CN106547866B - Fine-grained sentiment classification method based on random emotional word co-occurrence networks - Google Patents

Abstract

Based on a random fine-grained emotional sentiment classification word co-occurrence network, random network theory, the use of the phenomenon of co-occurrence words, after the label body emotional vocabulary lexicon, network model is formed based on a random sequence of words to a built emotion feature that emotional word co-occurrence network model, model reduction based on this, the longest term emotional matching method and algorithm combined with TC were SWLM-TC ​​classification unsupervised learning, emotional or further word longest matching method and HMM machine emotional learning algorithm in conjunction with the establishment of a fine-grained classification model and use the model to classify prediction; the present invention may be fine-grained text paragraph level sentiment classification, improve the accuracy of TC simple algorithm, so that more accurate classification using a sample set SWLM-TC and the test sample library emotion category followed by HMM model training, improve the automation of simple machine learning algorithms.

Description

Fine-grained sentiment classification method based on random emotional word co-occurrence networks

FIELD

[0001] The present invention belongs to the technical field of information retrieval, especially relates to fine-grained sentiment classification method based on random emotional word co-occurrence network.

Background technique

[0002] In recent years, with the rapid development of economy and information technology, the Internet profound impact on the development of social forms, and had a tremendous economic boost, the Internet population had a broad array of information, the mobile Internet speed up landing process, the popularity of a variety of smart mobile devices, so that information at a lower cost, faster spread on the Internet, different types of information will have different effects, negative comments will have a negative impact on Internet users, malignant community news and public events occurs, not only will affect the feelings of the individual, and even produce huge economic losses, emotional information mining became a pressing issue. Building emotional text corpus, currently available include Corpus Corpus Pang, Whissel 1 corpus, Berardinel I i Corpus movie reviews, product reviews Corpus, corpus annotation and emotional aspects of Chinese resources are relatively small, Tsinghua University marked some tourist attractions emotional corpus described, to assist speech synthesis, but the scale is relatively small. On the international Blog, forums and news commentary related text called the new text, the new text on the network such as our sentiment analysis provides data sources, analysis and processing of the new text is becoming a hot topic of current research . In information technology today, the network has become a part of life, sentiment analysis to become Internet users understand the true idea of ​​the important reference in the event of a public emergency management, the use of new text on the network users emotions become a new research direction.

[0003] For the current tendency has been more in-depth study of the text, product reviews and critics in more successful, given the complexity of the language, the expression of individual differences, and there is no systematic description of the formation of human emotion, emotion present, for fine-grained the analysis also few Chinese in the course of evolution, the free grammar, vocabulary, large, relatively free form, and many other causes of the differences in English and sentiment analysis, semantic analysis is often used in English in great difficulty in analysis of Chinese, It caused many difficulties. Analysis of the relationship between emotional and psychological inextricably linked, psychological studies have found that the relationship between words and human emotions are measurable, semantic orientation independent of words or phrases of human emotions is important to convey. Studies have shown that semantic orientation words and phrases are mainly two phenomena: 1) the same term emotional tendencies often occur simultaneously; 2) The term emotional opposite tendencies generally do not occur simultaneously. Because of these two phenomena, sentiment analysis can be simplified a lot of things, studies have shown that common words in English and Chinese text established co-occurrence networks to meet small world, and carried out on the basis of this network on text segmentation and topic Research extraction areas, studies have random network model for text analysis topics, studies the tendency for random network model analysis, but the random network theory is applied to fine-grained text sentiment analysis, the research currently no reported.

SUMMARY

[0004] In order to overcome the drawbacks of the prior art, an object of the present invention is to provide a method of fine-grained sentiment classification sentiment random word co-occurrence-based network, the longest match sentiment words SWLM (Sentimental Word Longest Match) and Machine Learning algorithm combine to provide the paragraph level of fine-grained text sentiment classification.

[0005] To achieve the above object, the technical solution adopted by the invention is:

[0006] A fine-grained random word sentiment sentiment classification method based on co-occurrence network, random network theory, the use of the phenomenon of co-occurrence words, after the label body emotional vocabulary lexicon, to form a construct based on emotion feature word sequence random network model, that is, the emotional word co-occurrence network model, model reduction based on this, the longest term emotional matching method (SWLM, Sentimental Word longest match) and TC algorithm combines were SWLM-TC ​​classification unsupervised learning, or further emotional word HMM longest matching method and machine learning algorithms combine to establish fine-grained sentiment classification model and use the model to classify forecast.

[0007] The word co-occurrence sentiment network model was constructed as follows:

[0008] D to obtain a set of ordered sentence Sl4 & amp each text clause perform operations; 4 ... 4Sn;

[0009] 2) for each of the 51 word sentence, as well as stop words filtered meaningless content words, ontology vocabulary using emotional stimulation was denoted T emotional words, to obtain an ordered set of emotional words Wl ^ W2 --- - ^ Wn;

[0010] 3) for each sentence, using WL (Word Long, window length, and generally 2) bit word sliding window extracted from sentences to <Wi, wj>, if

: To add a new node W Wi, Wi is the weight and the initial weight is set nwi 1; 1 nwi plus, if

> Add a new edge (Wi, Wj), and as (Wi, Wj) weight nwi, wj set to an initial value of E 1; otherwise Ilwi, wj plus 1;

