CN110704624A - A multi-level and multi-label classification method for geographic information service metadata text - Google Patents

Abstract

The invention discloses a multi-level and multi-label classification method for geographic information service metadata text, including: 1) acquiring a geographic information service metadata text set to perform text preprocessing, and dividing each data sample into text feature word combinations; 2) Set a first-level classification catalog, and generate a typical vocabulary related to the classification category semantics; 3) Screen the text feature words according to the typical vocabulary; 4) Select ML-KNN as a base model for collaborative training; 5) Establish The topic prediction model ML-CSW is used as another base model for collaborative training; 6) Design a collaborative mechanism to match multi-label topics for metadata texts as a first-level coarse-grained topic classification result; 7) Select the metadata corresponding to a certain classification label text, get different levels of fine-grained topic category catalogs. The method of the invention considers the domain characteristics and text semantics of the metadata of the geographic information service, only relies on a small number of labeled data samples, and the classification result has better overall performance than the traditional multi-label classification method.

Description

A multi-level and multi-label classification method for geographic information service metadata text

technical field

The invention relates to natural language processing technology, in particular to a multi-level and multi-label classification method for geographic information service metadata text.

Background technique

As an important means of data analysis, accurate text classification is the key to improving the retrieval quality of geographic information resources, and has a wide range of application scenarios. Most of the traditional classification methods are suitable for binary or single classification scenarios, and they rely too much on a large number of labeled samples to train the classification model, which limits the accuracy and comprehensiveness of text classification and the applicable scenarios of the model. Especially for the metadata of geographic information services, there is usually a lack of sample datasets to annotate topics, and the text content is complex. This makes the metadata text topics have multi-granularity and multi-category characteristics, which further increases the difficulty of topic classification. In response to the lack of training samples and the need for multi-category matching, some scholars have proposed semi-supervised and weakly supervised mechanisms to reduce the dependence of classifiers on training samples, and some scholars have used ML-KNN, BR-KNN and TSVM to achieve more text Label classification. However, these methods usually do not combine domain characteristics, do not consider the semantics of technical terms in the text, and cannot effectively fit the textual characteristics of geographic information service metadata.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a multi-level and multi-label classification method for geographic information service metadata text in view of the defects in the prior art.

The technical scheme adopted by the present invention to solve the technical problem is: a multi-level and multi-label classification method for geographic information service metadata text, comprising the following steps:

1) Acquire a geographic information service metadata text set containing unmarked samples and marked samples for text preprocessing, and divide each data sample into text feature word combinations;

2) Define a first-level taxonomy based on the domain application theme category of geographic information resources, and generate a typical vocabulary closely related to the semantics of the taxonomy category (hereinafter referred to as "theme");

3) Screen the text feature words according to the typical word list, filter out the features whose distance from the typical word is greater than the threshold, and obtain the feature subsets screened according to the topic classification;

4) Select the classic multi-label classification algorithm ML-KNN (Multi-label K Nearest Neighbors) as a base model H ₁ for collaborative training;

5) Calculate the semantic distance from the feature to the subject according to the corpus, establish a subject prediction model ML-CSW (Multi-label Classification based on SWEET & WordNet), and use this model as another base model H ₂ for collaborative training;

6) Design a collaborative mechanism based on the above two base models to match multi-label topics for metadata texts as a first-level coarse-grained topic classification result;

7) According to the first-level coarse-grained topic classification result, select the metadata text corresponding to a certain classification label, extract the text topic, as the fine-grained topic of the next level, and obtain the matching relationship between the metadata text and the two-level topic catalog;

8) Step 7) is repeated to obtain fine-grained subject category catalogues at different levels, as well as the matching relationship between the metadata text and the subject catalogue.

According to the above solution, in step 2), the first-level classification catalogue of the domain application subject category definition based on geographic information resources is based on the expansion of the SBAs for the social benefit field proposed by the International Earth Observation Organization for the field of geosciences to obtain the first-level classification.

According to the above scheme, the typical vocabulary list generation method in the step 2) is as follows:

Taking SBAs as the subject classification catalogue, the hypernyms, hyponyms and synonyms of the subjects in the SWEET and WordNet definitions are extracted as typical words related to the semantics of the subjects, and a typical vocabulary list is generated.

