Abstract
We present a new and realistic problem, open-categorical text classification, which requires us to classify documents without the categorization system known beforehand. To solve this problem, we propose a novel approach to construct the categorization system and classify documents based on multi-latent Dirichlet allocation (LDA) models. We cluster topics and extract topical keywords to help category annotation. Subsequently, the LDA models are applied to predict the categories of documents comprehensively. Our result, a macro-averaged F1 measure of 84.02 %, outperforms the state-of-the-art supervised and semi-supervised text classification methods.
Similar content being viewed by others
Explore related subjects
Discover the latest articles, news and stories from top researchers in related subjects.Avoid common mistakes on your manuscript.
1 Introduction
Text classification (TC) is a task consisting in labeling automatically a document with a certain category, based on its content. Traditional TC classifies documents into predefined categories. Many machine learning-based approaches are exploited to deal with this task, such as Naive Bayesian (Kim et al. 2002), k Nearest Neighbors (Danesh et al. 2007), and Sport Vector Machine (Joachims 1998; Lin 2002; Donghui and Zhijing 2010; Qin and Wang 2009; Fu and Lee 2012). These methods need manually labeled data with predefined categories to tune parameters. However, in some practical applications, we do not know the categories beforehand or only know a part of the categories. For example, given a set of documents, we may not know which and how many categories are included in it. Because of this, it is difficult to label training data for supervised machine learning methods.
In this paper, we propose a novel approach based on latent Dirichlet allocation (LDA) model to deal with this problem. LDA is a generative probabilistic model for collections of discrete data such as text corpora. It is a three-level hierarchical Bayesian model, in which each item of a collection is modeled as a finite mixture over an underlying set of topics. Each topic is, in turn, modeled as an infinite mixture over an underlying set of topic probabilities. First, we apply LDA on the whole data set and get the latent topics which are represented as an allocation of words in the text. Subsequently, we select good topics based on the entropy of word distributional probabilities in the topics. We then cluster the good topics and extract topical keywords to help the categorization system construction. Our approach is easier than labeling documents one-by-one. Using the latter approach, we can only label limited training data because this approach is costly and time consuming. Moreover, some minority categories may be missed because no instance of these categories may be labeled. On the contrary, our approach can cover all categories because the topics are generated from the whole data. When the topics are clustered and labeled with category tags, we can estimate the topic distributions of a given document based on the LDA model and classify the document into the category of the top one topic.
A deficiency of LDA is that the number of topics need be set beforehand. If the number is too small, different categories may be mixed together in a topic, which makes it difficult to assign a single proper category to the topic. Therefore, the model cannot distinguish these categories. If it is too big, one category may be split up into several topics. The performance of LDA will be harmed. To avoid this deficiency, we employ multi-LDA models and induce and map the topics from different LDA models onto unified categories. Then, we use the multi-models to classify documents together. Our experiments show the proposed approach outperforms other supervised or semi-supervised approaches.
Our contributions are as follows:
-
We propose a new and realistic problem, open-categorical text classification (OCTC). It is different from the traditional TC problem that OCTC requires us to classify documents without the categorization system known beforehand. Hence, it is difficult to label training data for the supervised machine learning method which are widely used for traditional TC.
-
We propose an approach based on multi-LDA models to construct the categorization system and classify documents. The experiments show that our approach is effective.
2 Related work
2.1 Text Classification
In the last decades, TC is solved using supervised and semi-supervised classification algorithms. Multiple approaches to TC were presented using some well-known classifiers. The most widely used model for text categorization is the vector space model (VSM) (Gerard Salton et al. 1975). Under the model, a feature space is extracted based on a set of unique, uncommon and frequent words which are evaluated for each document. Each document is represented by a real-valued vectors with dimensions corresponding to unique words. When some of the documents’ actual categories are known and used for training, many well-known classifiers in supervised machine learning, such as support vector machines (SVM) (Joachims 1998; Lin 2002; Donghui and Zhijing 2010; Qin and Wang 2009; Fu and Lee 2012), k Nearest Neighbors (kNN) (Danesh et al. 2007), Naive Bayes (NB) (Kim et al. 2002), Decision tree (DT) (Johnson et al. 2002; Vateekul and Kubat 2009) and Neural Network (Ng et al. 1997; Trappey et al. 2006; Li and Park 2009), can then be applied to categorize documents.
