![](bgb.png)
training increases. The average relative loss
improved from -7.66% for training on a single
domain to +0.8% for training on k-1 domains. The
reduction in relative loss suggests that the proposed
model handles new domains increasingly better as
its training corpus grows, which is the goal of our
domain independent sentiment polarity identification
approach.
7.2 Keyword Classifier
We chose between the classifier candidates offered
by Weka (Witten, et al., 2011). In terms of feature
vector, we performed four experiments:
E1: keep all words from the document as a
separate instance;
E2: collapse duplicates words into a single
instance;
E3: collapse synonyms into unique feature,
without semantic verification;
E4: collapse synonyms into unique feature,
only in case of similar semantic meaning.
All the four experiments use balanced datasets
with at least 200000 instances. They were created by
labelling the words from articles written in January
2014 on TechCrunch. The words were labelled in
keywords/non-keywords. For each experiment we
perform 10 fold cross-validations. The validation
results are presented in Table 10. We measure the
overall accuracy together with the precision and
recall for the positive and negative classes. We can
observe that the classifier J48 (decision tree)
performs better than Naïve Bayes (NB) for the small
set of meta-features in three of the experiments. In
the first experiment, Naïve Bayes performs slightly
better than J48 because this experiment considers all
the words from the documents (more than 600000
instances). The best results are obtained with the
setup of experiment 4 which combines the features
of words that have the same sense.
Table 10: Classifier comparison for Keyword detection.
Experiment
Classifier
Name
Accuracy
Precision
non-keyword
Precision
keyword
Recall
non-keyword
Recall
keyword
E1
NB 85.12 .836 .869 .827 .875
J48 85.08 .867 .836 .829 .873
E2
NB 82.65 .818 .835 .839 .814
J48 87.7 .919 .844 .828 .927
E3
NB 80.89 .900 .751 .695 .923
J48 89.18 .901 .883 .881 .903
E4
NB 82.14 .889 .774 .735 .908
J48 89.36 .901 .887 .884 .903
We are further interested in ranking each individual
feature with respect to their information gain. The
top 3 most information-bearing features are the part-
of-speech, tf-idf and the firstPosition.
8 CONCLUSIONS
This paper proposes a document sentiment polarity
identification approach based on an ensemble of
meta-features.
We propose the use of three meta-feature classes
that boost domain-independence increasing the
degree of generality. Sentiment lexicons provide a
basis for the analysis. Part-of-speech patterns reflect
syntactic constructs that are a good indicator of
polarity. Finally, polarity histograms provide an
insight in the distribution of polarized words within
the document. All three interact in order to associate
sentiment polarity to a document.
We incorporated sentiment detection into a
context-sensitive recommendation flow. The
language-agnostic input context is analysed and
reduced to a representative document. Based on its
identified sentiment orientation and thematic
distribution we recommend thematically similar
content with the same sentiment orientation.
We are currently integrating a more advanced
approach for negation detection leveraging typed
dependencies (Marneffe, et al., 2006). We also
consider exploring objectivity with the help of
undistinguishable sentiment lexicons and a third set
of part-of-speech patterns. Further efforts will be
focused on adapting our meta-feature approach to an
optimal dataset size for the problem of cross-domain
sentiment identification. We aim to shift towards an
unsupervised approach for sentiment detection.
REFERENCES
Baccianella, S., Esuli, A. & Sebastiani, F., 2010.
SentiWordNet 3.0: An Enhanced Lexical Resource for
Sentiment Analysis and Opinion Mining. Proceedings
of LREC.
Blei, D. M., Ng, A. Y. & Jordan, M. I., 2003. Latent
dirichlet allocation. The Journal of Machine Learning
Research, Volume 3, pp. 993-1022.
Blitzer, J., Dredze, M. & Pereira, F., 2007. Biographies,
Bollywood, Boom-boxes and Blenders: Domain
Adaptation for Sentiment Classification. Proceedings
of the 45th Annual Meeting of ACL.
Dinsoreanu, M., Macicasan, F. C., Hasna, O. L. & Potolea,
R., 2012. A Unified Approach for Context-sensitive
KDIR2014-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
136