According to the above scheme, in the step 3), the text feature words are screened according to the typical vocabulary list, and the details are as follows:

S31. Represent typical words and text feature words as two-dimensional space word vectors based on the Word2vec algorithm;

S32. Calculate the cosine distance between the typical word and the text feature word vector;

S33 , setting a distance threshold T, and filtering out text feature words whose cosine distance from typical words is greater than T.

According to the above scheme, the establishment method of the topic model in the step 5) is as follows:

According to the network definition of SWEET ontology library and WordNet English vocabulary network, calculate the semantic distance d _pi between the text feature f and each topic _pi

Find the minimum value of the semantic distance d _pi between the feature f and each topic pi, and find the maximum semantic correlation _{s f} between the text feature f and all topics P, where P is the set of all topics;

Define feature weights based on the shortest distance between text features and topics, establish topic prediction models, and predict multi-label topics for unlabeled samples;

Assuming that there are n text features in the training set, the vector S=[s ₁ ,s ₂ ,...,s _n ] can be calculated to obtain the maximum semantic relevance of all features in the training set to all topics, and the weight w of a single piece of data x can be calculated. (x) is defined as a 1×n vector, corresponding to the weights of n text features, if the feature f appears in the sample x, it is defined as s _f , otherwise it is defined as 0;

Establish a topic prediction model Y, where F is the adjustment vector of the feature, and α is the smoothing parameter. Based on the labeled sample data, BP neural network is used to iteratively optimize the training model Y, calculate the optimal solution of F and α under the condition of minimum loss, and obtain the final model, and predict the category set of unlabeled sample t according to the model;

Y=w(x)*F+α.

According to the above scheme, the step 6) design a collaborative mechanism to match the multi-label topic for the metadata text, as the first-level coarse-grained topic classification result; the details are as follows:

S61. Generate two subsets L ₁ and L ₂ according to the marked samples in the geographic information service metadata text set, which are respectively used as the training sets of the collaborative training base models H ₁ and H ₂ ;

S62, using the training set to train the base models H ₁ and H ₂ , and using the trained base model to predict the category vector of the unlabeled sample;

S63. Select the samples with the same prediction result by the classifiers H ₁ and H ₂ from the unlabeled samples and assign them to pseudo-labels, add the pseudo-labeled samples to the two training subsets L ₁ and L ₂ respectively, update the training set, and repeat the steps S62-S63, until the classification results of the two classifiers do not change significantly, obtain the category set of all unlabeled samples and the last updated training set;

S64: Train the classifier H ₁ based on all the labeled samples, and match the set of subject categories for the test samples.

According to the above scheme, in the described step 4), the classical multi-label classification algorithm ML-KNN is selected as a base model for collaborative training, and the details are as follows:

为1，

为0，反之则

为0，

为1；S41. Select the ML-KNN algorithm as the base model H ₁ for collaborative training, specify the number k of neighbor samples, and use N(x) to represent the set of k neighbor samples of the sample x in the training set, and count N(x) belonging to the subject category The number of samples c[j] of l, count the number of samples c'[j] in N(x) that do not belong to the topic category l. In the following formula, when the sample x belongs to the subject category l,

is 1,

0, otherwise

is 0,

is 1;

与后验概率其中b的取值为0和1，表示样本t属于主题类别l的事件，

表示样本t不属于主题类别l的事件，s为平滑参数，m为训练样本个数，

表示样本t的k个近邻样本中样本j属于类别l的事件；S42. Calculate the prior probability that the unlabeled sample t belongs to the topic category l

with the posterior probability where b takes the values 0 and 1, represents the event that the sample t belongs to the topic class l,

Indicates that the sample t does not belong to the event of the topic category l, s is the smoothing parameter, m is the number of training samples,

Represents the event that sample j belongs to category l in the k nearest neighbor samples of sample t;

S43. Predict the class set of the unlabeled sample t according to the maximized posterior probability and the Bayesian principle

According to the above scheme, the extraction of text topics in step 7) is based on the Latent Dirichlet Allocation (LDA) algorithm to extract text topics.