After [0011] 4) all text processing is complete, the network setup complete 1¾ type G;

[0012] where, S represents a sequence consisting of a plurality of sentences, W represents the extracted emotional words, we Σ, Σ is the set of Chinese vocabulary, Chinese vocabulary set is to remove stop words, nonsense words emotional content words and then through the body emotional body sets the word library labeled; W is the set of nodes of the network model G, W = IwiI ie [1, N]}, N is the number of the node G; E is the set of edges G network model, the network model G the number of side is M, E = {(wi, wj) | between wi, WjGW, Wi and Wj order and co-occurrence relation}, (wi, wj) represents from node points to the node Wi Wj directed edge; Nw is the weight of the network model G of nodes weight, Nw = {nwi I wiGW}; NE is the weight of the network model G of the edge weight represents the weight of the edges between nodes Wi and Wj weight, NE = {nWi, Wj I (wi, wj ) eE}.

[0013] The emotional vocabulary ontology database, is divided into seven categories 21 emotion subclasses, sentiment classification are happy music {(PA), Confidence (PE)}, {distinguished good (the PD), commends (PH) I believe (PG), like (PB), wish (PK)}, anger {anger (NA)}, sorrow {sadness (NB), down (NJ), guilt (NH), Si (PE)}, fear { : panic (NI), fear (NC), shame (NG)}, evil {bored (NE), hate (ND), derogatory responsible (NN), jealous (NK), suspected of (NL)}, panic {surprise ( PC)}; emotional intensity power into five files 1,3,5,7,9, 9 denotes the maximum intensity, minimum intensity represents 1, the type of emotion vocabulary speech body divided into a total of seven categories, namely, a noun (noun) verb (verb), adjectives (adj), adverb (adv), network words (nw), idioms (idiom), prepositional phrases (pr India), containing a total of 27466 sentiment words.

[0014] The network model in accordance with Le G, good, anger, sadness, fear, evil, shock seven kinds of emotions into 7 sub-networks, sub-networks resolution process, breakage occurs if the case, the use of that node with the highest weight fracture network sub-block, to construct a fine-grained seven sub-network can be used to calculate Gx IX = 0

[0015] The term emotional longest matching method emotional vocabulary words maximum weight for the longest match, and so do not use the disambiguation process for preventing noise, classification accuracy can be related to the emotional topic below, and by seven small classification model weight calculation, machine learning parameters can be obtained classification.

[0016] When performing the classification, have the following definitions:

[0017] The maximum weight matching path length (1_ "5): 614 = {1,2,3,4,5,6,7} network, if two sequential cover emotional words, directly connected to the edge is used to match the , if the two words present emotion interval network Gx network, the path is selected when the selection by matching node with the largest weight, that is, the length S is calculated as follows:

[0018]

[0019] wherein dmax (wi, wi + x) is the i-th word network to the maximum weight of the weight of the word i + x matching path;

[0020] emotional weight coefficient SW (Sentimetal weight): G network, the sentiment polarity proportion of their share of the seven sub-network, using this factor will become more apparent classification, reducing classification problems caused because of blurred, so emotional word network number of words is reproduced freq, strength of polarity is P, is calculated as follows:

[0024] WC is the sub-network where the value of each word Emotion, Wy of the emotional value of the sub-network, the SWx value for the SW sub-network X, i.e. emotional weight coefficient;

[0025] Category coefficient CC (Classification coefficient): After determining the maximum matching word path, Re and emotional intensity of the reproducing power of words on this path, assuming η word is calculated as follows:

[0028] where ^ 0 is a single word classification coefficient;

[0029] class prediction coefficient CPC (Classification prediction coefficient): prediction mechanism when using machine learning algorithms to classify, for the classification can not determine the sample taken; sorted by the SWx, if SW1 + SW2> 80%, SWi / SW2> 1.5 , under the classified SWi, if SWi + SW2> 80%, SWi / SW2 <= 1.5, in this case fall under sWdP SW2 two attributes; if sWi + swxso%, indicates the classification of the article comparison complex, classified under the corresponding coefficients classified according to the classification:

[0031] SWLM-TC ​​said method comprising the steps of:

[0032] 1) the classification of the articles were needed clause, sentence sequence divided into ...

[0033] 2) for each sentence word order, and removing the meaningless content words, auxiliary, and using the emotional vocabulary lexicon tagging body, selected words are marked, in order, i.e., ...

[0034] 3) The home network search corresponding marked words;

[0035] 4) of the word network path selection, if two adjacent words, then using the path directly connected; if two non-adjacent words, they are connected to the path is selected, the maximum weight word path, according to the above steps to find the maximum weight path, find dmax⑶;

[0036] 5) calculates the weight coefficient classification dmax⑶ maximum weight path the CC;

[0037] 6) Sort coefficient calculating respective home subnetwork the CC, comparison coefficients size, if the same, classified coefficient CO SW, SW of emotional weight coefficient (Sentimetal weight), i.e. classification semantic weight in seven sub-network weights, if not the same, according to the final ranking classification principles coefficient CC, if the first weight accounted for 80 percent ownership of the network to the appropriate mood, if not more than 80 percent this classification is attributed to the weight under the first two ranking emotional network;

[0038] 7) If you can not ensure that the classification under treatment classification text classification by category prediction prediction coefficient CPC.

[0039] The establishment of the model and fine-grained sentiment classification model to classify the prediction method is the use of:

[0040] 1) SWLM-TC ​​on a part of the text in all samples classified fine-grained, a sample concentration is calculated weights to each text belongs emotional weight coefficient SWX, as another part of the text classification verification experiment;

[0041] 2) For all samples SWLM-TC ​​classification of: calculating coefficients for each classification of the CC text, coefficients classified according to the classification step 6 SWLM-TC ​​algorithm, put this sample was added emotion category set corresponding X TSx lower (Train Set), if the classification is not the case the coefficient determined using SWLM-TC ​​performed using SWLM-TC ​​prediction algorithm in step 7, normalized to the corresponding classification;

[0042] 3) the sample data using the algorithm SWLM-TC ​​after completing calculating the emotional text, using the corresponding HMM trained classification text classification model, classification and then the HMM models trained:

[0043] a) For the test text using the HMM classification algorithm, if the classification is correct, it is determined that the sub-text sentiment classification;

[0044] b) For the classification result without text, using the prediction coefficients CPC classification classifies prediction.

[0045] HMM is a machine learning method, the first to use SWLM-TC ​​algorithm sample set of affective computing, emotion get classified into seven sub-text sample library, using the sample library training HMM model, using the trained HMM model, you can the remaining part of the library of text text text classification test verification.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] 1, the present method may be fine-grained emotion text classification, unlike conventional tendency calculating, has finer classification.

[0048] 2, improve the accuracy of TC simple algorithm, so that more accurate classification.

[0049] 3, and test sample library classification mood after use SWLM-TC ​​sample sets HMM model training, improve the automation of simple machine learning algorithms.

BRIEF DESCRIPTION

[0050] FIG. 1 is a flowchart showing the overall algorithm of the present invention.

[0051] FIG 2 is a flowchart of an algorithm SWLM-TC ​​present invention.

[0052] FIG. 3 is a flowchart SWLM-HMM algorithm of the present invention.

[0053] FIG. 4 is a line graph of experimental data TC labeling algorithm based on word frequency.

[0054] FIG. 5 is a line graph SWLM-TC ​​heuristic experimental data.

[0055] FIG. 6 is a line graph SWLM-HMM algorithm experimental data.

[0056] FIG. 7 is a schematic view of the experimental data of an average micro present invention.

[0057] FIG. 8 is a schematic view of a macro average data experiment of the present invention.

[0058] FIG. 9 is a classification of experimental data of the present invention in a schematic view of distribution (correctly classified).

[0059] FIG. 10 is a classification of experimental data of the present invention in a schematic view of distribution (error class assigned).

[0060] FIG. 11 is a classification of experimental data of the present invention in a schematic view of the distribution of (part of such misuse points).

Detailed ways

[0061] Next, embodiments of the present invention will be described in conjunction with the accompanying drawings and embodiments.

[0062] As shown in FIG. 1, the present invention is based on fine-grained random emotional sentiment classification word co-occurrence network, first, random network theory, the use of the phenomenon of co-occurrence words, through the emotional vocabulary lexicon marked body is formed emotion feature a random network model constructed based on the order of the words, i.e., word co-occurrence emotional network model, reduction model based on the emotional words longest matching method (SWLM, Sentimental word longest match) algorithm and TC binding be SWLM-TC ​​classification unsupervised learning, emotional or further word HMM longest matching method and machine learning algorithms combine to establish fine-grained sentiment classification model and use the model to classify forecast. Details are as follows:

[0063] 1 emotional word co-occurrence based on random network model

[0064] To facilitate the text paragraph level fine-grained, and found that the inherent rule of emotion between words, the present invention is improved by the literature [YANG Feng, PENG Qin-ke, XU Tao, Sentiment Classification for Comments Based on Random Network Theory , Acta Automatica Sinica, 2010 · 6 Vol · 36, No6] the proposed method, suitable for fine-grained sentiment analysis to build the emotional word co-occurrence network model.

[0065] 1.1 emotional vocabulary body vocabulary

[0066] Chinese emotional vocabulary ontology is Information Retrieval Laboratory, Dalian University of Technology through the efforts of all members of the Department of finishing and labeling of a Chinese bulk of resources under the guidance of Professor Lin Hongfei. Chinese vocabulary emotional sentiment classification system is built on the basis of body in a foreign country more influential Ekman classification system of six categories of emotion on. On the basis of Ekman on the emotional vocabulary body to join the category of "good" to compliment emotion in more detailed classification. The final terms of the emotional body is divided into 7 categories of 21 categories. Sentiment classification were happy {happy (PA), ease (PE)}, good {Dear (PD), praise (PH), I believe (PG), love (PB), wish (PK)}, anger {anger (NA )}, sorrow {sadness (NB), down (NJ), guilt (NH), Si (the PE)}, fear {: panic (the NI), fear (the NC), shame (the NG)}, evil {depressed (NE ), hate (ND), responsible banished (NN), jealous (of NK), suspected of (NL)}, {shock surprise (PC)}. Emotional intensity power into five files 1,3,5,7,9, 9 represents a maximum intensity, minimum intensity represents 1. Emotional Speech Vocabulary type body divided into a total of seven categories, namely, a noun (noun), verb (to verb), adjective (ADJ), adverb (ADV), the words in the network (NW), idioms (idiom), prepositional phrase (Prep ), containing a total of 27466 sentiment words.

[0067] 1.2 Network Model

[0068] The small-world network model Watts and Strogtz introduced [Watts DJ, Strogtz S !! Collective dynamics of 'small-world' networks.Nature, 1998,393 (6684):. 440-442], Barabasi proposed by Albert and scale-free network model [Barabasi AL, Albert R.Emergence of scaling in random networks · Science, 1999,286 (5439): 509-512] after the pioneering work of complex networks, compared to the regular network and random network, small world and scale-free networks: a small average path length, clustering coefficient larger constructed by the association between the word co-occurrence characteristic model of the network associated with small-world network. Can mean length and with a small coefficient of large aggregates on the meaning constructed network, fast emotional classification size characteristics.

[0069] 1.3 emotional word co-occurrence network model

[0070] Venant dogs [Shi Jing1Hu Ming, Dai Guo-Zhong · Topic analysis of Chinese text based on small world model Journal of Chinese Information Processing, 2007,21 (3):. 69-75 (Shi Jing, Ming Hu, . Dai Guozhong analysis model based on Chinese text small World theme Chinese information Technology, 2007,21 (3): 69-75)], the establishment of a random network model, the literature [YANG Feng based on common co-occurrence relations between words, PENG Qin-ke, XU Tao, Sentiment Classification for Comments Based on random network Theory, Acta Automatica Sinica, 2010.6Vol.36, No6] in accordance with the order co-occurrence relationships between words to establish random network model, this network model with Creating order words co-occurrence network, with a short comment SCP algorithm for incremental news, a great advantage for lack of a large network of short-term review, the algorithm for computing the tendency is more suitable, but for the more emotionally size classification Not suitable.

[0071] In order to construct a fine-grained analysis of emotion, the present invention employs stochastic network model constructing emotional word cooccurrence order, emotional word co-occurrence information sequence embodies semantic emotional words, before and after such modification, the co-occurrence of words emotional distance semantic relationships the words of a great relationship. According to the present invention is to establish a close emotional relationship word cooccurrence order model, i.e., co-occurrence window length WL small area (typically take 2), and taking into account co-occurrence of word order relationship.

[0072] In the emotional word co-occurrence network construction is completed, this large network model according to seven different emotional categories, divided into seven small emotional word network, then the related operations.

[0073] In order to describe the emotional word co-occurrence method for constructing the network model, used mathematical definition of the relevant:

[0074]

: Chinese vocabulary, vocabulary used in the invention to remove stop words, meaningless content words and then go through the emotional body after the emotional vocabulary word ontology annotation set;

[0075] w: extracted emotional words C

[0076] S: the sequence of a plurality of sentences;

[0077] N: number of the node G;

[0078] M: the number of edges of G;

[0079] W = {wi | ie [1, N]}: set the node G;

[0080] E = {(Wi, Wj) | wi, WjeW, and the sequence is present between Wi and Wj} co-occurrence relationship: G is the set of edges, wherein (Wi, Wj) indicates points to the node from the node Wi Wj are to the side;

[0081] Nw = {nwi I WiG w}: Right weight G of nodes;

[0082] NE = {nwi, wj (wi, wj) eE}: G in the weight of an edge, the edge weight represents the node between Wi and Wj.

[0083] The following presents a method to establish emotional word co-occurrence of the network model G:

[0084] D performed for each text clause obtained in an ordered set of sentences Sl4 & amp; 4 ... -Sn;

[0085] 2) for each of the 51 word sentence, as well as stop words filtered meaningless content words, ontology vocabulary using emotional stimulation was denoted T emotional words, to obtain an ordered set of emotional words Wl ^ W2 --- - ^ Wn;

[0086] 3) for each sentence Si, using WL (- generally take 2) bit word sliding window extracted from sentences to <Wi, Wj> if.

Add a new node to the W Wi, Wi is the weight and the initial weight is set nwi 1; otherwise nwi 1, the operation of Wi and Wj is similar to the old

Shu add a new edge (Wi, Wj), and as (Wi, Wj) weight nwi, wj set to an initial value of E 1; otherwise Ilwi, wj is incremented.

After [0087] 4) all text processing is complete, the network type G 1¾ established.

[0088] 5) according to the network model G seven kinds of emotions (Joy, good, anger, sadness, fear, evil, shock) classified into 7 sub-network. Subnetwork resolution process, if there event of breakage of, using a weight that the highest point of the fracture network of the sub-blocks are connected, it can be used to fine-grained seven sub-network computing (G1, G2, G3, G4, G5, G6, G7) build.

[0089] 2 fine-grained text feature classification for the emotional

Previous studies [0090], the emotion of the text as a tendency to focus on research, with the deepening of research in this area, fine-grained research value and uses it come out, fine-grained and tendentious analysis focus is different, fine-grained It is a multi-classification problem, and only need to calculate the propensity propensity to text, use the manual annotation dictionary is different, tendentious studies only need to mark out the word tendency, and fine-grained marked emotional vocabulary dictionary the ontology according to the emotional word speech characteristics were related type, intensity, polarity been labeled. HMM with the use of machine learning algorithms fine-grained sentiment classification.

[0091] emotion word longest match classification SWLM, the longest match by heavy emotional vocabulary words maximum weight, so do not use disambiguation and anti-noise processing, can be more accurate to classify under relevant to the emotional topic, and the Seven small classification model to calculate weights, HMM parameters may be used to obtain a machine learning classification.

[0092] The present invention is defined as follows:

[0093] Definition 1 (weight matching the longest path length dmax⑶)

[0094] Gx network, if two sequential cover emotional words, the edge directly connected using matched; emotional words, if two networks exist Gx interval network, path selection when choosing the maximum weight matching node by weight , is the longest matching path weights length S. Calculated as follows:

[0095]

[0096] wherein dmax (Wi, Wi + x) is the i-th word network to the maximum weight of the weight of the word i + x matching path.

[0097] Definition 2 (emotional weight coefficient SW (Sentimetal weight))

The number of reproducible [0098] G semantic network, the sentiment polarity proportion of their share of the seven sub-network, using this factor will become more apparent classification, reducing classification problems caused because of blurred, so emotional word for word network freq , the strength of polarity is P. Calculated as follows:

[0102] WC is the sub-network where the value of each word Emotion, Wy of the emotional value for the sub-network, the SWx value for the SW sub-network X, i.e. emotional weight coefficient.

Learning classification method [0103] 2.1 TC algorithms SWLM-TC ​​unsupervised

[0104] Definition 3 (Categories Coefficient CC (Classification coefficient))

[0105] Category coefficient SWLM-TC ​​when unsupervised classification algorithm defined classification coefficients, after the maximum matching word path is determined, to reproduce the words of Re on this path, and emotional intensity Power, assuming a η word is calculated as follows:

[0108] where ^ is a single word square classification coefficient.

[0109] Definition 4 (class prediction coefficient CPC (Classification prediction coefficient))

[0110] categorical predictor coefficients in the use of machine learning algorithms to classify the time, for time can not determine the classification of the samples, taken prediction mechanism. Sorted in accordance SWx, if SWdSW2MO%, wherein SWi / SWA 1.5, then classified SW1T, or fall under SWdPSW2; if sWi + swxso%, said that the article is more complex classification, the classification classified according to the respective coefficients the next classification.

[0112] Since the appearance of the text paragraph-level emotion words, the article presents the emotional context of the main line, by using a random network of co-occurrence, emotional context on such emotions are preserved, so by order of a random co-occurrence network has a good performance.

[0113] Referring to FIG 2, based on the use of emotional words SWLM TC labeling algorithm is the following processing steps:

[0114] 1) the classification of the articles were required clauses, Sk- ...

[0115] 2) for each sentence word order, and removing the meaningless content words, auxiliary, and using the emotional vocabulary lexicon tagging body, selected words are marked, in order, i.e., ...

[0116] 3) The home network search corresponding marked words;

[0117] 4) of the word network path selection, if the I two adjacent words, then using the path directly connected; II if two non-adjacent words, then the selected connected on their path through the maximum weight path words, according to the above steps to find the maximum weight path, find dmax⑶;

[0118] 5) calculated on the weight of the path coefficient classification dmax⑶ maximum weight CC, detailed calculation process defined 3;

[0119] 6) Sort coefficient calculating respective home subnetwork CC, comparison coefficients size, the I if the coefficient of the same size, classification coefficient CC * SW; II if the coefficient is not identical, step III, III in accordance with the final ordering principle CC If the first weight accounted for 80 percent ownership of the network to the appropriate mood, if not more than 80 percent this classification is attributed to the weight coefficient CC, under the first two categories.

[0120] 7) If you can not ensure that the classification under treatment classification text classification predicted by definition 4.

[0121] 2.2 SWLM-HMM classification algorithm based on supervised machine learning

[0122] In machine learning plays a big role in text classification, HMM algorithm has a very good performance in NLP, due to the simplicity of the HMM algorithm, less calculation for variable-length sequences of samples can be trained, in view of HMM Fine-grained classification emotional learning, to improve the accuracy SWLM-HMM classification.

[0123] using SWLM-HMM to classify the time, it can not be used directly for HMM training corpus, combined SWLM processing After treatment, reuse HMM training algorithm can improve the accuracy of the classification and speed up the classification speed.

[0124] Referring to Figure 3, the training data as follows:

[0125] (1) use a portion of the text sample database using SWLM-TC ​​classification of the fine-grained sample set, wherein SWLM-TC ​​classification process SWLM-TC ​​as shown, to calculate the weight of the sample belongs to the emotional the SWx weight coefficient;

[0126] (2) For all samples SWLM-TC ​​classification of: calculating coefficients for each classification of the CC text, coefficients classified according to the classification step 6 SWLM-TC ​​algorithm, put this sample into the respective emotion category set X under TSx (Train Set), if the classification is not the case the coefficient determined using SWLM-TC ​​performed using SWLM-TC ​​prediction algorithm in step 7, normalized to the corresponding classification;

[0127] (3) the portion of the text sample database text using training samples SWLM-TC ​​classification algorithm after Yoshiko emotion, obtained using the corresponding HMM trained classification text classification model; wherein each classification is characterized by good words emotion text has been marked, the form and order of the words in a text chain, the process uses HMM models trained, the classification of the emotional words and each word string of text as a parameter passed to the emotional well sub-HMM model, HMM model training carried out, all the samples are input to the HMM algorithm for training;

[0128] a) For the remainder of the text sample database, the HMM classification algorithm, if the classification is correct, the corresponding classification calculations.

[0129] b) For non-text classification result, using the defined class prediction 4.

[0130] 3 characterized by fine-grained sentiment classification experiment

[0131] 3.1 Data Categories

[0132] Experimental data the present invention is employed for data collection and CCF blog natural language processing and evaluation of the Chinese computing session NLP & amp; CC2014 data. Crawling Twitter data 7000, selected 4000 blog data therein, the NLP & amp; data CC2014 concentrating similarly 2000 topics were fused, final corpus data noted in the sample of about 6000, wherein all selected is Weibo contain emotions, no feelings for the micro-blog and blog data were removed, including the following data components:

[0133] I) TrainDataNet: 6000 used in which micro-blog data;

[0134] 2) TrainDataHMM: 5000 micro-blog using the 6000 data, in which the sample collection containing micro-blog 7 emotion data.

[0135] 3) TrainDataTest: 1000 using additional data other through the TrainDataHMM.

[0136] Table 1 data distribution

Two measurements [0139] precision and recall information retrieval and statistical classification applications most frequently used for the quality evaluation results.

[0140] The present invention uses experimental test data collection and Chinese NLP & amp; CC2014 in Chinese propensity evaluation data set, after testing the system, results are as follows.

[0141] 3.2 classification results

[0142] A. Experimental raw data

[0143] I) sentiment classification of experimental SWLM-TC

[0144] The validity verification using SWLM-TC ​​algorithm, correct, recall and F-measure the effectiveness of the classification.

[0145] The seven classification results in Table 2 show:

[0146] TABLE 2 SWLM-TC ​​classification algorithm results

[0147]

Sentiment classification experiment [0148] 2) SWLM-HMM's

[0149] The results SWLM-HMM algorithm shown in Table 3

[0150] Table 3 SWLM-HMM classification algorithm results

[0151]

Sentiment classification experiment [0153] 3) TC algorithms

[0154] TC verification test results shown in Table 4 below Algorithm

[0155] The results in Table 4 TC classification algorithm

[0156] _

[0157] B. accuracy, recall, Fl value

[0158] Data calculated above experimental data, the accuracy of the calculated SWLM-TC, recall and Fl data values ​​as follows:

[0159] I) sentiment classification of experimental SWLM-TC

[0160] P, R, and algorithm F1 SWLM-TC ​​values ​​in Table 5 below

[0161] TABLE 5 SWLM-TC ​​of P, R and Fl value

[0162]

Sentiment classification experiment [0163] 2) SWLM-HMM's

[0164] P, R and F1 values ​​SWLM-HMM algorithm is shown below in Table 6

[0165] TABLE 6 SWLM-HMM of P, R and Fl value

[0166]

Sentiment classification experiment [0168] 3) TC algorithms

[0169] P, R, and TC values ​​Fl algorithm is shown as follows in Table 7

[0170] P Table 7 TC algorithm, R value and Fl

[0171]

[0172] C. macro and micro Average Average

[0173] Using the algorithms and macro mean average of P, R and Fl values ​​shown in the following table 8

[0174] Table 8 mean the macro and micro-mean algorithm

[0175]

[0176] As shown in FIG 4, FIG 5 and FIG 6 based on a comparison of several line graph labeling algorithm TC word frequency, SWLM-TC ​​heuristic algorithm, SWLM-HMM algorithm, the experimental data.

[0177] Figures 4, 5 and 6 show the results of analysis using the same sentiment lexicon, wherein the generation of the algorithm to look at TC, TC basic emotion analysis algorithm SWLM-TC ​​algorithm, SWLM-HMM comparison algorithm results were as follows ,

[0178] 1) on the accuracy SWLM-HMM> SWLM-TC> TC, TC algorithm accuracy in the range of seven particle size was 52.42% -54.54% accuracy range SWLM-TC ​​algorithm on 7 granularity was 60.12% -65.73% accuracy range SWLM-HMM algorithm on 7 particle size was 69.60% -73.21%, which shows that, TC algorithm on the order of percentage behind SWLM-TC ​​and SWLM-HMM algorithm, SWLM-TC ​​on a percentage of the order behind SWLM-HMM.

[0179] 2) on the recall SWLM-HMM> SWLM-TC> TC, TC recall algorithm in the range of seven particle size is 60 · 14% -62 · 09%, SWLM-TC ​​algorithm granularity on 7 recall rate in the range of 31% -74 · 71 · 23% recall rate range SWLM-HMM algorithm on 7 particle size was 79.39% -83.88%, TC algorithm on the order of percentage behind SWLM-TC ​​and SWLM- HMM algorithm, SWLM-TC ​​on a percentage of the order behind SWLM-HMM.

[0180] 3) on the Fl value SWLM-HMM> SWLM-TC> TC, several evaluation criteria, SWLM-TC ​​and performance SWLM-HMM algorithm are better than TC, TC algorithm Fl value on 7 granularity in the range of 56.43% -57.66%, Fl range SWLM-TC ​​algorithm on 7 particle size was 66.10% -69.72%, Fl range SWLM-HMM algorithm on 7 particle size was 75.26% _ 77.90%, TC algorithm on the percentage of the order behind SWLM-TC ​​and SWLM-HMM algorithm, SWLM-TC ​​on a percentage of the order behind SWLM-HMMo

[0181] From the comparison of the data above, found that both the precision, recall rate, Fl value, the result SWLM-TC ​​algorithms and SWLM-HMM algorithm algorithms are better than TC, demonstrating, using only traditional TC algorithms performance in the calculation of the fine-grained relatively poor.

[0182] The average micro and macro-averaged data in Figures 7 and 8.

[0183] from the macro and micro average mean line graph can be seen that the performance SWLM-HMM and SWLM-TC ​​TC algorithm is better than the algorithm, SWLM-TC ​​and SWLM-HMM algorithm, the performance is better than SWLM-HMM SWLM -TC performance of the algorithm. The performance of the three algorithms on the same data set is almost an order of magnitude of 10%.

[0184] From the distribution data set, several draw bar graph table, three algorithms in three distribution Classification 9, 10 and 11:

[0185] From FIGS. 9, 10 and 11 can be seen in the correct classification, a minimum text TC correct classification algorithms, error bars assigned to the number of errors in the class also are more points belonging to such mistaken of this project, the number of TC algorithm is also the largest, are leading to TC algorithms SWLM-TC ​​and SWLM-HMM and the corresponding data.

[0186] 3.3 Experimental results

[0187] By the raw experimental data, and the accuracy of calculated recall and Fl value, and can draw the following conclusions views:

[0188] I) SWLM-TC ​​algorithm on a conventional high-accuracy algorithm TC -13.31% 7.7% percent, SWLM-HMM algorithm for high accuracy 9.48% -13.09% percentage points SWLM-TC ​​ratio algorithm. In the data described SWLM-X algorithm on the accuracy higher than the conventional algorithm because before calculation algorithm proposed by the invention compared to that classification, SWLM-TC ​​and emotional SWLM-HMM after random word co-occurrence network stage, co-occurrence word network in the present invention in role manifested verify the co-occurrence network based on emotion word original intention of the algorithm proposed by the invention is right; comparing SWLM-TC ​​and SWLM-HMM algorithm We found accuracy SWLM-TC ​​is lower than SWLM-HMM algorithm, because SWLM-HMM algorithm before classification algorithm using SWLM-TC ​​training training set, using the trained model to go in the late classify, but also in the late fuzzy feeling emotional text prediction, using these two strategies makes SWLM-HMM on the accuracy beyond the natural SWLM-TC ​​and TC algorithms.

[0189] emotion emotional network co-occurrence word on the missing text to be completed, as well as highlight text in a relatively large weight sentiment words appear on these two points, so that SWLM-X algorithm highlights the advantage in the fine-grained affective computing.

[0190] 2) in the recall, SWLM-TC ​​algorithm to be higher than the traditional algorithm TC -14.09% 11.17% percentage points, than SWLM-HMM algorithm for comparing SWLM-HMM algorithm percent 8.08% -12.57%, indicating the SWLM-TC ​​and when SWLM-HMM algorithm to classify certain types, the ability to correct classification than the traditional method, because the present invention proposed classification algorithm and a frame class, belonging to the class calibration of the efficiency and accuracy of specific emotions of the word is to be higher than the traditional method. Algorithm proposed in the present invention is the ability to highlight important emotional words than conventional algorithms, and the present invention is employed in a polar affective emotional vocabulary library body, this parameter is a good use of the text, by the co-occurrence frequency and polarity and emotion multiplying, making a strong emotional sentiment polarity word is more prominent, thus reducing interference with other emotions of the word, making it more accurate at the time of classification.

[0191] 3) comparing the value of Fl, SWLM-TC ​​algorithm is higher than 9.67% -13.29% percentage points higher than conventional algorithms, SWLM-HMM algorithm than SWLM-TC ​​algorithm to be higher than 9.16% -11.8% percentage points, integrated evaluation, SWLM-TC ​​and SWLM-HMM algorithm proposed by the invention are better than traditional algorithms, this respect, the algorithm proposed by the invention beyond the use of traditional methods in overall performance to be fine-grained sentiment classification algorithm.

[0192] 4) From the experimental point of view of the data, and algorithm computing framework proposed by the invention in the classification of related performance is good, but a lot of emotion and tendentious analysis of the differences, because the fine-grained computing not only in all aspects requirements of the algorithm can be fine-grained classification, anti-jamming performance is relatively high, for the emotional word appears in an article, use a calibrated approach is very prone to interference, causing unnecessary multi-classification problem, because of an article when emotional tendencies parity stable, under normal circumstances, the sentiment polarity and co-occurrence of nature is very important, this is not the general classification algorithm to process, so the text of such processing, the algorithm proposed by the invention on have a comparative advantage in the general text, the word sentiment polarity relatively strong and not strong co-occurrence of the text, the inventive algorithm and calibration algorithm based on the same when dealing with, but the difference is that in which the present invention use the emotional rather ambiguous text using predictive mechanism, so that the processing aspects of such a text would be better than no prediction mechanism of some . In the present invention employs a fine-grained SWLM algorithm computing framework, on the processing mechanism, absorb knowledge of complex networks, the performance of all aspects of the treatment of complement classification text words emotional aspects relatively good results, taking the above-described after the mechanism, fine-grained to fine-grained computational framework is calculated by experimental validation studies.

Claims (3)

1. A fine-grained random word sentiment sentiment classification method based on co-occurrence network, random network theory, the use of the phenomenon of co-occurrence words, after the label body emotional vocabulary lexicon, forming a random sequence based on the word emotion feature constructed network model, that is, the emotional word co-occurrence network model, model reduction based on this, the longest term emotional matching method (SWLM, Sentimental Word longest match) and TC algorithm combines were SWLM-TC ​​classification unsupervised learning, or further the longest matching method emotional words and HMM machine learning algorithms to establish a fine-grained binding emotional classification model and use the model to classify the prediction, wherein said emotional word existing network model were constructed as follows: 1) performing for each text sub sentence obtained in an ordered set of sentences S1 ^ S2H-HSn; 2) for each word sentences 31, was filtered off and stop words meaningless content words, using emotional words ontology emotional word tagging, to give a group emotional word sequence 3) for each sentence, using a sliding window WL bits extracted from the vocabulary sentence <wi, wj>, when W # W, and W is added to a The new node Wi, and as the weight Wi of weight nwi assumed that an initial value of 1; otherwise nwi plus 1, if (Wi, Wj) ί E, is added to E, a new edge (wi, wj), and as (Wi, of Wj ) weight nwi, wj assumed that an initial value of 1; otherwise nwi, wj plus 1; 4) after all text processing is complete, the network t Dae type G setup complete; wherein, S represents a sequence consisting of a plurality of sentences, W represents extraction the emotional words, we Σ, Σ for the Chinese vocabulary, Chinese vocabulary to remove stop words, meaningless content words and then go through the emotional body after the emotional vocabulary word set ontology annotation; W is a collection of nodes in the network model G, W = Iwi I ie [I, N]}, N is the number of nodes of G; E is the set of edges G network model, the number of edges of the network model G M, E = {(Wi, Wj) IWi, between WjGW, and Wi and Wj order co-occurrence relation}, (Wi, wj) represents from node Wi points to the node Wj directed edge; Nw is the weight of the network model G of nodes weight, Nw = {nwi I wiew}; NE is a weight network model G of the edge weight represents the weight of the edges between nodes Wi and Wj weight, NE = {nwi, wj I (wi, wj) eE}; the network model G according to music, good, anger, sadness, fear , evil, shock seven kinds of emotions into 7 Network, subnetwork resolution process, breakage occurs if the case, using the weight of the highest sub-block that network node connection and Fracture, seven sub-network construct can be used to fine-grained calculations Gxlx = wherein, after classification when, have the following definitions: maximum weight matching path length dmax (S): network Gx IX = {1,2,3,4,5,6,7}, if two sequential cover emotional words, directly connected using matching edges, two emotional words if there is a network in the network Gx in the interval, when the selected path is selected by matching node with the largest weight, that is, the length S is calculated as follows:

   

   Wherein dmax (Wi, Wi + x) is the network of the i-th word to the maximum weight of the i + x Words weight matching path; emotional weight coefficient SW (Sentimetal weight): a network G in seven sub-network are each occupied affections of gravity, this factor will become more apparent classification, classification reduced because blurred due, so reproducing the number of words in the word emotion FREQ network, strength of polarity is P, is calculated as follows:

   

   Where WC is the subnetwork emotional value of each word, Wy of the feeling and the value of the sub-network, the SWx for the SW value sub-network X, i.e. emotional weighting coefficient; classification coefficient CC (Classification coefficient): after the maximum matching word path determined , Re and emotional intensity of the reproducing power of words on this path, assuming η word is calculated as follows: CCi = Re Xpower

   

   Wherein the square ^ is a single word classification coefficient; class prediction coefficient CPC (Classification prediction coefficient): prediction mechanism when using machine learning algorithms to classify, for the classification can not determine the sample taken; sorted by the SWx, if SWdSWA 80%, SWi / SW2> 1 · 5, fall under the SWi, if SWi + SW2> 80%, SWi / SW2 <= 1 · 5, SWi and SW2 classified under two properties in this case; if SWi + SWXSO %, it means that the classification of this article is more complex, classified under the appropriate classification in accordance with the classification factor:

   
2. The fine-grained sentiment classification based on co-occurrence emotional word random network of claim 1, wherein said body emotional vocabulary database, is divided into seven categories 21 emotion subclasses, were classified music emotion {happy (PA), ease (PE)}, good {Dear (PD), praise (PH), I believe (PG), love (PB), wish (PK)}, anger {anger (NA)}, sorrow { sad (NB), down (NJ), guilt (NH), Si (PE)}, fear {: panic (NI), fear (NC), shame (NG)}, evil {: bored (NE), hate ( ND), responsible banished (NN), jealous (of NK), suspected of (NL)}, {shock surprise (PC)}; emotional intensity power into five files 1,3,5,7,9, 9 represents a maximum intensity, 1 represents the minimum intensity, the type of emotion vocabulary speech body divided into a total of seven categories, namely, a noun (noun), verb (to verb), adjective (ADJ), adverb (ADV), the words in the network (NW), idioms (idiom) , prepositional phrase (prep), containing a total of 27466 sentiment words.
3. The fine-grained sentiment classification based on co-occurrence emotional word random network of claim 1, wherein said longest word sentiment maximum weight matching method emotional vocabulary words the longest match so as not to eliminate use of discrimination and anti-noise processing, to accurately classified under related to the emotional topic, and heavy weights were calculated by seven small class model, machine learning can be derived parameter classification.