深度学习之（八）Twitter情感分析_twitter sentiment analysis.

链接：
https://github.com/ayushoriginal/Sentiment-Analysis-Twitter

原文链接：
http://weibo.com/5581058391/Ee7dUxZBv?from=page_1005055581058391_profile&wvr=6&mod=weibotime&type=comment

Sentiment-Analysis-Twitter

-Ayush Pareek

Microblogging today has become a very popular communication tool among Internet users. Millions of messages are appearing daily in popular web-sites that provide services for microblogging such as Twitter, Tumblr, Facebook. Authors of those messages write about their life, share opinions on variety of topics and discuss current issues. Because of a free format of messages and an easy accessibility of microblogging platforms, Internet users tend to shift from traditional communication tools (such as traditional blogs or mailing lists) to microblogging services. As more and more users post about products and services they use, or express their political and religious views, microblogging web-sites become valuable sources of people’s opinions and sentiments. Such data can be efficiently used for marketing or social studies.[1]

1 Introduction

1.1 Applications of Sentiment Analysis

Sentiment Analysis finds its application in a variety of domains.

A. Online Commerce

The most general use of sentiment analysis is in ecommerce activities. Websites allows their users to submit their experience about shopping and product qualities. They provide summary for the product and different features of the product by assigning ratings or scores. Customers can easily view opinions and recommendation information on whole product as well as specific product features. Graphical summary of the overall product and its features is presented to users. Popular merchant websites like amazon.com provides review from editors and also from customers with rating information. http://tripadvisor.inis a popular website that provides reviews on hotels, travel destinations. They contain 75 millions opinions and reviews worldwide. Sentiment analysis helps such websites by converting dissatisfied customers into promoters by analyzing this huge volume of opinions.

B. Voice of the Market (VOM)

Voice of the Market is about determining what customers are feeling about products or services of competitors. Accurate and timely information from the Voice of the Market helps in gaining competitive advantage and new product development. Detection of such information as early as possible helps in direct and target key marketing campaigns. Sentiment Analysis helps corporate to get customer opinion in real-time. This real-time information helps them to design new marketing strategies, improve product features and can predict chances of product failure. Zhang et al.proposed weakness finder system which can help manufacturers find their product weakness from Chinese reviews by using aspects based sentiment analysis. There are some commercial and free sentiment analysis services are available, Radiant6, Sysomos, Viralheat, Lexalytics, etc. are commercial services. Some free tools like www.tweettfeel.com, www.socialmention.com are also available.

C. Voice of the Customer (VOC)

Voice of the Customer is concern about what individual customer is saying about products or services. It means analyzing the reviews and feedback of the customers. VOC is a key element of Customer Experience Management. VOC helps in identifying new opportunities for product inventions. Extracting customer opinions also helps identify functional requirements of the products and some non-functional requirements like performance and cost.

D. Brand Reputation Management

Brand Reputation Management is concern about managing your reputation in market. Opinions from customers or any other parties can damage or enhance your reputation. Brand Reputation Management (BRM) is a product and company focused rather than customer. Now, one-to-many conversations are taking place online at a high rate. That creates opportunities for organizations to manage and strengthen brand reputation. Now Brand perception is determined not only by advertising, public relations and corporate messaging. Brands are now a sum of the conversations about them. Sentiment analysis helps in determining how company’s brand, product or service is being perceived by community online.

E. Government

Sentiment analysis helps government in assessing their strength and weaknesses by analyzing opinions from public. For example, “If this is the state, how do you expect truth to come out? The MP who is investigating 2g scam himself is deeply corrupt.”. this example clearly shows negative sentiment about government. Whether it is tracking citizens’ opinions on a new 108 system, identifying strengths and weaknesses in a recruitment campaign in government job, assessing success of electronic submission of tax returns, or many other areas, we can see the potential for sentiment analysis.

