CN104331506A - Multiclass emotion analyzing method and system facing bilingual microblog text - Google Patents

Abstract

The present invention relates to a multi-type sentiment analysis method and system for bilingual microblog texts, belonging to the technical field of microblog text sentiment analysis, comprising the following steps: (1) construction of a bilingual dictionary: firstly collect a certain scale of corpus with emotional tendencies, and Extract high-frequency vocabulary with emotional tendencies from the corpus; then use the existing knowledge base and vocabulary similarity calculation model to expand the emotional dictionary; finally add network language and emoticons to the emotional dictionary; (2) Text preprocessing: Segment the text to be recognized and remove stop words and English word form normalization; (3) text feature space representation: use the bilingual sentiment dictionary to vectorize the text; (4) use the multi-emotional classification model to realize the emotion of the corpus text Identify tasks. The accuracy rate and F1 value of the method of the invention are higher than those of traditional classification methods, especially the classification effect of the semi-supervised Gaussian mixture model classification algorithm in a small-scale training set is obviously better than other methods.

Description

A multi-category sentiment analysis method and system for bilingual microblog text

technical field

The invention relates to a sentiment analysis method and system, in particular to a multi-type sentiment analysis method and system for bilingual microblog texts, belonging to the technical field of microblog text sentiment analysis.

Background technique

With the rise of social media platforms and the widespread use of mobile devices, people have become accustomed to using 140 characters to express their appeals. Publishing Weibo has become an important means for individuals to express their emotions, so it is of great practical significance to analyze the sentiment tendency of Weibo texts. At present, Sina Weibo has become the main carrier of domestic Internet public opinion, and a large number of users interact with information and express emotions through Weibo. The development of an emotion classification system for user microblog texts and then the completion of emotion recognition has important reference significance in the fields of public opinion monitoring and product evaluation.

Most existing sentiment analysis systems classify microblog texts into positive and negative sentiments. But human emotions are complex and diverse. Positive emotions include emotions such as trust, gratitude, and happiness, while negative emotions include emotions such as pain, contempt, hatred, and jealousy. Simply dividing emotions into two categories cannot guarantee the accuracy of emotion discrimination. At present, there is still a lack of fine-grained emotional classification systems that can capture the attention of groups. The current Weibo sentiment analysis system mainly conducts statistical analysis on single-language texts, that is, Chinese sentiment tendency. However, in recent years, due to the improvement of the education level in mainland China and the influence of internationalization trends, Chinese and English collocations or pure English writing have gradually become popular. It has become an important form of individual emotional expression. This mix of Chinese and English microblog texts also brings new challenges to microblog sentiment analysis. Sentiment classification systems based on monolingual sentiment analysis methods are no longer suitable for the increasingly complex microblog language environment.

In addition, most of the current emotional vocabulary identification work uses machine translation to obtain emotional vocabulary, but for Weibo texts, due to its short text and 140-character limit, the vocabulary composition is relatively complicated, and the number of English slang and popular phrases on the Internet is increasing day by day. The quality of translations cannot be guaranteed.

Contents of the invention

The purpose of the present invention is to solve the problem that the classification granularity of the existing microblog sentiment analysis method is coarse, the quality of the microblog text analysis is not high for the Chinese and English mix and match, and the identification method of the emotional vocabulary is lagging behind. A Chinese-English bilingual sentiment dictionary construction method for microblog corpus, a bilingual dictionary-based microblog multi-category sentiment analysis method and a bilingual microblog text multi-category sentiment analysis system, thereby performing multi-category sentiment analysis on microblog text.

The idea of the technical solution of the present invention is to build a Chinese-English emotional dictionary by collecting a large amount of microblog text corpus with emotional tendencies, and use a semi-supervised and fully supervised hybrid model to construct a variety of emotional classifiers. After text processing of bilingual texts The spatial feature representation of the text is carried out according to the lexical emotion category, so that the emotional recognition task of the microblog text can be realized by using a variety of emotional classifiers constructed.

Concrete implementation steps of the present invention are as follows:

A Chinese-English bilingual emotion dictionary construction method, the method comprises the following steps:

Step 1. Grab the Weibo webpage, collect Chinese and English corpus with emotional tendency from the webpage, and extract high-frequency vocabulary with emotional tendency from the corpus and add it to the emotional dictionary database;

Step 2, applying the existing knowledge base to expand the sentiment dictionary;

Step 3, analyzing the captured microblog corpus, and adding new network languages and emoticons to the emotional dictionary.

