In this paper we describe the participation of the Laboratory of Language Technologies of INAOE at PAN 2014. We address the Author Profiling (AP) task finding and exploiting relationships among terms, documents, profiles and subprofiles. Our approach uses the idea of second order attributes (a lowdimensional and dense document representation) [4], but goes beyond incorporating information among each target profile. The proposed representation deepen the analysis incorporating information among texts in the same profile, this is, we focus in subprofiles. For this, we automatically find subprofiles and build document vectors that represent more detailed relationships of documents and subprofiles. We compare the proposed representation with the standard Bag-of-Terms and the best method in PAN13 using the PAN 2014 corpora for AP task. Results show evidence of the usefulness of intra-profile information to determine gender and age profiles. According to the PAN 2014 official results, the proposed method was one of the best three approaches for most social media domains. Particularly, it achieved the best performance in predicting age and gender profiles for blogs and tweets in English.
For several years the scientific community has been discussing the following basic
question: How much information can be known from an author’s document?, Commonly
known as the Author Profiling (AP) [
In this paper, we use well-known textual features for AP [
In this work, we generate new highly informative attributes that represents relationships among terms, documents, profiles and also subprofiles. In order to automatically generate the aforementioned subprofiles we propose dividing each target profile into several groups using a clustering algorithm. Then we build the final document representation on the top of the generated subprofiles, using them as the new target profiles. This approach improves even more the representation of documents in the AP task, and also mitigates common problems of other standard representations (e.g. the Bag-of-Terms, BOT), for example: i) high dimensionality, and ii) the sparseness of the representation. Results using the latter ideas also seem promising and competitive compared with other approaches and systems in PAN 2014 forum.
The rest of this paper is organized as follows: Section 2 introduces the proposed approach, Section 3 explains the evaluation and the obtained results, finally Section 4 outlines the conclusions. 2
The proposed method has three main stages to represent documents: i) representing terms in a space of profiles, ii) representing documents in a space of profiles, and iii) generating subprofiles and re-compute steps 1 and 2 using subprofiles as the new target profiles. The rest of this section explains the above steps in detail. 2.1
The intuitive idea is to capture the relationship of each term (i.e., a word, an n-gram, a punctuation mark, etc.) with each one of the target profiles. Let ft1; : : : ; tmg be the vocabulary in the corpus, and fp1; : : : ; png be the set of target profiles. We build term vectors ti = htpi1; : : : ; tpini, where tpij represents the relationship of the term ti with the profile pj . Equation 1 reflects the idea for computing tpij .
wij =
X k:dk2Pj log2 1 +
tf ki len(dk) (1) where Pj are training documents with profile pj , tfik is the term frequency of the term ti in the document dk, and len(dk) is the number of terms in the document dk. To avoid high cumulative term frequencies in high unbalanced data, a normalization that considers the proportion of each term in each profile is performed (Equations 2.1 and 2.2).
wij (2:1) tpij = T ERMS
X
wij i=1
wij (2:2) tpij = P ROF ILES
X j=1 wij 2.2
The intuitive idea is to capture the relationship of each document (i.e., a blog, a tweet, a review etc.) with each one of the target profiles. For this we use the previously computed term vectors to build document vectors in a space of profiles. Thus, to build the representation of each document we add its term vectors weighted by their frequency tfkj . Thus, we build document vectors dk = hdp1k; : : : ; dpnki, where n is number profiles, and dpik reflects the relationship of the document dk and the profile pi. Equation (3) expose the latter ideas. (2) (3) dk =
X ti Dk
tfik len(dk) ti where Dk is the set of terms that belongs to document dk 2.3
The latter ideas generate vectors where each value represents the relationship between
a document and each target profile. The intuitive idea is that the representation assumes
certain homogeneity in documents belonging to the same target profile, then a single
relationship value per profile is computed [
Generating subprofiles involves discovering natural subgroups among authors belonging to the same profile. In this regard, we decide to use the latter generated document representation to build document vectors in a space of profiles, then cluster documents in the same profile. The intuitive idea of this approach, is to use an appropriated base representation for AP to find documents similar to each other inside that space. Once a set of clusters (subprofiles) for each target profile are generated, we rebuilt the SOA using all found clusters as the new target profiles. In this way, as indicated in formula (3), we end up with a set of attributes that represents relationships between documents and detailed subprofiles. In order to build the aforementioned subprofiles, we have used the Expectation Maximization Clustering (EMC) algorithm.
We approached the AP problem in a separated way: i) age, and ii) gender prediction.
Thus, we have five age profile classes; 18-24, 25-34, 35-49, 50-64, and 65-more. Also
we have two gender profiling classes; male, female. Given this context, we build
subprofile attributes for age, and different subprofile attributes for gender, then we train two
classifiers, one for each representation. In order to evaluate and compare this proposal,
we have used the following experimental settings for the training dataset: i) the most
3,000 frequent terms as features, and ii) the standard LibLINEAR classifier without any
parameter optimization [
According to the PAN14 evaluation, the proposed attributes, get the best test-set accuracy performance for age and gender prediction in blogs and tweets domains for English language. Moreover, the reported results are in the top 3 positions for other social media domains. Thus, the approach presented in this paper is an effective alternative to address the AP task in different social media domains, where documents presents challenging difficulties hindered the accurate work of most natural language processing tools. 4
In this paper we presented a novel approach that considers the existing information among documents belonging to the same class. This is, even for authors belonging to the same target profiles (e.g., males), the approach look for more specific subgroups of authors (e.g., male employees and male gamers) in order to consider intra-profile information. To the best of our knowledge, this is the first time that AP is addressed using this kind of intra-class-relationships inside the target profiles. Such relationships help to achieve a better discrimination among several profiles. Using these automatically generated attributes, the classifier can keep good classification rates, even for imbalanced data. This is due to the relations among terms, documents and subprofiles, which provides few but more detailed predictive attributes. We have shown better experimental results than the standard BOT, the best method of PAN13, and most of the approaches participating at PAN14.
Acknowledgements: This work was partially funded by project CONACyT-Mexico 134186 and the program ECOS under the project M11-H04. López-Monroy also thanks for doctoral scholarship CONACyT-México 243957.