Figure: Sentiment Analysis can be useful to understand how the mood of the public affects election results

1.2 Characteristic features of Tweets

From the perspective of Sentiment Analysis, we discuss a few characteristics of Twitter:

Length of a Tweet The maximum length of a Twitter message is 140 characters. This means that we can practically consider a tweet to be a single sentence, void of complex grammatical constructs. This is a vast difference from traditional subjects of Sentiment Analysis, such as movie reviews.

Language used Twitter is used via a variety of media including SMS and mobile phone apps. Because of this and the 140-character limit, language used in Tweets tend be more colloquial, and filled with slang and misspellings. Use of hashtags also gained popularity on Twitter and is a primary feature in any given tweet. Our analysis shows that there are approximately 1-2 hashtags per tweet, as shown in Table 3 .

Data availability Another difference is the magnitude of data available. With the Twitter API, it is easy to collect millions of tweets for training. There also exist a few datasets that have automatically and manually labelled the tweets [2] [3].

Domain of topics People often post about their likes and dislikes on social media. These are not al concentrated around one topic. This makes twitter a unique place to model a generic classifier as opposed to domain specific classifiers that could be build datasets such as movie reviews.

2 Related Work

2.1 Go, Bhayani and Huang (2009)

They classify Tweets for a query term into negative or positive sentiment. They collect training dataset automatically from Twitter. To collect positive and negative tweets, they query twitter for happy and sad emoticons.

• Happy emoticons are different versions of smiling face, like ":)", ":-)", ": )", ":D", "=)" etc.
• Sad emoticons include frowns, like ":(", ":-(", ":(" etc.

They try various features – unigrams, bigrams and Part-of-Speech and train their classifier on various machine learning algorithms – Naive Bayes, Maximum Entropy and Scalable Vector Machines and compare it against a baseline classifier by counting the number of positive and negative words from a publicly available corpus. They report that Bigrams alone and Part-of-Speech Tagging are not helpful and that Naive Bayes Classifier gives the best results.

2.2 Pak and Paroubek (2010)

They identify that use of informal and creative language make sentiment analysis of tweets a rather different task . They leverage previous work done in hashtags and sentiment analysis to build their classifier. They use Edinburgh Twitter corpus to find out most frequent hashtags. They manually classify these hashtags and use them to in turn classify the tweets. Apart from using n-grams and Part-of-Speech features, they also build a feature set from already existing MPQA subjectivity lexicon and Internet Lingo Dictionary. They report that the best results are seen with n-gram features with lexicon features, while using Part-of-Speech features causes a drop in accuracy.

2.3 Koulompis, Wilson and Moore (2011)

They identify that use of informal and creative language make sentiment analysis of tweets a rather different task [5]. They leverage previous work done in hashtags and sentiment analysis to build their classifier. They use Edinburgh Twitter corpus to find out most frequent hashtags. They manually classify these hashtags and use them to in turn classify the tweets. Apart from using n-grams and Part-of-Speech features, they also build a feature set from already existing MPQA subjectivity lexicon and Internet Lingo Dictionary. They report that the best results are seen with n-gram features with lexicon features, while using Part-of-Speech features causes a drop in accuracy.

2.3 Saif, He and Alani (2012)

They discuss a semantic based approach to identify the entity being discussed in a tweet, like a person, organization etc. They also demonstrate that removal of stop words is not a necessary step and may have undesirable effect on the classifier.

All of the aforementioned techniques rely on n-gram features. It is unclear that the use of Part-of-Speech tagging is useful or not. To improve accuracy, some employ different methods of feature selection or leveraging knowledge about micro-blogging. In contrast, we improve our results by using more basic techniques used in Sentiment Analysis, like stemming, two-step classification and negation detection and scope of negation.

Negation detection is a technique that has often been studied in sentiment analysis. Negation words like “not”, “never”, “no” etc. can drastically change the meaning of a sentence and hence the sentiment expressed in them. Due to presence of such words, the meaning of nearby words becomes opposite. Such words are said to be in the scope of negation. Many researches have worked on detecting the scope of negation.

