Argumentation mining is a text mining task, which aims at automatically detecting the argumentative structure concealed in a huge amount of text data. Previous researches in this eld have focused on the classi cation of text sentences as arguments and the detection of relations between them. Di erent corpora have been used, such as newspaper articles and online debates. Due to the explosion of social networks, microblogging platforms like Twitter and Facebook have become interesting tools to evaluate public opinion on di erent domains. In this work, we propose a new pipeline process to achieve the goal of argumentation mining, based on 70 millions of twitter-microblogs released from MC2 CLEF-2018 lab dealing with cultural events. Our approach is based on Information Retrieval protocol combined with sentiment analysis and topic modeling using Latent Dirichlet Allocation (LDA).
Social network became an important source of information considering the number of users and the data exchanged. Twitter is one of the most famous platforms in the world, millions of people express themselves every minute by simply sharing a 140 characters short text (tweet ).SS The rst question which should be asked is can we use this large data ? The answer is yes, thanks to the TwitterAPI, people can have a free access to twitter's data. But that's not enough, unless we know what we are looking for.
Dealing with such data requires precautions : We must understand that a normal person can't analyse (manually) a huge volume of text data, that's why we need to ask computers to do it for us, using speci c methods and algorithms. The analysis of twitter text data must be di erent from the other types of text, because tweets are pressed, noisy and often unstructured.
Our purpose is to explore twitter data and to identify the argumentative parts
of it, but rst we need to know how can we consider a tweet as an argument.
Mochales Palau and Moens [
To achieve our goal, we create a pipeline containing three popular techniques : Information Retrieval, sentiment analysis and topic modeling. Each technique has it's own particularities; Information Retrieval, based on the indexing and the querying, allows a simple searching process of the whole data to get the result of the query. Sentiment analysis aims to categorize opinions about something (product, event, etc.). Topic modeling allows the discovery of hidden semantic topics into the text data.
We believe that combining these techniques can achieve the aim of argumentation mining, by identifying the most argumentative tweets within a 70 millions microblogs dataset of cultural events. The results con rm that our pipeline succeeded to identify the most 100 argumentative tweets compared to the baseline. 2
The MC2 CLEF2018's data contains 18 text les, each one represents a collection of tweets which contains 16 topics in di erent languages during one month. Each topic is a festival name, there are 12 festivals in english and 4 in french. Since our goal is to extract the most 100 argumentative tweets of each topic, we created a pipeline and we put these 18 les through it. This pipeline is composed of information retrieval, topic modeling and sentiment analysis. 2.1
Information retrieval is the process of obtaining relevant information about a collection of text data, we choose the Indri platform to make our IR process. A speci c format called TRECText (or TRECWeb) is required, any other format won't be recognized. As our input les are text documents, we need to get the correct format, that's why we have used Perl language to do the job. We wrote a program which accepts a .txt le as input and return a .trec le as output without changing the content of our tweets. Example of TRECText format :
Creating the index of our text data allows us to do the querying, for each
topic we wrote a lexicon-based query in Indri language, example :
Art festival # weight[m](1.0 # band(Art festival) 0.5 # or(abnormal aborted))
We demand the engine to return the best 500 tweets for each topic based on
their scores which have been calculated by default with Indri. It uses a query
likelihood function with Dirichlet [
score = log( numerator / denominator )
By default, mu is equal to 2500, which means that for the very small
documents you're using, the score di erences will be very small. [
We have made this process twice, one for english topics and one for french, so we have two result les which are considered as our baseline. 2.2
The baseline from the previous step does not con rm that the 500 resulting tweets are talking about the same topic because it is an index-quey process, the system is not intelligent enough to provide a list of tweets which have exactly the same topic simply by a lexicon query. That is why we need to use a machine learning technique to train a model that con rms the matching result.
Topic modeling is a modern technique frequently used in machine learning and natural language processing, it aims to extract topics from a collection of text documents. It helps in understanding our data, organizing it, and discovering hidden information into it. There are many methods that are used to obtain topic models, in this work we focused on Latent Dirichlet Allocation (LDA).
