CN105912625A - Linked data oriented entity classification method and system - Google Patents

Abstract

The invention discloses a link data-oriented entity classification method and system, aiming at the link data entity classification problem, including preprocessing, statistical classification and postprocessing; wherein, the preprocessing is through word segmentation of text description information in entity pages ; The word information obtained from the attribute name of the information box and word segmentation constitutes the feature of the entity page; the statistical classification process uses a variety of segmentation granularity to train the statistical classification model to classify the entity page, and obtains the preliminary prediction result of the entity category; Statistical classification results are corrected, including model fusion, language knowledge, link information, and the use of category-associated attribute information to correct the fused entity category. The technical solution of the invention is easy to implement, easy to debug, high in efficiency and good in precision, and is suitable for linking data for knowledge management; it can realize high-precision classification of entities.

Description

A Linked Data Oriented Entity Classification Method and System

technical field

The invention belongs to the field of information processing, relates to link data classification and search, in particular to a method and system for high-precision classification of entity pages in link data.

Background technique

Currently in the era of big data, how to maximize the use of data to help computers perform information processing has become the hottest research topic in the field of information processing. In recent years, with the advent of the Web 2.0 era, linked data (such as semantic web, knowledge graph, etc.) has attracted widespread attention because of its powerful relationship description capabilities. Linked data refers to data organization forms such as Baidu Encyclopedia and Wikipedia. In this kind of data, each page corresponds to an entity, and there are mutual links between entities, so it is called linked data. With the continuous increase of data scale, it is unrealistic to use manual methods to manage linked data, and there is an urgent need for efficient methods and systems for knowledge management of linked data.

Entity classification of linked data is an important technical issue in the field of knowledge management of linked data. Entity classification for linked data can effectively organize a large number of entity pages in linked data, thereby enhancing user search and reading experience.

At present, the common method of entity classification is to classify the descriptive text of the entity. However, this simple method cannot accurately analyze the category of entities in many cases, and its shortcomings are mainly manifested in:

(1) For humans, although it is easy to judge entity categories based on text descriptions, it is not realistic for current feature-based statistical classification methods to accurately judge entity categories through text descriptions ; for example, the text "X is an animation based on a famous game" and "A is a game based on a famous animation" have very similar representations at the lexical level, but the former is a description of an animated entity while the latter is a description of a game entity Description, which describes a completely different entity type. Therefore, the statistical classification method based solely on text features has insufficient recognition accuracy and cannot accurately obtain entity categories.

(2) Many entity pages do not have enough text description information. In this case, simply using text description information to classify entities will inevitably lead to classification errors, and the entity category cannot be obtained through text description.

Contents of the invention

In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a link data-oriented entity classification method and system, aiming at the entity classification problem of link data, the purpose of high-precision entity classification is achieved through statistical classification process and post-processing process; wherein , the statistical classification process classifies by modeling text information; the post-processing process uses rich resources (such as affix information, link data, etc.) to correct the results of entity statistical classification, including model fusion, language knowledge, link information, and utilization Methods such as modifying the fused entity category by category association attribute information.

Entity pages in Linked Data usually contain text descriptions and infoboxes. After the text description is segmented by the present invention, the attribute name of the infobox (infobox) and the word information obtained by segmenting are extracted as features and used as the feature representation of the entity page; Classify by granularity to obtain a preliminary prediction of the entity category; then post-process the obtained entity category to verify whether the classification result is reliable; post-processing specifically includes classification of classifiers trained using features of different segmentation granularities The results are fused; use the category-associated attribute information in the category attribute database to correct obvious prediction errors; deeply understand the first sentence of the text description, use grammatical analysis and other methods to analyze the sentence structure, and obtain entity category information to correct the previous prediction results. ; Preferably, the confusion matrix can also be used to identify categories that are difficult to classify correctly, and the prediction of categories that are difficult to correctly classify can be further verified, including using the category of the adjacent page linked by the entity page to correct the entity category and using the category of the entity page The affix information modifies the entity class.