Preferably, the emotional tendency includes five categories of social care, happiness, sadness, anger and fear.

Preferably, the knowledge base includes WordNet, NTUSD and HowNet.

Preferably, the expansion of the second step is to calculate the average similarity between the emotional vocabulary in each knowledge base and the emotional orientation vocabulary in the emotional dictionary, and expand the emotional words to the category of emotional orientation with the largest similarity.

Preferably, the emotional tendencies of the emerging online languages and emoticons are classified by raising hands and voting.

A method for multi-category sentiment analysis based on a bilingual dictionary, the method comprising the following steps:

Step 1, preprocessing the corpus text;

Step 2, performing feature space representation on the corpus text according to the Chinese-English bilingual sentiment dictionary;

Step 3: Carry out sentiment classification on the corpus text according to the established text sentiment classifier model.

Preferably, the preprocessing includes word segmentation and stop word removal, and for English text, it also includes morphological normalization.

Preferably, the text feature space representation is to represent each text in the corpus as a five-dimensional vector, and each element in the vector represents the number of sentiment words of the corresponding category contained in the Chinese-English bilingual sentiment dictionary.

Preferably, the emotion classifier model is a semi-supervised Gaussian mixture model classification algorithm (Semi-GMM) or a K-nearest neighbor algorithm based on symmetric relative entropy (KNN-KL).

Preferably, the semi-supervised Gaussian mixture model classification algorithm is to learn the Gaussian mixture model through the marked training corpus, and then use the model parameters and the probability distribution of the marked sample as the parameter initial value of the Gaussian mixture model. The test corpus is iteratively learned until the algorithm converges or the unlabeled set is empty.

Preferably, the K-nearest neighbor algorithm based on symmetric relative entropy uses relative entropy to measure the emotional similarity of texts to express the distance between texts, and determines the category of the sample to be classified according to the category of the adjacent samples.

Preferably, the relative entropy is calculated using the following formula:

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 5</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}<span\; class="notranslate">\; \&times;</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 5</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}}{\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 5</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}<span\; class="notranslate">\; )</span>}$

Among them, T _i is the normalized vector representation of the marked text, T _j is the normalized vector representation of the unmarked text, ω _ik and ω _jk represent the kth item of T _i and T _j respectively, and k is 1 to 5 Integer between.

A multi-category sentiment analysis system for bilingual microblog texts, including a Chinese-English bilingual sentiment dictionary, a corpus preprocessing module, a corpus text feature space representation module, and a sentiment classifier recognition module; the Chinese-English bilingual sentiment dictionary adopts the Chinese-English bilingual Sentiment dictionary construction method construction; corpus preprocessing module is used to perform word segmentation and stop word processing on the corpus text to be analyzed, and morphological normalization processing for English text; corpus text feature space representation module is used to preprocess the corpus The text processed by the module is vectorized, and the text is processed into a five-dimensional vector. The five elements in the vector represent the five categories of social care, happiness, sadness, anger and fear contained in the Chinese-English bilingual emotional dictionary. The number of emotional words; the emotion classifier identification module is used to adopt the emotion classifier model to carry out emotion recognition to the corpus text vector, and determine the emotion category to which the corpus text belongs.

Beneficial effect

Aiming at the field of microblog text sentiment analysis, the present invention constructs a fine-grained five-category Chinese-English bilingual sentiment dictionary by using Sina microblog message texts and existing knowledge bases, and constructs a K-nearest neighbor algorithm based on a semi-supervised Gaussian model and a symmetric relative entropy Bilingual Weibo Sentiment Multi-Classifier, which performs sentiment analysis on Chinese and English bilingual Weibo. Experimental results show that the accuracy and F1 value of the emotion classification method based on the bilingual emotion dictionary proposed by the present invention are higher than those of the traditional classification method. In particular, the classification effect of the semi-supervised Gaussian mixture model classification algorithm in a small-scale training set is significantly better than other methods.