The scope of negation of a cue can be taken from that word to the next following punctuation. Councill, McDonald and Velikovich (2010) discuss a technique to identify negation cues and their scope in a sentence. They identify explicit negation cues in the text and for each word in the scope. Then they find its distance from the nearest negative cue on the left and right.

3 Approach

We use different feature sets and machine learning classifiers to determine the best combination for sentiment analysis of twitter. We also experiment with various pre-processing steps like - punctuations, emoticons, twitter specific terms and stemming. We investigated the following features - unigrams, bigrams, trigrams and negation detection. We finally train our classifier using various machine-learning algorithms - Naive Bayes, Decision Trees and Maximum Entropy.

Figure 1: Schematic Block Representation of the Methodology

We use a modularized approach with feature extractor and classification algorithm as two independent components. This enables us to experiment with different options for each component.

3.1 Datasets

One of the major challenges in Sentiment Analysis of Twitter is to collect a labelled dataset. Researchers have made public the following datasets for training and testing classifiers.

3.1.1 Twitter Sentiment Corpus

This is a collection of 5513 tweets collected for four different topics, namely, Apple, Google, Microsoft, Twitter It is collected and hand-classified by Sanders Analytics LLC. Each entry in the corpus contains, Tweet id, Topic and a Sentiment label. We use Twitter-Python library to enrich this data by downloading data like Tweet text, Creation Date, Creator etc. for every Tweet id. Each Tweet is hand classified by an American male into the following four categories. For the purpose of our experiments, we consider Irrelevant and Neutral to be the same class. Illustration of Tweets in this corpus is show in Table 1 .

• Positive For showing positive sentiment towards the topic

• Positive For showing no or mixed or weak sentiments towards the topic

• Negative For showing negative sentiment towards the topic

• Irrelevant For non English text or off-topic comments

3.1.2 Stanford Twitter

This corpus of tweets, developed by Sanford’s Natural Language processing research group, is publically available. The training set is collected by querying Twitter API for happy emoticons like ":)" and sad emoticons like ":(" and labelling them positive or negative. The emoticons were then stripped and Re-Tweets and duplicates removed. It also contains around 500 tweets manually collected and labelled for testing purposes. We randomly sample and use 5000 tweets from this dataset. An example of Tweets in this corpus are shown in Table 2 .

3.2 Pre Processing

User-generated content on the web is seldom present in a form usable for learning. It becomes important to normalize the text by applying a series of pre-processing steps. We have applied an extensive set of pre-processing steps to decrease the size of the feature set to make it suitable for learning algorithms. Figure 2 illustrates various features seen in micro-blogging. Table 3 illustrates the frequency of these features per tweet, cut by datasets. We also give a brief description of pre-processing steps taken.

Figure 2: Illustration of a Tweet with various features

3.2.1 Hashtags

A hashtag is a word or an un-spaced phrase prefixed with the hash symbol (#). These are used to both naming subjects and phrases that are currently in trending topics. For example, #iPad, #news

Regular Expression: #(\w+)

Replace Expression: HASH_\1

3.2.2 Handles

Every Twitter user has a unique username. Any thing directed towards that user can be indicated be writing their username preceded by ‘@’. Thus, these are like proper nouns. For example, @Apple

Regular Expression: @(\w+)

Replace Expression: HNDL_\1

3.2.3 URLs

Users often share hyperlinks in their tweets. Twitter shortens them using its in-house URL shortening service, likehttp://t.co/FCWXoUd8 - such links also enables Twitter to alert users if the link leads out of its domain. From the point of view of text classification, a particular URL is not important. However, presence of a URL can be an important feature. Regular expression for detecting a URL is fairly complex because of different types of URLs that can be there, but because of Twitter’s shortening service, we can use a relatively simple regular expression.

Regular Expression: (http|https|ftp)://[a-zA-Z0-9\\./]+

Replace Expression: URL

3.2.4 Emoticons

Use of emoticons is very prevalent throughout the web, more so on micro- blogging sites. We identify the following emoticons and replace them with a single word. Table 4 lists the emoticons we are currently detecting. All other emoticons would be ignored.

3.2.5 Punctuations

Although not all Punctuations are important from the point of view of classification but some of these, like question mark, exclamation mark can also provide information about the sentiments of the text. We replace every word boundary by a list of relevant punctuations present at that point. Table 5 lists the punctuations currently identified. We also remove any single quotes that might exist in the text.

3.2.6 Repeating Characters

People often use repeating characters while using colloquial language, like "I’m in a hurrryyyyy", "We won, yaaayyyyy!" As our final pre-processing step, we replace characters repeating more than twice as two characters.

Regular Expression: (.)\1{1,}

Replace Expression: \1\1

Reduction in feature space

It’s important to note that by applying these pre-processing steps, we are reducing our feature set otherwise it can be too sparse. Table 6 lists the decrease in feature set due to processing each of these features.

3.3 Stemming Algorithms

All stemming algorithms are of the following major types – affix removing, statistical and mixed. The first kind, Affix removal stemmer, is the most basic one. These apply a set of transformation rules to each word in an attempt to cut off commonly known prefixes and / or suffixes [8]. A trivial stemming algorithm would be to truncate words at N-th symbol. But this obviously is not well suited for practical purposes.

J.B. Lovins described first stemming algorithm in 1968. It defines 294 endings, each linked to one of 29 conditions, plus 35 transformation rules. For a word being stemmed, an ending with a satisfying condition is found and removed. Another famous stemmer used extensively is described in the next section.

3.3.1 Porter Stemmer

Martin Porter wrote a stemmer that was published in July 1980. This stemmer was very widely used and became and remains the de facto standard algorithm used for English stemming. It offers excellent trade-off between speed, readability, and accuracy. It uses a set of around 60 rules applied in 6 successive steps [9]. An important feature to note is that it doesn’t involve recursion. The steps in the algorithm are described in Table 7 .

3.3.2 Lemmatization

Lemmatization is the process of normalizing a word rather than just finding its stem. In the process, a suffix may not only be removed, but may also be substituted with a different one. It may also involve first determining the part-of-speech for a word and then applying normalization rules. It might also involve dictionary look-up. For example, verb ‘saw’ would be lemmatized to ‘see’ and the noun ‘saw’ will remain ‘saw’. For our purpose of classifying text, stemming should suffice.

3.4 Features

A wide variety of features can be used to build a classifier for tweets. The most widely used and basic feature set is word n-grams. However, there's a lot of domain specific information present in tweets that can also be used for classifying them. We have experimented with two sets of features:

3.4.1 Unigrams

Unigrams are the simplest features that can be used for text classification. A Tweet can be represented by a multiset of words present in it. We, however, have used the presence of unigrams in a tweet as a feature set. Presence of a word is more important than how many times it is repeated. Pang et al. found that presence of unigrams yields better results than repetition [1]. This also helps us to avoid having to scale the data, which can considerably decrease training time [2]. Figure 3 illustrated the cumulative distribution of words in our dataset.

Figure 3: Cumulative Frequency Plot for 50 Most Frequent Unigrams

We also observe that the unigrams nicely follow Zipf’s law. It states that in a corpus of natural language, the frequency of any word is inversely proportional to its rank in the frequency table. Figure 4 is a plot of log frequency versus log rank of our dataset. A linear trendline fits well with the data.

3.4.2 N-grams

N-gram refers to an n-long sequence of words. Probabilistic Language Models based on Unigrams, Bigrams and Trigrams can be successfully used to predict the next word given a current context of words. In the domain of sentiment analysis, the performance of N-grams is unclear. According to Pang et al., some researchers report that unigrams alone are better than bigrams for classification movie reviews, while some others report that bigrams and trigrams yield better product-review polarity classification [1].