The technical scheme provided by the invention is:

A link data-oriented entity classification method, the link data is a plurality of entity pages, the entity pages include text descriptions and information boxes; the entity classification method includes a preprocessing stage, a statistical classification stage and a postprocessing stage, specifically Including the following steps:

1) In the preprocessing stage, word information is obtained by segmenting the text description information in the entity page; the attribute name of the information box and the word information form the characteristics of the entity page;

2) In the statistical classification stage, using the characteristics of the entity page, a variety of segmentation granularities are used to train the statistical classification model to classify the entity page, and obtain the preliminary prediction result of the entity category;

3) In the post-processing stage, the preliminary prediction result of the entity category is corrected to obtain the corrected entity classification category; the correction includes the following steps:

31) Through the multi-granularity model fusion method, the preliminary prediction results of the entity category obtained by the statistical classification model trained with different segmentation granularities are fused to obtain the fused entity category result;

32) Constructing a category attribute database, using the category-associated attribute information in the category-attribute database to correct the fused entity category, and obtain the entity category after the category-associated attribute correction;

33) Analyzing the sentence structure by means of grammatical analysis, and obtaining the entity category information corrected by the in-depth understanding of the first sentence through the corrected entity category of the category association attribute obtained in step 32) of the deep understanding of the first sentence of the text description.

For the above link data-oriented entity classification method, further, the word segmentation method in step 1) includes a front-back maximum matching method, a backward maximum matching method, and a statistical sequence-based tagging method.

For the above link data-oriented entity classification method, further, step 2) adopts two segmentation granularities, namely the segmentation granularity with named entity recognition and the segmentation granularity without named entity recognition.

For the above link data-oriented entity classification method, further, the statistical classification model is a maximum entropy model; step 31) the multi-granularity model fusion method is specifically calculated by formula 1 to obtain the probability distribution predicted by the fusion of different segmentation granularity classifiers , the maximum entropy classification model trained by multiple segmentation granularities is used to classify entity pages to obtain entity category results for fusion:

P _multi (y|x)＝λP _w (y|x)+(1-λ)P _n (y|x) (Formula 1)

In Equation 1, P _multi (y|x) is the probability distribution predicted by the fusion of different segmentation granularity classifiers; P _w (y|x) is the probability distribution predicted by the maximum entropy classification model for sample x using only word segmentation as a feature ; y is the sample category, x is the sample; P _n (y|x) is the probability distribution of the maximum entropy prediction based on word segmentation and adding named entity annotation as a feature; λ is a parameter to adjust the linear interpolation weight.

For the above-mentioned link data-oriented entity classification method, further, step 33) utilizes the grammatical analysis method to analyze the sentence structure, and obtains the entity category information after the depth understanding of the first sentence, which specifically includes the following steps:

331) Perform dependency syntactic analysis on the first sentence of the entity description, and identify whether the object of the first sentence belongs to the object of the judgment sentence;

332) Training Chinese word vectors on a large-scale unlabeled corpus, defining the lexical semantic similarity, calculating the lexical semantic similarity between the word vector and the judgment sentence object, and obtaining the word vector with the highest lexical semantic similarity;

333) The cosine similarity calculation method is used to set the cosine similarity threshold. When the cosine similarity between the object of the judgment sentence and the word vector of the most similar category is greater than the cosine similarity threshold, the category of the entity is corrected to the most similar category.

For the above link data-oriented entity classification method, further, in the post-processing stage, the preliminary prediction result of the entity category is corrected, and after the corrected entity classification category is obtained, the confusion matrix is used to identify the difficult entity category; Difficult entity category of the entity category, the entity category result is verified by the link analysis method and the affix analysis method; the confusion matrix identification method is specifically: on the verification set, when the prediction accuracy of the statistical classification model for a certain entity category y _i does not reach 90% of the time, class y _i is considered a difficult entity class.