Description of drawings

Fig. 1 is the flowchart of microblog text emotion analysis method in the embodiment of the present invention;

FIG. 2 is a schematic diagram of a bilingual microblog text example in an embodiment of the present invention;

Fig. 3 is the schematic flow chart of the emotion classification algorithm of semi-supervised Gaussian mixture model in the embodiment of the present invention;

Fig. 4 is a schematic diagram of accuracy comparison of various machine learning text emotion classification algorithms in the embodiment of the present invention;

Fig. 5 is a schematic diagram of comparison of accuracy rates of various emotion classification algorithms for bilingual microblog texts in an embodiment of the present invention.

FIG. 6 is a structural diagram of a multi-category sentiment analysis system for bilingual microblog texts in an embodiment of the present invention.

Detailed ways

FIG. 1 is a flowchart of a multi-category sentiment analysis method for bilingual microblog texts according to an embodiment of the present invention. The main workflow of text emotion recognition is as follows:

(1) Construction of a bilingual emotional dictionary: First, collect a certain scale of corpus with emotional tendencies, and extract high-frequency words with emotional tendencies from the corpus; then, use the existing knowledge base (WordNet and NTUSD, HowNet) and vocabulary similarity Computational models expand the emotional dictionary; finally, new online languages and emoticons are added to the emotional dictionary;

(2) Text preprocessing: segment the text to be recognized and remove stop words. Stop words are function words that human language contains without real meaning, such as determiners in English (“the”, “a”, “an”, “that”). On this basis, the English text also needs to perform lemmatization and stem extraction operations;

(3) Text feature space representation: use the constructed bilingual sentiment dictionary to extract the features of the text vocabulary, and represent the text in a five-dimensional vector according to the vocabulary emotion category;

(4) Use the multi-emotion classification model to realize the emotion recognition task of corpus text.

1. Construction of bilingual emotional dictionary:

Before introducing how to build a sentiment lexicon, we first introduce the classification of sentiment tendencies.

·Classification of text sentiment tendency:

Human emotions are complex and diverse, and venting emotions on Weibo has become a part of people's daily life. Simply dividing emotions into positive and negative emotions cannot guarantee the accuracy of emotion discrimination. In this embodiment, in order to better cover various emotions of microblog texts and further refine the emotion classification system, we divide the emotion categories into five categories: social care, happiness, sadness, anger, and fear.

·Emotional dictionary library construction:

The construction of the emotional dictionary depends on the microblog text collection. When artificially labeling texts with emotional tendencies, a few words with emotional tendencies in the text often play a decisive role. Sina Weibo has a large-scale microblog text collection, and users can immediately publish texts within 140 characters, so constructing a sentiment dictionary that satisfies the sentiment classification of microblog texts is the basis of this invention. At present, microblog research is mostly based on a single language corpus, but the combination of Chinese and English has become a popular trend for individual expression. Therefore, only constructing a Chinese sentiment dictionary cannot meet the needs of sentiment classification of Chinese-English mixed microblog texts. To further illustrate the necessity of adding an English emotional dictionary, we show two blog posts published by Weibo users in Figure 2. It can be seen from the figure that users with bilingual expressions habitually use English emotional words when talking about a certain topic Be emotional.

In order to establish a bilingual emotional dictionary, it is first necessary to collect a certain scale of corpus with emotional tendencies, and artificially extract a small number of high-frequency bilingual words with emotional tendencies (such as happy, sad, happy, beautiful, sad) from the corpus and add them to the emotional dictionary. Use these words as seed emotion words. In this embodiment, more than 7,000 microblog texts are collected from Sina Weibo as a corpus. We select some words from the corpus to manually label the five types of emotions as the seed word set seedset={PA, PB, PC, PD, PE}, where PA, PB, PC, PD, and PE represent various emotions (social care, happy, sad, angry, fearful).

Then, according to the selected seed word set, by calculating the similarity between the seed emotional vocabulary and the emotional vocabulary in each knowledge base, the emotional vocabulary in the knowledge base with a large similarity with the seed emotional vocabulary is expanded to the five types of seed emotions that have been selected. in vocabulary. In this embodiment, we use the existing knowledge bases WordNet, NTUSD and HowNet to expand the sentiment dictionary in turn. Select an appropriate number of words from the knowledge base for operation according to specific needs. For example, most of the words in HowNet and WordNet are mainly selected in this example. In the existing knowledge base, each vocabulary v _k can be described by multiple concepts, and each concept C _k is defined based on the sememe through the knowledge base expression language, and each concept contains multiple sememe interpretations.

