CN110377744A - A kind of method, apparatus, storage medium and the electronic equipment of public sentiment classification - Google Patents
A kind of method, apparatus, storage medium and the electronic equipment of public sentiment classification Download PDFInfo
- Publication number
- CN110377744A CN110377744A CN201910683305.XA CN201910683305A CN110377744A CN 110377744 A CN110377744 A CN 110377744A CN 201910683305 A CN201910683305 A CN 201910683305A CN 110377744 A CN110377744 A CN 110377744A
- Authority
- CN
- China
- Prior art keywords
- text
- training
- comprehensive
- entity
- target
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 52
- 239000013598 vector Substances 0.000 claims abstract description 95
- 238000012545 processing Methods 0.000 claims abstract description 24
- 238000012549 training Methods 0.000 claims description 191
- 238000005457 optimization Methods 0.000 claims description 37
- 238000013145 classification model Methods 0.000 claims description 31
- 238000000605 extraction Methods 0.000 claims description 22
- 230000008569 process Effects 0.000 claims description 15
- 239000000284 extract Substances 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 7
- 230000000694 effects Effects 0.000 description 7
- 230000011218 segmentation Effects 0.000 description 5
- 238000011144 upstream manufacturing Methods 0.000 description 5
- 230000001364 causal effect Effects 0.000 description 4
- 238000004891 communication Methods 0.000 description 3
- 238000003058 natural language processing Methods 0.000 description 3
- 238000006467 substitution reaction Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 239000000243 solution Substances 0.000 description 2
- 230000008094 contradictory effect Effects 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 239000011550 stock solution Substances 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
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
- G06F16/35—Clustering; Classification
-
- 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
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
- G06F40/295—Named entity recognition
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Audiology, Speech & Language Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention provides method, apparatus, storage medium and the electronic equipments of a kind of classification of public sentiment, wherein this method comprises: obtaining text to be identified, and extracts target entity；A target entity is chosen as effective target entity and generates comprehensive text；Coded treatment is carried out to comprehensive text, the text vector of comprehensive text is generated, and public sentiment classification processing is carried out according to text vector, determines public sentiment classification corresponding to effective target entity；Next target entity is chosen later as effective target entity, is proceeded as described above, until traversing all target entities in text to be identified.Method, apparatus, storage medium and the electronic equipment of the public sentiment classification provided through the embodiment of the present invention can determine the public sentiment classification of each entity included in text by way of combining text and entity and generating comprehensive text；And can more targetedly determine the public sentiment classification of the entity, the recognition accuracy of public sentiment classification is higher.
Description
Technical Field
The invention relates to the technical field of public opinion classification, in particular to a public opinion classification method, a public opinion classification device, a storage medium and electronic equipment.
Background
At present, public opinion classification is one of the most common tasks in public opinion analysis, and plays an important role in public opinion monitoring. The most common public opinion classification is the commendative and derogatory binary classification, and analysis such as looking at the sky and looking up can reflect the tendency of subjective public opinion. The public opinion classification may also be a more detailed multi-element classification, such as expressing a description of a document about the status of factual information of a financial entity based on words or phrases that the user may think or see (e.g., debt, profit increase, tax evasion, technical innovation, etc.).
The traditional public opinion classification scheme is to classify the problem into a text classification problem and then use a classifier to solve the problem. However, for a piece of text, the conventional classifier can generally recognize only one label, i.e., only one kind of public opinion category. However, text in the real world may often contain multiple financial entities, and the public opinion type corresponding to each financial entity is likely to be different or even contradictory. For example, there are two financial entities in the text "recruit bank earns 500 ten thousand dollars by reducing hundredths of stocks": the public opinion categories corresponding to the public opinion categories are 'good operation' and 'stock banning or reduction', and are two labels with completely different positive and negative faces. Obviously, such text cannot be correctly recognized by using only text information.
Disclosure of Invention
To solve the above problems, embodiments of the present invention provide a method, an apparatus, a storage medium, and an electronic device for public opinion classification.
In a first aspect, an embodiment of the present invention provides a public opinion classification method, including:
acquiring a text to be recognized, and extracting one or more target entities in the text to be recognized;
selecting one target entity as an effective target entity, and generating a comprehensive text by combining the effective target entity and the text to be recognized;
coding the comprehensive text to generate a text vector of the comprehensive text, and carrying out public sentiment classification processing according to the text vector to determine the public sentiment category corresponding to the current effective target entity;
and then selecting the next target entity as an effective target entity, and continuing the process of determining the public sentiment category corresponding to the effective target entity until all the target entities in the text to be recognized are traversed.
In a second aspect, an embodiment of the present invention further provides a public opinion classifying device, including:
the system comprises an acquisition module, a recognition module and a recognition module, wherein the acquisition module is used for acquiring a text to be recognized and extracting one or more target entities in the text to be recognized;
the text combination module is used for selecting one target entity as an effective target entity and combining the effective target entity and the text to be recognized to generate a comprehensive text;
the public opinion classification module is used for coding the comprehensive text to generate a text vector of the comprehensive text, and performing public opinion classification according to the text vector to determine the public opinion category corresponding to the current effective target entity;
and the traversing module is used for selecting the next target entity as an effective target entity, and continuing the process of determining the public sentiment category corresponding to the effective target entity until all the target entities in the text to be recognized are traversed.
In a third aspect, the embodiment of the invention further provides a computer storage medium, where computer-executable instructions are stored, and the computer-executable instructions are used in any one of the above public opinion classification methods.
In a fourth aspect, an embodiment of the present invention further provides an electronic device, including:
at least one processor; and the number of the first and second groups,
a memory communicatively coupled to the at least one processor; wherein,
the memory stores instructions executable by the at least one processor to enable the at least one processor to perform any one of the methods for public opinion classification.