Further, the link analysis method specifically includes: setting the category prediction made by the classifier for the entity page e as y', recording the set of entity pages linked by the entity page e as N(e), and finding N For the pages marked with categories in (e), the category with the most pages marked with categories in N(e) is counted, which is recorded as y*; when the category y* is inconsistent with the category prediction y', use y* to correct y' As a result, the category of entity page e is y*.

For the above link data-oriented entity classification method, further, the affix analysis method is specifically: for the entity category whose entity name ends with a fixed Chinese character, using the affix information associated with the entity type learned from large-scale unlabeled data, through Perform frequency statistics on the affixes of the most similar vocabulary to obtain the affixes associated with the difficult entity category, and obtain the category of the entity by analyzing the affixes.

The present invention also provides a link data-oriented entity classification system realized by the link data-oriented entity classification method, including a preprocessing module, a statistical classification module and a postprocessing module; the preprocessing module is used to describe the text in the entity page The information is word-segmented, and the word information obtained by the information frame attribute name and word segmentation is extracted as a feature, and used as a feature representation of the entity page; the statistical classification module trains the classification model by using the maximum entropy classification algorithm, and utilizes the description of the entity in the entity page The entity category is obtained through information identification; the post-processing module is used to correct the entity category obtained by the statistical classification module by adopting multi-granularity model fusion, category association attributes, and first sentence in-depth understanding to obtain the corrected entity category.

In the above link data-oriented entity classification system, the word segmentation tool is the Stanford CoreNLP toolkit; the classification model uses the maximum entropy classifier software package Maxent.

Compared with prior art, the beneficial effect of the present invention is:

The present invention provides a link data-oriented entity classification method and system. Aiming at the link data entity classification problem, the purpose of high-precision entity classification is achieved through a statistical classification process and a post-processing process. Among them, on the basis of the basic classification of the text, the result of the classification of the entity description text is corrected, and the methods used include:

(1) Using a multi-granularity word segmentation model fusion method to overcome the defects of single segmentation granularity in text feature extraction;

(2) Use the category association attribute information to correct the fused entity category to achieve the purpose of correcting obvious errors;

(3) Through in-depth understanding of the first sentence, the effect of reducing text noise is achieved;

(4) Be able to identify difficult samples, and use methods such as link analysis and affixes to verify the identification results.

Compared with the existing technology, the current existing entity classification method does not process any more, and the result cannot be corrected for the possible error of the entity recognition classification; and the present invention corrects the possible error of the text-based statistical classification module through the post-processing process . The technical solution proposed by the invention is easy to implement, easy to debug, high in efficiency and good in precision, and is very suitable for enterprises to link data for knowledge management; it can classify entities with high precision. In the JIST2015 entity classification evaluation competition, the accuracy rate of the scheme of the present invention was 98.6%, which was the classification scheme with the highest accuracy rate in the current evaluation competition.

Description of drawings

Fig. 1 is a flow chart of the link data-oriented entity classification method provided by the present invention.

Fig. 2 is a structural block diagram of a linked data-oriented entity classification system provided by an embodiment of the present invention.

Fig. 3 is a flow chart of the first sentence in-depth understanding step in the method provided by the present invention.

detailed description

Below in conjunction with accompanying drawing, further describe the present invention through embodiment, but do not limit the scope of the present invention in any way.

The present invention provides a link data-oriented entity classification method and system. Aiming at the link data entity classification problem, the purpose of high-precision entity classification is achieved through a statistical classification process and a post-processing process; The post-processing process utilizes rich resources (such as affix information, link data and other information) to correct the results of entity statistical classification. Fig. 1 is a flow chart of the entity classification method for link data provided by the present invention. As shown in FIG. 1 , the method of the present invention includes a preprocessing process, a statistical classification process and a postprocessing process; firstly, word segmentation feature extraction is performed on the entity page, and then the statistical classification model is trained using the extracted features. For the classification results, we first use multi-granularity model fusion to correct single-model prediction errors, and then use category-associated attribute information to correct the fused entity categories to correct some obvious mispredictions. Sentence descriptions are analyzed in depth to determine their categories. For samples of some categories that are difficult to classify correctly, the present invention can re-correct their categories through link analysis and affix analysis methods. Specific steps include:

