Author profiling is the task of determining the age, gender or type of the author's personality by studying their sociolect aspect, that is, how the language is shared by people. This paper presents the COMSATS Institute of Information Technology, Lahore entry for the PAN 2015 competition on Author Profiling task. Our proposed system is based on stylometry features. We implemented 29 different stylistic features, many of which are language independent. Since the training data was available in multiple languages, one of our main objectives was to explore which language independent features are most effective. The problem of author profiling was casted as a supervised document classification task. Results showed that features (Percentage of Question Sentences, Average Sentence Length, Percentage of Punctuations, Percentage of Comma and Percentage of Full stops) were most effective multilingual features.
Personality in Encyclopedia of Psychology is defined as “individual differences in characteristic patterns of thinking, feeling and behaving.” [1] These differences can be reflected through one’s speech, writing, images etc. Authorship analysis deals to find out the methods to identify these differences and use them to know profile of author such as gender, age, native language, education, profession or personality type. So author profiling can simply be defined as: given the set of texts, you need to identify age, gender, profession, education, native language and similar personality traits. Authorship analysis has attracted much attention in recent years due to the rapid increase of electronic text and the need for expert systems able to handle this information. Like from a marketing viewpoint, companies may be interested in knowing about the trends regarding their products, on the basis of the analysis of blogs and online product reviews, what types of people like or dislike their products, what is required by the customer and which customer category/ class they can attract more. Similarly from a forensic viewpoint, determining the linguistic profile of a person i.e. who wrote a "suspicious text"’ may provide valuable background information. In this paper we explore how different stylistic features help in conveying about the personality of writer. Moreover, we also tried to find out how the use of different stylistic features affect multilingual results. We used the datasets provided by PAN organizers, and applied different machine learning practices for predicting writer’s traits.
The rest of this paper is organized as follows: Section 2 describes related work. Section 3 presents the approach used to identify personality traits of author. Section 4 describes the results obtained on training data. Section 5 presents the results obtained on test data and section 6 concludes the paper. 2
With the advancement in web technology, there is an abundant increase in use of electronic media. Along with useful information there is also heap of anonymous data available so the need to identify “who is who” has become important. A lot of establishments have been done in this field so far, Pennebaker et al. [2] joined dialect utilization with identity qualities, examining how the variety of semantic attributes in a content can give data in regards to the gender and age of its author. Argamon et al. [3] analyzed formal written texts extracted from the British National Corpus [4]1 combining function words with part-of-speech features. Koppel studied the problem of automatically determining an author’s gender by proposing combinations of simple lexical and syntactic features.
Holmes and Meyerhoff [7], Burger and Henderson [6] have also investigated obtaining age and gender information from formal texts.
Seifeddine and Maher [8], focus is on author’s discussions, they combine content based approach and statistic approach. They show a hemi- strategy that merges the analysis of information in writings with a machine learning technique. To get a higher organization of these statistics, they depended on the utilization of the “Decision table calculation".
The approaches commonly used for the automatic identification of an author’s
personality traits from text can be categorized into three broad categories:
1 www.natcorp.ox.ac.uk
(
Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes
Yes Yes Yes Yes Yes Yes
Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes ---------Yes ------------------Yes Yes ---Yes Yes ---------------Yes Yes Yes ---------Yes ------------------Yes Yes ---Yes Yes ---------------
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
---------Yes
------------------Yes
Yes
---Yes
Yes
---------------The system presented in this PAN Author Profiling Competition is based on stylometry.
Table 3.1 shows the list of 29 stylistic features used for the development of our proposed
author profiling detection system. As can be noted that these features aim to extract
different stylistic information from a single document, which can be helpful in
identifying the age, gender, personality type i.e. stable, open, extroverted, agreeable and
conscientious .In this year’s training data, the tweets are available for four different
languages: (
Our proposed system was trained using the PAN 2015 Author Profiling training data,
which consists of Twitter tweets in four different languages: (
The task of identifying an author’s profile from text was treated as a supervised machine learning problem. N-fold-cross-validation was used. Due to difference in the sizes of training data for different languages, we used 5-fold cross validation for English language corpus, 4-fold cross validation for Dutch and 3-fold cross validation for the Italian and Spanish corpora respectively.
We applied a range of machine learning algorithms on the training data including Navies Bayes, Support Vector Machine, Random Forest, J48 and Logistics. The WEKA’s2 [9] implementation of these algorithms was used. The scores generated using the stylistic features (see Table 3.1) were used as “input features” to the machine learning algorithms.
We used WEKA’s “attribute selection” approach for selecting “best features” from the complete feature set. For that purpose we explored CFsSubSetEval, Filtered Attribute, 2 mailto:https://weka.wikispaces.com/ SVM Attribute and Classifier Subset as evaluators, and Best first and ranker as search methods.
As recommended by PAN 2015 organizers, the performance of the proposed system for age and gender personality traits was measured using accuracy, whereas for other personality traits (stable, open, extroverted, agreeable and conscientious) average Root Mean Squared Error was used. 4
We carried out three sets of experiments: (