As we all know, supervised methods need labeled corpora for training. If there is not enough training data, their numerous parameters cannot be learned well. Therefore some researchers propose semi-supervised methods to obtain competitive models with limited labeled data with a large amount of unlabeled data, such as bootstrapping methods. Bootstrapping methods are used to train a model on a small number of labeled data and predict the labels of unlabeled data. They iteratively expand the set of labeled data using high-confidence results and improve the performance by training a new model on the larger labeled data. Such methods have shown promise in applications such as web page classification (Blum and Mitchell 1998), named entity classification (Collins and Singer 1999), parsing (McClosky et al. 2006), machine translation (Ueffing 2006), and information extraction (Carlson et al. 2010).
Cheng et al. (2013) propose a method to improve SVM by bootstrapping unlabeled data with self-training. The SVM classifier is iteratively refined through the augmentation of the training set.
Supervised or semi-supervised methods need some labeled data with predicted categories. However, in our task, we do not know categories beforehand. Therefore, it is difficult to apply the methods to the task directly.
2.2 Topic models
Topic models are a type of statistical models for discovering the abstract topics that occur in a collection of documents. Well-known topic models include probabilistic latent semantic analysis (pLSA) (Hofmann 1999), latent Dirichlet allocation (LDA) (Blei et al. 2003b) and their varieties (Blei et al. 2003a; Blei and McAuliffe 2007; Petinot et al. 2011; Mao et al. 2012).
LDA is widely used for identifying the topics in a set of documents, building on previous work about pLSA by Hofmann (1999). It is an unsupervised algorithm. In this model, the topic distribution is assumed to have a Dirichlet prior. Each document is represented as a mixture of a mixed number of topics, with topic \(z\) receiving weight \(\theta _z^d\) in document \(d\). Each topic is a probability distribution over a finite vocabulary of words, with word \(w\) having probability \(\phi _w^z\) in topic \(z\).
In this paper, we apply LDA to help category system construction and document categorization.
3 Our approach
In this section, we first define the task formally. Then, we elaborate on our proposed approach composed of three major steps, namely, topic discovery, category system construction, and document categorization (Fig. 1).
3.1 Task definition
The goal of OCTC task is to classify documents without the categories known beforehand. The input of the task is only a set of unlabeled documents \(D=\{d_1, d_2,\ldots , d_M\}\). There, \(M\) denotes the total number of documents. We need first abstract and construct the categorization system \(C=\{c_1,c_2,\ldots , c_K\}\) from \(D\). \(K\) is the number of categories in all. When a new document \(d_i\) is given, the final goal is to assign a proper category \(c_j\) to it.
3.2 Topic discovery
To discover topics from messages, we choose to directly apply LDA. The basic idea of LDA is that documents are represented as random mixtures over latent topics, where each topic is characterized by a distribution over words. Our experiments show that we can obtain meaningful topics from our data set using standard LDA. We set multi-numbers of topics and ran 1,000 iterations of Gibbs sampling using the GibbsLDA++ toolkit.Footnote 1
3.3 Category system construction
After we obtain topics, we induce and annotate the categories using the distributions of words in topics (Fig. 1). In LDA, each topic is represented as a distribution over words as Table 1 shows. We apply multi-LDA models with different numbers of topics. However, manually annotating so many topics one-by-one is still time consuming. To solve the problem, we first cluster the topics and then extract keywords of each cluster to help category annotation. Specifically, to measure the similarity between two topics \(t\) and \(t^\prime \), we use JS-divergence between the word distributions of the two topics, denoted as \(\mathbf P _t\) and \(\mathbf P _{t^\prime }\). Each distribution vector is composed of distributional probabilities of words in the corresponding topic.