The beneficial effects of the present invention are as follows: the present invention proposes a new multi-level and multi-label classification process for OGC network map service WMS and other geographic information network resource metadata texts. This process introduces the geoscience ontology library SWEET and the general English vocabulary network WordNet into the classification process, and combines the traditional classification algorithm ML-KNN and the classification algorithm ML-CSW, which closely fits the domain characteristics and text semantics, for collaborative training to obtain geographic information service metadata. The matching relationship between the text and the multi-level topic directory. The method of the present invention considers the domain characteristics and text semantics of the metadata of the geographic information service, and only relies on a small number of labeled data samples; at the same time, compared with traditional multi-label classification algorithms such as classifier chains and voting classifiers, the classification results of the method of the present invention are as a whole. perform better.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings and embodiments, in which:

Fig. 1 is the method flow chart of the embodiment of the present invention;

Fig. 2 is the method flow chart of the embodiment of the present invention;

Fig. 3 is the typical word example diagram of the embodiment of the present invention;

4 is an example diagram of calculating the shortest distance between text features and topics in the ML-CSW algorithm according to an embodiment of the present invention;

Fig. 5 is the classification result of the example text of the embodiment of the present invention;

Fig. 6 is the classification result comparison of the different classification algorithms of the embodiment of the present invention;

FIG. 7 is a comparison of classification results based on different feature selection algorithms according to an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

There are currently 46,000 Web Map Service (WMS) text data, of which 400 are marked with SBAs topics, and each topic is evenly distributed. The text content comes from the URL, Abstract, Keywords and Title fields in the Service tag in the WMS GetCapability capability document. Due to the variety of text content and different lengths of texts, a single piece of data corresponds to multiple topic categories, and the amount of sample data for labeling topics is small, it is difficult for traditional multi-label classification algorithms to classify accurately and comprehensively, and it is impossible to obtain multi-level topics. match results.

Combined with the theoretical basis of collaborative training in semi-supervised learning, the invention introduces a geoscience ontology database and a general English vocabulary network to design a basic classification model that fits the characteristics of the geoscience field. In the classification process, the widely used classical multi-label classification model is used for collaborative training, and multi-level fine-grained topics are extracted to match multi-level multi-label topics for WMS metadata text.

Below in conjunction with the accompanying drawings in the present invention, the algorithm process of the present invention will be described in detail, as follows:

As shown in Figure 1 and Figure 2, a multi-level and multi-label classification method for geographic information service metadata text includes the following steps:

1) Perform text preprocessing on all WMS metadata, including three steps of word segmentation, removal of stop words and morphological restoration, and divide each text into text feature word combinations;

2) The first-level classification is based on the expansion of the social benefit areas (SBAs) proposed by the International Earth Observation Organization (Group on Earth Observations, GEO) for the field of geosciences. SBAs include 9 major topics of interest, including agriculture (Agriculture) , Biodiversity, Climate, Disaster, Ecosystem, Energy, Health, Water and Weather, etc. SBAs are based on international earth observation The social benefit areas (SBAs) proposed by the Group on Earth Observations (GEO) for the field of geosciences, including 9 major topics of interest, including Agriculture, Biodiversity, Climate, Disasters (Disaster), Ecosystem, Energy, Health, Water and Weather, etc. The subject classification catalogue of this embodiment is extended on the basis of SBAs, and Geology is added as the tenth subject, so all subject classification catalogues and first-level subject classification catalogues involved in this embodiment refer to this 10 themes.

Taking SBAs as the subject category catalogue, the hypernyms, hyponyms and synonyms of the subjects in the definitions of SWEET and WordNet are extracted as typical words related to the semantics of the subject, and a typical vocabulary is generated. Figure 3(a) is the subject extracted from SWEET" An example of a typical word corresponding to "Agriculture", Figure 3(b) is an example of a typical word corresponding to the topic "Agriculture" extracted from WordNet, and different colors represent different sets of word meanings;

3) The CBOW model based on the Word2vec algorithm expresses typical words and text feature words as two-dimensional space word vectors, and calculates the cosine distance between the typical words and text feature word vectors;