1) Preprocessing the entity page, including Chinese word segmentation (typical word segmentation methods include front and rear maximum matching, backward maximum matching, and methods based on statistical sequence annotation), feature extraction (extracting word features and entity information box attribute name features to the page to represent), etc., to obtain the entity page features;

2) Using the features of the entity page extracted in step 1), the maximum entropy model is used to classify the entity page with multiple segmentation granularities, and a preliminary prediction of the entity category is obtained;

In the embodiment of the present invention, the maximum entropy model is used to train two classifiers; the feature representation of one classifier uses the word+infobox attribute with named entity recognition granularity segmentation; the other classifier uses no named The word and infobox attributes generated by the segmentation performed by entity recognition.

3) Post-processing the entity category obtained in step 2), verifying whether the classification result is reliable; specifically including the following steps:

31) Fusing classification results of classifiers trained using features of different segmentation granularities;

In the embodiment of the present invention, two segmentation granularities are adopted, respectively referring to segmentation with named entity recognition and without named entity recognition;

32) Constructing a category attribute database in advance, using the category-associated attribute information in the category attribute database to correct obvious prediction errors;

33) Use a syntax analyzer to deeply understand the first sentence of the text description, and use syntax analysis and other methods to analyze the sentence structure, so as to obtain entity category information to correct the previous prediction results;

34) Use the confusion matrix to identify categories that are difficult to classify correctly, and further verify the predictions of this category, including:

341) Correcting entity categories using categories of adjacent pages to which the entity page links;

342) Correcting the entity category by using the affix information of the entity page.

Fig. 2 is a structural block diagram of a linked data-oriented entity classification system provided by an embodiment of the present invention. The entity classification system of linked data includes a preprocessing module, a statistical classification module and a postprocessing module; further descriptions for each module are as follows:

preprocessing module

Entity pages in Linked Data usually contain text descriptions and infoboxes.

In the preprocessing module, we use the Stanford CoreNLP toolkit to segment the text description information in the entity page. In this embodiment, we adopt two different segmentation granularities: named entity recognition and non-named entity recognition. For example, under segmentation with named entity recognition, "Times Square, New York" will be considered as a word, while under segmentation without named entity recognition, the word will be segmented into "New York", "Times", "Square" three words.

After segmenting the Chinese text, we extract the attribute name of the infobox (infobox) together with the segmented word information as a feature, and use it as a feature representation of the entity page.

Statistical Classification Module

The present invention mainly uses the description information of entities in entity pages as the basis for judging entity categories. The present invention adopts the logarithmic linear model commonly used in the field of natural language processing—the maximum entropy classification algorithm to train the classification model. As mentioned in the preprocessing module, the features used in the statistical classification module include word features and information box attribute features; word features are classic word bag model feature representations; information box attribute features play a very important role in identifying the category of entities. For example, "Date of Birth" might also be associated with an entity of type Person.

In the text classification module, we use different granularity of word segmentation to train the text classification model, because in some cases, a segmentation granularity cannot meet the classification requirements. For example, "New York Times Square" as a named entity is not as useful for classification as splitting it into "New York", "Times" and "Square", because the word "Square" is crucial to the category. influences. On the other hand, if we do not perform named entity recognition, then something like "Zhang Yishan" will be divided into "Zhang", "Yi" and "Mountain", which will also affect the classification results. Therefore, in the statistical classification module, the embodiment of the present invention uses the maximum entropy classifier software package Maxent (the maximum entropy classifier software package can be downloaded by the following link website: http://homepages.inf.ed.ac.uk/lzhang10/maxent_toolkit .html) trained two classification models, one is fine-grained segmentation with named entity recognition, and the other is simple coarse-grained word segmentation.

