In this paper, we present results of our study on age-based categorization of short texts as 85 words per author. We introduce a novel set of features that will reliably work with short texts, and is easy to extract from the text itself without any outside databases. These features were formerly known as variables in readability formulas. We tested datasets presented two age groups children and teens up to age 15 and adults 20 years and older. Besides readability features, we also tested widely used n-gram features. Models trained on readability features performed better or as well as models trained on n-gram features. Model generated by Support Vector Machine with readability features yield to f-score 0.953.
With a wide spread of social media, growing problem is related to false identities. Younger people might pretend adults to access adult sites, and older people might pretend youngsters to communicate with youngsters. As we can imagine, this might lead to serious threats, as for pedophilia or other criminal activities. Thus, automatic age detection has serious practical application in social media.
While many works are published on text authorship profiling, social media poses two problems that are not solved this far.
The first problem is related to the amount of the text needed to make predictions.
Usually a large training data sets and long texts per author are used [
The second problem is related to the cost of feature extraction. Most of the recent
studies [
In this paper, we propose a novel set of features for author’s age based profiling that solves both previously mentioned problems. We call these new features the readability features. These features can be easily extracted using client side JavaScript and they make at least as best classifiers as widely used n-gram based features.
We suppose that authors reading skills and writing skills are correlated, and by
analyzing author’s text readability, we can conclude about his/her education level,
which at least to the particular age is correlated with the actual age of an author.
Therefore, we can employ old readability formulas that were developed already
before computerized era. Automated Readability Index [
ARI = 4.71× chwaroarcdtsers sentences − 21.43
+ 0.5 × words GFI = 0.4 × words + 100 × complexwords sentences words (1) (2)
As readability indexes are developed for texts with about 100 words, these are good candidates for our task. 2
Methodology
We collected short texts, average 85 words long, from different social media sources like Facebook, Blog comments, and Internet forums. All authors were identified, and they used in their texts Estonian language. We chose balanced and stratified dataset with 400 instances and with different age groups: 7-15 and 20-48.
We used three types of features in our training datasets: readability features, character n-grams and word n-grams.
Readability features are quantitative data about texts, as for instance an average number of characters in a word, syllables in word, etc. All together 14 different features were extracted from each text as shown in Table 1.
feature CPW WPS CPS explanation average number of characters per word average number of words per sentence average number of commas per sentence average number of syllables per word
Complex word in our feature set, is a loan from Gunning Fog Index [
Another type of features we used, are character n-grams. We extracted all occurred character bigrams and trigrams and using Χ2 attribute evaluation, we selected 119 character bigrams and 576 character trigrams.
Similarly, we extracted all occurred word unigrams, bigrams and trigrams and using Χ2 attribute evaluation, we selected as features 100 word unigrams, 30 word bigrams and 6 word trigrams.
We made four different datasets: with readability features, with character n-grams,
with word n-grams, and with all features combined. The models were generated using
Support Vector Machine, Logistic Regression and Naïve Bayes algorithm. Motivation
of using these algorithms comes from the literature [
As shown in Table 2, readability features trained a better classifier with Support Vector Machines and Logistic Regression, yielding to f-scores 0.953, and 0.95 accordingly. Naïve Bayes performed better with n-gram features. Combined feature sets did not improve the models.
Employing relations between reading and writing skills, and using features from old readability formulas proved to be an effective way to predict author age class. Readability features are in many ways favorable. First, they are easy to extract, they are self sufficient, and can be computed without any extra help. Syllable counting is problematic for some languages, but maybe it can be omitted, as syllable count is also not used in all readability indexes.
Secondly, when dealing with short texts, content-based features, as n-grams tend to be very context dependent, the topic can cause a rise of frequency of some words that can be associated to a particular age group. It seems, that how we write depends less on the context than what we write.
However, we have to address limitations of the current study. First, it is obvious, that we cannot use readability features to categorize older age groups. For most of the people, reading and writing skills will not improve continuously during the whole life.
Secondly, it is possible that good age based categorization results are caused by some specific property of Estonian language. For example, Estonian language has many agglutinative inflectional suffixes, and therefore grammatical richness yield directly to more syllables and longer words. Therefore, we look forward to test how readability features work with other agglutinative and inflectional languages.