有限公司 | 一种心理健康测评智能预警方法及存储介质 |
CN116071759A (zh) * | 2023-03-06 | 2023-05-05 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | 一种融合gpt2预训练大模型的光学字符识别方法 |
CN116204642A (zh) * | 2023-03-06 | 2023-06-02 | 上海阅文信息技术有限公司 | 数字阅读中角色隐式属性智能识别分析方法、***和应用 |
CN116204642B (zh) * | 2023-03-06 | 2023-10-27 | 上海阅文信息技术有限公司 | 数字阅读中角色隐式属性智能识别分析方法、***和应用 |
CN116071759B (zh) * | 2023-03-06 | 2023-07-18 | 合肥综合性国家科学中心人工智能研究院(安徽省人工智能实验室) | 一种融合gpt2预训练大模型的光学字符识别方法 |
CN116304745B (zh) * | 2023-03-27 | 2024-04-12 | 济南大学 | 基于深层次语义信息的文本话题匹配方法及*** |
CN116304745A (zh) * | 2023-03-27 | 2023-06-23 | 济南大学 | 基于深层次语义信息的文本话题匹配方法及*** |
CN117271438A (zh) * | 2023-07-17 | 2023-12-22 | 乾元云硕科技(深圳)有限公司 | 用于大数据的智能存储***及其方法 |
CN116629275B (zh) * | 2023-07-21 | 2023-09-22 | 北京无极慧通科技有限公司 | 一种基于大数据的智能决策支持***及方法 |
CN116629275A (zh) * | 2023-07-21 | 2023-08-22 | 北京无极慧通科技有限公司 | 一种基于大数据的智能决策支持***及方法 |
CN116680590A (zh) * | 2023-07-28 | 2023-09-01 | 中国人民解放军国防科技大学 | 基于工作说明书解析的岗位画像标签提取方法及装置 |
CN116680590B (zh) * | 2023-07-28 | 2023-10-20 | 中国人民解放军国防科技大学 | 基于工作说明书解析的岗位画像标签提取方法及装置 |
CN116822495B (zh) * | 2023-08-31 | 2023-11-03 | 小语智能信息科技(云南)有限公司 | 基于对比学***行句对抽取方法及装置 |
CN116822495A (zh) * | 2023-08-31 | 2023-09-29 | 小语智能信息科技(云南)有限公司 | 基于对比学***行句对抽取方法及装置 |
CN117590944A (zh) * | 2023-11-28 | 2024-02-23 | 上海源庐加佳信息科技有限公司 | 实体人对象和数字虚拟人对象的绑定*** |
CN117390141B (zh) * | 2023-12-11 | 2024-03-08 | 江西农业大学 | 一种农业社会化服务质量用户评价数据分析方法 |
CN117390141A (zh) * | 2023-12-11 | 2024-01-12 | 江西农业大学 | 一种农业社会化服务质量用户评价数据分析方法 |
CN117556027A (zh) * | 2024-01-12 | 2024-02-13 | 一站发展(北京)云计算科技有限公司 | 基于数字人技术的智能交互***及方法 |
CN117556027B (zh) * | 2024-01-12 | 2024-03-26 | 一站发展(北京)云计算科技有限公司 | 基于数字人技术的智能交互***及方法 |
CN117633518B (zh) * | 2024-01-25 | 2024-04-26 | 北京大学 | 一种产业链构建方法及*** |
CN117633518A (zh) * | 2024-01-25 | 2024-03-01 | 北京大学 | 一种产业链构建方法及*** |
CN117669593A (zh) * | 2024-01-31 | 2024-03-08 | 山东省计算中心(国家超级计算济南中心) | 基于等价语义的零样本关系抽取方法、***、设备及介质 |
CN117669593B (zh) * | 2024-01-31 | 2024-04-26 | 山东省计算中心(国家超级计算济南中心) | 基于等价语义的零样本关系抽取方法、***、设备及介质 |
CN117744787A (zh) * | 2024-02-20 | 2024-03-22 | 中国电子科技集团公司第十研究所 | 一阶研判规则知识合理性的智能度量方法 |
CN117744787B (zh) * | 2024-02-20 | 2024-05-07 | 中国电子科技集团公司第十研究所 | 一阶研判规则知识合理性的智能度量方法 |
CN117874209A (zh) * | 2024-03-12 | 2024-04-12 | 深圳市诚立业科技发展有限公司 | 基于nlp的诈骗短信监控告警*** |
CN117874209B (zh) * | 2024-03-12 | 2024-05-17 | 深圳市诚立业科技发展有限公司 | 基于nlp的诈骗短信监控告警*** |