4) Set the distance threshold, screen the text feature words based on the distance threshold, and filter out the features whose distance from the typical word is greater than the threshold, thereby obtaining a feature subset that contributes more to the topic classification, as the model input of the classification algorithm;

5) Design a multi-label classification algorithm ML-CSW that fits the characteristics of the WMS domain and considers text semantics, as a collaborative training base model H ₁ , and uses the corpus to calculate the semantic correlation between text features and topics as feature weights to train the topic prediction model:

5.1) The network definition of SWEET is the main, and WordNet is the auxiliary to calculate the shortest semantic distance between the text feature and the topic;

If the text feature word is included in SWEET, the shortest distance between the feature word and the topic is crawled according to the network definition of SWEET, as shown in Figure 4(a), the feature "Glacier (glacier)" to the topic "Water (water)" The distance between them is 3;

If the text feature is not included in SWEET, look up the hypernym in WordNet layer by layer as the substitute word for the text feature, until the substitute word included in SWEET is found, and calculate the shortest distance D ₁ from the feature to the substitute word in the WordNet definition. , as shown in Figure 4(b), the substitute for the feature "Neve (grain snow)" is "Ice (ice)", and the shortest distance is 1. Then, according to the network definition of SWEET, the shortest distance D ₂ between the substitute word and the topic is calculated based on the Dijkstra algorithm. As shown in Figure 4(b), the shortest distance from the substitute word "Ice" to the topic "Water" is 2 . The final distance between the text feature and the topic is the sum of the distance from the text feature to the substitute word and the distance from the substitute word to the topic, that is, D=D ₁ +D ₂ , as shown in Figure 4(b), the feature "Neve (grain snow)" to The shortest distance for the topic "Water" is 3.

5.2) Define feature weights based on the shortest distance between text features and topics, establish a topic prediction model, and predict multi-label topics for unlabeled samples;

对最短距离求导作为文本特征f与所有主题P的最大语义相关度s_f，其中P为所有主题集合；a) According to step 5.1), the semantic distance between the text feature f and each topic p _i can be calculated

The derivation of the shortest distance is taken as the maximum semantic relevance s _f between the text feature f and all topics P, where P is the set of all topics;

b) If all texts contain n text features, the maximum semantic relevance vector S=[s ₁ , s ₂ ,...,s _n ] from all features in the training set to all topics can be calculated. The weight w(x) of a single piece of data x is defined as a 1×n vector, corresponding to the weights of n text features respectively. If the feature f appears in the sample x, it is defined as s _f , otherwise it is defined as 0.

c) Establish a topic prediction model Y, where F is the adjustment vector of the feature, and α is the smoothing parameter. Based on the labeled sample data, the BP neural network is used to iteratively optimize the training topic prediction model, and the optimal solution of F and α under the condition of minimum loss is calculated to obtain the final model. According to the model, the category set of the unlabeled sample t can be predicted;

Y=w(x)*F+α

6) Select the widely used classical multi-label classification algorithm ML-KNN as the collaborative training base model H ₂ :

为1，

为0，当样本x不属于主题类别l时，

为0，

为1；Specify the number of neighbor samples k, and use N(x) to represent the k neighbor sample set of the sample x in the training set L ₁ , count the number of samples c[j] belonging to the topic category l in N(x), and count N(x) The number of samples c'[j] that do not belong to topic category l in . In the following formula, when the sample x belongs to the subject category l,

is 1,

is 0, when the sample x does not belong to the subject category l,

is 0,

is 1;

与后验概率

其中s为平滑参数，m为训练样本个数，表示事件样本t属于主题类别l，

表示事件样本t不属于主题类别l，

表示事件样本t的k个近邻样本中实例j属于类别l；Calculate the prior probability that an unlabeled sample t belongs to topic class l

with the posterior probability

where s is the smoothing parameter, m is the number of training samples, means that the event sample t belongs to the topic category l,

indicates that the event sample t does not belong to the topic category l,

Indicates that instance j in the k nearest neighbor samples of event sample t belongs to category l;

Predict the class set of the unlabeled sample t according to the maximized posterior probability and Bayes' principle

7) 80% repeated random sampling of all labeled samples is divided into two subsets, L ₁ and L ₂ , which are used as the training sets of classifiers H ₁ and H ₂ respectively, and the two classifiers are used to predict the category set of all unlabeled samples;

8) Select the samples with the same prediction results of the classifiers H ₁ and H ₂ to give pseudo-labels, add the pseudo-labeled samples to the two training subsets L ₁ and L ₂ respectively, update the training set, and repeat 7) until two classifications There is no obvious change in the classification results of the classifier, and the category set of unlabeled samples is obtained.

