[go: up one dir, main page]

CN111160031A - Social media named entity identification method based on affix perception - Google Patents

Social media named entity identification method based on affix perception Download PDF

Info

Publication number
CN111160031A
CN111160031A CN201911289215.9A CN201911289215A CN111160031A CN 111160031 A CN111160031 A CN 111160031A CN 201911289215 A CN201911289215 A CN 201911289215A CN 111160031 A CN111160031 A CN 111160031A
Authority
CN
China
Prior art keywords
word
representation
character
text
level
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201911289215.9A
Other languages
Chinese (zh)
Inventor
蔡毅
吴志威
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
South China University of Technology SCUT
Original Assignee
South China University of Technology SCUT
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 South China University of Technology SCUT filed Critical South China University of Technology SCUT
Priority to CN201911289215.9A priority Critical patent/CN111160031A/en
Publication of CN111160031A publication Critical patent/CN111160031A/en
Priority to PCT/CN2020/112923 priority patent/WO2021114745A1/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/31—Indexing; Data structures therefor; Storage structures
    • G06F16/316—Indexing structures
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Machine Translation (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本发明公开了一种基于词缀感知的社交媒体命名实体识别方法，包括步骤：采集已标注出命名实体的社交媒体数据集；捕捉词的嵌入表示、字符级别表示和词缀特征表示，并将词嵌入表示、字符级别表示和词缀特征表示进行融合，作为词的最终表示；将得到的词的最终表示输入到双向卷积神经网络以及条件随机场中，预测标签序列并计算损失值；根据得到的损失值，采用随机梯度下降算法对模型进行训练；将文本输入训练好的模型中，识别出文本中的命名实体。本发明丰富了词的语义表示、缓解了社交媒体数据中未登录词的问题，提高了命名实体识别的效果。

Figure 201911289215

The invention discloses a social media named entity recognition method based on affix perception, comprising the steps of: collecting social media data sets that have marked named entities; capturing word embedding representation, character level representation and affix feature representation, and embedding the word The representation, character level representation and affix feature representation are fused as the final representation of the word; the final representation of the word is input into the bidirectional convolutional neural network and the conditional random field, the label sequence is predicted and the loss value is calculated; according to the obtained loss value, the stochastic gradient descent algorithm is used to train the model; the text is input into the trained model, and the named entities in the text are identified. The present invention enriches the semantic representation of words, alleviates the problem of unregistered words in social media data, and improves the effect of named entity recognition.