CN117910460A (zh) * | 2024-03-18 | 2024-04-19 | 国网江苏省电力有限公司南通供电分公司 | 一种基于bge模型的电力科研知识关联性构建方法及*** |
CN117910460B (zh) * | 2024-03-18 | 2024-06-07 | 国网江苏省电力有限公司南通供电分公司 | 一种基于bge模型的电力科研知识关联性构建方法及*** |
CN118093791A (zh) * | 2024-04-24 | 2024-05-28 | 北京中关村科金技术有限公司 | 结合云计算的ai知识库生成方法及*** |
CN118132683A (zh) * | 2024-05-07 | 2024-06-04 | 杭州海康威视数字技术股份有限公司 | 文本抽取模型的训练方法、文本抽取方法和设备 |
CN118153553A (zh) * | 2024-05-09 | 2024-06-07 | 江西科技师范大学 | 基于多任务的社交网络用户心理危机原因抽取方法与*** |
CN118193743A (zh) * | 2024-05-20 | 2024-06-14 | 山东齐鲁壹点传媒有限公司 | 一种基于预训练模型的多层级文本分类模型 |
Also Published As
Publication number | Publication date |
---|---|
CN111310438B (zh) | 2021-06-08 |
CN111310438A (zh) | 2020-06-19 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021164199A1 (zh) | 基于多粒度融合模型的中文句子语义智能匹配方法及装置 | |
WO2021164200A1 (zh) | 一种基于深度分层编码的智能语义匹配方法和装置 | |
CN111259127B (zh) | 一种基于迁移学习句向量的长文本答案选择方法 | |
CN111310439B (zh) | 一种基于深度特征变维机制的智能语义匹配方法和装置 | |
CN112347268A (zh) | 一种文本增强的知识图谱联合表示学习方法及装置 | |
CN110019732B (zh) | 一种智能问答方法以及相关装置 | |
WO2021204014A1 (zh) | 一种模型训练的方法及相关装置 | |
CN111159485B (zh) | 尾实体链接方法、装置、服务器及存储介质 | |
CN113377897B (zh) | 基于深度对抗学习的多语言医疗术语规范标准化***及方法 | |
CN110032635A (zh) | 一种基于深度特征融合神经网络的问题对匹配方法和装置 | |
CN111597314A (zh) | 推理问答方法、装置以及设备 | |
WO2024131111A1 (zh) | 一种智能写作方法、装置、设备及非易失性可读存储介质 | |
TW201841121A (zh) | 一種自動生成語義相近句子樣本的方法 | |
CN107180026B (zh) | 一种基于词嵌入语义映射的事件短语学习方法及装置 | |
CN113392209B (zh) | 一种基于人工智能的文本聚类方法、相关设备及存储介质 | |
CN111222330B (zh) | 一种中文事件的检测方法和*** | |
CN111241303A (zh) | 一种大规模非结构化文本数据的远程监督关系抽取方法 | |
CN112926324A (zh) | 融合词典与对抗迁移的越南语事件实体识别方法 | |
CN113204611A (zh) | 建立阅读理解模型的方法、阅读理解方法及对应装置 | |
CN112581327B (zh) | 基于知识图谱的法律推荐方法、装置和电子设备 | |
CN116821373A (zh) | 基于图谱的prompt推荐方法、装置、设备及介质 | |
CN112417170B (zh) | 面向不完备知识图谱的关系链接方法 | |
CN117828024A (zh) | 一种插件检索方法、装置、存储介质及设备 | |
CN110826341A (zh) | 一种基于seq2seq模型的语义相似度计算方法 | |
WO2023130688A1 (zh) | 一种自然语言处理方法、装置、设备及可读存储介质 |
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: 20920016 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: 20920016 Country of ref document: EP Kind code of ref document: A1 |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20920016 Country of ref document: EP Kind code of ref document: A1 |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 15/03/2023) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20920016 Country of ref document: EP Kind code of ref document: A1 |