where word distributions are represented as vectors:
Each dimension corresponds to the distributional probability of a separate word on topic \(t\). \(L\) is the number of words in the vocabulary (the number of distinct words occurring in the corpus). \(\mathbf P _\mathrm{a}\) is the average word distribution of topic \(t\) and \(t^\prime \). That is \(p(w_i|a)=\frac{1}{2}p(w_i|t)+\frac{1}{2}p(w_i|t^\prime )\) for \(i\in \{1,2,\ldots ,L\}\). \(D_\mathrm{KL}\) is the KL-divergence which is a non-symmetric measure of the difference between two probability distributions.
While the JS-divergence has the advantage that it is symmetric.
However, not all of the topics are meaningful. Some of them are noisy topics such as the last example in Table 1. To remove them, we exploit an observation that most meaningful topics focus on a single theme. If the words in a topic distributed in a wide range, it is likely a noisy topic. We, therefore, defined a measure called topic entropy (TE) as follows:
The larger the \(\mathrm{TE}(t)\), the more likely \(t\) is a noisy topic. We remove topics whose \(\mathrm{TE}(t)\) is larger than a threshold (see 5.1.1).
We then apply K-means algorithm to cluster the topics from multi-LDA models. We choose the number of clusters, \(K\), using the method proposed by Pham et al. (2005). The method takes into account information reflecting the performance of the K-means algorithm.
Subsequently, we extract keywords from each cluster to help category construction. For example, the word “hospital” infers the current cluster being close to the category “health care”. To the contrary, the word “expert” does not. We rank the words in a cluster by their average distributional probabilities over the topics in the cluster. From the ranked words, we select the top \(N\) as the category tag candidates of the cluster. Finally, we manually check the candidates and merge similar clusters to get the final categories. Comparing to the approach annotating instances one-by-one, our approach can greatly improve work efficiency of category construction.
3.4 Document classification
After we annotate the category tags of topics, we can use the multi-LDA models to classify new documents. When a piece of text is given, we predict the topic distributions based on LDA models. Then, we compute the category distributions based on the topic distributions. Equation 5 shows a measure, called average measure, which simply computes average probability of all topics in a cluster as the probability of the category.
where \(p(c_j|d)\) denotes the probability that the given document \(d\) belongs to the jth category (a cluster) \(c_j \cdot t\) denotes a topic. \(Z\) is a normalizing factor ensuring that the result is a probability.
In the average measure, all topics in a category (a cluster) have equal weights. However, the distance between a topic and the centroid of a cluster can reflect the probability that it belongs to the cluster. Closer distance means higher probability. Therefore, we also propose a modified measure called weighted measure as follows:
where \(\mathrm{ID}_\mathrm{JS}(\mathbf V _t \parallel \mathbf V _{\overline{c}_j})\) denotes the inverse JS-divergence between \(\mathbf V _t\) and \(\mathbf V _{\overline{c}_j} \cdot \overline{c}_j\) denotes the centroid of cluster \(c_j \cdot Z^\prime \) is also a normalizing factor.
Finally, we select the category with the highest probability as the result of classification.
4 Experimental setup
4.1 Experimental data
In this work, the data are collected from the introductions of subscription accounts on WeChat,Footnote 2 a mobile text and voice messaging communication service in China. Subscription accounts are a kind of accounts which provide services and can be subscribed by users. Table 2 shows some instances.
We combine the title and introduction of an account as a short document, based on which we classify the account. Each document contains 11.74 words on average. We totally collect 985,397 subscription accounts for LDA learning and category construction. We then label 1,500 accounts with the constructed categories and randomly split them into 1/5 for development and 4/5 for testing.
The Chinese segmentation is provided by an open-source Chinese language processing platform LTP (Che et al. 2010).Footnote 3
4.2 Evaluation metrics
Two widely used metrics in TC to evaluate classifiers’ accuracy are the macro-averaged F1 score ( Yang 1999) and the micro-averaged F1 score ( Sebastiani 2002).