9) Using 10% of all labeled samples as test samples, use the trained classifier to match the set of subject categories for the test samples, as shown in Figure 5, the SBA category labels of the sample text include Biodiversity, Climate, Disaster, Ecosystem, Water and Weather.

10) Specify the number of topic layers N. For each layer, select the metadata text of a single topic category, and extract the fine-grained topics of the text based on the Latent Dirichlet Allocation (LDA) algorithm until an N-layer topic directory is generated, which is: The WMS metadata text matches N-layer themes. As shown in Figure 5, the secondary themes corresponding to Biodiversity are wildlife, specie and diversity, the secondary themes corresponding to Climate are forest and meteorology, the secondary themes corresponding to Disaster are pollution, and the secondary themes corresponding to Ecosystem The first-level themes are habitat, resource, and conserve, the second-level theme corresponding to Water is rain, and the second-level theme corresponding to Weather is meteorology.

The present invention considers the domain characteristics and textual semantics of geographic information service metadata, and only relies on a small number of labeled data samples; as shown in FIG. 6, compared with traditional multi-label classification algorithms such as classifier chains and voting classifiers, the method of the invention has The classification results perform better overall.

As shown in FIG. 7 , compared with the chi-square test and the WordNet-based feature selection method, the text feature selection process of the present invention can filter out features that do not contribute to the classification result. The method of the invention can be applied to geographic information portals and data catalog services, and assists retrieval and discovery of various geographic information resources.

It should be understood that, for those skilled in the art, improvements or changes can be made according to the above description, and all these improvements and changes should fall within the protection scope of the appended claims of the present invention.

Claims (8)

1. a multi-level multi-label classification method for geographic information service metadata text, is characterized in that, comprises the following steps: 1) Acquire a geographic information service metadata text set containing unmarked samples and marked samples for text preprocessing, and divide each data sample into text feature word combinations; 2) Set a first-level classification directory based on the domain application theme category of geographic information resources, obtain the classification category, that is, the theme, and then generate a typical vocabulary related to the classification category semantics; 3) Screen the text feature words according to the typical word list, filter out the features whose distance from the typical word is greater than the threshold, and obtain the feature subsets screened according to the topic classification; 4) Select the classic multi-label classification algorithm ML-KNN as a base model for collaborative training, denoted as H ₁ ; 5) Calculate the semantic distance from the feature to the topic according to the corpus, establish a topic prediction model ML-CSW, and use this model as another base model for collaborative training, denoted as H ₂ ; 6) Design a collaborative mechanism based on the above two base models to match multi-label topics for metadata texts as a first-level coarse-grained topic classification result; 7) Select the metadata text corresponding to a certain classification label, extract the text subject as the fine-grained subject of the next level, and obtain the matching relationship between the metadata text and the double-layer subject directory; 8) Step 7) is repeated to obtain fine-grained subject category catalogues at different levels, as well as the matching relationship between the metadata text and the subject catalogue. 2. the multi-level multi-label classification method of geographic information service metadata text according to claim 1, is characterized in that, in the described step 2), based on the domain application subject category definition first-level classification of geographic information resources is based on international earth The observation organization expands the SBAs proposed in the field of geosciences to the social benefit field and obtains a first-level classification. 3. The multi-level multi-label classification method of geographic information service metadata text according to claim 1, is characterized in that, in described step 2), typical vocabulary list generation mode is as follows: Taking SBAs as the subject classification catalogue, the hypernyms, hyponyms and synonyms of the subjects in the SWEET and WordNet definitions are extracted as typical words related to the semantics of the subjects, and a typical vocabulary list is generated. 4. The multi-level and multi-label classification method of geographic information service metadata text according to claim 1, is characterized in that, in described step 3), according to typical word list, text feature word is screened, specifically as follows: S31. Represent typical words and text feature words as two-dimensional space word vectors based on the Word2vec algorithm; S32. Calculate the cosine distance between the typical word and the text feature word vector; S33 , setting a distance threshold T, and filtering out text feature words whose cosine distance from typical words is greater than T. 5. the multi-level multi-label classification method of geographic information service metadata text according to claim 1, is characterized in that, the establishment method of topic model in described step 5) is specifically as follows: S51. According to the network definition of the SWEET ontology library and the WordNet English vocabulary network, calculate the semantic distance between the text feature f and each topic p _i