Post-processing module

Based on the fact that the text statistical classification module may be wrong, the present invention uses a post-processing module to make corrections. The post-processing module can perform the following processes:

31) Multi-granularity model fusion process

Although model fusion is widely used in the field of machine learning, most of the methods of model fusion are aimed at the fusion of different machine learning models. For natural language (especially Chinese), the difference in segmentation granularity will have an impact on the effect of the entire model. Aiming at the advantages and disadvantages of different segmentation granularities, the present invention proposes a method of model fusion to "learn from each other" the classification models obtained by various segmentation granularities.

We define P _w (y|x) as the probability distribution of category y predicted by the maximum entropy classification model for sample x using only word segmentation as a feature, and P _n (y|x) as adding named entity annotations based on word segmentation Probability distribution for maximum entropy predictions as features. We fuse the results of these two classifiers in the following way:

P _multi (y|x)＝λP _w (y|x)+(1-λ)P _n (y|x) (Formula 1)

In Equation 1, P _multi (y|x) is the probability distribution predicted by the fusion of different segmentation granularity classifiers; P _w (y|x) is the probability distribution predicted by the maximum entropy classification model for sample x using only word segmentation as a feature ; y is a sample category, and x is a sample; P _n (y|x) is the probability distribution of adding named entity labels as the maximum entropy prediction of features on the basis of word segmentation; λ is a parameter for adjusting the linear interpolation weight, the present embodiment , let λ=0.5.

32) Category association attribute correction prediction

This module fixes some obviously wrong class predictions by exploiting class association properties. This module exploits mainly the category specificity of the infobox attributes. As shown in Table 1, for some attributes, it is impossible for them to be associated with certain categories. For example, it is impossible for a "game platform" to be associated with a city entity. Thus, by exploiting the specificity of these attributes, apparent prediction errors of the classifier can be corrected. The present invention artificially establishes a category attribute database for the predefined entity type to correct the prediction.

Table 1 Example of category association attributes

33) Deeply understand the first sentence of the entity description through the dependency syntax analyzer, and further accurately identify the entity category;

The first sentence of the entity page description in linked data (eg: Wikipedia, Baidu Baike, etc.) is usually a qualitative description of the entity (eg: Shamiliujia is a poker game popular in Tianjin). If you can deeply understand the first sentence of the entity description, it will be of great help to accurately identify the entity category.

Fig. 3 is a flow chart of the first sentence in-depth understanding step in the method provided by the present invention. The present invention first utilizes the dependency syntax analyzer to find out the judgment sentence object in the first sentence of the entity page text description, and then utilizes the judgment sentence object to analyze the category of the entity page; specifically comprises the following steps:

331) Judgment sentence object recognition

The invention utilizes the dependency syntax analyzer of Stanford University to analyze the dependency syntax of the first sentence described by the entity, and analyze the subject, predicate and object in the first sentence. If the object of the first sentence obtained by dependency syntax has a direct dependence relationship with "is", then the object is called "the object of the judgment sentence"; otherwise, the object is called "the object of the non-judgment sentence".

If the object of the first sentence described in the entity text is the object of the judgment sentence, we can use the object as a clue to determine the category of the entity, thereby verifying whether the result predicted by the classifier is accurate. If the result predicted by the classifier contradicts the conclusion drawn by the sentence-segmented object, the result is used to correct the prediction of the classifier. If there is no judgment sentence object in the first sentence, then skip this step and enter 34).

For example, in the sentence "Smaliujia is a poker game popular in Tianjin", the result of dependency parsing analysis shows that "game" is the object of the sentence, and "game" has a direct dependency relationship with "is", then " Game" is the object of the sentence. If "game" is a predefined entity category in the entity taxonomy, then we use that as the entity's category.