The macro-averaged F1 (\(F1_\mathrm{macro}\)) is computed locally over each category. It can be computed as a weighted average of the macro precision (\(P_\mathrm{macro}\)) and the macro recall (\(R_\mathrm{macro}\)).
where \(q\) denotes the total number of categories.
The micro-averaged F1 (\(F1_\mathrm{micro}\)) is computed globally over all category decisions.
where \(P_\mathrm{micro}\) and \(R_\mathrm{micro}\) denote the micro-averaged precision and recall, respectively. Actually, every instance in the test set needs to be classified. Therefore, the number of total instances predicted equals the number of instances in the test set. That is to say, \(P_\mathrm{micro}\) is the same as \(R_\mathrm{micro}\). \(F1_\mathrm{micro}\) is actually the same as accuracy (A).
For the sake of simplicity, we use the accuracy measure instead of micro-averaged F1 in the experiments.
To evaluate topical keywords extraction, we use Precision@N (P@N) for the ranked keyword lists.
In our example, \(N\) is set as 1, 3, 5, 10, and 30.
5 Results and discussion
In this section, we evaluate two aspects of our approach as follows: categorization system construction and document classification.
5.1 Evaluation of categorization system construction
5.1.1 Threshold of filtering out noisy topics
As Sect. 3.3 proposes, we use topic entropy to filter out noisy topics. We remove topics while their topic entropy is greater than a threshold \(\delta \). To tune \(\delta \), we ask two volunteers to distinguish whether the topics are noisy. We measure the inter-annotator agreement using the kappa coefficient. The kappa value is 0.81, which indicates a good strength of agreement. Figure 2 shows the evaluation results when we vary \(\delta \). When we increase \(\delta \), the F-scores of ”good” topics increase until \(\delta \) reaches 10.7. When \(\delta \) reaches 10.7, we obtain the best performance of topic filtering under different LDA models. It shows that the entropy of word distribution under a topic can reflect its quality. After topic filtering, about 60 % high-quality topics remain in our experiments, which can help to construct the categories.
5.1.2 Quality of topical keywords
To reduce the manual labor of category annotation, we cluster all high-quality topics from multi-topics and extract topical keywords for each cluster. These keywords can help annotators label categories of clusters easily. In our experiment, we use K-means algorithm to cluster topics and compute the average distribution of words in each cluster. Subsequently, we rank words in each cluster based on their distributional probabilities. The top words are selected as the topical keywords of the cluster.
When the two annotators label the category tags, we also ask them to judge whether the top keywords are helpful. We compute kappa value, a statistical measure of inter-annotator agreement for categorical items (Carletta 1996), on the evaluation data. The kappa value is 0.75, which also indicates a good strength of agreement. As Table 3 shows, 81.30 % of the top one keywords are helpful. When we look down to top ten keywords, the precision still reaches about 70 %. With the help of the keywords, annotators can induce categories more easily.
5.1.3 Number of LDA models
In the section, we analyze how many LDA models is proper for our task. We apply different numbers of LDA models to construct categories. The number of topics \(K\) is set from 150 to 600. The interval is 50. We thus train 10 models totally and label the category for each high-quality topic after noise filtering. Then, we try to combine the categories from different models gradually. For example, in Fig. 3, the 2-model combination means combining the categories from the LDA models with 150 and 200 topics. The 3-model combination means combining the categories from the LDA models with 150, 200 and 250 topics, and so on.
The result shows that the number of categories reaches the highest point when we combine eight LDA models. After that, the number or the coverage of categories do not increase while adding new models. We thus select eight models in the remaining experiments.
5.1.4 Distribution of categories
We annotate 83 categories in all (see “Appendix”). Table 4 shows their distributions on the test corpus, which are sorted from high to low. The distributions are non-uniform. The most frequent 20 categories cover nearly a half of documents (48.55 %) in the test corpus.
Moreover, about 12.36 % documents cannot be classified into one of the categories. That is because these documents are too short to contain enough information for classification.