As the order of the n-grams increases, they tend to be more and more sparse. Based on our experiments, we find that number of bigrams and trigrams increase much more rapidly than the number of unigrams with the number of Tweets. Figure 4 shows the number of n-grams versus number of Tweets. We can observe that bigrams and trigrams increase almost linearly where as unigrams are increasing logarithmically.

Figure 4: Number of n-grams vs. Number of Tweets

Because higher order n-grams are sparsely populated, we decide to trim off the n-grams that are not seen more than once in the training corpus, because chances are that these n-grams are not good indicators of sentiments. After the filtering out non-repeating n-grams, we see that the number of n-grams is considerably decreased and equals the order of unigrams, as shown in Figure 5 .

Figure 5: Number of repeating n-grams vs. Number of Tweets

3.4.3 Negation Handling

The need negation detection in sentiment analysis can be illustrated by the difference in the meaning of the phrases, "This is good" vs. "This is not good" However, the negations occurring in natural language are seldom so simple. Handling the negation consists of two tasks – Detection of explicit negation cues and the scope of negation of these words.

Councill et al. look at whether negation detection is useful for sentiment analysis and also to what extent is it possible to determine the exact scope of a negation in the text [7]. They describe a method for negation detection based on Left and Right Distances of a token to the nearest explicit negation cue.

Detection of Explicit Negation Cues

To detect explicit negation cues, we are looking for the following words in Table 8 . The search is done using regular expressions.

Scope of Negation

Words immediately preceding and following the negation cues are the most negative and the words that come farther away do not lie in the scope of negation of such cues. We define left and right negativity of a word as the chances that meaning of that word is actually the opposite. Left negativity depends on the closest negation cue on the left and similarly for Right negativity. Figure 7 illustrates the left and right negativity of words in a tweet.

Figure 7: Scope of Negation

4 Experimentation

We train 90% of our data using different combinations of features and test them on the remaining 10%. We take the features in the following combinations

• only unigrams, unigrams + filtered bigrams and trigrams, unigrams + negation, unigrams + filtered bigrams and trigrams + negation. We then train classifiers using different classification algorithms - Naive Bayes Classifier and Maximum Entropy Classifier.

The task of classification of a tweet can be done in two steps - first, classifying "neutral" (or "subjective") vs. "objective" tweets and second, classifying objective tweets into "positive" vs. "negative" tweets. We also trained 2 step classifiers. The accuracies for each of these configuration are shown in Figure 8 , we discuss these in detail below.

Figure 8: Accuracy for Naive Bayes Classifier

4.1 Naive Bayes

Naive Bayes classifier is the simplest and the fastest classifier. Many researchers [2], [4] claim to have gotten best results using this classifier.

For a given tweet, if we need to find the label for it, we find the probabilities of all the labels, given that feature and then select the label with maximum probability.

The results from training the Naive Bayes classifier are shown below in Figure 8 . The accuracy of Unigrams is the lowest at 79.67%. The accuracy increases if we also use Negation detection (81.66%) or higher order n-grams (86.68%). We see that if we use both Negation detection and higher order n-grams, the accuracy is marginally less than just using higher order n-grams (85.92%). We can also note that accuracies for double step classifier are lesser than those for corresponding single step.

We have also shown Precision versus Recall values for Naive Bayes classifier corresponding to different classes – Negative, Neutral and Positive in Figure 9 . The solid markers show the P-R values for single step classifier and hollow markers show the affect of using double step classifier. Different points are for different feature sets. We can see that both precision as well as recall values are higher for single step than that for double step.

Figure 9: Precision vs. Recall for Naive Bayes Classifier

4.2 Maximum Entropy Classifier

This classifier works by finding a probability distribution that maximizes the likelihood of testable data. This probability function is parameterized by weight vector. The optimal value of which can be found out using the method of Lagrange multipliers.

The results from training the Maximum Entropy Classifier are shown below in Figure 10 . Accuracies follow a similar trend as compared to Naive Bayes classifier. Unigram is the lowest at 79.73% and we see an increase for negation detection at 80.96%. The maximum is achieved with unigrams, bigrams and trigrams at 85.22% closely followed by n-grams and negation at 85.16%. Once again, the accuracies for double step classifiers are considerably lower.

Precision versus Recall map is also shown for maximum entropy classifier in Figure 10 . Here we see that precision of "neutral" class increase by using a double step classifier, but with a considerable decrease in its recall and slight fall in precision of "negative" and "positive" classes.

Figure 10: Precision vs. Recall for Maximum Entropy Classifier

5 Future Work

Investigating Support Vector Machines Several papers have discussed the results using Support Vector Machines (SVMs) also. The next step would be to test our approach on SVMs. However, Go, Bhayani and Huang have reported that SVMs do not increase the accuracy [2].

Building a classifier for Hindi tweets There are many users on Twitter that use primarily Hindi language. The approach discussed here can be used to create a Hindi language sentiment classifier.

Improving Results using Semantics Analysis Understanding the role of the nouns being talked about can help us better classify a given tweet. For example, "Skype often crashing: microsoft, what are you doing?" Here Skype is a product and Microsoft is a company. We can use semantic labellers to achieve this. Such an approach is discussed by Saif, He and Alani [6].

6 Conclusion

In this paper, we create a sentiment classifier for twitter using labelled data sets. We also investigate the relevance of using a double step classifier and negation detection for the purpose of sentiment analysis.

Our baseline classifier that uses just the unigrams achieves an accuracy of around 80.00%. Accuracy of the classifier increases if we use negation detection or introduce bigrams and trigrams. Thus we can conclude that both Negation Detection and higher order n-grams are useful for the purpose of text classification. However, if we use both n-grams and negation detection, the accuracy falls marginally. We also note that Single step classifiers out perform double step classifiers. In general, Naive Bayes Classifier performs better than Maximum Entropy Classifier.

We achieve the best accuracy of 86.68% in the case of Unigrams + Bigrams + Trigrams, trained on Naive Bayes Classifier.

References

[1] Pak, Alexander, and Patrick Paroubek. "Twitter as a Corpus for Sentiment Analysis and Opinion Mining." LREc. Vol. 10. 2010.

[2] Alec Go, Richa Bhayani, and Lei Huang. Twitter sentiment classification using distant supervision. Processing, pages 1-6, 2009.

[3] Niek Sanders. Twitter sentiment corpus. http://www.sananalytics.com/lab/twitter-sentiment/. Sanders Analytics.

[4] Alexander Pak and Patrick Paroubek. Twitter as a corpus for sentiment analysis and opinion mining. volume 2010, pages 1320-1326, 2010.

[5] Efthymios Kouloumpis, Theresa Wilson, and Johanna Moore. Twitter sentiment analysis: The good the bad and the omg!ICWSM, 11:pages 538-541, 2011.

[6] Hassan Saif, Yulan He, and Harith Alani. Semantic sentiment analysis of twitter. In The Semantic Web-ISWC 2012, pages 508-524. Springer, 2012.

[7] Isaac G Councill, Ryan McDonald, and Leonid Velikovich. What's great and what's not: learning to classify the scope of negation for improved sentiment analysis. In Proceedings of the workshop on negation and speculation in natural language processing, pages 51-59. Association for Computational Linguistics, 2010.

[8] Ilia Smirnov. Overview of stemming algorithms. Mechanical Translation, 2008.

[9] Martin F Porter. An algorithm for suffix stripping. Program: electronic library and information systems, 40(3):pages 211-218, 2006.

[10] Balakrishnan Gokulakrishnan, P Priyanthan, T Ragavan, N Prasath, and A Perera. Opinion mining and sentiment analysis on a twitter data stream. In Advances in ICT for Emerging Regions (ICTer), 2012 International Conference on. IEEE, 2012.

[11] John Ross Quinlan. C4. 5: programs for machine learning, volume 1. Morgan kaufmann, 1993.

[12] Steven Bird, Ewan Klein, and Edward Loper. Natural language processing with Python. " O'Reilly Media, Inc.", 2009.