If the feature f is included in SWEET, the semantic distance between the feature f and each topic p _i is obtained directly based on the Dijsktra algorithm according to the SWEET network. If the feature f is not included in SWEET, then look up the hypernym that is included in SWEET as the substitute word for the feature f, and compare the distance between the feature f and the substitute word in WordNet and the distance between the substitute word in SWEET and each topic p _i sum, as the semantic distance between feature f and each topic _pi S52. Calculate the semantic distance between the feature f and each topic p _i The minimum value of , and find the maximum semantic relevance s _f between the text feature f and all topics P, where P is the set of all topics;

S53. Define feature weights based on the shortest distance between text features and topics, establish a topic prediction model, and predict multi-label topics for unlabeled samples; S54. Assuming that the training set contains _n text features in total, the vector S=[s ₁ , s ₂ , . The weight w(x) is defined as a 1×n vector, corresponding to the weights of n text features respectively. If the feature f appears in the sample x, it is defined as s _f , otherwise it is defined as 0; S55. Establish a topic prediction model Y, where F is the adjustment vector of the feature, and α is a smoothing parameter; based on the labeled sample data, the BP neural network is used to iteratively optimize the training model Y, and the optimal solutions of F and α under the condition of minimum loss are calculated and obtained. The final model predicts the category set of unlabeled samples t according to the model; Y=w(x)*F+α. 6. The multi-level and multi-label classification method of geographic information service metadata text according to claim 1, is characterized in that, described step 6) design collaboration mechanism, for metadata text matching multi-label theme, as one-level coarse-grained theme Classification results; details are as follows: S61. Generate two subsets L ₁ and L ₂ according to the marked samples in the geographic information service metadata text set, which are respectively used as the training sets of the collaborative training base models H ₁ and H ₂ ; S62, using the training set to train the base models H ₁ and H ₂ , and using the trained base model to predict the category vector of the unlabeled sample; S63. Select the samples with the same prediction result by the classifiers H ₁ and H ₂ from the unlabeled samples and assign them to pseudo-labels, add the pseudo-labeled samples to the two training subsets L ₁ and L ₂ respectively, update the training set, and repeat the steps S62-S63, until there is no obvious change in the classification results of the two classifiers, obtain the category set of all unlabeled samples; S64: Train the classifier H ₁ based on all the labeled samples, and match the set of subject categories for the test samples. 7. The multi-level and multi-label classification method of geographic information service metadata text according to claim 1, is characterized in that, in described step 4), select classic multi-label classification algorithm ML-KNN as a base model of collaborative training, specifically as follows: S41. Specify the number k of neighboring samples, use N(x) to represent the set of k neighboring samples of the sample x in the training set, count the number of samples c[j] belonging to the topic category l in N(x), and count N(x) The number of samples c'[j] that do not belong to the subject category l in the following formula; when the sample x belongs to the subject category l,

is 1,

0, otherwise

is 0,

is 1; S42. Calculate the prior probability that the unlabeled sample t belongs to the topic category l

with the posterior probability

where b takes the values 0 and 1,

represents the event that the sample t belongs to the topic class l,

Indicates that the sample t does not belong to the event of the topic category l, s is the smoothing parameter, m is the number of training samples, Represents the event that sample j belongs to category l in the k nearest neighbor samples of sample t;

S43. Predict the class set of the unlabeled sample t according to the maximized posterior probability and the Bayesian principle

8 . The multi-level and multi-label classification method for geographic information service metadata text according to claim 1 , wherein the extraction of text topics in the step 7) is based on the hidden Dirichlet distribution algorithm to extract text topics. 9 .

Images