332) Correcting Category Predictions Using Judgmental Sentence Objects

In some cases, even if we find out the object of the judgment sentence, it cannot be used to correct the prediction at will, because it may introduce some unnecessary errors. At the same time, in many cases, the object of the judgment sentence does not exactly match the category name. For example: "Masako Nozawa is a famous voice actor in Japan", although the dependency syntax analysis can get that "voice actor" is the judgment object of this sentence, but there may not be a category of "voice actor" in the predefined entity categories. In this regard, the present invention defines correction conditions, and utilizes lexical semantic similarity, that is, the cosine similarity between word vectors, to find the most similar category of judgment sentence objects from large-scale unlabeled corpus to reliably perform category correction.

In the field of natural language processing, cosine similarity is usually regarded as the semantic similarity of words. Specifically, the embodiment of the present invention first utilizes the word2vec toolkit (https://word2vec.googlecode.com/svn/trunk/) to train the Chinese word vector on the Gigaword Chinese corpus (Chinese Gigaword is a public data set), Use the word vectors obtained from the training to find the category name that is most similar to the semantics of the object of the judgment sentence. If the cosine similarity of the word vector of the judgment sentence object and its most similar category is greater than the preset threshold (in the embodiment of the present invention, by calculating the cosine similarity method, the cosine similarity threshold is set to 0.9), the entity The category of is corrected to the most similar category.

To this end, we define the object of the judgment sentence described in the first sentence of the entity page as w ₀ , the category word as y∈Y (Y is the set of entity categories), and sim(w ₁ , w ₂ ) as the words w ₁ and w ₂ Cosine similarity of vectors. Then the correction condition is formula 2 as follows:

y ^* ＝ _{argmax y∈Y} sim(w ₀ ,y)∧sim(w ₀ ,y ^* )>0.9 (Formula 2)

In formula 2, ^ represents the and (and) relationship; the content of the first part of ^ (the left item) indicates that y* is the category with the highest semantic similarity, and the content of the latter part of ^ (the right item) indicates that the similarity between y* and w0 requires It is higher than 0.9; the correction condition (Formula 2) is satisfied before the correction is made, that is, only when y* is the category with the highest semantic similarity and the similarity between y* and w0 needs to be higher than 0.9, use y* to correct the original category predict.

In the above example (“Masako Nozawa is a famous voice actor in Japan”), we can find out that the category most similar to “voice actor” is “actor” (as shown in Table 2, Table 2 is a word trained from Chinese gigaword Some words that are most similar to the category calculated by the vector, where the words in bold indicate that the similarity between these words and the category is above 0.9), and it is found that the semantic similarity between "actor" and "seiyuu" is above 0.9, so the " Nozawa Masako" entity page category was corrected to "actor".

Table 2 The most similar vocabulary of categories

34) Identifying Difficult Samples Using a Confusion Matrix

In practical applications, we often encounter some types of samples that are difficult to distinguish, and such samples are called difficult samples. For example, for two categories of entities "city" and "attraction", the classifier tends to make wrong predictions because the description and infobox attributes of these two types of entities are very similar. In order to improve the accuracy of classification, the present invention uses a confusion matrix to find out sample categories where classifiers are prone to errors. Specifically, if the prediction accuracy of the statistical classification model for a certain entity category y _i does not reach 90% on the validation set, the category y _i is regarded as a difficult sample category. For example, on the validation set, the statistical classification model predicts 18 entity pages as the "city" category, but only 15 of them are indeed the "city" category, so the prediction accuracy of the statistical classification model in the "city" category is only 83.33% % (15/18), the 'urban' category was identified as a hard sample category. For those samples that are predicted by the statistical classification model as the hard sample class, we call them hard samples.

For the identified difficult samples, we used the following two methods to verify the results.

341) Link Analysis

For difficult samples, relying solely on the content on the entity page may not be enough to make a correct judgment. Therefore, the present invention uses a link analysis method to verify the classification results of difficult samples.

