Much attention in recent years has been attracted by the process exploring useful information from a large amount of textual data produced by microblogging services such as Twitter. A very important preprocessing step is to convert natural language texts of microblog text mining into proper numerical representations. The short-length characteristics of microblog texts result in using the term frequency vectors to represent microblog texts and it will cause “sparse data” problem. Finding proper representations for microblog texts is a challenging issue.In the previous paper, they applied deep networks so that they can map the high-dimensional representations to low-dimensional representations.The retweet and hashtags have been used as the semantic similarity. They used two types of approaches which includes modifying the training data and modifying the training objective. They have also shown that deep models perform better than traditional methods such as latent Dirichlet allocation topic model and latent semantic analysis.
In the last few years, microblogging services such as Twitter has gained great popularity among the Internet users. The high volume of textual data produced by the microblogging services is very attractive to the researchers in the text mining field. Transforming natural language texts into numerical vectors is an important preprocessing step for many text mining tasks, such as cluster analysis and sentiment classification. The most widely adopted model for text representation is vector space model 1, where each document in a corpus is represented by a vector with each dimension corresponding to a separate term and the elements denoting the frequencies of the terms. An important issue that needs to be dealt with carefully when using term frequency vectors to represent texts is the “sparse data” problem.
Exploring potentially useful information from the huge amount of textual data produced by microblogging has gained much attention. Transforming natural language texts into numerical vectors. We apply deep networks to map the high-dimensional representations of microbiology. Creating proper low-dimensional feature space for microblog texts and also investigate how to utilize the expansion approach to learn better features. For the classification of documents, given the document training data, the most well-known approaches start by assessing the words’ co-occurrence matrix versus documents. Researchers have proposed many approaches to enhance the representations of short text, such as expanding the original short document by adding semantically related terms 2, 3, or mapping a high-dimensional term frequency vector to a low-dimensional feature vector via latent semantic analysis (LSA) 4. There are a few approaches specifically proposed for dimensionality reduction of tweets. Instead, low dimensional representations of tweets are usually obtained as the by-products of topic modeling 5, 6, 7.
For the classification of documents when the document training data is given the most well-known approach is to start assessing the words that are co-occurrence matrix versus documents. It is understood that in any case, the count matrices have a very high tendency to be noisy and sparse especially when training data is moderately small13. Usually, the documents are presented in a high dimensional feature space, this is long way best for algorithms that are used for classification. A known methodology to reduce dimensionality is the feature selection. In this one has to choose a subset of words using some pre-defined paradigm and as features, it is able to use only the words that are chosen for classification. In the information retrieval, other similar strategies are Latent Semantic Indexing (LSI) and Probabilistic LSI, LDA. Also, there is a methodology to minimize the feature dimensionality by collecting ‘similar’ words into a much smaller number of groups and then use these groups as features14.
2.1 Mapping Messages
Tweets and retweets of a user’s followers appear alongside the user’s own tweets in reverse chronological order. People often have only enough patience to skim through the first 20 – 50 messages. When the messages become overwhelming, it is impractical for a user to quickly gauge the main subjects from their followers’ posts. To make a large collection of microblogging messages accessible to users, current web systems need to provide not only accurate clusters for subtopics in microblogging messages, but also meaningful labels for each cluster2. Enhancing the accessibility of microblogging messages entails two tasks: (a) cluster microblogging messages into manageable categories, and (b) assign readable and meaningful labels for each cluster of messages. Unlike standard text with many sentences or paragraphs, microblogging messages are noisy and short4. In addition, microbloggers, when composing a message, may use or coin new abbreviations or acronyms that are uncommon in conventional text documents.5 Furthermore, these short messages do not provide enough information so that they can capture the semantic meanings.
2.2 Traditional text mining
Traditional text mining methods, when applied to microblogging messages directly, lead to unsatisfactory results. In this paper, they present a framework to enhance the accessibility of microblogging message. The proposed framework helps in improving the message representation by mapping the messages from an feature space which is unstructured to a semantically meaningful knowledge space8. First, in order to reduce the noise yet keep the key information as expressed in each message, we propose to use natural language processing (NLP) techniques to analyze the message and extract informative words and phrases. Then, to overcome the extreme sparsity of microblogging messages, we map the selected terms to structured concepts derived from external knowledge bases that are semantically rich9. Because of conducting feature selection to refine the feature space they were able to cluster all message accurately and generate humanly understandable labels efficiently from the related concept.
2.3 Data Sparseness
Most of the recent related work concentrates on the elimination of the problem of data sparseness. One instinctive solution to solve this is to increase the short text with additional information to make it like a large document of text. Then the algorithms for clustering or classification can be applied easily to it. As discussed in 2,3, 11 , their primary focus is on combining short text messages with web search engines such as Bing, Google to extract more information about the short text. In 12 Poschko said that if two hashtags co-occur in a tweet, then they are similar. With co-occurrence frequency as a distance measure, he created a clustered graph. The other works on clustering text -related entities typically focus on a bag –of – words(BOW). BOW model takes all the words of entity followed by reduction of dimensionality. It makes clustering computationally feasible as in 15,16. As said in 4,5,13, more recent works focused on web searches and instead used data repositories. One of the biggest data sources is Wikipedia. Short text messages can be strengthened with additional semantic knowledge by combining knowledge accessible within the Wikipedia For the classification of documents, given the document training data, the most well-known approaches start by assessing the words’ co-occurrence matrix versus documents. It is surely understood, in any case, such count matrices have the tendency to be excessively noisy and sparse, particularly when training data is moderately small. Subsequently, usually, the documents are depicted in a feature space that is high dimensional, which is the long way from optimal for algorithms of classification.
2.4 Feature Dimensionality
A standard methodology to reduce the dimensionality of features is feature selection. In this methodology, one chooses a subset of words, utilizing some pre-defined paradigm, and as features, it uses only the chosen words for classification. In the task of information retrieval, other similar strategies for dimensionality reduction are Latent Semantic Indexing (LSI) and Probabilistic LSI. An alternative methodology is to minimize the feature dimensionality by gathering ‘similar’ words into a much smaller number of word groups and utilize these groups as features.
2.5 Probabilistic Topic Models
To capture semantic similarity among words from the message they derived a model of documents which learns word representations which is probabilistic. This component does not require labeled data and shares its foundation with probabilistic topic models such as LDA. The sentiment component of their model use sentiment annotations such that the words expressing a similar sentiment to have similar representations. It can ef?ciently learn parameters for the joint objective function using alternating maximization. The model which is presented does not capture information from the sentiments from the message.By applying the algorithm to documents will produce representations, where which are occurring together in documents, will be given similar representations. This approach which is unsupervised has no way of capturing which words are predictive of sentiment.1718 Much previous work in natural language processing achieves better representations by learning from multiple tasks (Collobert and Weston, 2008; Finkel and Manning, 2009). Following this theme, they introduced a second task to use labeled documents so that improve their model’s word representation.
2.6 Hierarchical Dirichlet process (HDP)
Each and every group of data being modeled with a mixture with the number of components which are open-ended and they are inferred automatically by the model. Further, the components can be shared across various groups and allow the dependencies across groups to be modeled effectively as well as conferring generalization to new groups. Such clustering problems occur in practice, e.g. in the problem of topic discovery in document corpora. We report experimental results on three text corpora showing the effective and superior performance of the HDP over previous models. They considered the application which has hierarchical Bayesian ideas to a problem in “multi-task learning” where the “tasks” are clustering problems, and their goal is to share clusters among multiple, related clustering problems. We are motivated by the task of discovering topics in document corpora 19. A topic (i.e., a cluster) is a distribution across words while documents are viewed as distributions across topics.They wanted to discover various topics which are common across various multiple documents in the same corpus and also as across multiple corpora. Their work is based on a tool which is from nonparametric Bayesian analysis also called (DP) mixture model 20. Skirting technical de?nitions for now, “nonparametric”.
Bag Of Word: The bag-of-words model is a representation which is used in NLP and IR. It is also known as the vector space model in which a text is represented as the bag of its words.
The following shows a text document by using BOW.
Given two text documents:
(1) John likes to play games. Mary likes games too.
(2) John also likes to watch series.
The list constructed given below.
Microblog: It is a blog that lets user to publish short text updates.
Deep networks: It’s an artificial network with multiple hidden layers between the input and output layers.
Dimensionality reduction: It means converting high dimensional representation to low dimensional representation. Also, it is the process of reducing random variables by obtaining variables that are known as principal variables.
Term Frequency Vector: It means how important a word is to a document.
Latent Semantic Analysis: It is related to analyzing the relationship between term and document by obtaining the set of related concepts.
Hashtag: Hashtags are mostly used on social media sites and used to draw attention, to organize and to promote. Hashtags are words(multi-word phrases) preceded by the # symbol.
Sentiment Analysis: Classifying the polarity of a given text or sentence –whether the expressed opinion in a text or sentences is positive, negative or neutral.It is also known as opinion mining.
Microblogging: It is a combination of blogging and instant messaging. E.g. Twitter, Tumblr, Facebook, Instagram, and Vine.Twitter has become the most popular forms of this new type of blogging. These short messages can come in the form of text, images, audio, video, and hyperlinks.
4.Conclusion and Result
In this, They investigated how to apply deep networks to perform dimensionality reduction on microblog texts. A priori knowledge about semantic similarity, which is derived from retweet relationships and hashtags, was explored to train the deep networks. Two types of approaches, namely modifying the training data and modifying the objective of ?ne-tuning, were proposed to utilize such prior knowledge. Experiment results validated that deep models can learn better representations than LSA and LDA, and the use of microblog-speci?c information can further improve the performance of deep models. DBN models. The evaluation results also demonstrate that the proposed modi?cations of training data can help to learn better representations. Next, they give some detailed explanations to the evaluation results.
In this, they have proposed Real-Time Big data analytical Architecture for emotion extraction from social networks tweets. The proposed architecture by them was capable of efficiently processing and analyzing real-time and offline data and classification of tweets in positive and negative .We can use the big data technologies to classify text in a distributed manner and can decrease the execution time. LMClassifier, OpenNLPClassifer andNaiveBayesclassifer these classifiers can be used with MapReduce approach to reducing execution time for a large amount of text data.If we use NaiveBayesclassifer then we can achieve the highest accuracy with less execution time because in normal Naivebayesapprocah it took more time to execute than OpenNLPClassifer andLMClassifer.They can use MapReduce with all three classifiers to reduce the execution time. Text data can arrive with high velocity, veracity, Volume, variety so in these conditions, They can use MapReduce approach because it handles all these scenarios in execution. Execution time is decreasing when size is increasing in MapReduce approach. In conventional use of classifiers execution time increases with an increment of the size of dataset So applying classification techniques in distributed environment provides better execution time.
In the first Map-Reduce pass, the mapper takes the labeled tweets from the training data and outputs category and word as a key-value pair. The Reducer then sums up all instances of the words for each category and outputs category and word-count pair as a key value. The Map-Reduce then deals with the making of model for the classifier. The next Map-Reduce pass does the classification by calculating the conditional probability of each word (i.e. feature) and outputs category and the conditional probability of each word as a key-value pair. Then final reducer calculates the final probability of each category to which the tweet may belong to and outputs the predicted category and its probability value as key-value pair The final step after preprocessing of tweets is that they label the tweets based on categories namely politics, sports, and technology. It is proposed to stream real-time live tweets from Twitter using Twitter API, and the large volume of data makes the application suitable for Big Data Analytics. The method to predict the location of a tweet based on the tweet’s information and the user’s information can be found in the future.
In this proposes a semi-supervised learning called Active Deep Networks algorithm (ADN) to solve the sentiment classification problem with a very few number of labeled data. ADN can choose the proper training data to be labeled manually, and fully exploits the embedding information from a large amount of unlabeled data to improve the robustness of the classifier. We propose a new architecture to guide the output vector of samples belong to different regions of new Euclidean space and use an exponential loss function to maximize the separability of labeled data in global refinement for better discriminability. The ADN can make a right decision about which training data should be labeled based on the data which is not labeled and labeled data. By using the both supervised and unsupervised learning , ADN is able to choose the training data to be labeled and train the architecture at the same time. The deep architecture is then re-trained using the labeled data and all the data which is not labeled. They also conduct experiments to verify that ADN method is effective with the various number of labeled data, and also demonstrated that ADN is able to reach very competitive performance for classification just by using few labeled data. These shows that the ADN method which needs only a fewer manual labeled reviews so that it can reach a higher accuracy also can be used to train a high-performance sentiment classification system.
This is one of the directions studies in the future. Another limitation of the current method is that the number of topics is predetermined. We planned to look into methods that allow the appearance and disappearance of topics along with the timeline. There are also other properties, such as mass data, real-time updates, and so on. make the detection method can not be applied to microblog topic detection. But, on the other hand, the microblog structure gives us a new idea to improve the performance of the microblog topic detection. The best of these structural properties, and propose an effective microblog topic detection method which aims at the above three properties. In future studies, to optimize the learning of representations towards speci?c microblog mining tasks, such as sentiment classi?cation. And more types of meta-information contained in tweets, such as emoticons and the embedded hyperlinks, will be explored.
To address the sentiment classi?cation problem with a small number of labeled reviews. We studied the problem of ?nding bursty topics from the text streams on microblogs. Another limitation of the current method is that the number of topics is predetermined. separate topics for positive words and a separate one for negative words links.Are spammers more likely to link to each other than regular people? Is content similarity more important than linguistic similarity when predicting links? representations towards speci?c microblog mining tasks, such as sentiment classi?cation. And more types of meta-information contained in tweets, such as emoticons and the embedded hyperlinks, will be explored.
1 Lei Xu, Chunxiao Jiang,”Microblog Dimensionality Reduction—A Deep Learning Approach,” Ieee Transactions On Knowledge And Data Engineering, Vol. 28, No. 7, July 2016.
2 Zhi-Qiang Xian , “Sentiment Analysis of Chinese Micro-blog Using Vector Space Model,” APSIPA,2014.
3 Amit mittal , “Social Networking text Classification in Big Data Environment,” IJlEET, 2016
4 X. Yan and H. Zhao, “Chinese microblog topic detection based on the latent semantic analysis and structural property,” J. Netw., vol. 8, pp. 917–9233, no. 4, 2013.
5 D. Ramage, S. T. Dumais, and D. J. Liebling, “Characterizing microblogs with topic models,” in Proc. 4th Int. Conf. Weblogs Social Media, pp. 130–137, 2010.
6 O. Jin, N. N. Liu, K. Zhao, Y. Yu, and Q. Yang, “Transferring topical knowledge from auxiliary long texts for short text clustering,” in Proc. 20th ACM Int. Conf. Inf. Knowl. Manag., pp. 775–784, 2011.
7 Q. Diao, J. Jiang, F. Zhu, and E.-P. Lim, “Finding bursty topics from microblogs,” in Proc. 50th Annu. Meet. Assoc. Comput. Linguistics: Long Papers-Vol. 1. , pp. 536–544, 2012.
8 M. A. Ranzato and M. Szummer, “Semi-supervised learning of compact document representations with deep networks,” in Proc. 25th Int. Conf. Mach. Learning, pp. 792–799, 2008.
9 G. Hinton and R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,” Sci., vol. 313, no. 5786, pp. 504–507, Jul. 2006.
10 R. Salakhutdinov and G. Hinton, “Semantic hashing,” Int. J. Approx. Reasoning, vol. 50, no. 7, pp. 969–978, Jul. 2009.
11 M. A. Ranzato and M. Szummer, “Semi-supervised learning of compact document representations with deep networks,” in Proc. 25th Int. Conf. Mach. Learning, pp. 792–799, 2008.
12 S. Zhou, Q. Chen, and X. Wang, “Active deep learning method for semi-supervised sentiment classi?cation,” Neurocomputing, vol. 120, pp. 536–546, 2013.
13 M. R. Min, L. Maaten, Z. Yuan, A. J. Bonner, and Z. Zhang, “Deep supervised t-distributed embedding,” in Proc. 27th Int. Conf. Mach. Learn. , pp. 791–798, 2010.
14 D. Blei, A. Y. Ng, and M. I. Jordan, “Latent dirichlet allocation,” J. Mach. Learning Res., vol. 3, pp. 993–1022, 2001.
15 T. K. Landauer, P. W. Foltz, and D. Laham, “An introduction to latent semantic analysis,” Discourse Processes, vol. 25, pp. 259–284, 1998.
16 T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Ef?cient estimation of word representations in vector space,” CoRR, vol. abs/1301.3781, 2013.
17 A. L. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y. Ng, and C. Potts, “Learning word vectors for sentiment analysis,” in Proc. 49th Annu. Meet. Assoc. Comput. Linguistics: Human Language Technol.-Volume 1., pp. 142–150,2011.
18 J. Tang, X. Wang, H. Gao, X. Hu, and H. Liu, “Enriching short text representation in microblog for clustering,” Frontiers Comput. Sci., vol. 6, no. 1, pp. 88–101, 2012.
19 Y. W. Teh, M. I. Jordan, M. J. Beal, and D. M. Blei, “Sharing clusters among related groups: Hierarchical dirichlet processes,” in Proc. Int. Conf. Neural Information Processing Syst, pp. 1385– 1392, 2004.
20 C. E. Grant, C. P. George, C. Jenneisch, and J. N. Wilson, “Online topic modeling for real-time twitter search,” in Proc. Text Retrieval Conf. , pp. 1–9, 2011.
21 X. Wang, F. Zhu, J. Jiang, and S. Li, “Real time event detection in twitter,” in Proc. 14th Int. Conf. Web-Age Inf. Manag., pp. 502–513, 2013.
22 D. Yu and L. Deng, “Deep learning and its applications to signal and information processing,” IEEE Signal Process. Mag., vol. 28, no. 1, pp. 145–154, Jan. 2011.
23 Y. Bengio, A. C. Courville, and P. Vincent, “Unsupervised feature learning and deep learning: A review and new perspectives,”CoRR, vol. abs/1206.5538, 2012.
24 R. Collobert and J. Weston, “A unified architecture for natural language processing: Deep neural networks with multitask learning,” in Proc. 25th Int. Conf. Mach. Learning, pp. 160–167,2008.
25 J. P. Turian, L.-A. Ratinov, and Y. Bengio, “Word representations:A simple and general method for semi-supervised learning,” in Proc. 48th Annu. Meet. Assoc. Comput. Linguistics, pp. 384–394, 2010.