For the semantic similarity between two Chinese vocabulary, the present invention adopts HowNet vocabulary similarity computing method, and its definition is as shown in following formula (1) and formula (2):

$\mathrm{<span\; class="notranslate">\; similarity</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; v</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; max</span>}}\mathrm{<span\; class="notranslate">\; similarity</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

$\mathrm{<span\; class="notranslate">\; similarity</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; p</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; \&times;</span>\frac{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

In Formula 1, similarity(v ₁ ,v ₂ ) represents the similarity between two words. Vocabulary _v1 has n concepts, and each concept contains p1 sememes. Vocabulary _v2 has m concepts and each concept has p2 sememes. C _1i represents the i-th concept of vocabulary _v1 , and C _2j represents the j-th concept of vocabulary _v2 .

In formula 2, p1 and p2 respectively represent the number of sememes contained in the two concepts, Then represents the weight of the i-th sememe in concept _C1 . And select the maximum sememe similarity between two words as the similarity of two words. α is a positive variable parameter, and d represents the distance of the sememe in the HowNet sememe tree.

For the semantic similarity between English words, the Lesk method in Wordnet is used to measure the degree of association between words. Every concept in Wordnet, namely word sense, is defined by a short annotation. The Lesk method calculates the similarity (v ₁ , v ₂ ) between the two words by finding and calculating the intersection of annotations of two concepts. Since each English word has multiple morphological representations, for example, happy, happiness and other words are related to happy, so it is necessary to merge the different forms of a word, remove the affixes to get the root, that is, normalize the word form, so as to improve the efficiency of text processing . The present invention standardizes English vocabulary by adopting two morphological restoration methods of Lancaster and WordNet Lemmatizer provided in NLTK.

After determining the similarity, we use the following formula to calculate the similarity between the emotional vocabulary in the knowledge base and the vocabulary in the seed word set and extend the emotional vocabulary in the knowledge base to the corresponding emotional orientation classification of the emotional dictionary:

$\mathrm{<span\; class="notranslate">\; \&piv;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; PA</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; PB</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; PC</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; PD</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; PF</span>}<span\; class="notranslate">\; \}</span>}{\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; similarity</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; PA</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; similarity</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; PB</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$

$\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; similarity</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; PC</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 3</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 4</span>}}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 4</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; similarity</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; PD</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 4</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 5</span>}}}\underset{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 5</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; similarity</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; PE</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; 5</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>$

Among them, N ₁ , N ₂ , N ₃ , N ₄ , and N ₅ are the number of seed words in each emotion subset. ω(v) represents the emotional category of the non-seed vocabulary, which depends on the maximum value of the average similarity between it and various emotional subsets.

In addition to traditional emotional words, more and more emerging online languages and emoticons are widely used by users in the emotional expression of Weibo texts. Therefore, in addition to the vocabulary expansion based on the existing knowledge base, we have observed and summarized a large number of microblog texts, artificially introduced network language and emoticons into the emotional dictionary, and used the method of many people to raise their hands to vote for their emotional tendencies. Classification.

To sum up, there are 7,590 Chinese emotional vocabulary, 421 English emotional vocabulary, 613 online vocabulary and 101 commonly used emoticons constructed in this embodiment. It covers 971 words of "social care", 2731 words of "happiness", 2289 words of "sadness", 1458 words of "anger" and 1276 words of "fear".

2. Text preprocessing:

Segment the microblog message text. The present invention adopts the ICTCLAS word segmentation system to carry out vocabulary recognition on Chinese texts, and carries out vocabulary recognition on the basis of blank spaces for English texts. After word segmentation of a Weibo message text, it is processed by removing stop words, such as: "的", "a", "the", etc. On this basis, the English text also needs to be morphologically restored and stemmed. The specific operations are consistent with the processing of the English text during the construction of the emotional dictionary.

3. The feature space representation of the text:

Let D={d ₁ ,d ₂ ,...,d _n } represent the collection of all microblog message texts, so each microblog message d _i can be represented by a five-dimensional vector, and each item represents the corresponding The number of vocabulary for sentiment classification.

4. Use the emotion classification model to realize the emotion recognition task of microblog text:

There are many existing methods or models that can implement sentiment analysis on texts. In this invention, we only introduce two types of microblog text multi-category sentiment analysis models. Among them, the semi-supervised Gaussian mixture model sentiment classification algorithm needs iterative training of the model, and the K-nearest neighbor sentiment classification algorithm based on symmetric relative entropy only needs to input the marked text vector without learning.