In linked data, an entity page usually links to other entity pages related to it. Generally speaking, the category of other entity pages it links to is very likely to be the same as its own category. Therefore, using the categories of other entity pages linked to by an entity page can help the system better judge the category of the entity.

Specifically, for a certain entity page e, we analyze the entity pages linked by e, and its set is denoted as N(e). Some of the pages in N(e) have category annotation information. The present invention finds the pages marked with categories in N(e), counts the category y* with the most pages, and judges whether the category is consistent with the category prediction y' made by the classifier for e. If the results are inconsistent, use y* to correct the result of y'.

342) Affix Analysis

For some samples that are difficult to distinguish, the present invention also utilizes the affix analysis method to verify the classification results. For some categories, their entity names usually end with fixed Chinese characters. For example, a "City" attribute usually ends with "City, County", a "Attraction" attribute usually ends with "Lake, Mountain", etc. Table 3 lists examples of common entity affixes for categories.

Table 3 Category Affixes of Common Entities

attractions city Mountains, lakes, cities, roads, scenery, rivers, ridges, caves district, city, county

The present invention first proposes to use large-scale unlabeled data to learn the affix information associated with entity types. Specifically, we use the word vector toolkit word2vec to train word vectors on the Chinese Gigaword data set, and then find the Get the words with the most similar semantics in each category (words with word vector cosine similarity above 0.7). Then, by performing frequency statistics on the affixes of the most similar vocabulary of the two scenic spots, the affixes associated with the difficult sample category can be obtained, and the category to which it belongs can be determined by analyzing the affixes. Specifically, if the frequency of an entity page affix s in a category y ₁ is significantly higher (more than 2 times) in another category y ₂ , then we take y ₁ as the entity category to modify the original forecast result. For example, for the "Lushan Fairy Cave" entity page, the frequency of the affix "cave" appearing in the category of "attraction" is significantly higher than that in the category of "city", so the predicted category of the entity is corrected to "attraction" .

It should be noted that the purpose of the disclosed embodiments is to help further understand the present invention, but those skilled in the art can understand that various replacements and modifications are possible without departing from the spirit and scope of the present invention and the appended claims of. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the protection scope of the present invention is subject to the scope defined in the claims.

Claims (10)