In the solution provided by the first aspect of the embodiments of the present invention, the entities in the text are identified in advance, and the public opinion category of each entity included in the text can be determined by combining the text and the entities to generate a comprehensive text; even different entities can be normally identified for totally different public opinion categories, which are negative. Meanwhile, the entity in the text is extracted in advance, the public sentiment category of the entity can be determined more pertinently, and the recognition accuracy of the public sentiment category is higher. During coding processing, characters are used as the minimum semantic unit, the word segmentation task at the upstream does not need to be relied on, and rare words can be well processed.
In order to make the aforementioned and other objects, features and advantages of the present invention comprehensible, preferred embodiments accompanied with figures are described in detail below.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.
Fig. 1 is a flowchart illustrating a method for public opinion classification according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of a public opinion classification model in the method for public opinion classification according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of a language prediction model in the public opinion classification method according to an embodiment of the present invention;
fig. 4 is a schematic structural diagram illustrating a public opinion classifying apparatus according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of an electronic device for performing a method for public opinion classification according to an embodiment of the present invention.
Detailed Description
In the description of the present invention, it is to be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", and the like, indicate orientations and positional relationships based on those shown in the drawings, and are used only for convenience of description and simplicity of description, and do not indicate or imply that the device or element being referred to must have a particular orientation, be constructed and operated in a particular orientation, and thus, should not be considered as limiting the present invention.
Furthermore, the terms "first", "second" and "first" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the present invention, "a plurality" means two or more unless specifically defined otherwise.
In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "secured," and the like are to be construed broadly and can, for example, be fixedly connected, detachably connected, or integrally connected; can be mechanically or electrically connected; they may be connected directly or indirectly through intervening media, or they may be interconnected between two elements. The specific meanings of the above terms in the present invention can be understood by those skilled in the art according to specific situations.
The public opinion classification method provided by the embodiment of the invention is shown in the figure 1 and comprises the following steps:
step 101: and acquiring a text to be recognized, and extracting one or more target entities in the text to be recognized.
In an embodiment of the present invention, the text to be recognized is a text that needs to recognize a public opinion category, wherein the text to be recognized may include a plurality of entities, and each entity may have the same or different public opinion categories in the text to be recognized. The "entity" in this embodiment refers to a named entity, which may specifically be a company name, a person name, or the like. After the text to be recognized is determined, entities in the text to be recognized can be recognized based on a company name recognition system or a named entity recognition system, and then, part or all of the entities are selected as target entities, so that the public opinion category of each target entity is determined. For example, the text to be recognized is "bank a earns 500 ten thousand dollars by reducing stock of company B", wherein the entities include "bank a" and "company B", and both entities can be target entities in the present embodiment.
Step 102: and selecting a target entity as an effective target entity, and combining the effective target entity and the text to be recognized to generate a comprehensive text.
In the embodiment of the present invention, the target entity is an entity that needs to determine a public opinion category, one text to be recognized may include one or more target entities, and in the embodiment, the public opinion category corresponding to each target entity is recognized. Specifically, in this embodiment, one target entity is sequentially selected as an effective target entity, and then a comprehensive text is generated by combining the effective target entity and the complete text to be recognized. For example, the text to be recognized is "bank a earns 500 ten thousand dollars by deducting stock of company B", the target entities are "bank a" and "company B", at this time, "bank a" may be selected as an effective target entity, and then a comprehensive text including the effective target entity and the text to be recognized is generated, where the comprehensive text is, for example: "Bank A earns 500 ten thousand dollars by reducing the stock of company B".
Step 103: and coding the comprehensive text to generate a text vector of the comprehensive text, and carrying out public opinion classification processing according to the text vector to determine the public opinion category corresponding to the current effective target entity.
In the embodiment of the invention, after the comprehensive text is determined, the comprehensive text can be converted into a code which can be identified by a machine, namely, a text vector of the comprehensive text is generated through coding processing; and then, carrying out public opinion classification processing on the text vector by using a trained classifier so as to determine a corresponding public opinion category. The classifier can adopt a traditional classifier, and is different in that samples adopted by the classifier in training also take 'entity + text' as input of the classifier.
Optionally, the step 103 of performing encoding processing on the synthesized text to generate a text vector of the synthesized text includes: determining characters contained in the comprehensive text, determining a character vector of each character based on a trained encoder, and taking a vector sequence formed by all the character vectors as a text vector of the comprehensive text.
In the embodiment of the invention, before the encoding processing, the input comprehensive text is firstly processed into the characters, namely the characters contained in the input comprehensive text are determined. The characters are used as the minimum semantic unit, the word segmentation task at the upstream does not need to be relied on, and therefore rare words can be well processed. For example, if the integrated text is a Chinese text, each Chinese character can be regarded as a character.
After the characters in the synthesized text are determined, encoding processing can be performed based on an encoder to generate a character vector of each character. The encoder may be a recurrent neural network, or an encoder portion of a converter model, and the like, which is not limited in this embodiment. And then, all the character vectors can be formed, so that the generated vector sequence is used as a text vector of the comprehensive text.
Optionally, the step of determining a character vector of each character based on the trained encoder specifically includes: determining the query code, key code and value code of each character, and determining the character vector c of each character according to the query code, key code and value codei；
And,wherein, ciA character vector representing the ith character, n representing the number of characters, qiRepresenting the query code, k, of the ith characteriKey code representing the ith character, s (q)i,ki) Representing query codes qiAnd key code kiSimilarity between, viA value code representing the ith character.
In this embodiment, after determining the characters in the synthesized text, the trained transformation matrix may be used to transform the synthesized text into a new textThe character is converted into a corresponding query code (query), key code (key) and value code (value); then calculating the similarity between the query code and the key code, i.e. calculating s (q)i,ki) And determining a weight of each character based on the similarityThereby generating a character vector for each characterThe similarity between the query code and the key code may be calculated by using a similarity function, such as a dot product, a perceptron, and the like, which is not limited in this embodiment. Further, optionally, the key code may be the same as the value code, i.e., ki＝vi. After each character vector is determined, a corresponding text vector [ c ] may be generated1,c2,…,ci,…,cn]。
In addition, since the synthesized text contains the effective target entity and the text to be recognized, in order to distinguish the effective target entity from the text to be recognized, a separator may be added between the effective target entity and the text to be recognized, and at this time, the separator also needs to be a character of the synthesized text, that is, a character vector of the separator also needs to be determined. For example, the text to be recognized is "bank a obtains 500 ten thousand dollars through reducing company B, the effective target entity is" bank a ", and the integrated text is" bank a obtains 500 ten thousand dollars through reducing company B [ SEP ] bank a ", where" [ SEP ] "is a separator, and at this time, the character vector of the separator also needs to be determined.