5.2 Classification
5.2.1 Comparison between average distribution and weighted average distribution
In Sect. 3.4, we propose two methods, the average distribution (\(M_\mathrm{AD}\)) and the weighted average distribution (\(M_\mathrm{WAD}\)) method, to compute category distribution based on topic distribution for a given document. We compare the two methods in Table 4. \(M_\mathrm{WAD}\) outperforms \(M_\mathrm{AD}\) significantly (\(p < 0.01), \) Footnote 4 which shows that different topics have different importance in a category. The topics which are close to the centers of clusters have large weights, and therefore are important for classification. In contrast, the topics far away from the centers have little importance for classification.
5.2.2 Comparison to SVM
SVM is a representative supervised machine learning method for text classification. If we use it to our task, we should first annotate enough data for training. It is impossible to annotate all of the data because it is a costly job. However, the categories are distributed unevenly as Fig. 4 shows. If we randomly sample a subset of the data, few or even no samples of the minority categories may be selected. It may lead that SVM cannot handle these long-tailed categories.
Clustering data and then sampling instances from each cluster, respectively, may solve this problem. We apply the K-means method to cluster the 0.98 million documents. Each document is represented as the vector of words occurring in it. Each dimension corresponds to a separate word. Its value is term frequency inverse document frequency (TF-IDF) weighting of the corresponding word. We annotate the samples with the categories constructed based on LDA. However, in a real situation, the annotation is more difficult because the categories are unknown. Finally, spending the same amount of time with LDA topic annotation, we annotate about 30 thousand documents. Nevertheless, there are still nine categories absent in the training data. We train an SVM on these documents and tune the parameters on the development data. To avoid sparse data problems, we use the results of Brown clustering as features instead of words. Brown clustering is a form of hierarchical clustering of words based on the contexts in which they occur (Brown et al. 1992; Turian et al. 2010). The results in Table 4 shows that our multi-LDA-based approach outperforms SVM.
We see that the supervised SVM method can only annotate limited documents and cannot construct the complete category system. Moreover, because the documents in our data set are short, there are few features which can be used for classification. But our proposed multi-LDA approach can handle more unlabeled data to build complete category system and improve document classification.
5.2.3 Comparison to semi-supervised SVM
We also apply the semi-supervised SVM proposed by Cheng et al. (2013) to our task. We exploit 140 thousand unlabeled documents to augment training data and refine the SVM iteratively. We use the SVM to classify the unlabeled documents and select some instances to augment the training data for SVM retraining. In each iteration, 5 % instances with highest confidence are selected and added into SVM training data. The process stops until the performance does not been improved in the development set.
We can see from Table 4 that the semi-supervised methods can improve the performance of SVM. However, the best result is still significantly worse than the result of our proposed multi-LDA method. The reason may be that the semi-supervised method can augment training data but cannot supplement new categories. That is to say, if we have a small training data with 74 categories initially, we will also label new data with the 74 categories. While, our proposed method can construct a more complete category system.
6 Conclusion
In this paper, we propose a novel approach based on multi-LDA models to solve the new problem named open-categorical text classification. Because we do not know the categorization system beforehand, we first apply multi-LDA models to a large scale unlabeled corpus and obtain the topics. Then, we cluster the topics and extract topical keywords for each cluster. In the top ten extracted keywords, nearly seven are useful in average for categorization system construction as our experiments show. After we construct the categorization system, we actually get a projection from topics to category tags. Finally, we apply the multi-LDA models to classify documents. The experiments show that our approach outperforms the state-of-the-art supervised and semi-supervised SVM.
In the future, we will study how to add new categories to the old categorization system expediently, because some new topics maybe occur with the passing of time.
Notes
The p value is the probability of obtaining a test statistic result at least as extreme as the one that was actually observed, assuming that the null hypothesis is true. We use the significance testing method proposed by Zhang et al. (2004).
These categories are constructed using our proposed semi-automatic approach based on multi-LDA models. Totally, we obtain 83 categories.
