This paper focuses on the author profiling task published in the FIRE 2019 (Forum for Information Retrieval Evaluation), which includes automatic identification of the age, gender, and language variety of Arabic tweets. We think the author profiling task as a multi-Classification problem. We have used word and character based on TFIDF features, learned the logistic regression classifier to predict the labels. In the final results, our proposed method shows a good performance in terms of age prediction, the accuracy rate is 0.6250. Additionally, we have obtained 0.5111 and 0.9604 accuracy for gender and language variety classifications respectively. In the experiment, We have used the different feature combination and adjusted the feature parameters to test the system. The combination of word and character features can improve the prediction accuracy and enhance the system performance significantly.
With the continuous development of social media, the research of author profiling
task has significant progress that has been made [
features from bigram to 4-gram and its combinations as a standard of the label predictions. The final evaluation results show that in the model the age prediction accuracy can reach 0.6250. For the language variety, it is 0.9694, and 0.5111 accuracy for the gender.
Firstly, we read and parsed all .xml documents, and combined the each author's tweets into a single text. Secondly, in this paper we have proposed the method based on text vocabulary to extract the corresponding features. So we filtered out the non-text content in document, such as @, emoticons and URL. Thirdly, we normalized the text content, removed the unnecessary spaces, tabs and punctuations. 2.2
Following the successful of author profiling system[
We compared to the Logistic Regression classifier and Linear SVC classifier, found the performance of LR classification is more stable and simple. So we chose LR as the final classifier. In terms of feature selection, we have used the character features from bigram to 4-gram ,word feature from unigram and its combinations of the features, exploited TFIDF method to extract more representative features. Giving a tern, to calculate its TF, IDF and DF values, combine TF and IDF as features and remove features of DF value which is lower than predefined minimum and higher than the predefined maximum.
The process of our method is shown in Fig. 1. Firstly, we preprocessed the training data set. Secondly, based on TF and IDF values, we extracted the classification features. In the experiment, we combined the character features from bigram to 4-gram with the word feature from unigram. Thirdly, we learned the LR classifier (default setting all parameters) based on the filtered features. Finally, we used the learning classifier to predict the test data set.
The corpus2 of this task is consists of Arabic tweets, sign with age, gender, language variety labels. Data set divided into five groups, each group contains three languages, all of which belong to Arabic. In gender classification, the label including two types: male and female. Age label divided into three types: Under(< 25), Between(25-35) and Above(>=35).
Through by analyzing the corpus, we found the number of types is same in the labels and it is a balance state. 3.2
The performance of author profiling approach is evaluated by the joint accuracy. The accuracy is defined as the ratio of the predicted correct number Pc to the total predicted number Pt. 3.3
We split the training data, 80% for training and 20% for testing, to observe the different effect in the feature combinations. The experimental results are shown in Table 1. 2 https://www.autoritas.net/APDA/corpus/ Team DBMS-KU.2(Top 1) Nayel.1(Top 2) Nayel.3(Top 3) Yutong.2(Our team)
Parameter Combination
max_df = 0.7 min_df=4 max_df = 0.8 min_df=5 This paper presents the method based on multi-classification with word and character features for author profiling in Arabic tweets. In our method, we have chose word and character and their combinations as the features and classified the LR classifier. The final evaluation results show that the best performance is the combination features of gender and language (word unigram) + age (word + char-bigram-trigram). We have obtained 0.6250, 0.5111 and 0.9604 accuracy for age, gender and language variety classifications respectively. In the future work, we will consider the feature extraction of non-text content, and further improve the experimental performance.
This research was supported by the Social Science Fund of Heilongjiang Province of China (No.18TQB103).