Step 104: and then selecting the next target entity as an effective target entity, and continuing the process of determining the public sentiment category corresponding to the effective target entity until all the target entities in the text to be recognized are traversed.
In the embodiment of the present invention, after determining the public sentiment category of the currently selected target entity, other target entities may be selected as valid target entities, and then the above steps 102 to 103 are repeated until all target entities in the text to be recognized are traversed to determine the public sentiment categories of all target entities. For example, the text to be recognized is "Bank A earns 500 ten thousand dollars by reducing the stock of company B", and the target entities include "Bank A" and "company B"; at this time, the target entity 'bank a' can be selected as an effective target entity, and based on the above process, it can be determined that the public opinion category corresponding to the target entity 'bank a' is 'profit' or 'good interest', etc.; then, the target entity company B is selected as an effective target entity, and the public opinion categories corresponding to the target entity, namely 'reduction support', 'stock solution forbidding' or 'free space', are determined. In the embodiment, the public opinion category of each entity contained in the text can be determined in a mode of combining the text and the entity to generate the comprehensive text; meanwhile, the entity in the text is extracted in advance, the public sentiment category of the entity can be determined more pertinently, and the category identification accuracy is higher. Meanwhile, for the same text, different comprehensive texts can be generated when the same text is combined with different entities, namely, the finally determined public opinion category can also change along with the change of the entities. As in the above example, for the same text to be recognized, the public opinion category corresponding to the entity "bank a" is "good" and the public opinion category of the entity "company B" is "empty" in the exact opposite.
The method for classifying the public sentiments provided by the embodiment of the invention has the advantages that the entities in the text are identified in advance, and the public sentiment category of each entity contained in the text can be determined in a mode of combining the text and the entities to generate a comprehensive text; even different entities can be normally identified for totally different public opinion categories, which are negative. Meanwhile, the entity in the text is extracted in advance, the public sentiment category of the entity can be determined more pertinently, and the recognition accuracy of the public sentiment category is higher. During coding processing, characters are used as the minimum semantic unit, the word segmentation task at the upstream does not need to be relied on, and rare words can be well processed.
On the basis of the above embodiment, the method further includes a model training process, and the embodiment trains the model in a parameter sharing manner to improve the accuracy of model classification. Specifically, before "acquiring a text to be recognized" in step 101, the method further includes:
step A1: the method comprises the steps of determining a training text, and predetermining a target training entity in the training text, a first public sentiment category corresponding to the target training entity and a first public sentiment keyword.
In the embodiment of the invention, a training text is firstly determined, and a target training entity in the training text is determined; meanwhile, each target training entity has a corresponding public sentiment category, i.e., a first public sentiment category. In addition, the current public opinion classification model only gives the final classification result, but does not relate to the reason why the classification result is obtained, namely, the current public opinion classification model does not show the causal relationship of classification; in this embodiment, keywords related to the public sentiment category, i.e. the first public sentiment keyword, are extracted and trained, so as to highlight the causal relationship of the public sentiment classification, and further improve the accuracy of the public sentiment classification model.
For example, the training text is "great wall securities withholding hundred", the target training entity is "hundred", the public sentiment category of the target training entity is "Likong", and the corresponding public sentiment keyword is "withholding". Target training entities, public opinion categories, public opinion keywords and the like in the training texts can be determined in a manual labeling mode. Specifically, available data, such as news data, financial field self-media data, company bulletin text, research report text, and the like, may be collected via the internet or the like, and the collected text including the company name is used as a training text, and then the model in the present embodiment is trained.
Step A2: and generating a comprehensive training text by combining the training text and the target training entity, and coding the comprehensive training text based on the encoder to be trained to generate a training text vector of the comprehensive training text.
In this embodiment, similar to the step 103, in the training stage, it is still necessary to generate a comprehensive text by combining the text and the entity for encoding, that is, generating a comprehensive training text by combining the training text and the target training entity, and then performing encoding processing based on an encoder. In the training phase, the encoder is an encoder that needs to be trained, that is, parameters (such as weights) of the encoder need to be adjusted based on the training result.
Step A3: taking the training text vector as the input of a classification model to be trained, and determining a second public opinion category corresponding to the comprehensive training text based on the classification model; meanwhile, the training text vector is used as the input of a keyword extraction model to be trained, and a second public opinion keyword in the comprehensive training text is extracted based on the keyword extraction model.
In an embodiment of the present invention, after the encoder-based generation of the training text vector, on the one hand, the training text vector is input into the classification model; on the other hand, the training text vector is input into a keyword extraction model; based on the training model, the public opinion category (i.e. the second public opinion category) and the public opinion keyword (i.e. the second public opinion keyword) of the training text recognized by the model can be determined respectively.
Specifically, the model structure in this embodiment can be seen in fig. 2. In fig. 2, the training text is "great wall securities withholding hundred degrees", the labeled target training entity is "hundred degrees", and then the synthetic training text is subjected to a tokenization process to generate a tokenized synthetic training text: "[ CLS ] great wall securities subtract hundred [ SEP ]". Wherein, [ CLS ] represents a classifier for later classification, [ SEP ] is a separator, which is arranged at the head position of the target entity and is used for distinguishing the text from the target entity. After the symbolized comprehensive training text is determined, the comprehensive training text can be coded and corresponding training text vectors are generated; fig. 2 illustrates an example of an encoder as a converter (transforms). And then, the training text vectors generated by the encoder can be respectively transmitted to the classification model and the keyword extraction model. The classification model is a "classification layer" in fig. 2, which may be specifically a multi-layer neural network; the keyword extraction model is a "traceability layer" in fig. 2, and two traceability layers capable of outputting the interval start position are used as the keyword extraction model in fig. 2. Specifically, for each character, a classifier is used for calculating the probability that the character is an interval starting word, the probability is described by a traceability layer-1 in a schematic diagram, and the 'minus' in fig. 2 is the start of the interval; meanwhile, another classifier is used for calculating the probability that the word is the interval end word, and the tracing layer-2 in the schematic diagram is used for describing that the 'holding' in fig. 2 is the end of the interval. Based on the two traceability layers, a required interval can be extracted, namely the keyword 'withholding'.