·Semi-supervised Gaussian mixture sentiment classifier model:

The sentiment classification algorithm process of the semi-supervised Gaussian mixture model is shown in Figure 3. Gaussian mixture models mostly use the expectation-maximization algorithm (EM) for parameter estimation. The process of training the mixed Gaussian model (GMM) refers to estimating the probability density distribution of the sample, and the estimated model is the weighted sum of multiple Gaussian models, where each Gaussian model represents a class. In this embodiment, for five emotion classifications, we use five Gaussian models for training and learning. Gaussian mixture model learning is the process of maximum likelihood estimation for each Gaussian model plus probability density estimation and weight (π _i ). π _i depends on the proportion of texts of various emotional categories in the training model. During the iterative process, π _i will change with the addition of samples in the test set. The initial value is equal to the proportion of the texts of the five emotional categories in the training set. is 1/5.

The semi-supervised Gaussian mixture model is a self-training algorithm, so in this embodiment, the marked microblog text is artificially divided into two parts, one part is the training set, and the other part is the test set, so as to realize the In the process, texts are selected from the test set and added to the training set to achieve the purpose of self-training. In each iteration process, compare the probability values of the text in the test set in the Gaussian mixture model of the five emotion categories, select the emotion category with the highest probability value in the Gaussian mixture model as the text emotion category, and select the emotion category from the corresponding Select a text with the highest probability value from all the texts with correct category classification, add it to the training set, and then continue to learn the mixed Gaussian model according to the new training set. The algorithm stops until the classification effect of the model after iterative training on the test set is the same as that before the iteration or the difference is negligible, or the test set is empty.

φ(u _j |θ _k ) is a Gaussian probability density function, The model parameters are obtained after training and learning, where μ _k represents the mean value of each Gaussian mixture model, Represents the variance of each Gaussian mixture model. The present invention first learns the Gaussian mixture model through the marked microblog message text, and the input is the marked text vector in the training set. Then, the model parameters and the probability distribution of the marked samples are used as the initial parameters of the Gaussian mixture model to iteratively learn the existing model, and the final output is the mean and variance of five Gaussian mixture models, representing five types of emotions. Using the learned model parameters, the unlabeled text vector can be input, compared according to the five output probability values, and finally classified according to the emotion category with the largest probability value.

· K-nearest neighbor sentiment classifier model based on symmetric relative entropy:

The K-nearest neighbor classification algorithm means that the category of a sample depends on the k most similar samples in the specific space where the sample is located, that is, most of the categories of the nearest neighbor samples in the feature space. In the classification decision, this method only determines the category of the sample to be classified according to the category of the nearest one or several samples.

The choice of k value, the measurement of distance and the decision rule of classification constitute the three basic elements of the K-nearest neighbor algorithm. In the present invention, we use relative entropy to measure text sentiment similarity. Relative entropy is a measure of the asymmetry of two probability distributions (of X and Y) in the same event space, denoted as D(X||Y). Therefore, the text vector representation is normalized, and the normalized text vector is denoted as T _i .

The distance between the microblog message text T _i and T _j is defined as follows:

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 5</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}$

ω _ik and ω _jk represent the kth item of T _i and T _j respectively, and k is an integer between 1 and 5. Due to the asymmetry of relative entropy, when measuring the difference between probability distribution X and Y, X represents the real distribution of data, and Y represents the approximate distribution of X. Therefore, when calculating the distance between texts, T _i is the normalized vector representation of the marked text, and T _j is the normalized vector representation of the unmarked text. But this form of asymmetric calculation ignores the approximate distribution of X with respect to Y. In order to improve the asymmetry of traditional relative entropy calculation, the symmetric relative entropy calculation formula is defined as follows:

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 5</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}<span\; class="notranslate">\; \&times;</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 5</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}}{\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 5</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; jk</span>}}}{{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; ik</span>}}}<span\; class="notranslate">\; )</span>}$

The input is the vector representation of the labeled normalized microblog text as the training set, and the vector representation of the unlabeled normalized microblog text. The asymmetric relative entropy distance measurement algorithm or the symmetric relative entropy distance measurement algorithm stipulated in the above model can be used to calculate the distance between the unlabeled text vector and the marked microblog text in the training set, that is, the similarity calculation of the microblog text. Select the emotion category with the largest proportion among the K nearest neighbors of the text vector to be classified as the emotion category of the text.