References
Blei DM, Griffiths TL, Jordan MI, Tenenbaum JB (2003a) Hierarchical topic models and the nested Chinese restaurant process. In: NIPS, vol 16
Blei DM, Ng AY, Jordan MI (2003b) Latent dirichlet allocation. J Mach Learn Res 3:993–1022
Blei DM, McAuliffe JD (2007) Supervised topic models. NIPS 7:121–128
Blum A, Mitchell T (1998) Combining labeled and unlabeled data with co-training. In: Proceedings of COLT, pp 92–100
Brown PF, Desouza PV, Mercer RL, Della Pietra VJ, Lai JC (1992) Class-based n-gram models of natural language. Comput Linguist 18(4):467–479
Carletta J (1996) Assessing agreement on classification tasks: the kappa statistic. Comput Linguist 22(2):249–254
Carlson A, Betteridge J, Wang RC, Hruschka Jr ER, Mitchell TM (2010) Coupled semi-supervised learning for information extraction. In: Proceedings of the third ACM international conference on Web search and data mining, pp 101–110
Che W, Li Z, Liu T (2010) Ltp: a Chinese language technology platform. In: Coling 2010: demonstrations, pp 13–16
Cheng SJ, Huang QC, Liu JF, Tang XL (2013) A novel inductive semi-supervised SVM with graph-based self-training. In: Intelligent science and intelligent data engineering. Springer, Berlin Heidelberg, pp 82–89
Collins M, Singer Y (1999) Unsupervised models for named entity classification. In: Proceedings of EMNLP, pp 100–110
Danesh A, Moshiri B, Fatemi O (2007) Improve text classification accuracy based on classifier fusion methods. 10th international conference on information fusion, pp 1–6
Donghui C, Zhijing L (2010) A new text categorization method based on HMM and SVM. In: 2nd international conference on computer engineering and technology (ICCET), vol 7, pp 383–386
Fu JH, Lee SL (2012) A multi-class SVM classification system based on learning methods from indistinguishable chinese official documents. Expert Syst Appl 39(3):3127–3134
Hofmann T (1999) Probabilistic latent semantic indexing. In: Proceedings of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval, ACM, pp 50–57
Joachims T (1998) Text categorization with support vector machines: learning with many relevant features. In: Proceedings of ECML-98, 10th European conference on machine learning (Chemnitz, DE), pp 137–142
Johnson DE, Oles FJ, Zhang T, Goetz T (2002) A decision-tree-based symbolic rule induction system for text categorization. IBM Syst J 41(3):428–437
Kim S-B, Rim H-C, Yook DS, Lim H-S (2002) Effective methods for improving naive bayes text classifiers. LNAI 2417:414–423
Li CH, Park SC (2009) n efficient document classification model using an improved back propagation neural network and singular value decomposition. Expert Syst Appl 36(2):3208–3215
Lin Y (2002) Support vector machines and the Bayes rule in classification. Data Min Knowl Discov 6:259–275
Mao X-L, Ming Z-Y, Chua T-S, Li S, Yan H, Li X (2012) SSHLDA: a semi-supervised hierarchical topic model. In: Proceedings of EMNLP-CoNLL, pp 800–809
McClosky D,Charniak E, Johnson M (2006) Effective self-training for parsing. In: Proceedings of NAACL, pp 152–159
Ng HT, Goh WB, Low KL (1997) Feature selection, perception learning, and a usability case study for text categorization. In: Proceedings of the 20th annual international ACM SIGIR conference on Research and development in information retrieval, Philadelphia PA, pp 67–73
Petinot Y, McKeown K, Thadani K (2011) A hierarchical model of web summaries. Proc ACL HLT Short Pap Vol 2:670–675
Pham DT, Dimov SS, Nguyen CD (2005) Selection of K in K-means clustering. Proc Inst Mech Eng Part C J Mech Eng Sci 219(1):103–109
Qin Y-P, Wang X-K (2009) Study on multi-label text classification based on SVM. Sixth international conference on fuzzy systems and knowledge discovery, pp 300–304
Salton G, Wong A, Yan C-S (1975) A vector space model for automatic indexing. Commun ACM 18(11):613–620
Sebastiani F (2002) Machine learning in automated text categorization. ACM Comput Surv (CSUR) 34(1):1–47
Trappey AJC, Hsu F-C, Trappey CV, Lin C-I (2006) Development of a patent document classification and search platform using a back-propagation network. Expert Syst Appl 31(4):755–765
Turian J, Ratinov L, Bengio Y (2010) Word representations: a simple and general method for semi-supervised learning. In: Proceedings of the ACL, pp 384–394
Ueffing N (2006) Self-training for machine translation. In: Proceedings of NIPS workshop on machine learning for multilingual information access
Vateekul P, Kubat M (2009) Fast induction of multiple decision trees in text categorization from large scale, imbalanced, and multi-label data. IEEE International Conference on Data Mining Workshops, pp 320–325
Yang Y (1999) An evaluation of statistical approaches to text categorization. Inf Retr 1(1–2):69–90
Zhang Y, Vogel S, Waibel A (2004) Interpreting bleu/nist scores: how much improvement do we need to have a better system. In: Proceedings of the 2004 international conference on language resources and evaluation. pp 2051–2054
Acknowledgments
This work is supported by National Natural Science Foundation of China (NSFC) via Grant 61133012, 61273321 and the National 863 Leading Technology Research Project via grant 2012AA011102. Special thanks to Jianfei Guo and Xiaocheng Feng for their help in the experiments..
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by L. Xie.
Appendix: the categorization system of WeChat subscription accounts
Appendix: the categorization system of WeChat subscription accounts
-
finance and economics
-
1.
banking institutions
-
2.
business
-
3.
financing
-
4.
insurance
-
5.
marketing
-
6.
realty
-
7.
start-ups
-
shopping
-
-
8.
automobile
-
9.
commodity
-
10.
decoration
-
11.
discount shopping
-
12.
dresses
-
13.
electronic products
-
14.
luxuries
-
15.
online shopping
-
16.
purchasing agents
-
17.
sports equipments
-
18.
wholesale
-
19.
health care
-
20.
maternal and infant
-
21.
nourishing of life
-
22.
dating
-
communication platform
-
-
23.
friends making
-
24.
job hunting
-
education
-
-
25.
art schools
-
26.
business administration
-
27.
driving schools
-
28.
foreign language training
-
29.
raining for study abroad
-
30.
tutoring
-
military affairs
-
-
31.
military affairs
-
science and technology
-
-
32.
IT
-
33.
mobile internet applications
-
media
-
-
34.
news media
-
35.
print media
-
36.
TV and radio
-
37.
we-media
-
38.
cosmetic surgery
-
39.
hairdressing
-
40.
skin protection
-
food and drink
-
-
41.
green food
-
42.
restaurants
-
43.
tea
-
44.
western-style pastry
-
45.
wine
-
services for life
-
-
46.
air tickets booking
-
47.
Campus
-
48.
car rental
-
49.
community
-
50.
design
-
51.
emotion
-
52.
environmental protection
-
53.
Express delivery
-
54.
homemaking
-
55.
hot lines
-
56.
hotel booking
-
57.
law works
-
58.
life assistants
-
59.
lotteries
-
60.
public good
-
61.
recharging
-
62.
tourism
-
63.
weddings
-
culture
-
-
64.
art
-
65.
culture
-
66.
originality
-
67.
popularization of science
-
68.
reading
-
entertainment
-
-
69.
adult entertainment
-
70.
caricatures
-
71.
entertainment stars
-
72.
entertainment venues
-
73.
fashion
-
74.
games
-
75.
image show
-
76.
jokes
-
77.
movies
-
78.
music
-
79.
pets
-
sports
-
-
80.
sports clubs
-
81.
sports news
-
others
-
-
82.
brand
-
83.
government
Rights and permissions
About this article
Cite this article
Fu, R., Qin, B. & Liu, T. Open-categorical text classification based on multi-LDA models. Soft Comput 19, 29–38 (2015). https://doi.org/10.1007/s00500-014-1374-x
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00500-014-1374-x