Step A4: and taking the difference between the minimized first public opinion category and the minimized second public opinion category as a classification optimization target, taking the difference between the minimized first public opinion keyword and the minimized second public opinion keyword as a keyword extraction optimization target, and training based on the classification optimization target and the keyword extraction optimization target.
In the embodiment of the invention, on the basis of the public sentiment classification task, the task of public sentiment keyword extraction is added, and the two tasks can share the encoder part, namely the two tasks share the parameters of the encoder part, so that the encoder can be better trained in the unified training. Specifically, the encoder and the classification model in this embodiment form a complete public opinion classification model, and the encoder and the keyword extraction model form a complete keyword model, at this time, the difference between the minimized first public opinion category and the minimized second public opinion category can be used as an optimization target of the complete public opinion classification model, that is, a classification optimization target; meanwhile, the difference value between the minimized first public opinion keyword and the minimized second public opinion keyword is used as the optimization target of the complete keyword model, namely the keyword extraction optimization target, and the classification optimization target and the keyword extraction optimization target are maintained simultaneously in the training process, so that the public opinion keywords with causal relation can be transmitted to the classification task, and the classification effect of the classification model is improved. Meanwhile, when the public opinion category of the text to be recognized is determined, the causal relationship of the corresponding category can be output for subsequent use or viewing by the user.
Optionally, in this embodiment, public sentiment classification may be performed based on discrete features in the text, so as to further improve the public sentiment classification effect. Wherein, the step 103 of performing the public opinion classification processing according to the text vector includes:
extracting discrete features in the comprehensive text, wherein the discrete features comprise one or more of word frequency of public sentiment keywords in the comprehensive text, positions of the public sentiment keywords in the comprehensive text, and incidence relation between the public sentiment keywords and a target entity; and carrying out public opinion classification processing according to the discrete features and the text vectors.
In the embodiment of the invention, in the training stage, discrete features in the training text, such as the word frequency of the public sentiment keyword in the training text, the position of the public sentiment keyword in the training text, the association relationship between the public sentiment keyword and the target training entity, and the like, can be automatically extracted in advance or manually extracted, and the discrete features are also used as the input of the classification model to train the classification model in combination with the training text vector. When the public sentiment category of the text to be recognized needs to be determined, as shown in the above steps, discrete features in the comprehensive text can be extracted, and the discrete features and the text vector are combined to perform public sentiment classification processing. In fig. 2, a policy network is used as a network for extracting discrete features. The discrete characteristics are added during public opinion classification, classification accuracy can be further improved, and later maintenance and artificial rule addition are facilitated.
On the basis of the embodiment, in the process of model training, the task of language model class can be added, parameter sharing can be realized, and the accuracy of model classification can be improved. Specifically, before "acquiring a text to be recognized" in step 101, the method further includes:
step B1: determining a training text and a first public opinion category corresponding to the training text, and predetermining a target training entity in the training text.
In the embodiment of the present invention, similar to the above step a1, the training text and the target training entity and the public opinion category therein are determined during training, except that the public opinion keyword therein does not need to be determined.
Step B2: one or more characters are randomly selected from the training texts to be used as first mask characters, and the training texts with the first mask characters deleted are used as mask training texts.
In this embodiment, the deleted word in the predicted text is used as a new language model task. Specifically, part of the characters in the training text are deleted randomly, the deleted characters are first mask characters, and the training text from which the first mask characters are deleted is a mask training text. For example, the training text is "great wall securities minus Mask hundred", the characters "certificate" and "hold" are randomly deleted, at this time, the corresponding substitute Mask can be used for substitution, and the Mask training text is "great wall Mask ticket minus Mask hundred". Optionally, characters related to the public sentiment keywords may be preferentially deleted.
Step B3: generating a comprehensive training text by combining the training text and a target training entity, and coding the comprehensive training text based on a coder to be trained to generate a training text vector of the comprehensive training text; and generating a mask comprehensive training text by combining the mask training text and the target training entity, and coding the mask comprehensive training text based on the same coder to be trained to generate a mask training text vector of the mask comprehensive training text.
In the embodiment of the invention, in the training stage, two vectors are respectively input under the condition of not changing the parameters of the encoder. Specifically, as in the above step a2, the integrated training text is input to an encoder for encoding, so as to generate a corresponding training vector; in addition, in this embodiment, the mask training text and the target training entity are further combined to generate a mask integrated training text, and a mask training text vector is generated based on an encoder having the same parameters. The above-mentioned "training text vector" and "mask training text vector" correspond to different inputs, but are generated by the same encoder.
Step B4: taking the training text vector as the input of a classification model to be trained, and determining a second public opinion category corresponding to the comprehensive training text based on the classification model; and meanwhile, the mask training text vector is used as the input of a language prediction model to be trained, and the deleted second mask characters in the mask training text are determined based on the language prediction model.
In this embodiment, similar to step a3 described above, the training text vector is used to train the classification model, and the mask training text vector in this embodiment is used to train the language prediction model in this embodiment, i.e., to predict the deleted character in the mask training text, i.e., the second mask character. Referring to fig. 3, fig. 3 illustrates a process of performing language prediction based on the mask training text, where the "language prediction model" in fig. 3 is the "language prediction model" in step B4.
Step B5: and taking the difference between the minimized first public sentiment category and the minimized second public sentiment category as a classification optimization target, taking the difference between the minimized first mask character and the minimized second mask character as a mask prediction optimization target, and training based on the classification optimization target and the mask prediction optimization target.