As shown in Figure 6, it is a kind of multi-category sentiment analysis system facing bilingual microblog text according to the embodiment of the present invention, including a Chinese-English bilingual sentiment dictionary, a corpus preprocessing module, a corpus text feature space representation module, and an emotion classifier recognition module; The Chinese-English bilingual sentiment dictionary adopts the construction method of the Chinese-English bilingual sentiment dictionary to construct; the corpus preprocessing module is used to carry out word segmentation and de-stop word processing for the corpus text to be analyzed, and also to carry out morphological normalization processing for the English text; The feature space representation module is used to vectorize the text processed by the corpus preprocessing module, and process the text into a five-dimensional vector, and the five elements in the vector represent the social values included in the text in the Chinese-English bilingual sentiment dictionary. The number of the five types of emotional words of caring, joy, sadness, anger and fear; the emotion classifier identification module is used to adopt the emotion classifier model to carry out emotion recognition to the corpus text vector, and determine the emotion category to which the corpus text belongs.

Apply the multi-category sentiment analysis system of the embodiment of the present invention, input a bilingual microblog text, and output the emotional category to which the text belongs after being processed by the system, the output A indicates that the microblog belongs to the social care category, and the output B indicates that the microblog belongs to For the happy category, the output C indicates that the Weibo belongs to the sad category, the output D indicates that the Weibo belongs to the angry category, and the output E indicates that the Weibo belongs to the fear category.

Evaluation indicators:

The invention invites students who study natural language to manually label the text captured by the API provided by Sina according to five types of emotions. Some of them are used as training set and some are used as test set. After the model is trained on the training set, the accuracy of the microblog text classification on the test set can be used to evaluate the quality of the model.

data set:

When comparing machine learning algorithms, the data set we selected was 7170 pieces of Chinese Weibo text information captured by Sina API as experimental data. And invited 25 students in the direction of natural language to manually label the text according to the five types of emotions, so that the emotional category of the text depends on the emotional category selected by most people. The distribution of corpus in each emotion category is shown in Table 1:

Table 1. Distribution of Weibo texts in 5 sentiment categories

Similarly, we used the Sina API to capture 7,000 bilingual microblog texts and invited 25 students who studied natural language to manually label the texts according to five types of emotions. The distribution of corpus in each emotion category is shown in Table 2:

Table 2 Distribution of microblog texts in 5 emotional categories

Experimental results:

We selected 3170 microblogs from 7170 Chinese microblogs as the test set (see Table 1), among which 500 microblog texts expressing social care, 1300 microblog texts expressing happiness, and 540 microblog texts expressing sadness , 510 Weibo texts expressing anger, and 320 Weibo texts expressing fear. For the training set, 1,000 to 4,000 microblogs were selected from the remaining 4,000.

We first compare the K-nearest neighbor classification algorithm based on asymmetric relative entropy and the K-nearest neighbor classification algorithm based on symmetric relative entropy. The experimental results are shown in Table 3:

Table 3 Comparison of the accuracy of the K-nearest neighbor classification algorithm based on different distance measurement algorithms under different training set sizes

The results show that the accuracy of the K-nearest neighbor classification algorithm based on symmetric relative entropy is slightly higher. Therefore, in the comparison of various machine learning classification algorithms, we only choose the K-nearest neighbor classification algorithm based on symmetric relative entropy to participate in the comparison.

We select majority voting algorithm (Majority vote), support vector machine algorithm (SVM), the K nearest neighbor classification algorithm (KNN-Cosine) based on cosine distance and the semi-supervised Gaussian mixture model classification algorithm (Semi-GMM) that proposes in the present invention and The K-nearest neighbor algorithm (KNN-KL) based on symmetric relative entropy was compared. The comparison results are shown in Figure 4. It can be seen from the figure that when the training set text size is 4000, the accuracy rate of KNN-KL can reach up to 85.1%. When the same k-nearest neighbor algorithm is selected, it is better to use symmetric relative entropy for text distance measurement than to use cosine distance for text distance measurement classification.

When the training set text size is less than 3000, Semi-GMM has better performance. Compared with KNN-Cosine and KNN-KL, Semi-GMM has better stability when the number of training texts decreases. As the number of texts in the training set drops to 1000, the accuracy of KNN-KL drops by 8.9%, while that of Semi-GMM drops only by 2.9%. This further confirms that Semi-GMM is more suitable for use when the training set size is small, while KNN, a fully supervised learning algorithm, is easily selected with a neighbor number of around k, which affects the classification effect.

SVM is mostly used for binary classification problems. Although there are multi-class support classifiers, their accuracy depends on the quality of the training data. And SVM has high complexity and is not good at dealing with large-scale classification problems. Since the cost of obtaining large-scale high-quality training data in text classification is too high, we believe that when the unlabeled text set is too large, it is more suitable to use Semi-GMM to classify text sentiment.

For the five types of emotions, when the size of the text training set is different, the accuracy of KNN-KL is shown in Table 4. We can see that when the number of texts in the training set drops to 1000, the classification effect of microblog texts expressing social care and sadness also drops sharply. Compared with the classification effect when the number of texts in the training set was 4000, 79 microblog texts expressing social care were misclassified, of which 64 were identified as happy, 11 as sad, and 4 as angry . For the sad Weibo texts, 60 were misclassified, 8 of which were identified as social care, 28 were identified as happy, 13 were identified as angry, and 11 were identified as fear.

Table 4 Under different training set sizes, text classification accuracy based on Semi-GMM and KMM-KL

Under different text training set sizes, the F1 values of Semi-GMM and KNN-KL are shown in Table 5, which further confirms the classification advantage of Semi-GMM in small-scale training sets.

Table 5 F1 value of text classification based on Semi-GMM and KMM-KL under different training set sizes

For the bilingual microblog text sentiment classification experiment, we selected 3000 microblogs from 7000 bilingual microblogs as the test set (see Table 2), and selected 1000 to 4000 microblogs from the remaining 4000 for the training set.

We use the semi-supervised Gaussian mixture model classification algorithm (Semi-GMM(Ch.)) which only uses the Chinese sentiment dictionary for sentiment word recognition and the K-nearest neighbor algorithm (KNN-KL(Ch.)) based on symmetric relative entropy. The majority vote algorithm (Majority vote (Ch.+Eng.)), SVM (Ch.+Eng.) algorithm, K-nearest neighbor classification algorithm based on cosine distance (KNN-Cosine (Ch.+ Eng.)) and the semi-supervised Gaussian mixture model classification algorithm (Semi-GMM (Ch.+Eng.)) proposed by the present invention and the K nearest neighbor algorithm (KNN-KL (Ch.+Eng.)) based on symmetric relative entropy are carried out Compare. The comparison results are shown in Figure 5. It can be seen from the figure that the accuracy rate of the text sentiment classification algorithm using the combination of Chinese and English sentiment words for sentiment word recognition is significantly higher than that of the single text sentiment classification algorithm using the Chinese sentiment dictionary for sentiment word recognition, which further confirms the bilingual The effectiveness of a lexicon of sentiment words. When the number of microblog texts in the training set drops to 1000, the classification accuracy of Semi-GMM (Ch.+Eng.) reaches 68.3%.

Table 6 Under different training set sizes, text classification accuracy based on Semi-GMM and KMM-KL

Table 7 F1 value of text classification based on Semi-GMM and KMM-KL under different training set sizes

Table 6 and Table 7 show the accuracy of Semi-GMM and KNN-KL for 5 types of emotion recognition for text when the size of the text training set is different. When the size of the text training set drops to 1000, the F1 value of Semi-GMM is greater than that of KNN-KL, which further confirms that the presence of different languages in the text will not affect the stability of Semi-GMM, and in Semi-GMM has more classification advantages under the small-scale training set.

Therefore, the bilingual dictionary-based microblog multi-category sentiment analysis method proposed by the present invention is of great practical application value.

In order to illustrate the content and implementation method of the present invention, this specification provides a specific embodiment. The purpose of introducing details in the examples is not to limit the scope of the claims, but to facilitate the understanding of the method described by the invention. Those skilled in the art should understand that various modifications, changes or substitutions to the steps of the preferred embodiment are possible without departing from the spirit and scope of the present invention and its appended claims. Therefore, the present invention should not be limited to what is disclosed in the preferred embodiments and drawings.

Claims (10)

1. a bilingual Chinese-English sentiment dictionary building method, is characterized in that: comprise the following steps:

Step one, capture microblogging webpage, from webpage, collect the Chinese and English language material with Sentiment orientation, and from corpus, extract the high frequency vocabulary with Sentiment orientation add sentiment dictionary storehouse;

Step 2, the existing knowledge base of application are expanded described sentiment dictionary;

The microblogging language material that step 3, analysis capture, adds described sentiment dictionary by emerging for network language and emoticon.

2. the bilingual Chinese-English sentiment dictionary building method of one according to claim 1, is characterized in that: described Sentiment orientation comprises the social help, happiness, sadness, indignation and frightened 5 classes.

3. the bilingual Chinese-English sentiment dictionary building method of one according to claim 1, it is characterized in that: the expansion of step 2 is the average similarity by calculating each Sentiment orientation vocabulary in emotion vocabulary and sentiment dictionary in each knowledge base respectively, and emotion word is extended in the maximum Sentiment orientation classification of similarity; Described knowledge base comprises WordNet, NTUSD and HowNet.

4., according to the arbitrary described bilingual Chinese-English sentiment dictionary building method of one of claim 1-3, it is characterized in that: adopt the mode of many people handshow to classify to its Sentiment orientation to described netspeak and emoticon.

5., based on a multiclass sentiment analysis method for bilingual dictionary, the method comprises the following steps:

Step one, pre-service is carried out to language material text;

Step 2, according to described bilingual Chinese-English sentiment dictionary, feature space expression is carried out to described language material text;

The many disaggregated models of text emotion that step 3, basis have been set up carry out emotional semantic classification to language material text.

6. a kind of microblogging multiclass sentiment analysis method based on bilingual dictionary according to claim 5, is characterized in that: described pre-service comprises participle further and removes stop words, also comprises morphology standardization for English text.

7. a kind of microblogging multiclass sentiment analysis method based on bilingual dictionary according to claim 5, it is characterized in that: described text feature space representation is that each text in language material is expressed as five dimensional vectors, and in vector, each element represents the number of the emotion word of corresponding classification in the described bilingual Chinese-English sentiment dictionary comprised respectively.

8. a kind of microblogging multiclass sentiment analysis method based on bilingual dictionary according to claim 5, is characterized in that:

The many disaggregated models of described emotion are semi-supervised gauss hybrid models sorting algorithm or the k nearest neighbor algorithm based on symmetric relative entropy;

Described semi-supervised gauss hybrid models sorting algorithm is the corpus collection study gauss hybrid models by having marked, then iterative learning is carried out to the testing material collection marked, until algorithm convergence or unlabeled set are combined into sky using the probability distribution of this model parameter and marker samples as the initial parameter values of gauss hybrid models;

The described k nearest neighbor algorithm based on symmetric relative entropy is the distance adopting relative entropy to measure to express text to text emotion similarity, and the classification according to adjacent sample decides sample generic to be sorted.

9. a kind of microblogging multiclass sentiment analysis method based on bilingual dictionary according to claim 8, is characterized in that: described relative entropy adopts following formula to calculate:

Wherein, T _ifor the normalized vector of retrtieval represents, T _jfor the normalized vector of unmarked text represents, ω _ik, ω _jkrepresent T respectively _i, T _jkth item, k is the integer between 1 to 5.

10. towards a multiclass sentiment analysis system for bilingual microblogging text, it is characterized in that: comprise bilingual Chinese-English sentiment dictionary, language material pretreatment module, language material text feature space representation module and emotion classifiers identification module four modules;

Bilingual Chinese-English sentiment dictionary adopts bilingual Chinese-English sentiment dictionary building method as claimed in claim 1 to build;

Language material pretreatment module is used for carrying out participle herein to language material to be analyzed and going stop words process, also will carry out morphology standardization processing for English text;

Language material text feature space representation module is used for carrying out vectorization expression to the text after the process of language material pretreatment module, be five dimensional vectors by text-processing, five elements in vector represent in text the number being included in the social help in described bilingual Chinese-English sentiment dictionary, happiness, sadness, indignation and frightened five class emotion word respectively;

Emotion classifiers identification module carries out emotion recognition for adopting emotion classifiers model as claimed in claim 8 to language material text vector, determines the emotion classification belonging to language material text.