LDA is a popular method of topic modeling which considers each document as a blend of topics, and each topic as a blend of words. We imagine that we have the following tweets (documents): { Tweet 1 : Jazz festival is the best, i can't wait for the next year. { Tweet 2 : Rock festival is always amazing !! { Tweet 3 : Jazz music is relaxing !! { Tweet 4 : I don't like rock music.
The LDA model discover the topics in these tweets, it may produce these results : { Topic 1 : 50 % Festival, 15% Jazz, 15% Rock.. (This topic deals with festivals) { Topic 2 : 50 % Music, 15% Jazz, 15% Rock..(This topic deals with music) { Tweet 1 and 2 : 100% topic 1 { Tweet 3 and 4 : 100% topic 2
The hard part when working with LDA is how to nd the number of topics.
In general, the use of metrics, like Gri ths2004 [
The preprocessing phase is needed because of the noise of tweets, we have removed punctuations, URLs, stop words (english and french), twitter identi ers and we have done the stemming process. The gure below presents a wordcloud of our tweets after cleaning, and it contains the most frequently words.
After cleaning our tweets, we have used the LDA function provided by the topicmodels package in R. This function was fed with our baseline results.
As LDA supports two methods, VEM and Gibbs Sampling, we need to choose which ts our model.
The estimates for the topic distributions for the documents are included
which are the estimates of the corresponding variational parameters for the VEM
algorithm and the parameters of the predictive distributions for Gibbs sampling.
In additional slots the objects contain the assignment of terms to the most
likely topic and the log-likelihood which is log p(w j ; ) for LDA with VEM
estimation and log p(wj z) for LDA using Gibbs sampling.[
The estimation using Gibbs sampling requires speci cation of values for the parameters of the prior distributions, that's why we choose VEM.
The aim is to score each tweet by how much it belongs to his expected topic. We can examine the per-document-per-topic probabilities called Gamma. This will result k number of clusters or groups of words, we manually annotate these groups to do the matching between score and topic for every single tweet. 2.3
Sentiment analysis is a text classi cation tool that uses NLP and machine learning to analyse text and extract the writer's sentiment about a given topic, it is widely used in online debates and social media platforms. Since our data is a collection of tweets dealing with cultural events, we can apply the sentiment analysis on it to discover what people feel about each topic.
The same baseline is considered as our input, and since we have cleaned the tweets (see previous section), the data is ready to be analysed. We have used the R sentiment package, which perform sentiment analysis by scoring individual words based on the a nity score de ned by NRC 3. 3 The NRC Emotion Lexicon is a list of English words and their associations with eight basic emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust) and two sentiments (negative and positive).
To get a sentiment score of each tweet in our corpus, we applied the get nrc sentiment() function. This function doesn't stop at providing a positive or negative score for a tweet, it goes deeper to tell us the other emotions that a tweet contains (fear, joy, ..), in addition, it counts the frequency of each emotion word. With these sentiments, we can calculate a new score which re ects how the writer of the tweet have felt. 3
After nishing the process, we have three scores for each tweet, as our goal is not particularly one of them, we need to put these scores together. The new calculated score is the key to get a new ranking.
Score = IRscore + LDAscore + Sentimentscore Based on this score, we re-ranked the tweets. To test our approach, we need to compare the new results with the baseline. We have picked the top 3 tweets from Toronto festival, and located the same tweets in the baseline ranking.
The whole ranking is changing after the process, besides w didn't show the 500 tweets of each topic, we just take the best 100 between them. It's remarkable that the top ranked tweets have become more argumentative than before.
Argumentation mining is a new research eld which aims to automatically extract the argumentation structure hidden in a collection of text data. Previous studies have been through the classi cation of sentences into argument and non-argument, based on a manually annotated corpus and using static machine learning algorithms.
In this paper, we tried to achieve the same goal with a di erent style, because we think that annotating a corpus of tweets is a hard task, besides it is not accurate due to human errors. We tried to combine computational techniques that made our results more speci c, especially with the ranking method. Our model provides an easy result list which contains a summary of ranked tweets according to their scores.