1. A link data-oriented entity classification method, the link data is a plurality of entity pages, and the entity page includes text description and information box; the entity classification method includes a preprocessing stage, a statistical classification stage and a postprocessing stage , including the following steps: 1) In the preprocessing stage, word information is obtained by segmenting the text description information in the entity page; the attribute name of the information box and the word information form the characteristics of the entity page; 2) In the statistical classification stage, using the characteristics of the entity page, a variety of segmentation granularities are used to train the statistical classification model to classify the entity page, and obtain the preliminary prediction result of the entity category; 3) In the post-processing stage, the preliminary prediction result of the entity category is corrected to obtain the corrected entity classification category; the correction includes the following steps: 31) Through the multi-granularity model fusion method, the preliminary prediction results of the entity category obtained by the statistical classification model trained with multiple segmentation granularities are fused to obtain the fused entity category result; 32) Constructing a category attribute database, using the category-associated attribute information in the category-attribute database to correct the fused entity category, and obtain the entity category after the category-associated attribute correction; 33) Analyzing the sentence structure by means of grammatical analysis, and obtaining the entity category information corrected by the in-depth understanding of the first sentence through the corrected entity category of the category association attribute obtained in step 32) of the deep understanding of the first sentence of the text description. 2. The linked data-oriented entity classification method according to claim 1, wherein the word segmentation method in step 1) includes a front-to-back maximum matching method, a backward maximum matching method, and a statistical sequence labeling method. 3. The linked data-oriented entity classification method as claimed in claim 1, characterized in that, step 2) adopts two kinds of segmentation granularities, which are respectively the segmentation granularity with Named Entity Recognition and the segmentation without Named Entity Recognition. Sub-granularity. 4. the link data-oriented entity classification method as claimed in claim 1, is characterized in that, described statistical classification model is maximum entropy model; Step 31) described multi-granularity model fusion method specifically calculates by formula 1 and obtains and merges different segmentations The probability distribution predicted by the granular classifier is combined with the maximum entropy classification model trained by multiple segmentation granularity to classify the entity page to obtain the entity category result: P _multi (y|x)＝λP _w (y|x)+(1-λ)P _n (y|x) (Formula 1) In Equation 1, P _multi (y|x) is the probability distribution predicted by the fusion of different segmentation granularity classifiers; P _w (y|x) is the probability distribution predicted by the maximum entropy classification model for sample x using only word segmentation as a feature ; y is the sample category, x is the sample; P _n (y|x) is the probability distribution of the maximum entropy prediction based on word segmentation and adding named entity annotation as a feature; λ is a parameter to adjust the linear interpolation weight. 5. as claimed in claim 1, the entity classification method facing linked data is characterized in that, step 33) described utilizing syntax analysis method to analyze the sentence structure, obtaining the entity category information after first sentence depth understanding correction, specifically comprising the following steps: 331) Perform dependency syntactic analysis on the first sentence of the entity description, and identify whether the object of the first sentence belongs to the object of the judgment sentence; 332) Train Chinese word vectors on a large-scale unlabeled corpus, define lexical semantic similarity, calculate the lexical semantic similarity between the word vector and the judgment sentence object, and obtain the word vector with the highest lexical semantic similarity; 333) Through the cosine similarity calculation method, the cosine similarity threshold is set, and when the cosine similarity between the object of the judgment sentence and the word vector of the most similar category is greater than the cosine similarity threshold, the category of the entity is corrected to the most similar category. 6. as claimed in claim 1, the linked data-oriented entity classification method is characterized in that, in the post-processing stage, the preliminary prediction result of the entity category is corrected, and after obtaining the corrected entity classification category, the confusion matrix is used to identify Difficult entity category; for the identified difficult entity category, the entity category result is verified by link analysis method and affix analysis method; the confusion matrix identification method is specifically: on the verification set, when the statistical classification model for a certain entity category When the prediction accuracy of y _i does not reach 90%, the class y _i is regarded as a difficult entity class. 7. The link data-oriented entity classification method according to claim 6, wherein the link analysis method specifically comprises: setting the category prediction made by the classifier to the entity page e as y', and setting the entity page e The set of linked entity pages is recorded as N(e), find out the pages with category labels in N(e), and get the category with the most category labels in N(e), which is recorded as y*; when the category When y* is inconsistent with the category prediction y', use y* to correct the result of y', and obtain the category of entity page e as y*. 8. The link data-oriented entity classification method according to claim 6, wherein the affix analysis method is specifically: for the entity category whose entity name ends with a fixed Chinese character, the entity type obtained by using large-scale unlabeled data learning For the associated affix information, the affixes associated with the difficult entity category are obtained by performing frequency statistics on the affixes of the most similar vocabulary, and the category of the entity is obtained by analyzing the affixes. 9. The link data-oriented entity classification system realized by the link data-oriented entity classification method according to claims 1 to 8, is characterized in that, comprising a preprocessing module, a statistical classification module and a postprocessing module; The preprocessing module is used to perform word segmentation on the text description information in the entity page, and extract the word information obtained by the information box attribute name and word segmentation as a feature, and use it as a feature representation of the entity page; The statistical classification module trains the classification model by using the maximum entropy classification algorithm, and obtains the entity category by using the description information of the entity in the entity page to identify; The post-processing module is used to correct the entity category obtained by the statistical classification module by adopting multi-granularity model fusion, category association attributes and deep understanding of the first sentence to obtain the corrected entity category. 10. The linked data-oriented entity classification system of claim 9, wherein the word segmentation tool is a Stanford CoreNLP toolkit; and the classification model adopts a maximum entropy classifier software package Maxent.