Figure 201911289215

Description

Social media named entity identification method based on affix perception
Technical Field
The invention relates to the technical field of natural language processing, in particular to a social media named entity recognition method based on affix perception.
Background
In the world today, with the explosion of mobile internet, people are not publishing information on social media all the time, and the information constitutes a huge amount of social media data. Compared with the traditional news private line manuscript data, the data on the social media are more time-efficient, contain rich information and gradually become a plurality of application potential information sources, such as news hotspot tracking, user public opinion analysis, potential violence incident early warning and the like. Therefore, how to mine potential information from social media data becomes an important task. The entity extraction is a basic task of information extraction, and a powerful entity extraction system is indispensable to the construction of the applications, and has extremely high social and economic values.
In recent years, with the rise of deep neural network models, end-to-end neural network-based models have become the mainstream method for named entity recognition. These methods can be broadly classified into the following categories: word-based representations, character-based representations, phrase-based representations, or any combination of the foregoing. Although these methods have achieved good performance on news text, the performance of such methods is dramatically reduced when confronted with social media data due to the inherent features of social media, such as informal expressions, irregular noun abbreviations, non-grammatical expressions, more unknown words, etc.
Applicants have discovered that affixes, as morphemes with certain semantics, can assist in identifying to some extent whether a word is part of an entity. For simplicity, only the most common prefixes and suffixes of affixes are considered. Two benefits can be brought about by introducing affix-feature representations, one is that words with the same affix tend to have similar meanings, and introducing affix representations can enrich semantic representations of words, e.g., "autopen", "automat", etc., words with the same prefix "auto-" all have "automatic" meaning; the second is that some affixes themselves have the semantics of named entities, for example, the suffix "-ie" is derived from ancient english and is commonly found in related names, names of people, children's words, or colloquial languages, so that the word whose "-ie" ends is likely to be a name of a person.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a social media named entity identification method based on affix perception. The method can combine word embedding, word character level representation, word prefix characteristic representation and word suffix characteristic representation, capture word affix characteristic representation by utilizing the bidirectional cyclic neural network, well enrich semantic representation of words by introducing the affix characteristic representation, relieve the problem of unknown words in social media data and improve the effect of named entity recognition. The method has certain generalization and is also suitable for named entity identification in the fields of news and the like.
The purpose of the invention can be realized by the following technical scheme:
a social media named entity recognition method based on affix perception comprises the following steps:
collecting a social media data set marked with named entities, wherein each piece of data comprises an original text and is marked with the named entities;
preprocessing texts in the data set, and constructing index vector representation of the texts at a word level and index vector representation of the texts at a character level;
capturing embedding representation, character level representation and affix characteristic representation of a word by adopting a recurrent neural network and a word embedding technology, and fusing the word embedding representation, the character level representation and the affix characteristic representation to be used as final representation of the word;
inputting the final representation of the obtained word into a bidirectional convolution neural network and a conditional random field, predicting a tag sequence and calculating a loss value;
training the model by adopting a random gradient descent algorithm according to the obtained loss value;
and inputting the text into the trained model, and identifying the named entity in the text.
Compared with the prior art, the invention has the following beneficial effects:
the method for recognizing the named entity of the social media based on affix perception introduces affix characteristic representation of words on the basis of word embedding and character level representation of the words, enriches semantic representation of the words, relieves the problem of unknown words in social media data, improves the effect of recognizing the named entity, has certain generalization, and is also suitable for recognizing the named entity in the fields of news and the like.
Drawings
Fig. 1 is a flowchart of a social media named entity recognition method based on affix sensing according to the present invention.
Fig. 2 is a schematic diagram of a model used for extracting affix features in this embodiment.
Detailed Description
The present invention will be described in further detail with reference to examples and drawings, but the present invention is not limited thereto.
Examples
The embodiment provides a method for identifying a social media named entity based on affix perception, and a flowchart of the method is shown in fig. 1, and the method comprises the following steps:
(1) and collecting a social media data set marked with the named entity, wherein each piece of data comprises original text and is marked with the named entity.
The collected social media data set is used as a training set.
(2) And preprocessing the texts in the data set, and constructing the index vector representation of the texts at the word level and the index vector representation of the texts at the character level.
Specifically, the pretreatment comprises:
replacing all lower case letters in the text with corresponding upper case letters;
all the numbers in the text are replaced by 0.
Specifically, the constructing the index representation of the text at a word level and a character level comprises:
(2-1) traversing all texts of the social media data set, and constructing a word dictionary and a character dictionary;
specifically, the word dictionary is constructed by traversing each word of each text in the data set, adding the word to a word list when different words are encountered, and assigning an index to each word according to the adding sequence, wherein the index value is 0, 1, 2 and so on. And the vocabulary obtained after traversing is a word dictionary.
The method of character dictionary construction is as above except that each character of each word of each text is traversed.
And (2-2) using the word dictionary and the character dictionary obtained in the step (2-1) to sequence the text at a word level and a character level.
The text is serialized at a word level and a character level, namely, each word in each sentence is subjected to one-hot coding, and corresponding vectors are formed according to the word level and the character level.
(3) Capturing embedded representation, character level representation and affix characteristic representation of a word by adopting a recurrent neural network and a word embedding technology, and fusing the embedded representation, the character level representation and the affix characteristic representation of the word to be used as final representation of the word, wherein the steps comprise:
(3-1) the serialization of texts at the word level is expressed as s ═ w1,w2,…,wnWhere n denotes the number of words in the text,
Figure BDA0002317242800000041
representing the one-hot encoding of the ith word of the sentence, v being the number of words in the word dictionary. Inputting s into the word embedding layer to obtain a corresponding word embedding representation:
Figure BDA0002317242800000042
wherein,
Figure BDA0002317242800000043
representing trainable parameters of the word embedding layer and d representing the dimensions of the word embedding vector.
(3-2) let the character serialization of the ith word in the text be represented as column wi＝{ci,1,ci,2,…,ci,mWhere m denotes the number of characters contained in the ith word,
Figure BDA0002317242800000044
the ith word in the sentenceOne-hot encoding of j characters, vcIs the number of characters in the character dictionary. Extracting character level representation of word by bidirectional cyclic neural network, firstly inputting each character of word into character embedding layer to obtain corresponding character embedding
Figure BDA0002317242800000051
Wherein, WcParameter matrix representing character embedding layer with size dc×vc. Then the character is embedded and input into a bidirectional cyclic neural network to respectively obtain the forward hidden state vector of each character
Figure BDA0002317242800000052
And reverse implicit state vector
Figure BDA0002317242800000053
Finally, the last implicit state vector of the forward cyclic neural network and the last implicit state vector of the reverse cyclic neural network are spliced to represent the character-level representation of the word
Figure BDA0002317242800000054
(3-3) for simplicity, the first t characters of each word are considered as the prefix of the word, and similarly, the last t words of each word are considered as the suffix of the word, t being the hyperparameter. Let the character serialization of the ith word in the text be represented as column wi＝{ci,1,ci,2,…,ci,mTherein of
Figure BDA0002317242800000055
One-hot encoding of the jth character representing the ith word in a sentence, vcIs the character dictionary size. Extracting character level representation of the word by adopting a bidirectional cyclic neural network, firstly inputting each character of the word into a character embedding layer to obtain corresponding character embedding
Figure BDA0002317242800000056
Wherein, WcMoments of parameters representing embedded layers of charactersMatrix size dc×vc. Then the character is embedded and input into a bidirectional cyclic neural network to respectively obtain the forward hidden state vector of each character
Figure BDA0002317242800000057
And reverse implicit state vector
Figure BDA0002317242800000058
Finally, the implicit state vectors of the first k characters are spliced to obtain a matrix, and the dimensionality of the matrix is dvX t, this matrix contains prefix information. Specially, if the length of a word is less than t, the hidden state vectors of all time steps are spliced together, and the dimension of the obtained matrix is dvX m. In order to ensure consistent dimensionality of prefix features of all words, an averaging operation is performed on the implicit state matrices, namely, averaging is performed on the second dimensionality of the matrices, and finally prefix feature representations are obtained
Figure BDA0002317242800000059
Similarly, the implicit state vector corresponding to the last t characters of the word is operated in the same way to obtain the suffix feature representation of the word
Figure BDA00023172428000000510
(3-4) splicing the word embedded representation, the character level representation, the prefix feature representation and the suffix feature representation of the word obtained in the steps (3-1) - (3-3) to obtain a final representation of the word
Figure BDA0002317242800000061
Figure BDA0002317242800000062
(4) Inputting the final representation of the obtained word into a bidirectional convolutional neural network and a conditional random field, predicting a label sequence and calculating a loss value, wherein the steps comprise:
(4-1) inputting the final representation of the word obtained in step (3) intoIn the bidirectional cyclic neural network, the obtained forward hidden state and the reverse hidden state are spliced to obtain word sequence representation
Figure BDA0002317242800000063
(4-2) inputting the word sequence representation obtained in the step (4-1) into a full-connection layer, and obtaining the score of each word on all labels: pi＝Whi+ b, where W and b are trainable parameters;
(4-3) y ═ y1,y2,…,ynDenotes the predicted tag sequence corresponding to the input text s, y(s) denotes the set of all possible tag sequences for the input sentence s. The P obtained in the step (4-2)iWhen the random field is input into the conditional random field, the score of each possible sequence is calculated according to the following formula:
Figure BDA0002317242800000064
wherein A represents a state transition score matrix, the size of A is k × k, Ai,jRepresenting a score for a transition from label i to label j,
Figure BDA0002317242800000067
indicating the tag y in the predicted sequenceiFollowed by label yi+1Score (likelihood), yiRepresenting the ith tag in the predicted tag sequence y.
Figure BDA0002317242800000068
The label of the i-th word representing the input text s is yiScore (likelihood).
Figure BDA0002317242800000069
Represents the predicted tag sequence, with the highest scoring tag sequence as the final prediction:
Figure BDA0002317242800000065
finally, the loss value is calculated as follows:
Figure BDA0002317242800000066
Figure BDA0002317242800000071
(5) training the model by adopting a random gradient descent algorithm according to the loss value obtained in the step (4) to obtain a trained model;
when the loss value of the model is not reduced any more, the training is completed.
(6) And (5) inputting the text into the trained model obtained in the step (5), and identifying the named entity in the text.
Fig. 2 is a schematic diagram of a model used for extracting the affix feature representation.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims (10)

1.一种基于词缀感知的社交媒体命名实体识别方法，其特征在于，包括步骤：1. a social media named entity recognition method based on affix perception, is characterized in that, comprises the steps: 采集已标注出命名实体的社交媒体数据集，每条数据均包含有原始的文本并已标注命名实体；Collect social media datasets with named entities marked, each piece of data contains the original text and named entities marked; 对数据集中的文本进行预处理，构建文本在词级别的索引向量表示和文本在字符级别的索引向量表示；Preprocess the text in the dataset, and construct the index vector representation of the text at the word level and the index vector representation of the text at the character level; 采用循环神经网络以及词嵌入技术捕捉词的嵌入表示、字符级别表示和词缀特征表示，并将词嵌入表示、字符级别表示和词缀特征表示进行融合，作为词的最终表示；Using recurrent neural network and word embedding technology to capture word embedding representation, character level representation and affix feature representation, and fuse word embedding representation, character level representation and affix feature representation as the final representation of the word; 将得到的词的最终表示输入到双向卷积神经网络以及条件随机场中，预测标签序列并计算损失值；Input the final representation of the obtained word into a bidirectional convolutional neural network and a conditional random field, predict the label sequence and calculate the loss value; 根据得到的损失值，采用随机梯度下降算法对模型进行训练；According to the obtained loss value, the stochastic gradient descent algorithm is used to train the model; 将文本输入训练好的模型中，识别出文本中的命名实体。Feed the text into the trained model and identify the named entities in the text. 2.根据权利要求1所述的方法，其特征在于，所述预处理包括：2. The method according to claim 1, wherein the preprocessing comprises: 将文本中的小写字母全部替换为相应的大写字母；Replace all lowercase letters in the text with the corresponding uppercase letters; 将文本中的数字全部替换为0；Replace all numbers in the text with 0; 遍历社交媒体数据集的所有文本，构建单词词典以及字符词典；Traverse all texts of social media datasets, build word dictionaries and character dictionaries; 利用得到的单词词典和字符词典，将文本在词级别和字符级别进行序列化。Using the resulting word dictionary and character dictionary, serialize the text at the word level and at the character level. 3.根据权利要求1所述的方法，其特征在于，所述构建文本在词级别的索引向量表示和文本在字符级别的索引向量表示的步骤中，包括：3. The method according to claim 1, wherein, in the step of constructing the index vector representation of the text at the word level and the index vector representation of the text at the character level, comprising: 遍历社交媒体数据集的所有文本，构建单词词典以及字符词典；Traverse all texts of social media datasets, build word dictionaries and character dictionaries; 利用得到的单词词典和字符词典，将文本在词级别和字符级别进行序列化。Using the resulting word dictionary and character dictionary, serialize the text at the word level and at the character level. 4.根据权利要求3所述的方法，其特征在于，单词词典的构建方法为：4. method according to claim 3, is characterized in that, the construction method of word dictionary is: 遍历数据集中每个文本的每个词，遇到不同的词则将其添加到词表中，并按照添加的顺序为每个词赋予一个索引，索引值为0，1，2以此类推，遍历完成后得到的词表即为单词词典；Traverse each word of each text in the data set, add it to the vocabulary when encountering different words, and assign an index to each word in the order of addition, the index value is 0, 1, 2 and so on, The word list obtained after the traversal is completed is the word dictionary; 字符词典构建的方法与单词词典构建方法相同，不同的是遍历的是每个文本的每个词的每个字符；The construction method of character dictionary is the same as that of word dictionary construction, the difference is that each character of each word of each text is traversed; 序列化方法为：The serialization method is: 文本在词级别和字符级别进行序列化，即对每个句子中的每个词进行独热编码，并分别按词级别和字符级别组成相应的向量。The text is serialized at the word level and character level, that is, each word in each sentence is one-hot encoded, and the corresponding vectors are formed at the word level and character level, respectively. 5.根据权利要求1所述的方法，其特征在于，所述捕捉词的嵌入表示、字符级别表示和词缀特征表示，并将词嵌入表示、字符级别表示和词缀特征表示进行融合，作为词的最终表示的步骤中，包括：5. The method according to claim 1, characterized in that, the embedding representation, character level representation and affix feature representation of the words are captured, and the word embedding representation, character level representation and affix feature representation are fused, as the word embedding representation, character level representation and affix feature representation. The steps in the final representation include: 将文本在词级别的序列化表示输入到词嵌入层中得到相应的词嵌入表示；Input the serialized representation of the text at the word level into the word embedding layer to obtain the corresponding word embedding representation; 采用双向循环神经网络提取词的字符级别表示：首先将词的每个字符输入到字符嵌入层中得到相应的字符嵌入；然后将该字符嵌入输入到双向循环神经网络中，分别得到每个字符的前向隐含状态向量和反向隐含状态向量，最后再将将前向循环神经网络的最后一个隐含状态向量和反向循环神经网络的最后一个隐含状态向量拼接起来表示词的字符级别表示；A bidirectional recurrent neural network is used to extract the character-level representation of a word: first, each character of the word is input into the character embedding layer to obtain the corresponding character embedding; then the character embedding is input into the bidirectional recurrent neural network to obtain the The forward hidden state vector and the reverse hidden state vector, and finally the last hidden state vector of the forward recurrent neural network and the last hidden state vector of the reverse recurrent neural network are spliced together to represent the character level of the word express; 采用双向循环神经网络提取词的字符级别表示：首先将词的每个字符输入到字符嵌入层中得到对应的字符嵌入，然后将该字符嵌入输入到双向循环神经网络中，分别得到每个字符的前向隐含状态向量和反向隐含状态向量，最后将前t个字符的隐含状态向量拼接起来得到一个矩阵；对该词的后t个字符对应的隐含状态向量进行同样的操作，得到该词的后缀特征表示；The character-level representation of a word is extracted by a bidirectional recurrent neural network: first, each character of the word is input into the character embedding layer to obtain the corresponding character embedding, and then the character embedding is input into the bidirectional recurrent neural network to obtain the corresponding character embedding of each character. Forward implicit state vector and reverse implicit state vector, and finally concatenate the implicit state vectors of the first t characters to obtain a matrix; perform the same operation on the implicit state vector corresponding to the last t characters of the word, Get the suffix feature representation of the word; 将得到的词的词嵌入表示、字符级别表示、前缀特征表示以及后缀特征表示拼接起来，得到词的最终表示。The word embedding representation, character-level representation, prefix feature representation, and suffix feature representation of the obtained word are spliced together to obtain the final representation of the word. 6.根据权利要求5所述的方法，其特征在于，在提取词的字符级别表示中，如果词的长度小于t，则将所有时间步的隐含状态向量拼接在一起；为了保证所有词的前缀特征的维度一致，在这些隐含状态矩阵上执行了取均值的操作，即在矩阵的第二个维度上取均值，最后得到前缀特征表示。6. The method according to claim 5, wherein, in the character level representation of the extracted word, if the length of the word is less than t, the hidden state vectors of all time steps are spliced together; The dimensions of the prefix features are the same, and the mean value operation is performed on these hidden state matrices, that is, the mean value is taken on the second dimension of the matrix, and finally the prefix feature representation is obtained. 7.根据权利要求1所述的方法，其特征在于，所述将得到的词的最终表示输入到双向卷积神经网络以及条件随机场中，预测标签序列并计算损失值的步骤中，包括：7. The method according to claim 1, wherein the final representation of the obtained word is input into a bidirectional convolutional neural network and a conditional random field, and in the step of predicting the label sequence and calculating the loss value, comprising: 将得到的词的最终表示输入到双向循环神经网络中，并将得到的前向隐含状态和反向隐含状态拼接起来，得到词序列表示；The final representation of the obtained word is input into the bidirectional recurrent neural network, and the obtained forward hidden state and reverse hidden state are spliced together to obtain the word sequence representation; 将词序列表示输入到全连接层中，计算每个词在所有标签上的得分；The word sequence representation is input into the fully connected layer, and the score of each word on all labels is calculated; 将得到的词序列表示输入到条件随机场中，计算每个可能序列的得分；Input the resulting word sequence representation into a conditional random field and calculate the score for each possible sequence; 取得分最高的标签序列作为最终的预测结果，并计算其损失值。The label sequence with the highest score is obtained as the final prediction result, and its loss value is calculated. 8.根据权利要求7所述的方法，其特征在于，每个词在所有标签上的得分的计算公式为：8. method according to claim 7 is characterized in that, the calculation formula of the score of each word on all labels is: Pi＝Whi+bP i =Wh i +b 其中，W和b是可训练的参数。where W and b are trainable parameters. 9.根据权利要求7所述的方法，其特征在于，每个可能序列的得分的计算公式为：9. The method according to claim 7, wherein the calculation formula of the score of each possible sequence is:
Figure FDA0002317242790000031
Figure FDA0002317242790000031
 其中，A表示状态转移得分矩阵，A的大小是k×k，Ai，j表示从标签i转移到标签j的得分；，
Figure FDA0002317242790000032
表示预测序列中标签yi后接标签yi+1的得分，yi表示预测标签序列y中的第i个标签，
Figure FDA0002317242790000033
表示输入文本s的第i个词的标签为yi的得分；Among them, A represents the state transition score matrix, the size of A is k × k, A i,j represents the score of the transition from label i to label j;
Figure FDA0002317242790000032
Represents the score of label y i followed by label y i+1 in the predicted sequence, y i represents the ith label in the predicted label sequence y,
Figure FDA0002317242790000033
Represents the score of the i -th word of the input text s with the label yi; 得分最高的标签序列作为最终的预测结果表示为：The label sequence with the highest score is expressed as the final prediction result as:
Figure FDA0002317242790000034
Figure FDA0002317242790000034
 10.根据权利要求9所述的方法，其特征在于，损失值计算公式为：10. The method according to claim 9, wherein the loss value calculation formula is:
Figure FDA0002317242790000041
Figure FDA0002317242790000041
Figure FDA0002317242790000042
Figure FDA0002317242790000042
CN201911289215.9A 2019-12-13 2019-12-13 Social media named entity identification method based on affix perception Pending CN111160031A (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN201911289215.9A CN111160031A (en) 2019-12-13 2019-12-13 Social media named entity identification method based on affix perception
PCT/CN2020/112923 WO2021114745A1 (en) 2019-12-13 2020-09-02 Named entity recognition method employing affix perception for use in social media

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911289215.9A CN111160031A (en) 2019-12-13 2019-12-13 Social media named entity identification method based on affix perception

Publications (1)

Publication Number Publication Date
CN111160031A true CN111160031A (en) 2020-05-15

Family

ID=70557141

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911289215.9A Pending CN111160031A (en) 2019-12-13 2019-12-13 Social media named entity identification method based on affix perception

Country Status (2)

Country Link
CN (1) CN111160031A (en)
WO (1) WO2021114745A1 (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112270194A (en) * 2020-11-03 2021-01-26 沈阳雅译网络技术有限公司 Named entity identification method based on gradient neural network structure search
WO2021114745A1 (en) * 2019-12-13 2021-06-17 华南理工大学 Named entity recognition method employing affix perception for use in social media
CN113033206A (en) * 2021-04-01 2021-06-25 重庆交通大学 Bridge detection field text entity identification method based on machine reading understanding
CN113609857A (en) * 2021-07-22 2021-11-05 武汉工程大学 Legal named entity identification method and system based on cascade model and data enhancement
CN115757325A (en) * 2023-01-06 2023-03-07 珠海金智维信息科技有限公司 Intelligent conversion method and system for XES logs

Families Citing this family (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113468330B (en) * 2021-07-06 2023-04-28 北京有竹居网络技术有限公司 Information acquisition method, device, equipment and medium
CN113449816B (en) * 2021-07-20 2024-10-01 恒安嘉新（北京）科技股份公司 Website classification model training, website classification method, device, equipment and medium
CN113849597B (en) * 2021-08-31 2024-04-30 艾迪恩(山东)科技有限公司 Illegal advertising word detection method based on named entity recognition
CN113886522B (en) * 2021-09-13 2022-12-02 苏州空天信息研究院 Discontinuous entity identification method based on path expansion
CN113869037B (en) * 2021-09-27 2024-06-21 北京航空航天大学 Learning method of topic label representation based on content-enhanced network embedding
CN113901218B (en) * 2021-10-08 2025-02-07 国家电网有限公司客户服务中心 A method and device for extracting basic rules of audit business
CN114021549B (en) * 2021-10-15 2024-10-22 华中科技大学 Chinese named entity recognition method and device based on vocabulary enhancement and multi-features
CN113886516B (en) * 2021-10-29 2025-05-06 陕西师范大学 An automatic labeling system and method for cultural resource entity recognition training data
CN114048750B (en) * 2021-12-10 2024-06-28 广东工业大学 A named entity recognition method integrating high-level information features
CN113963358B (en) * 2021-12-20 2022-03-04 北京易真学思教育科技有限公司 Text recognition model training method, text recognition device and electronic equipment
CN114297079B (en) * 2021-12-30 2024-04-02 北京工业大学 XSS fuzzy test case generation method based on time convolution network
CN114398076B (en) * 2022-01-18 2024-12-17 北京工业大学 Object-oriented program method naming peculiar smell detection method based on deep learning
CN114638214B (en) * 2022-03-18 2024-11-29 中国人民解放军国防科技大学 Medical field Chinese named entity recognition method
CN114781382B (en) * 2022-04-25 2025-01-21 苏州科技大学 Medical named entity recognition system and method based on RWLSTM model fusion
CN115130468B (en) * 2022-05-06 2023-04-07 北京安智因生物技术有限公司 Myocardial infarction entity recognition method based on word fusion representation and graph attention network
CN114912453B (en) * 2022-05-20 2025-04-29 大连大学 Named Entity Recognition Method for Chinese Legal Documents Based on Enhanced Sequence Features
CN115017144B (en) * 2022-05-30 2024-03-29 北京计算机技术及应用研究所 Judicial document case element entity identification method based on graphic neural network
CN115329769B (en) * 2022-07-05 2025-11-21 中国电子科技集团公司电子科学研究院 Platform enterprise network public opinion emotion analysis method based on semantic enhancement network
CN115329766B (en) * 2022-08-23 2023-04-18 中国人民解放军国防科技大学 Named entity identification method based on dynamic word information fusion
CN115544251B (en) * 2022-09-14 2024-09-17 山东科技大学 Named entity identification method for bursty news text
CN116108850B (en) * 2023-02-14 2025-10-31 杭州电子科技大学 Small sample learning named entity identification method integrating entity tag coding
CN116579343B (en) * 2023-05-17 2024-06-04 成都信息工程大学 Named entity identification method for Chinese text travel class
CN116432655B (en) * 2023-06-12 2023-12-08 山东大学 Method and device for identifying named entities with few samples based on language knowledge learning
CN117933380B (en) * 2024-01-31 2024-10-29 国网江苏省电力有限公司南通供电分公司 A method and system for extracting scientific research knowledge in the power industry
CN120745632B (en) * 2025-08-21 2025-11-14 兰州交通大学 A method for naming and recognizing geographic entities

Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106569998A (en) * 2016-10-27 2017-04-19 浙江大学 Text named entity recognition method based on Bi-LSTM, CNN and CRF
US20180131645A1 (en) * 2016-09-29 2018-05-10 Admit Hub, Inc. Systems and processes for operating and training a text-based chatbot
CN108460013A (en) * 2018-01-30 2018-08-28 大连理工大学 A kind of sequence labelling model based on fine granularity vocabulary representation model
CN108959252A (en) * 2018-06-28 2018-12-07 中国人民解放军国防科技大学 Semi-supervised Chinese named entity recognition method based on deep learning
US10169315B1 (en) * 2018-04-27 2019-01-01 Asapp, Inc. Removing personal information from text using a neural network
CN109325243A (en) * 2018-10-22 2019-02-12 内蒙古大学 Character-level sequence model-based Mongolian word segmentation method and word segmentation system
CN109493265A (en) * 2018-11-05 2019-03-19 北京奥法科技有限公司 A kind of Policy Interpretation method and Policy Interpretation system based on deep learning
CN109871535A (en) * 2019-01-16 2019-06-11 四川大学 A French Named Entity Recognition Method Based on Deep Neural Network
CN110096713A (en) * 2019-03-21 2019-08-06 昆明理工大学 A kind of Laotian organization names recognition methods based on SVM-BiLSTM-CRF
CN110110042A (en) * 2019-03-21 2019-08-09 昆明理工大学 Laotian complexity name place name entity recognition method based on CNN+BLSTM+CRF
CN110162749A (en) * 2018-10-22 2019-08-23 哈尔滨工业大学（深圳） Information extracting method, device, computer equipment and computer readable storage medium
CN110223737A (en) * 2019-06-13 2019-09-10 电子科技大学 A kind of chemical composition of Chinese materia medica name entity recognition method and device

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359293B (en) * 2018-09-13 2019-09-10 内蒙古大学 Mongolian name entity recognition method neural network based and its identifying system
CN110008469B (en) * 2019-03-19 2022-06-07 桂林电子科技大学 Multilevel named entity recognition method
CN110083710B (en) * 2019-04-30 2021-04-02 北京工业大学 A Word Definition Generation Method Based on Recurrent Neural Network and Latent Variable Structure
CN111160031A (en) * 2019-12-13 2020-05-15 华南理工大学 Social media named entity identification method based on affix perception

Patent Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180131645A1 (en) * 2016-09-29 2018-05-10 Admit Hub, Inc. Systems and processes for operating and training a text-based chatbot
CN106569998A (en) * 2016-10-27 2017-04-19 浙江大学 Text named entity recognition method based on Bi-LSTM, CNN and CRF
CN108460013A (en) * 2018-01-30 2018-08-28 大连理工大学 A kind of sequence labelling model based on fine granularity vocabulary representation model
US10169315B1 (en) * 2018-04-27 2019-01-01 Asapp, Inc. Removing personal information from text using a neural network
CN108959252A (en) * 2018-06-28 2018-12-07 中国人民解放军国防科技大学 Semi-supervised Chinese named entity recognition method based on deep learning
CN109325243A (en) * 2018-10-22 2019-02-12 内蒙古大学 Character-level sequence model-based Mongolian word segmentation method and word segmentation system
CN110162749A (en) * 2018-10-22 2019-08-23 哈尔滨工业大学（深圳） Information extracting method, device, computer equipment and computer readable storage medium
CN109493265A (en) * 2018-11-05 2019-03-19 北京奥法科技有限公司 A kind of Policy Interpretation method and Policy Interpretation system based on deep learning
CN109871535A (en) * 2019-01-16 2019-06-11 四川大学 A French Named Entity Recognition Method Based on Deep Neural Network
CN110096713A (en) * 2019-03-21 2019-08-06 昆明理工大学 A kind of Laotian organization names recognition methods based on SVM-BiLSTM-CRF
CN110110042A (en) * 2019-03-21 2019-08-09 昆明理工大学 Laotian complexity name place name entity recognition method based on CNN+BLSTM+CRF
CN110223737A (en) * 2019-06-13 2019-09-10 电子科技大学 A kind of chemical composition of Chinese materia medica name entity recognition method and device

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
CHANGMENG ZHENG: "A Boundary-aware Neural Model for Nested Named Entity Recognition", 《PROCEEDINGS OF THE 2019 CONFERENCE ON EMPIRICAL METHODS IN NATURAL LANGUAGE PROCESSING AND THE 9TH INTERNATIONAL JOINT CONFERENCE ON NATURAL LANGUAGE PROCESSING》 *
SOBHANA N.V: "Conditional Random Field Based Named Entity Recognition in Geological text", 《2010 INTERNATIONAL JOURNAL OF COMPUTER APPLICATIONS》 *
VIKAS YADAV: "Deep Affix Features Improve Neural Named Entity Recognizers", 《PROCEEDINGS OF THE 7TH JOINT CONFERENCE ON LEXICAL AND COMPUTATIONAL SEMANTICS》 *
买买提阿依甫: "基于 BiLSTM-CNN-CＲF 模型的维吾尔文命名实体识别", 《计算机工程》 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021114745A1 (en) * 2019-12-13 2021-06-17 华南理工大学 Named entity recognition method employing affix perception for use in social media
CN112270194A (en) * 2020-11-03 2021-01-26 沈阳雅译网络技术有限公司 Named entity identification method based on gradient neural network structure search
CN112270194B (en) * 2020-11-03 2023-07-18 沈阳雅译网络技术有限公司 A Named Entity Recognition Method Based on Gradient Neural Network Structure Search
CN113033206A (en) * 2021-04-01 2021-06-25 重庆交通大学 Bridge detection field text entity identification method based on machine reading understanding
CN113609857A (en) * 2021-07-22 2021-11-05 武汉工程大学 Legal named entity identification method and system based on cascade model and data enhancement
CN113609857B (en) * 2021-07-22 2023-11-28 武汉工程大学 Legal named entity recognition method and system based on cascade model and data enhancement
CN115757325A (en) * 2023-01-06 2023-03-07 珠海金智维信息科技有限公司 Intelligent conversion method and system for XES logs
CN115757325B (en) * 2023-01-06 2023-04-18 珠海金智维信息科技有限公司 Intelligent conversion method and system for XES log

Also Published As

Publication number Publication date
WO2021114745A1 (en) 2021-06-17

Similar Documents

Publication Publication Date Title
CN111160031A (en) Social media named entity identification method based on affix perception
CN110119765B (en) A keyword extraction method based on Seq2seq framework
CN114757182B (en) A BERT short text sentiment analysis method with improved training method
CN111291195B (en) Data processing method, device, terminal and readable storage medium
Jia Sentiment classification of microblog: A framework based on BERT and CNN with attention mechanism
CN109977416A (en) A kind of multi-level natural language anti-spam text method and system
Zhang et al. Learning Chinese word embeddings from stroke, structure and pinyin of characters
CN114912453B (en) Named Entity Recognition Method for Chinese Legal Documents Based on Enhanced Sequence Features
CN114595338B (en) Entity relationship joint extraction system and method based on hybrid feature representation
CN113961666A (en) Keyword recognition method, apparatus, device, medium, and computer program product
CN105893362A (en) A method for acquiring knowledge point semantic vectors and a method and a system for determining correlative knowledge points
TW201826145A (en) Method and system for knowledge extraction from Chinese corpus useful for extracting knowledge from source corpuses mainly written in Chinese
CN111476036A (en) A Word Embedding Learning Method Based on Chinese Word Feature Substrings
CN116595023A (en) Method and device for updating address information, electronic device, and storage medium
Mhamed et al. A deep CNN architecture with novel pooling layer applied to two Sudanese Arabic sentiment data sets
CN114065749A (en) Text-oriented Guangdong language recognition model and training and recognition method of system
CN114064901B (en) Book comment text classification method based on knowledge graph word meaning disambiguation
CN118711198A (en) Information identification method and device
CN104317882B (en) Decision-based Chinese word segmentation and fusion method
CN119272772A (en) Method and system for named entity recognition in social media texts integrating multi-scale features and syntactic information
CN111159405A (en) Irony detection method based on background knowledge
CN115017404A (en) Target news topic abstracting method based on compressed space sentence selection
CN114757184B (en) Method and system for realizing knowledge question and answer in aviation field
CN114169325B (en) Webpage new word discovery and analysis method based on word vector representation
CN112749566A (en) English writing auxiliary oriented semantic matching method and device

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20200515