In this embodiment, similar to the above step a4, a word prediction task is added on the basis of a public sentiment classification task, and although the inputs of the two tasks are different, the two tasks can still share the same encoder, i.e. the two tasks (the public sentiment classification task and the word prediction task) can still share the parameters of the encoder part, so that the encoder can be better trained in the unified training. Similarly, the encoder and the language prediction model in this embodiment form a complete language model, and the minimum difference between the first mask character and the second mask character is used as an optimization target of the complete language model, that is, a mask prediction optimization target, and the classification optimization target and the mask prediction optimization target are maintained simultaneously in the training process, so that the encoder can more effectively identify a text, and the classification effect of the classification model is further improved. Meanwhile, a large amount of labeled data is needed in the traditional public sentiment classification training process, and the language prediction model in the embodiment can realize word prediction without labeled data, so that the training can be carried out by utilizing huge labeled-free data outside.
It will be appreciated by those skilled in the art that the above steps A1-A4 and steps B1-B5 may be further combined, i.e., three tasks are simultaneously established during the training phase: the method comprises a public opinion classification task, a public opinion keyword extraction task and a word prediction task, wherein the three tasks are executed by utilizing a classification model, a keyword extraction model and a language prediction model respectively, optimization targets of the three tasks are maintained simultaneously, and classification effects of the classification model are improved. In this embodiment, based on a plurality of natural language processing tasks, an addition to the classification effect can be obtained, and the accuracy of public opinion classification can be improved.
The method for classifying the public sentiments provided by the embodiment of the invention has the advantages that the entities in the text are identified in advance, and the public sentiment category of each entity contained in the text can be determined in a mode of combining the text and the entities to generate a comprehensive text; even different entities can be normally identified for totally different public opinion categories, which are negative. Meanwhile, the entity in the text is extracted in advance, the public sentiment category of the entity can be determined more pertinently, and the recognition accuracy of the public sentiment category is higher. During coding processing, characters are used as the minimum semantic unit, the word segmentation task at the upstream does not need to be relied on, and rare words can be well processed. A plurality of natural language processing tasks are established in a training stage, the plurality of tasks can share parameters, and optimization targets of all the tasks need to be maintained simultaneously during training and interact with each other, so that a coder can effectively recognize texts, and the classification effect of a classification model is improved.
The above describes the flow of the public opinion classification method in detail, and the method can also be implemented by a corresponding device, and the structure and function of the device are described in detail below.
An embodiment of the present invention provides a public opinion classifying device, as shown in fig. 4, including:
the acquiring module 41 is configured to acquire a text to be recognized and extract one or more target entities in the text to be recognized;
the text combining module 42 is configured to select one target entity as an effective target entity, and combine the effective target entity and the text to be recognized to generate a comprehensive text;
a public opinion classification module 43, configured to perform coding processing on the comprehensive text to generate a text vector of the comprehensive text, perform public opinion classification processing according to the text vector, and determine a public opinion category corresponding to the current effective target entity;
and the traversing module 44 is configured to select a next target entity as an effective target entity, and continue the above-mentioned process of determining the public opinion category corresponding to the effective target entity until all the target entities in the text to be recognized are traversed.
On the basis of the above embodiment, the public opinion classification module 43 performs encoding processing on the integrated text to generate a text vector of the integrated text, including:
determining characters contained in the comprehensive text, determining a character vector of each character based on a trained encoder, and taking a vector sequence formed by all the character vectors as a text vector of the comprehensive text.
On the basis of the above embodiment, the determining, by the public opinion classification module 43, the character vector of each character based on the trained encoder includes:
determining a query code, a key code and a value code for each character, and determining a character vector c for each character according to the query code, the key code and the value codei；
Wherein, ciA character vector representing the ith character, n representing the number of characters, qiRepresenting the query code, k, of the ith characteriKey code representing the ith character, s (q)i,ki) Representing query codes qiAnd key code kiSimilarity between, viA value code representing the ith character.
On the basis of the embodiment, the device also comprises a training module;
before the obtaining module 41 obtains the text to be recognized, the training module is configured to:
determining a training text, and predetermining a target training entity in the training text, a first public sentiment category and a first public sentiment keyword corresponding to the target training entity;
generating a comprehensive training text by combining the training text and the target training entity, and coding the comprehensive training text based on a coder to be trained to generate a training text vector of the comprehensive training text;
the training text vector is used as the input of a classification model to be trained, and a second public opinion category corresponding to the comprehensive training text is determined based on the classification model; meanwhile, the training text vector is used as the input of a keyword extraction model to be trained, and a second public opinion keyword in the comprehensive training text is extracted based on the keyword extraction model;
and taking the minimized difference value between the first public opinion category and the second public opinion category as a classification optimization target, taking the minimized difference value between the first public opinion keyword and the second public opinion keyword as a keyword extraction optimization target, and training based on the classification optimization target and the keyword extraction optimization target.
On the basis of the embodiment, the device also comprises a training module;
before the obtaining module 41 obtains the text to be recognized, the training module is configured to:
determining a training text, and predetermining a target training entity in the training text and a first public opinion category corresponding to the target training entity;
randomly selecting one or more characters from the training text as first mask characters, and using the training text with the first mask characters deleted as a mask training text;
generating a comprehensive training text by combining the training text and the target training entity, and coding the comprehensive training text based on a coder to be trained to generate a training text vector of the comprehensive training text; generating a mask comprehensive training text by combining the mask training text and the target training entity, and coding the mask comprehensive training text based on the same encoder to be trained to generate a mask training text vector of the mask comprehensive training text;
the training text vector is used as the input of a classification model to be trained, and a second public opinion category corresponding to the comprehensive training text is determined based on the classification model; meanwhile, the mask training text vector is used as the input of a language prediction model to be trained, and the deleted second mask characters in the mask training text are determined based on the language prediction model;
and taking the difference between the first public sentiment category and the second public sentiment category which is minimized as a classification optimization target, taking the difference between the first mask character and the second mask character which is minimized as a mask prediction optimization target, and training based on the classification optimization target and the mask prediction optimization target.
On the basis of the above embodiment, the public opinion classification module 43 performing public opinion classification processing according to the text vector includes:
extracting discrete features in the comprehensive text, wherein the discrete features comprise one or more of word frequency of public sentiment keywords in the comprehensive text, positions of the public sentiment keywords in the comprehensive text, and incidence relation between the public sentiment keywords and the target entity;
and carrying out public opinion classification processing according to the discrete features and the text vectors.
The public opinion classification device provided by the embodiment of the invention identifies the entities in the text in advance, and can determine the public opinion category of each entity contained in the text in a mode of combining the text and the entities to generate a comprehensive text; even different entities can be normally identified for totally different public opinion categories, which are negative. Meanwhile, the entity in the text is extracted in advance, the public sentiment category of the entity can be determined more pertinently, and the recognition accuracy of the public sentiment category is higher. During coding processing, characters are used as the minimum semantic unit, the word segmentation task at the upstream does not need to be relied on, and rare words can be well processed. A plurality of natural language processing tasks are established in a training stage, the plurality of tasks can share parameters, and optimization targets of all the tasks need to be maintained simultaneously during training and interact with each other, so that a coder can effectively recognize texts, and the classification effect of a classification model is improved.
The embodiment of the invention also provides a computer storage medium, wherein the computer storage medium stores computer-executable instructions, which comprise a program for executing the public opinion classification method, and the computer-executable instructions can execute the method in any method embodiment.
The computer storage medium may be any available medium or data storage device that can be accessed by a computer, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO), etc.), optical memory (e.g., CD, DVD, BD, HVD, etc.), and semiconductor memory (e.g., ROM, EPROM, EEPROM, nonvolatile memory (NANDFLASH), Solid State Disk (SSD)), etc.
Fig. 5 shows a block diagram of an electronic device according to another embodiment of the present invention. The electronic device 1100 may be a host server with computing capabilities, a personal computer PC, or a portable computer or terminal that is portable, or the like. The specific embodiment of the present invention does not limit the specific implementation of the electronic device.
The electronic device 1100 includes at least one processor (processor)1110, a Communications Interface 1120, a memory 1130, and a bus 1140. The processor 1110, the communication interface 1120, and the memory 1130 communicate with each other via the bus 1140.
The communication interface 1120 is used for communicating with network elements including, for example, virtual machine management centers, shared storage, etc.
Processor 1110 is configured to execute programs. Processor 1110 may be a central processing unit CPU, or an Application Specific Integrated Circuit (ASIC), or one or more Integrated circuits configured to implement embodiments of the present invention.
The memory 1130 is used for executable instructions. The memory 1130 may comprise high-speed RAM memory, and may also include non-volatile memory (non-volatile memory), such as at least one disk memory. The memory 1130 may also be a memory array. The storage 1130 may also be partitioned and the blocks may be combined into virtual volumes according to certain rules. The instructions stored in the memory 1130 are executable by the processor 1110 to enable the processor 1110 to perform the method for public opinion classification in any of the above-described method embodiments.
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. A public opinion classification method is characterized by comprising the following steps:
acquiring a text to be recognized, and extracting one or more target entities in the text to be recognized;
selecting one target entity as an effective target entity, and generating a comprehensive text by combining the effective target entity and the text to be recognized;
coding the comprehensive text to generate a text vector of the comprehensive text, and carrying out public sentiment classification processing according to the text vector to determine the public sentiment category corresponding to the current effective target entity;
and then selecting the next target entity as an effective target entity, and continuing the process of determining the public sentiment category corresponding to the effective target entity until all the target entities in the text to be recognized are traversed.
2. The method according to claim 1, wherein said encoding said synthesized text to generate a text vector of said synthesized text comprises:
determining characters contained in the comprehensive text, determining a character vector of each character based on a trained encoder, and taking a vector sequence formed by all the character vectors as a text vector of the comprehensive text.
3. The method of claim 2, wherein determining the character vector for each of the characters based on the trained encoder comprises:
determining a query code, a key code and a value code for each character, and determining a character vector c for each character according to the query code, the key code and the value codei；
Wherein, ciA character vector representing the ith character, n representing the number of characters, qiRepresenting the query code, k, of the ith characteriKey code representing the ith character, s (q)i,ki) Representing query codes qiAnd key code kiSimilarity between, viA value code representing the ith character.
4. The method according to claim 1, wherein before the obtaining the text to be recognized, further comprising:
determining a training text, and predetermining a target training entity in the training text, a first public sentiment category and a first public sentiment keyword corresponding to the target training entity;
generating a comprehensive training text by combining the training text and the target training entity, and coding the comprehensive training text based on a coder to be trained to generate a training text vector of the comprehensive training text;
the training text vector is used as the input of a classification model to be trained, and a second public opinion category corresponding to the comprehensive training text is determined based on the classification model; meanwhile, the training text vector is used as the input of a keyword extraction model to be trained, and a second public opinion keyword in the comprehensive training text is extracted based on the keyword extraction model;
and taking the minimized difference value between the first public opinion category and the second public opinion category as a classification optimization target, taking the minimized difference value between the first public opinion keyword and the second public opinion keyword as a keyword extraction optimization target, and training based on the classification optimization target and the keyword extraction optimization target.
5. The method according to claim 1, wherein before the obtaining the text to be recognized, further comprising:
determining a training text, and predetermining a target training entity in the training text and a first public opinion category corresponding to the target training entity;
randomly selecting one or more characters from the training text as first mask characters, and using the training text with the first mask characters deleted as a mask training text;
generating a comprehensive training text by combining the training text and the target training entity, and coding the comprehensive training text based on a coder to be trained to generate a training text vector of the comprehensive training text; generating a mask comprehensive training text by combining the mask training text and the target training entity, and coding the mask comprehensive training text based on the same encoder to be trained to generate a mask training text vector of the mask comprehensive training text;
the training text vector is used as the input of a classification model to be trained, and a second public opinion category corresponding to the comprehensive training text is determined based on the classification model; meanwhile, the mask training text vector is used as the input of a language prediction model to be trained, and the deleted second mask characters in the mask training text are determined based on the language prediction model;
and taking the difference between the first public sentiment category and the second public sentiment category which is minimized as a classification optimization target, taking the difference between the first mask character and the second mask character which is minimized as a mask prediction optimization target, and training based on the classification optimization target and the mask prediction optimization target.
6. The method according to any one of claims 1-5, wherein the performing the public opinion classification process according to the text vector comprises:
extracting discrete features in the comprehensive text, wherein the discrete features comprise one or more of word frequency of public sentiment keywords in the comprehensive text, positions of the public sentiment keywords in the comprehensive text, and incidence relation between the public sentiment keywords and the target entity;
and carrying out public opinion classification processing according to the discrete features and the text vectors.
7. The utility model provides a public opinion classification's device which characterized in that includes:
the system comprises an acquisition module, a recognition module and a recognition module, wherein the acquisition module is used for acquiring a text to be recognized and extracting one or more target entities in the text to be recognized;
the text combination module is used for selecting one target entity as an effective target entity and combining the effective target entity and the text to be recognized to generate a comprehensive text;
the public opinion classification module is used for coding the comprehensive text to generate a text vector of the comprehensive text, and performing public opinion classification according to the text vector to determine the public opinion category corresponding to the current effective target entity;
and the traversing module is used for selecting the next target entity as an effective target entity, and continuing the process of determining the public sentiment category corresponding to the effective target entity until all the target entities in the text to be recognized are traversed.
8. The apparatus of claim 7, wherein the public opinion classification module encodes the integrated text to generate a text vector of the integrated text, and comprises:
determining characters contained in the comprehensive text, determining a character vector of each character based on a trained encoder, and taking a vector sequence formed by all the character vectors as a text vector of the comprehensive text.
9. A computer storage medium storing computer-executable instructions for performing the method for public opinion classification of any one of claims 1-6.
10. An electronic device, comprising:
at least one processor; and the number of the first and second groups,
a memory communicatively coupled to the at least one processor; wherein,
the memory stores instructions executable by the at least one processor to cause the at least one processor to perform the method for public opinion classification of any of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910683305.XA CN110377744B (en) | 2019-07-26 | 2019-07-26 | Public opinion classification method and device, storage medium and electronic equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910683305.XA CN110377744B (en) | 2019-07-26 | 2019-07-26 | Public opinion classification method and device, storage medium and electronic equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110377744A true CN110377744A (en) | 2019-10-25 |
| CN110377744B CN110377744B (en) | 2022-08-09 |
Family
ID=68256435
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910683305.XA Active CN110377744B (en) | 2019-07-26 | 2019-07-26 | Public opinion classification method and device, storage medium and electronic equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110377744B (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110851598A (en) * | 2019-10-30 | 2020-02-28 | 深圳价值在线信息科技股份有限公司 | Text classification method and device, terminal equipment and storage medium |
| CN111046172A (en) * | 2019-10-30 | 2020-04-21 | 北京奇艺世纪科技有限公司 | Public opinion analysis method, device, equipment and storage medium |
| CN112667779A (en) * | 2020-12-30 | 2021-04-16 | 北京奇艺世纪科技有限公司 | Information query method and device, electronic equipment and storage medium |
| CN112784612A (en) * | 2021-01-26 | 2021-05-11 | 浙江香侬慧语科技有限责任公司 | Method, apparatus, medium, and device for synchronous machine translation based on iterative modification |
| CN113434688A (en) * | 2021-08-23 | 2021-09-24 | 南京擎盾信息科技有限公司 | Data processing method and device for public opinion classification model training |
| WO2021208612A1 (en) * | 2020-04-13 | 2021-10-21 | 华为技术有限公司 | Data processing method and device |
| CN113743117A (en) * | 2020-05-29 | 2021-12-03 | 华为技术有限公司 | Method and device for entity marking |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104615687A (en) * | 2015-01-22 | 2015-05-13 | 中国科学院计算技术研究所 | Entity fine granularity classifying method and system for knowledge base updating |
| US20170192958A1 (en) * | 2015-12-31 | 2017-07-06 | Accenture Global Solutions Limited | Input entity identification from natural language text information |
| CN107526819A (en) * | 2017-08-29 | 2017-12-29 | 江苏飞搏软件股份有限公司 | A kind of big data the analysis of public opinion method towards short text topic model |
| CN108133038A (en) * | 2018-01-10 | 2018-06-08 | 重庆邮电大学 | A kind of entity level emotional semantic classification system and method based on dynamic memory network |
| CN108628974A (en) * | 2018-04-25 | 2018-10-09 | 平安科技（深圳）有限公司 | Public feelings information sorting technique, device, computer equipment and storage medium |
| CN109213868A (en) * | 2018-11-21 | 2019-01-15 | 中国科学院自动化研究所 | Entity level sensibility classification method based on convolution attention mechanism network |
| CN109446300A (en) * | 2018-09-06 | 2019-03-08 | 厦门快商通信息技术有限公司 | A kind of corpus preprocess method, the pre- mask method of corpus and electronic equipment |
| CN109857868A (en) * | 2019-01-25 | 2019-06-07 | 北京奇艺世纪科技有限公司 | Model generating method, file classification method, device and computer readable storage medium |
-
2019
- 2019-07-26 CN CN201910683305.XA patent/CN110377744B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104615687A (en) * | 2015-01-22 | 2015-05-13 | 中国科学院计算技术研究所 | Entity fine granularity classifying method and system for knowledge base updating |
| US20170192958A1 (en) * | 2015-12-31 | 2017-07-06 | Accenture Global Solutions Limited | Input entity identification from natural language text information |
| CN107526819A (en) * | 2017-08-29 | 2017-12-29 | 江苏飞搏软件股份有限公司 | A kind of big data the analysis of public opinion method towards short text topic model |
| CN108133038A (en) * | 2018-01-10 | 2018-06-08 | 重庆邮电大学 | A kind of entity level emotional semantic classification system and method based on dynamic memory network |
| CN108628974A (en) * | 2018-04-25 | 2018-10-09 | 平安科技（深圳）有限公司 | Public feelings information sorting technique, device, computer equipment and storage medium |
| CN109446300A (en) * | 2018-09-06 | 2019-03-08 | 厦门快商通信息技术有限公司 | A kind of corpus preprocess method, the pre- mask method of corpus and electronic equipment |
| CN109213868A (en) * | 2018-11-21 | 2019-01-15 | 中国科学院自动化研究所 | Entity level sensibility classification method based on convolution attention mechanism network |
| CN109857868A (en) * | 2019-01-25 | 2019-06-07 | 北京奇艺世纪科技有限公司 | Model generating method, file classification method, device and computer readable storage medium |
Non-Patent Citations (4)
| Title |
|---|
| JAVIER ZAMBRANO FERREIRA: "Multi-Entity Polarity Analysis in Financial Documents", 《WEBMEDIA "14: PROCEEDINGS OF THE 20TH BRAZILIAN SYMPOSIUM ON MULTIMEDIA AND THE WEB》 * |
| SBJ123456789: "从attention到self-attention", 《HTTPS://WWW.CNBLOGS.COM/SBJ123456789/P/10750819.HTML》 * |
| ZHE YE: "Encoding Sentiment Information intoWord Vectors for Sentiment Analysis", 《PROCEEDINGS OF THE 27TH INTERNATIONAL CONFERENCE ON COMPUTATIONAL LINGUISTICS》 * |
| 林海伦: "面向网络大数据的知识融合方法综述", 《计算机学报》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110851598A (en) * | 2019-10-30 | 2020-02-28 | 深圳价值在线信息科技股份有限公司 | Text classification method and device, terminal equipment and storage medium |
| CN111046172A (en) * | 2019-10-30 | 2020-04-21 | 北京奇艺世纪科技有限公司 | Public opinion analysis method, device, equipment and storage medium |
| CN111046172B (en) * | 2019-10-30 | 2024-04-12 | 北京奇艺世纪科技有限公司 | Public opinion analysis method, device, equipment and storage medium |
| WO2021208612A1 (en) * | 2020-04-13 | 2021-10-21 | 华为技术有限公司 | Data processing method and device |
| CN113743117A (en) * | 2020-05-29 | 2021-12-03 | 华为技术有限公司 | Method and device for entity marking |
| CN113743117B (en) * | 2020-05-29 | 2024-04-09 | 华为技术有限公司 | Method and device for entity labeling |
| CN112667779A (en) * | 2020-12-30 | 2021-04-16 | 北京奇艺世纪科技有限公司 | Information query method and device, electronic equipment and storage medium |
| CN112667779B (en) * | 2020-12-30 | 2023-09-05 | 北京奇艺世纪科技有限公司 | Information query method and device, electronic equipment and storage medium |
| CN112784612A (en) * | 2021-01-26 | 2021-05-11 | 浙江香侬慧语科技有限责任公司 | Method, apparatus, medium, and device for synchronous machine translation based on iterative modification |
| CN112784612B (en) * | 2021-01-26 | 2023-12-22 | 浙江香侬慧语科技有限责任公司 | Method, device, medium and equipment for synchronous machine translation based on iterative modification |
| CN113434688A (en) * | 2021-08-23 | 2021-09-24 | 南京擎盾信息科技有限公司 | Data processing method and device for public opinion classification model training |
| CN113434688B (en) * | 2021-08-23 | 2021-12-21 | 南京擎盾信息科技有限公司 | Data processing method and device for public opinion classification model training |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110377744B (en) | 2022-08-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110377744B (en) | Public opinion classification method and device, storage medium and electronic equipment | |
| CN112084337B (en) | Training method of text classification model, text classification method and equipment | |
| CN106951422B (en) | Webpage training method and device, and search intention identification method and device | |
| CN109992664B (en) | Dispute focus label classification method and device, computer equipment and storage medium | |
| CN107229627B (en) | Text processing method and device and computing equipment | |
| CN111858843B (en) | Text classification method and device | |
| CN111753082A (en) | Text classification method and device based on comment data, equipment and medium | |
| CN114330354A (en) | Event extraction method and device based on vocabulary enhancement and storage medium | |
| CN110909540B (en) | Method and device for identifying new words of short message spam and electronic equipment | |
| CN111462752A (en) | Client intention identification method based on attention mechanism, feature embedding and BI-L STM | |
| CN112732871A (en) | Multi-label classification method for acquiring client intention label by robot | |
| CN110796171A (en) | Unclassified sample processing method and device of machine learning model and electronic equipment | |
| CN113220999B (en) | User characteristic generation method and device, electronic equipment and storage medium | |
| CN108205524B (en) | Text data processing method and device | |
| CN114997169A (en) | Entity word recognition method and device, electronic equipment and readable storage medium | |
| CN112084779B (en) | Entity acquisition method, device, equipment and storage medium for semantic recognition | |
| CN111612284A (en) | Data processing method, device and equipment | |
| CN113420119B (en) | Intelligent question-answering method, device, equipment and storage medium based on knowledge card | |
| CN114611625A (en) | Language model training method, language model training device, language model data processing method, language model data processing device, language model data processing equipment, language model data processing medium and language model data processing product | |
| CN114266252A (en) | Named entity recognition method, device, equipment and storage medium | |
| US11922515B1 (en) | Methods and apparatuses for AI digital assistants | |
| CN117807482A (en) | Method, device, equipment and storage medium for classifying customs clearance notes | |
| CN111797194A (en) | Text risk detection method and device, electronic equipment and storage medium | |
| CN111274382A (en) | Text classification method, device, equipment and storage medium | |
| CN113051396B (en) | Classification recognition method and device for documents and electronic equipment |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |