CN103207855A - Fine-grained sentiment analysis system and method specific to product comment information - Google Patents
Fine-grained sentiment analysis system and method specific to product comment information Download PDFInfo
- Publication number
- CN103207855A CN103207855A CN2013100360341A CN201310036034A CN103207855A CN 103207855 A CN103207855 A CN 103207855A CN 2013100360341 A CN2013100360341 A CN 2013100360341A CN 201310036034 A CN201310036034 A CN 201310036034A CN 103207855 A CN103207855 A CN 103207855A
- Authority
- CN
- China
- Prior art keywords
- emotion
- module
- user
- model
- entity
- 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
Images
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses a fine-grained sentiment analysis system and method specific to product comment information. The system comprises a user interface, a product comment information training sample data base, a loading module of relevant dictionaries such as sentiment dictionaries, a text preprocessing module, a characteristic extraction module, a sentiment analysis model training module, a sentiment tendency judging module and a feedback module. A user can store and manage various tagged product comment information training samples and perform processes of sentence segmentation, word segmentation, part-of-speech tagging, syntactic analysis and the like by means of the system. The user can also add user-defined sentiment dictionaries to perform characteristic extraction and vectorization on well processed text information, train sentiment analysis models and use the sentiment analysis models to perform sentiment tendency judgment on vectorized text information. The system feeds sentiment analysis results back to the user to support the user to correct and store the analysis results.
Description
Technical field
The present invention relates to Chinese text emotion analysis field, particularly a kind of entity level fine granularity emotion analytic system and method at product review information.
Background technology
Product review information typically refers to evaluation or the suggestion that user or consumer deliver with regard to a certain attribute of a certain product or this product.These review information often have stronger subjective emotion color, have embodied the subjective feeling of user for product or its a certain attribute.
One side is along with Web2.0's is flourish, and the Internet user changes to initiatively creating internet information by receiving internet information passively; Electricity merchant's the shopping that is changing people that develops rapidly is accustomed on the other hand, beginning is by transferring on the line under the line, shopping at network universal brings be this series products review information rapid expansion, these magnanimity informations are to make shopping for businessman's design improvement product or consumer to select that immense value is all arranged, yet only depend on artificial method to be difficult to tackle the processing of magnanimity information.Therefore press for automatic emotion recognition technology.To be that tendentiousness that whether text message is had emotion color and an emotion color is namely positive or negative make judgement to the main task of emotion recognition technology.
At present existing emotion analytic system and technology mainly concentrate on the emotion analysis of chapter rank and sentence level from the granularity of analyzing, and the emotion analytical technology of the entity level of only a few with Entity recognition and emotion analysis be divided into two independently task carry out.To pay close attention to the analysis of social public sentiment at review information such as news, microbloggings from present system and the technology of analyzing of object.
Present existing chapter rank and sentence level emotion analytical technology mainly contain: the application number of Northwestern Polytechnical University is that CN200910219161.9, denomination of invention are the patent of " based on the WEB text emotion subject identifying method of mixture model "; The application number of Inst. of Computing Techn. Academia Sinica is that CN200910083522.1, denomination of invention are the patented claim of " method for analyzing emotion tendentiousness of text "; The application number of Institute of Automation Research of CAS is that CN201210088366.X, denomination of invention are the patented claim of " a kind of emotion analytical approach towards the microblogging short text "; The application number of Fujitsu Ltd. is that CN201010157784.0, denomination of invention are the patented claim of " emotional orientation analytical method and device ".
Above-mentioned emotion analytical technology mainly comprises training and two steps of emotion judgement, " based on the WEB text emotion subject identifying method of mixture model " of thinking Northwestern Polytechnical University below introduced it in the key step of training and emotion judgement for example, and all the other correlation techniques are similar substantially.This method mainly comprises following step: 1, the text in the training set is carried out manual mark, estimate two class emotion models: " commendation " model and " derogatory sense " model; According to the language performance mode of different themes text, estimate all kinds of topic language models respectively simultaneously; 2, emotion model and the topic model that adopts maximal possibility estimation (MLE) method to set up for step 1 carries out parameter estimation respectively; 3, for pending text, calculate the distance of its language model and two class emotion models, thereby emotion tendency and the theme of text are judged.
Present emotion tendency technology mainly concentrates on chapter rank and sentence level, the emotion tendency technology of entity level is still very few, and it is that CN200910086542.4, patent name are the patented claim of " a kind of emotion tendency analysis system based on news comment webpage " that such technology has only the application number of Peking University at present.Introduce the basic procedure of this technology below: 1, the set of input news web page and emotion analysis rule; 2, from news comment webpage, extract effective entity, and set up entity level graph of a relation; 3, position reference entity in entity level graph of a relation is set up the entity emotion relationship graph of being made up of emotion relationship tree; 4, obtain intermediate entities in the news comment webpage, and the emotion of output intermediate entities is analyzed data; 5, judge whether intermediate entities is present in entity emotion relationship graph and the emotion relationship tree, take corresponding strategy output emotion judged result according to judged result.
Mainly there is the deficiency of several aspects in existing emotion analytical technology: A) granularity of emotion analysis is bigger, only the emotion tendency of whole piece sentence or entire article is made analysis and judgement, these class methods have directly been ignored more fine-grained information in the sentence, cause in the text in a large number according to the losing of valuable information, can not take full advantage of the information that contains in the text; B) the emotion analytical technology of the entity level of existing minority is divided into independently two steps with the judgement of the identification of entity and emotion tendency and carries out, and has ignored contacting between the judgement of Entity recognition and emotion tendency, makes that final associating precision is not high; C) need manually the concentrated text of training data to be marked, this can expend a large amount of time and human resources, and because the tagger's is cognitive different, conflicting mark result can appear inevitably also in the mark process, the effect of influence training; D) lack feedback mechanism and automatic study mechanism, existing system and technology are directly exported to the user for judged result, and do not do follow-up processing, lack the process that the judged result of mistake is learnt again; E) lack the research that concentrates on the analysis of product review information emotion, also very easily obtain and this category information is very abundant on network, also contain simultaneously huge commercial value and business opportunity, rationally fully using these information to create huge value and income as businessman and society.
Summary of the invention
In order to address the above problem, the present invention proposes a kind of fine granularity emotion analytic system and method that is directed to product review information.
Emotion analytic system of the present invention comprises: user interface, be used for the mutual of system and user, and the user can submit the product review information set to by this module; Product review information is climbed the delivery piece, this module can be regularly carried out product review information to the shopping website of appointments such as store, Jingdone district, Amazon and is climbed and get, product review information on these websites or be divided into favorable comment and difference is commented or star mark is arranged, the training sample data that are converted into mark according to these information are stored in the database; Product review information training sample database is used for the various product review information training samples that marked of storage; Text pretreatment module, this module are used for text message and the original training sample data of user's input are carried out pre-service work such as subordinate sentence, participle, part-of-speech tagging and syntactic analysis; The dictionary load-on module is used for loading related resources such as emotion word dictionary in feature extraction or preprocessing process; The feature extraction module is used for pretreated text message is carried out feature extraction, with the text message vectorization, is converted into the training data form of regulation in conjunction with the resources such as emotion word dictionary that load; Emotion analytical model training module uses the training data adjustment model parameter after transforming, and the emotion analytical model is trained the emotion analytical model that output trains; The emotion analysis module carries out emotion tendency in the emotion analytical model that the input of the text message of vectorization has been trained and judges output emotion analysis result; Feedback module, the user can the correction analysis result after obtaining analysis result, and feeds back to system by this module, confirms that through keeper's desk checking the back adds the training sample database.
The invention allows for a kind of fine-grained emotion analytical approach, this method comprises: 1) climb the product review information of getting band star mark on the network, reduce manually mark; 2) with the word be the minimum granularity of analyzing, constructed a kind of DCRF model of the Two-Level of having structure, carry out Entity recognition and emotion tendency simultaneously and judge two tasks, can introduce the raising that abundant characteristic information helps analytical effect, and can identify new emotion word and matched combined, upgrade dictionary data automatically; 3) introduce feedback mechanism, model can be learnt wrongheaded sample data.
Use System and method for of the present invention to have the advantage of several respects: 1) granularity of emotion analysis is little.Can carry out the fine granularity emotion analytical work of entity level, to text carry out more comprehensively, accurate, careful analysis, fully obtain the information in the comment text, increased the quantity of information of obtaining greatly, improved value and the authenticity of information; 2) the accuracy rate height of emotion analysis.In the emotion analytical work, introduce the DCRF model with Two-Level structure on the one hand first, realization Entity recognition work and emotion analytical work are carried out simultaneously, its advantage is to set up the contact of two workplaces, and between word relation information and the sentence between contextual information, these enrich the introducing of effective information, realized to a certain extent from semantically carrying out the emotion analysis, can help to improve the precision that emotion is analyzed.And these valuable information can not effectively be utilized under existing method frame, often are left in the basket.On the other hand, can effectively identify new emotion word and matched combined, add few manual examination and verification of trying one's best, dictionary resources is upgraded, guarantee the real-time of these resources; 3) manual intervention is few.By grasping the similar information such as review information that have star on the net, reduce the subjective factor of people in the artificial mark on the one hand to the influence of the mark of emotion word, reduced the artificial mark of training data on the other hand, save the time and reduced human cost, and can upgrade corpus termly; 4) system has automatic study mechanism.The introducing of feedback mechanism can help the wrongheaded sample information of model learning, makes model constantly to learn, and improves precision, reaches and more uses effect more accurately.
Description of drawings
Fig. 1 is the fine granularity emotion analytic system Organization Chart at product information of the present invention;
Fig. 2 is the process flow diagram of fine granularity emotion analytical approach of the present invention;
Fig. 3 is the realization schematic diagram that the training sample data are collected step in the fine granularity emotion analytical approach of the present invention;
Fig. 4 is the realization schematic diagram of emotion analytical model training step in the fine granularity emotion analytical approach of the present invention;
Fig. 5 is for being the graph structure of the Linear-CRF model of example with the named entity task;
Fig. 6 is the graph structure exemplary plot of the Two-Level CRF model that adopts in the fine granularity emotion analytical approach of the present invention;
Fig. 7 is the realization schematic diagram of product review information emotion analytical procedure in the fine granularity emotion analytical approach of the present invention.
Specific embodiments
Many detailed descriptions with reference to the accompanying drawings, and the model of introducing emotion analytical algorithm and use in detail is for making the purpose, technical solutions and advantages of the present invention clearer, below in conjunction with specific embodiment, and with reference to accompanying drawing, the present invention is described in more detail.
Fig. 1 shows the Organization Chart of the fine granularity emotion analytic system that the present invention is directed to product information.
With reference to Fig. 1, emotion analytic system of the present invention comprises that user interface, product review information climb delivery piece, product review information training sample database, dictionary database, text pretreatment module, dictionary load-on module, feature extraction module, emotion analytical model training module, emotion analysis module, administrator interface, system management module and database interface.
User interface be used for to realize that the emotion analytic system communicates by letter with the various of user, comprises obtaining product review relevant textual information that the user imports and information being passed to the text pretreatment module; The emotion analysis result that the emotion analysis module is finally obtained returns to the user; Analysis result is wrong if the user thinks emotion, and interface will pass to system management module to the correction result of user feedback allows the keeper examine.
Product review information is climbed the delivery piece, be used for climbing and getting having similar webpages with emotion tendency markup information such as star mark on the large-scale shopping website such as Jingdone district, Amazon at interval according to certain hour, extract product review information wherein and align negative information and put in order, connect with the foundation of product review information training data sample database by database interface, the formatted data of handling well is deposited in the training sample database.
The text pretreatment module, connect with the foundation of product review information training sample database by database interface, obtain the training sample data, and obtain the text data of user input from user interface, these text datas are carried out pre-service such as participle, POS mark (part-of-speech tagging), stop words processing and syntactic analysis, and the data of handling well are passed to the word load-on module.
The dictionary load-on module connects by database interface and dictionary database, obtains dictionary data such as emotion dictionary, matched combined dictionary, negative word dictionary, is used for the feature extraction of feature extraction module.
The feature extraction module, data after the dictionary data that loads of load-on module pair and the processing are carried out the extraction of pre-defined feature with the help of a dictionary, with text vectorization, be converted into the form that emotion analytical model training module can be handled, and pass to emotion analytical model training module.
Emotion analytical model training module is used at interval the emotion analytical model of native system core being trained by certain hour.Obtain from the feature extraction module and to be converted into the training data that requires form, use the Two-Level DCRF model training of L-BFGS algorithm to making up according to training data.The Two-Level DCRF model that the present invention uses is to develop in Linear CRF (linear conditions random field) model based, is a kind of in CRF (condition random field) model, is to use at the emotion analysis field first time.Method is in the past generally artificially independently got up this two parts work, ignored contact between the two, this model is by the structure double-layer structure, evaluation object and the identification of emotion word are carried out in a model simultaneously with the emotion tendency judgement is unified, realized the intercommunication of two workplace information, introduce contact details between the two, helped the raising of final precision.This module is given the emotion analysis module with the Model Transfer that trains.
The emotion analysis module loads the emotion analysis module that trains, and the user input text information after the format conversion is carried out fine-grained emotion analysis, namely draws the emotion tendency of specified evaluation object is judged.For example: " touch-screen is very cruel, and sound is also very clear, and just battery is not durable " the words is (touch-screen, front), (sound, front) and (battery, negative) with the emotion analysis result that obtains, and analysis result is offered the user by user interface.Simultaneously non-existent emotion word and matched combined in the dictionary of identification being passed to system management module examines for the keeper, the emotion word that identifies in the last example is (cruel, the front), (clear, front) and (is not durable, negative), the matched combined that identifies is (touch-screen, cruel, front), (sound, clear, positive) and (battery is not durable, and is negative).
Administrator interface is used for the system manager the new emotion word of emotion analysis module identification and the error analysis result of matched combined and user feedback is carried out the manual examination and verification affirmation.
System management module is used for the system manager and connects by database interface and database, more new database.If new emotion word or matched combined are correctly then deposit it in corresponding dictionary database, otherwise then give up; If the same revised feedback result of user correctly then will be deposited in the training sample database, otherwise then give up.
Database interface, unified interface and the access rights control of database manipulations such as the access of realization training sample data, dictionary data, renewal.
Product review information training sample database is used for the storage products review information and climbs the format training sample data that the delivery piece transmits.
Dictionary database is used for dictionary data such as storage emotion word dictionary, matched combined dictionary.
To sum up, anti-rubbish mail gateway of the present invention is climbed delivery piece, product review information training sample database, dictionary database, text pretreatment module, dictionary load-on module, feature extraction module, emotion analytical model training module, emotion analysis module, administrator interface, system management module and database interface etc. by user interface, product review information and is partly formed.Above-mentioned module is finished fine-grained emotion analysis, the collection of user profile feedback information, training text automated data acquiistion and dictionary database data in real time together and is upgraded this four functions.In fine-grained emotion analytic function, the Two-Level CRF model of emotion analytic system of the present invention by using training module to train, evaluation object in the text message of in the emotion analysis module user being imported, evaluation word and user identify the emotion tendency of each evaluation object and judge, judged result is offered the user by user interface; In the field feedback collecting function, the user passes to system management module with feedback information, and the keeper carries out depositing training sample data Kucheng in by database interface after the manual examination and verification to feedback information and is new learning sample in system management module; In training text automated data acquiistion function, emotion analytic system of the present invention is climbed the delivery piece by information the product review information of the similar marks such as band star mark on the network is collected, format is handled, and deposits in the training sample database by database interface; In dictionary database data in real time update functions, after the emotion word that the keeper does not include and matched combined are carried out manual examination and verification, deposit dictionary database in by database interface in the dictionary database of information management module to the output of emotion analysis module.
The present invention adopts has the algorithm of supervision that text emotion is carried out fine-grained analysis.And the algorithm that supervision is arranged needs a large amount of labeled data as training sample, and artificial mark need expend a large amount of manpowers and time and bring the subjective factor in the mark process to influence, and this also is to hinder the main cause that has supervise algorithm to use in reality.The review information that native system has star mark by automatic collection and extraction has reduced artificial intervention and cost, and can regularly effectively upgrade training data as corpus.
The present invention introduces feedback mechanism error analysis information is learnt.Existing method does not generally deal with for mistake branch result, but these feedback informations have comprised a large amount of useful informations, how can take full advantage of these information and become system to realize the key of self-teaching.The introducing of feedback mechanism makes model to learn again the result of error analysis, and the system that makes uses more accurate and more accurate.
Fig. 2 is the process flow diagram of the fine granularity emotion analytical approach of the present invention's proposition.
With reference to Fig. 2, this method comprises step: 1. product review information is climbed delivery piece response message and is climbed the request of getting, regularly obtain product review information and carry out information extraction from network, and obtain field feedback, store these information into the training sample sample database by connecting with the training sample data; 2. the response model train request connects with the training sample data, obtains training data, training data is carried out pre-service such as subordinate sentence, participle and part-of-speech tagging; 3. to pretreated data, carry out feature extraction by dictionary load-on module and feature extraction module, be converted to the vectorization data; 4. utilize after the feature extraction characteristic to the present invention propose the emotion analytical model---Two-Level CRF Model trains; 5. obtain the user and import product review information to be analyzed, and carry out pre-service and the feature extraction work identical with the 2-3 step; 6. load the emotion analytical model that trains characteristic is carried out the emotion analysis; 7. connect with user interface, the emotion analysis result is exported to the user.
In sum, this method has mainly comprised the collection of training sample data step, emotion analytical model training step and product review information emotion analytical procedure.
Fig. 3 collects the realization schematic diagram of step for the training sample data.With reference to Fig. 3, this step realizes obtaining the product review information training sample from network and these two sources of user feedback, and these sample datas are climbed by product review information respectively enters the product review information sample database after delivery piece and system management module are handled.In the sorting algorithm of supervision was arranged, training data had tremendous influence to the final effect of model, and the method that tradition manually marks training data needs a large amount of manpowers and time, and the feasibility in real world applications is not high.Therefore in this step, the system product review information is climbed the delivery piece to the mark of the band star on the network or is had review information that the good job assessment of bids annotates and climbs and get and extract on the one hand; The keeper carries out the correctness audit to user's feedback information in system management module on the other hand, and rational feedback information is stored as the training sample data.By the work of this two aspect, realize training data is collected comprehensively and effectively automatically.
Fig. 4 is the realization schematic diagram of emotion analytical model training step.With reference to Fig. 4, in this step, system at first extracts the training sample data in nearest a period of time from the training sample database, the pretreatment module of system, dictionary load-on module and feature extraction module are carried out the characteristic that a series of processing obtain vectorization to these training sample data then, and output to the training that the model training module is carried out the emotion analytical model.
The present invention carries out the analysis of fine-grained entity level emotion by the DCRF model of constructing a kind of Two-Level of having structure to review information first according to characteristics such as its composition structures after the large-tonnage product review information is analyzed.This model is the key point of fine granularity emotion analytical approach of the present invention, therefore, will introduce structure, principle and the advantage of Two-Level CRF Model below in detail.
Fine-grained entity level emotion analytical work purpose is to analyze in the text message emotion tendency at concrete object.Therefore just must relate to identification and the work of emotional orientation analysis two parts of entity.Entity level emotion analytical work is in the past regarded above-mentioned two parts work independently as usually, after namely the entity in the first distich is identified, the emotion of concrete entity is analyzed again, and has ignored contact between the two.Two-Level CRF not only can sentence between word structure carry out modeling, and two parts working relation is got up, carry out simultaneously, by the final associating precision of mutual raising of information between the two.
Two-Level CRF is a kind of special CRF model.CRF is a kind of non-directed graph model, and it carries out modeling to the conditional probability distribution of sequence mark on given characteristic set basis.Be example with the most basic Linear-CRF, under the condition of given observation sequence, the conditional probability of flag sequence can formalized description be following form:
Wherein, ψ
iBe the potential function in the non-directed graph model concept,
Be length be I the regularization factor under might flag sequence.Potential function ψ
iCan be decomposed into following form, wherein f
kFundamental function for definition.
Its corresponding graph model structure is example with the named entity recognition task as shown in Figure 5 here, imports pretreated text message, sets up its corresponding Linear-CRF model.Different with traditional sorting technique such as naive Bayesian, Logic Regression Models etc., Linear-CRF regards classification problem as feature that the sequence mark problem namely not only can utilize the traditional classification model to adopt, also by doing suitable Markov hypothesis, introduce different classes of position feature information, for example in this example, named entity occurs near the emotion word usually.And being traditional classification models, these contact details of different classes of are difficult to show.Linear-CRF directly carries out modeling to the conditional probability of flag sequence simultaneously, is different from digraph model such as HMM (latent horse model), and it does not need just can introduce abundant feature to doing independence assumption between feature; On the other hand, it also can regard the MEMM (maximum entropy Markov model) of overall regularization as, and has avoided the marking bias problem among the MEMM.Therefore, Linear-CRF no matter compare traditional classification model or digraph model, can both obtain better effect when solving the identification of sequence mark problem such as named entity.
Two-Level CRF can be regarded as the combination of two Linear-CRF.Shown in the graph model of Two-Level CRF among Fig. 6, its structure has comprised linear chain and the observation sequence of two marks, and the flag node of while in the different levels of identical time point interconnects.In Fig. 6 example, given one section pretreated product review example, we are node with the word, are launched into corresponding Two-Level CRF.Product entity and emotion word in the ground floor flag sequence distich are identified, namely to three kinds of mark T, S and O being arranged, difference representative products physical name, emotion word and other words.In second layer flag sequence, the emotion tendency of entity and the emotion tendency of emotion word are analyzed, namely to three kinds of mark P, N and O should be arranged, represented positive emotion, negative emotion respectively and do not have emotion.As can be seen, Two-Level CRF not only has the characteristics of Linear-CRF, also different markers work is merged, and introduces the contact details of different markers work, and this also has at present and is difficult in the method accomplish.
The formalized description of Two-Level CRF is as follows:
Wherein, Ψ
1Be illustrated in the potential function on the same flag sequence, φ
1Represent the potential function between two flags sequence.T represents the node number on the same flag sequence, the number of L expressive notation sequence, L=2 in model of the present invention.Same potential function can be expressed as following two kinds of forms, wherein f respectively
k(y
L, t, y
L, t+1, x, t) and f
k(y
L, t, y
L+1, t, x t) is respectively the fundamental function that is defined between same flag sequence and different flags sequence:
Ψ
l(y
l，t，y
l，t+1，x，t)＝exp{Σ
kλ
k*f
k(y
l，t，y
l，t+1，x，t)}
φ
l(y
l，t，y
l+1，t，x，t)＝exp{Σ
kλ
k*f
k(y
l，t，y
l+1，t，x，t)}
Contrast the formalized description of Linear-CRF and Two-Level CRF as can be seen, the difference that both ask is potential function φ among the Two-Level CRF
l(y
L, t, y
L+1, t, x, introducing t).Potential function φ
l(y
L, t, y
L+1, t, x t) is the formalized description that contacts between Entity recognition and two tasks of emotional orientation analysis.With Linear-CRF in that do Markov hypothesis with the mark on one deck diverse location similar, Two-Level CRF is on the basis of Linear-CRF, further the mark of identical timing node position in the different layers is done the Markov hypothesis, the not contact details between isolabeling on the different flags sequence have been introduced, according to maximum entropy criterion, be potential function φ with this contact details formalized description again
l(y
L, t, y
L+1, t, x, t), it is the weighting of fundamental function on the identical timing node for being defined in different layers.This is the key point of Two-Level CRF just, by introducing potential function φ
l(y
L, t, y
L+1, t, x t), has set up the information interaction between different flags sequence, will think separate in the former studies, not have two flags sequence of information interaction effectively to link together.Therefore, Two-Level CRF not only has the advantage of above-mentioned Linear-CRF, also further introduced the feature between the different flags sequence that enrich more, ignored by former studies or be difficult to utilize on its basis, especially the problem that needs to carry out twice sequence mark in the sequence mark problem, for example in the emotion problem analysis of the fine granularityization of the present invention's solution, finally can obtain the better effect than Linear-CRF.
In the model training module, we use the Two-Level CRF model training of L-BFGS algorithm to launching, the parameter lambda in the learning model
k
Fig. 7 is the realization schematic diagram of product review information emotion analytical procedure.With reference to Fig. 7, in this step, user interface transmits the product review information of user's input, through pretreatment module, dictionary load-on module and feature extraction resume module, the characteristic of output vectorization is to the emotion analysis module, call the emotion analytical model that trains in the emotion analysis module entity in the data and emotion are analyzed, and the result is offered the user by user interface.In this module, we use the TRP algorithm that the mark in the flag sequence is derived.The new emotion word of identifying in emotion word identifying will be examined the back by database interface the system management module keeper, deposit in the dictionary database, guarantee the real time automatic update of dictionary resources.
Above-described specific embodiment; purpose of the present invention, technical scheme and beneficial effect are further described; institute is understood that; the above only is specific embodiments of the invention; be not limited to the present invention; within the spirit and principles in the present invention all, any modification of making, be equal to replacement, improvement etc., all should be included within protection scope of the present invention.
Claims (8)
1. fine granularity emotion analytic system at product review information, it comprises:
User interface is used for the mutual of system and user, and the user can submit the product review information set to by this module;
Product review information training sample database is used for the various product review information training samples that marked of storage, and regularly upgrades;
Relevant dictionary load-on modules such as emotion dictionary are used for loading corresponding dictionary resources in feature extraction or preprocessing process;
The text pretreatment module is used for text message and the training sample data of user's input are carried out pre-service work such as subordinate sentence, participle, part-of-speech tagging and syntactic analysis;
The feature extraction module is used for pretreated text message is carried out feature extraction, with the text message vectorization;
Emotion analytical model training module is used for according to the existing corpus training of database emotion analytical model;
The emotion tendency judge module carries out emotion tendency in the emotion analytical model that the input of the text message of vectorization has been trained and judges;
Feedback module is used for the emotion analysis result is fed back to the user, and the user can revise judged result simultaneously, and deposits revised result in the training sample database.
2. the fine granularity emotion analytic system at product review information as claimed in claim 1 is characterized in that realizing simultaneously a stage identification and the emotional orientation analysis of entity, and sets up the relation between entity and the fine granularity emotion tendency.
3. the fine granularity emotion analytic system at product review information as claimed in claim 1 is characterized in that, by the DCRF model of constructing a kind of Two-Level of having structure review information is carried out the analysis of fine-grained entity level emotion.
4. Two-Level CRF as claimed in claim 3 is characterized in that its structure has comprised linear chain and the observation sequence of two marks, and the flag node of while in the different levels of identical time point interconnects.Wherein the product entity in the ground floor flag sequence distich and emotion word are identified, namely to three kinds of mark T, S and O being arranged, difference representative products physical name, emotion word and other words.In second layer flag sequence, the emotion tendency of entity and the emotion tendency of emotion word are analyzed, namely to three kinds of mark P, N and O should be arranged, represented positive emotion, negative emotion respectively and do not have emotion.
5. Two-Level CRF as claimed in claim 3 is characterized in that its formalized description is as follows:
Wherein, Ψ
lBe illustrated in the potential function on the same flag sequence, φ
lRepresent the potential function between two flags sequence., f wherein
k(y
L, t, y
L, t+1, x, t) and f
k(y
L, t, y
L+1, t, x t) is respectively the fundamental function that is defined between same flag sequence and different flags sequence:
Ψ
l(y
l，t，y
l，t+1，x，t)＝exp{Σ
kλ
k*f
k(y
l，t，y
l，t+1，x，t)}
φ
l(y
l，t，y
l+1，t，x，t)＝exp{Σ
kλ
k*f
k(y
l，t，y
l+1，t，x，t)}
6. fine granularity emotion analytical approach at product review information, the method comprising the steps of:
1) climbs the product review information of getting band star mark on the network, reduce manually mark;
2) adopting sample data, is the minimum granularity of analyzing with the word, has trained a kind of DCRF model of the Two-Level of having structure, carries out Entity recognition and emotion tendency simultaneously and judges two tasks
3) product review adopts the DCRF model of the Two-Level structure that trains to realize that entity information is identified and the emotion tendency is judged through carrying out subordinate sentence, participle, part-of-speech tagging and syntactic analysis
4) introduce feedback mechanism, model can be learnt wrongheaded sample data.
7. method as claimed in claim 6 is characterized in that analyzing the DCRF model that granularity is trained the Two-Level structure with the minimum of word, comprises in the training process to carry out two tasks of Entity recognition and emotion tendency judgement simultaneously.
8. method as claimed in claim 7 is characterized in that, at the data of carrying out after subordinate sentence, participle, part-of-speech tagging and the syntactic analysis, uses the DCRF model of the Two-Level structure that trains to realize that entity information identification and emotion tendency judge.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310036034.1A CN103207855B (en) | 2013-04-12 | 2013-04-12 | For the fine granularity sentiment analysis system and method for product review information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310036034.1A CN103207855B (en) | 2013-04-12 | 2013-04-12 | For the fine granularity sentiment analysis system and method for product review information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103207855A true CN103207855A (en) | 2013-07-17 |
| CN103207855B CN103207855B (en) | 2019-04-26 |
Family
ID=48755080
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310036034.1A Active CN103207855B (en) | 2013-04-12 | 2013-04-12 | For the fine granularity sentiment analysis system and method for product review information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103207855B (en) |
Cited By (46)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103559174A (en) * | 2013-09-30 | 2014-02-05 | 东软集团股份有限公司 | Semantic emotion classification characteristic value extraction method and system |
| CN104268197A (en) * | 2013-09-22 | 2015-01-07 | 中科嘉速(北京)并行软件有限公司 | Industry comment data fine grain sentiment analysis method |
| CN104484437A (en) * | 2014-12-24 | 2015-04-01 | 福建师范大学 | Network brief comment sentiment mining method |
| CN104765733A (en) * | 2014-01-02 | 2015-07-08 | 华为技术有限公司 | Method and device for analyzing social network event |
| CN105005560A (en) * | 2015-08-26 | 2015-10-28 | 苏州大学张家港工业技术研究院 | Maximum entropy model-based evaluation type emotion sorting method and system |
| CN105787461A (en) * | 2016-03-15 | 2016-07-20 | 浙江大学 | Text-classification-and-condition-random-field-based adverse reaction entity identification method in traditional Chinese medicine literature |
| CN105868185A (en) * | 2016-05-16 | 2016-08-17 | 南京邮电大学 | Part-of-speech-tagging-based dictionary construction method applied in shopping comment emotion analysis |
| CN105930503A (en) * | 2016-05-09 | 2016-09-07 | 清华大学 | Combination feature vector and deep learning based sentiment classification method and device |
| TWI553573B (en) * | 2014-05-15 | 2016-10-11 | 財團法人工業技術研究院 | Aspect-sentiment analysis and viewing system, device therewith and method therefor |
| CN106021391A (en) * | 2016-05-11 | 2016-10-12 | 广东工业大学 | Product comment information real-time collection method based on Storm |
| CN106062809A (en) * | 2014-03-10 | 2016-10-26 | Kddi株式会社 | Device, program, and method for analyzing transition in psychological state of poster on basis of comment text |
| CN106127507A (en) * | 2016-06-13 | 2016-11-16 | 四川长虹电器股份有限公司 | A kind of commodity the analysis of public opinion method and system based on user's evaluation information |
| CN106649270A (en) * | 2016-12-19 | 2017-05-10 | 四川长虹电器股份有限公司 | Public opinion monitoring and analyzing method |
| CN106874363A (en) * | 2016-12-30 | 2017-06-20 | 北京光年无限科技有限公司 | The multi-modal output intent and device of intelligent robot |
| CN107038609A (en) * | 2017-04-24 | 2017-08-11 | 广州华企联信息科技有限公司 | A kind of Method of Commodity Recommendation and system based on deep learning |
| CN107066449A (en) * | 2017-05-09 | 2017-08-18 | 北京京东尚科信息技术有限公司 | Information-pushing method and device |
| CN107102980A (en) * | 2016-02-19 | 2017-08-29 | 北京国双科技有限公司 | The extracting method and device of emotion information |
| CN107451116A (en) * | 2017-07-14 | 2017-12-08 | 中国地质大学（武汉） | Raw big data statistical analysis technique in a kind of Mobile solution |
| CN107562816A (en) * | 2017-08-16 | 2018-01-09 | 深圳狗尾草智能科技有限公司 | User view automatic identifying method and device |
| CN107895027A (en) * | 2017-11-17 | 2018-04-10 | 合肥工业大学 | Individual feelings and emotions knowledge mapping method for building up and device |
| CN107918633A (en) * | 2017-03-23 | 2018-04-17 | 广州思涵信息科技有限公司 | Sensitive public sentiment content identification method and early warning system based on semantic analysis technology |
| CN108073703A (en) * | 2017-12-14 | 2018-05-25 | 郑州云海信息技术有限公司 | A kind of comment information acquisition methods, device, equipment and storage medium |
| CN108549692A (en) * | 2018-04-13 | 2018-09-18 | 重庆邮电大学 | The method that sparse multivariate logistic regression model under Spark frames classifies to text emotion |
| CN108763210A (en) * | 2018-05-22 | 2018-11-06 | 华中科技大学 | A kind of sentiment analysis and forecasting system based on automated data collection |
| CN109214008A (en) * | 2018-09-28 | 2019-01-15 | 珠海中科先进技术研究院有限公司 | A kind of sentiment analysis method and system based on keyword extraction |
| CN109460940A (en) * | 2018-11-26 | 2019-03-12 | 北京香侬慧语科技有限责任公司 | A kind of method for early warning and device based on sentiment analysis |
| CN109857837A (en) * | 2019-01-16 | 2019-06-07 | 苏宁易购集团股份有限公司 | A kind of dictionary loading method and device that can customize |
| CN110069625A (en) * | 2017-09-22 | 2019-07-30 | 腾讯科技（深圳）有限公司 | A kind of content categorizing method, device and server |
| CN110134765A (en) * | 2019-05-05 | 2019-08-16 | 杭州师范大学 | A kind of dining room user comment analysis system and method based on sentiment analysis |
| CN110309959A (en) * | 2019-06-19 | 2019-10-08 | 广州市高速公路有限公司营运分公司 | A kind of emergency event processing method, system and storage medium |
| CN110413773A (en) * | 2019-06-20 | 2019-11-05 | 平安科技（深圳）有限公司 | Intelligent text classification method, device and computer readable storage medium |
| CN110717325A (en) * | 2019-09-04 | 2020-01-21 | 北京三快在线科技有限公司 | Text emotion analysis method and device, electronic equipment and storage medium |
| CN110766435A (en) * | 2018-12-19 | 2020-02-07 | 北京嘀嘀无限科技发展有限公司 | Vector training method and device, electronic equipment and computer readable storage medium |
| CN110990565A (en) * | 2019-11-20 | 2020-04-10 | 广州商品清算中心股份有限公司 | Extensible text analysis system and method for public sentiment analysis |
| CN111091000A (en) * | 2019-12-24 | 2020-05-01 | 深圳视界信息技术有限公司 | Processing system and method for extracting user fine-grained typical opinion data |
| CN111143559A (en) * | 2019-12-24 | 2020-05-12 | 北京明略软件***有限公司 | Triple-based word cloud display method and device |
| CN111241842A (en) * | 2018-11-27 | 2020-06-05 | 阿里巴巴集团控股有限公司 | Text analysis method, device and system |
| CN111339253A (en) * | 2020-02-25 | 2020-06-26 | 中国建设银行股份有限公司 | Method and device for extracting article information |
| US10824812B2 (en) | 2016-06-07 | 2020-11-03 | International Business Machines Corporation | Method and apparatus for informative training repository building in sentiment analysis model learning and customization |
| CN112069311A (en) * | 2020-08-04 | 2020-12-11 | 北京声智科技有限公司 | Text extraction method, device, equipment and medium |
| CN112364605A (en) * | 2020-11-27 | 2021-02-12 | 智业软件股份有限公司 | Text labeling method based on double-array Trie, terminal equipment and storage medium |
| CN112667886A (en) * | 2020-12-02 | 2021-04-16 | 浙江学海教育科技有限公司 | Method, device, equipment and medium for detecting improper comments |
| CN113343714A (en) * | 2021-07-02 | 2021-09-03 | 马上消费金融股份有限公司 | Information extraction method, model training method and related equipment |
| CN113627969A (en) * | 2021-06-21 | 2021-11-09 | 杭州盟码科技有限公司 | Product problem analysis method and system based on E-commerce platform user comments |
| CN114510555A (en) * | 2022-02-24 | 2022-05-17 | 平安普惠企业管理有限公司 | Method and device for making business strategy and related equipment |
| WO2022267454A1 (en) * | 2021-06-24 | 2022-12-29 | 平安科技（深圳）有限公司 | Method and apparatus for analyzing text, device and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7792841B2 (en) * | 2006-05-30 | 2010-09-07 | Microsoft Corporation | Extraction and summarization of sentiment information |
| CN101876985A (en) * | 2009-11-26 | 2010-11-03 | 西北工业大学 | WEB text sentiment theme recognizing method based on mixed model |
| CN101882136A (en) * | 2009-05-08 | 2010-11-10 | 中国科学院计算技术研究所 | Method for analyzing emotion tendentiousness of text |
| CN102663046A (en) * | 2012-03-29 | 2012-09-12 | 中国科学院自动化研究所 | Sentiment analysis method oriented to micro-blog short text |
-
2013
- 2013-04-12 CN CN201310036034.1A patent/CN103207855B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7792841B2 (en) * | 2006-05-30 | 2010-09-07 | Microsoft Corporation | Extraction and summarization of sentiment information |
| CN101882136A (en) * | 2009-05-08 | 2010-11-10 | 中国科学院计算技术研究所 | Method for analyzing emotion tendentiousness of text |
| CN101876985A (en) * | 2009-11-26 | 2010-11-03 | 西北工业大学 | WEB text sentiment theme recognizing method based on mixed model |
| CN102663046A (en) * | 2012-03-29 | 2012-09-12 | 中国科学院自动化研究所 | Sentiment analysis method oriented to micro-blog short text |
Non-Patent Citations (4)
| Title |
|---|
| CHARLES SUTTON等: "Dynamic Conditional Random Fields: Factorized Probabilistic Models for Labeling and Segmenting Sequence Data", 《PROCEEDING ICML’04 PROCEEDINGS OF THE TWENTY-FIRST INTERNATIONAL CONFERENCE ON MACHINE LEARNING》 * |
| 刘宁: "客户评价挖掘算法研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 史鹏治: "基于CRFs的产品评论情感分类", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 来亮等: "文本情感分析综述", 《计算机光盘软件与应用》 * |
Cited By (61)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104268197A (en) * | 2013-09-22 | 2015-01-07 | 中科嘉速(北京)并行软件有限公司 | Industry comment data fine grain sentiment analysis method |
| CN104268197B (en) * | 2013-09-22 | 2017-11-07 | 中科嘉速(北京)并行软件有限公司 | A kind of industry comment data fine granularity sentiment analysis method |
| CN103559174A (en) * | 2013-09-30 | 2014-02-05 | 东软集团股份有限公司 | Semantic emotion classification characteristic value extraction method and system |
| CN103559174B (en) * | 2013-09-30 | 2016-03-09 | 东软集团股份有限公司 | Semantic emotion classification characteristic value extraction and system |
| CN104765733A (en) * | 2014-01-02 | 2015-07-08 | 华为技术有限公司 | Method and device for analyzing social network event |
| CN104765733B (en) * | 2014-01-02 | 2018-06-15 | 华为技术有限公司 | A kind of method and apparatus of social networks event analysis |
| CN106062809A (en) * | 2014-03-10 | 2016-10-26 | Kddi株式会社 | Device, program, and method for analyzing transition in psychological state of poster on basis of comment text |
| TWI553573B (en) * | 2014-05-15 | 2016-10-11 | 財團法人工業技術研究院 | Aspect-sentiment analysis and viewing system, device therewith and method therefor |
| CN104484437B (en) * | 2014-12-24 | 2018-07-20 | 福建师范大学 | A kind of network short commentary emotion method for digging |
| CN104484437A (en) * | 2014-12-24 | 2015-04-01 | 福建师范大学 | Network brief comment sentiment mining method |
| CN105005560A (en) * | 2015-08-26 | 2015-10-28 | 苏州大学张家港工业技术研究院 | Maximum entropy model-based evaluation type emotion sorting method and system |
| CN107102980A (en) * | 2016-02-19 | 2017-08-29 | 北京国双科技有限公司 | The extracting method and device of emotion information |
| CN105787461A (en) * | 2016-03-15 | 2016-07-20 | 浙江大学 | Text-classification-and-condition-random-field-based adverse reaction entity identification method in traditional Chinese medicine literature |
| CN105787461B (en) * | 2016-03-15 | 2019-07-23 | 浙江大学 | Document adverse reaction entity recognition method based on text classification and condition random field |
| CN105930503A (en) * | 2016-05-09 | 2016-09-07 | 清华大学 | Combination feature vector and deep learning based sentiment classification method and device |
| CN106021391A (en) * | 2016-05-11 | 2016-10-12 | 广东工业大学 | Product comment information real-time collection method based on Storm |
| CN106021391B (en) * | 2016-05-11 | 2019-06-21 | 广东工业大学 | Product review information real-time collecting method based on Storm |
| CN105868185A (en) * | 2016-05-16 | 2016-08-17 | 南京邮电大学 | Part-of-speech-tagging-based dictionary construction method applied in shopping comment emotion analysis |
| US10824812B2 (en) | 2016-06-07 | 2020-11-03 | International Business Machines Corporation | Method and apparatus for informative training repository building in sentiment analysis model learning and customization |
| CN106127507A (en) * | 2016-06-13 | 2016-11-16 | 四川长虹电器股份有限公司 | A kind of commodity the analysis of public opinion method and system based on user's evaluation information |
| CN106649270A (en) * | 2016-12-19 | 2017-05-10 | 四川长虹电器股份有限公司 | Public opinion monitoring and analyzing method |
| CN106874363A (en) * | 2016-12-30 | 2017-06-20 | 北京光年无限科技有限公司 | The multi-modal output intent and device of intelligent robot |
| CN107918633A (en) * | 2017-03-23 | 2018-04-17 | 广州思涵信息科技有限公司 | Sensitive public sentiment content identification method and early warning system based on semantic analysis technology |
| CN107918633B (en) * | 2017-03-23 | 2021-07-02 | 广州思涵信息科技有限公司 | Sensitive public opinion content identification method and early warning system based on semantic analysis technology |
| CN107038609A (en) * | 2017-04-24 | 2017-08-11 | 广州华企联信息科技有限公司 | A kind of Method of Commodity Recommendation and system based on deep learning |
| CN107066449A (en) * | 2017-05-09 | 2017-08-18 | 北京京东尚科信息技术有限公司 | Information-pushing method and device |
| CN107451116A (en) * | 2017-07-14 | 2017-12-08 | 中国地质大学（武汉） | Raw big data statistical analysis technique in a kind of Mobile solution |
| CN107451116B (en) * | 2017-07-14 | 2020-05-22 | 中国地质大学（武汉） | Statistical analysis method for mobile application endogenous big data |
| CN107562816A (en) * | 2017-08-16 | 2018-01-09 | 深圳狗尾草智能科技有限公司 | User view automatic identifying method and device |
| CN110069625B (en) * | 2017-09-22 | 2022-09-23 | 腾讯科技（深圳）有限公司 | Content classification method and device and server |
| CN110069625A (en) * | 2017-09-22 | 2019-07-30 | 腾讯科技（深圳）有限公司 | A kind of content categorizing method, device and server |
| CN107895027A (en) * | 2017-11-17 | 2018-04-10 | 合肥工业大学 | Individual feelings and emotions knowledge mapping method for building up and device |
| CN108073703A (en) * | 2017-12-14 | 2018-05-25 | 郑州云海信息技术有限公司 | A kind of comment information acquisition methods, device, equipment and storage medium |
| CN108549692A (en) * | 2018-04-13 | 2018-09-18 | 重庆邮电大学 | The method that sparse multivariate logistic regression model under Spark frames classifies to text emotion |
| CN108549692B (en) * | 2018-04-13 | 2021-05-11 | 重庆邮电大学 | Method for classifying text emotion through sparse multiple logistic regression model under Spark framework |
| CN108763210A (en) * | 2018-05-22 | 2018-11-06 | 华中科技大学 | A kind of sentiment analysis and forecasting system based on automated data collection |
| CN109214008A (en) * | 2018-09-28 | 2019-01-15 | 珠海中科先进技术研究院有限公司 | A kind of sentiment analysis method and system based on keyword extraction |
| CN109460940A (en) * | 2018-11-26 | 2019-03-12 | 北京香侬慧语科技有限责任公司 | A kind of method for early warning and device based on sentiment analysis |
| CN111241842B (en) * | 2018-11-27 | 2023-05-30 | 阿里巴巴集团控股有限公司 | Text analysis method, device and system |
| CN111241842A (en) * | 2018-11-27 | 2020-06-05 | 阿里巴巴集团控股有限公司 | Text analysis method, device and system |
| CN110766435A (en) * | 2018-12-19 | 2020-02-07 | 北京嘀嘀无限科技发展有限公司 | Vector training method and device, electronic equipment and computer readable storage medium |
| CN109857837A (en) * | 2019-01-16 | 2019-06-07 | 苏宁易购集团股份有限公司 | A kind of dictionary loading method and device that can customize |
| CN110134765A (en) * | 2019-05-05 | 2019-08-16 | 杭州师范大学 | A kind of dining room user comment analysis system and method based on sentiment analysis |
| CN110309959A (en) * | 2019-06-19 | 2019-10-08 | 广州市高速公路有限公司营运分公司 | A kind of emergency event processing method, system and storage medium |
| CN110413773B (en) * | 2019-06-20 | 2023-09-22 | 平安科技（深圳）有限公司 | Intelligent text classification method, device and computer readable storage medium |
| CN110413773A (en) * | 2019-06-20 | 2019-11-05 | 平安科技（深圳）有限公司 | Intelligent text classification method, device and computer readable storage medium |
| CN110717325A (en) * | 2019-09-04 | 2020-01-21 | 北京三快在线科技有限公司 | Text emotion analysis method and device, electronic equipment and storage medium |
| CN110990565B (en) * | 2019-11-20 | 2023-12-08 | 广州商品清算中心股份有限公司 | Extensible text analysis system and method for public opinion analysis |
| CN110990565A (en) * | 2019-11-20 | 2020-04-10 | 广州商品清算中心股份有限公司 | Extensible text analysis system and method for public sentiment analysis |
| CN111143559A (en) * | 2019-12-24 | 2020-05-12 | 北京明略软件***有限公司 | Triple-based word cloud display method and device |
| CN111091000A (en) * | 2019-12-24 | 2020-05-01 | 深圳视界信息技术有限公司 | Processing system and method for extracting user fine-grained typical opinion data |
| CN111339253A (en) * | 2020-02-25 | 2020-06-26 | 中国建设银行股份有限公司 | Method and device for extracting article information |
| CN112069311A (en) * | 2020-08-04 | 2020-12-11 | 北京声智科技有限公司 | Text extraction method, device, equipment and medium |
| CN112069311B (en) * | 2020-08-04 | 2024-06-11 | 北京声智科技有限公司 | Text extraction method, device, equipment and medium |
| CN112364605A (en) * | 2020-11-27 | 2021-02-12 | 智业软件股份有限公司 | Text labeling method based on double-array Trie, terminal equipment and storage medium |
| CN112667886A (en) * | 2020-12-02 | 2021-04-16 | 浙江学海教育科技有限公司 | Method, device, equipment and medium for detecting improper comments |
| CN113627969A (en) * | 2021-06-21 | 2021-11-09 | 杭州盟码科技有限公司 | Product problem analysis method and system based on E-commerce platform user comments |
| WO2022267454A1 (en) * | 2021-06-24 | 2022-12-29 | 平安科技（深圳）有限公司 | Method and apparatus for analyzing text, device and storage medium |
| CN113343714B (en) * | 2021-07-02 | 2022-06-07 | 马上消费金融股份有限公司 | Information extraction method, model training method and related equipment |
| CN113343714A (en) * | 2021-07-02 | 2021-09-03 | 马上消费金融股份有限公司 | Information extraction method, model training method and related equipment |
| CN114510555A (en) * | 2022-02-24 | 2022-05-17 | 平安普惠企业管理有限公司 | Method and device for making business strategy and related equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103207855B (en) | 2019-04-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103207855A (en) | Fine-grained sentiment analysis system and method specific to product comment information | |
| CN111737495B (en) | Middle-high-end talent intelligent recommendation system and method based on domain self-classification | |
| CN110633409B (en) | Automobile news event extraction method integrating rules and deep learning | |
| CN112417880B (en) | Automatic case information extraction method for court electronic files | |
| CN109446331B (en) | Text emotion classification model establishing method and text emotion classification method | |
| CN104391942B (en) | Short essay eigen extended method based on semantic collection of illustrative plates | |
| CN111325029B (en) | Text similarity calculation method based on deep learning integrated model | |
| CN109684440A (en) | Address method for measuring similarity based on level mark | |
| CN103440287B (en) | A kind of Web question and answer searching system based on product information structure | |
| CN108182295A (en) | A kind of Company Knowledge collection of illustrative plates attribute extraction method and system | |
| CN109543034B (en) | Text clustering method and device based on knowledge graph and readable storage medium | |
| CN107220237A (en) | A kind of method of business entity's Relation extraction based on convolutional neural networks | |
| CN103176963B (en) | Chinese sentence meaning structure model automatic labeling method based on CRF ++ | |
| CN112149421A (en) | Software programming field entity identification method based on BERT embedding | |
| CN104809176A (en) | Entity relationship extracting method of Zang language | |
| CN106599032A (en) | Text event extraction method in combination of sparse coding and structural perceptron | |
| CN102314417A (en) | Method for identifying Web named entity based on statistical model | |
| WO2020010834A1 (en) | Faq question and answer library generalization method, apparatus, and device | |
| CN104484380A (en) | Personalized search method and personalized search device | |
| CN111177322A (en) | Ontology model construction method of domain knowledge graph | |
| CN112199512B (en) | Scientific and technological service-oriented case map construction method, device, equipment and storage medium | |
| CN110532398A (en) | Family's map method for auto constructing based on multitask united NNs model | |
| CN114462556B (en) | Enterprise association industry chain classification method, training method, device, equipment and medium | |
| Miao et al. | A dynamic financial knowledge graph based on reinforcement learning and transfer learning | |
| CN111710428A (en) | Biomedical text representation method for modeling global and local context interaction |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |