Sentiment analysis for Serbian language
2 likes2,127 views
This document discusses sentiment analysis of sentences in Serbian language. It notes the need for sentiment analysis tools in Serbian due to a lack of existing natural language processing tools and the industrial need for automated market research and customer satisfaction analysis. It then describes the Serbian language and an overview of the sentiment analysis workflow including tokenization, preprocessing, stemming, sentiment analysis using naive Bayes classification, and a web API implementation. Evaluation results show the Serbian stemmer achieves 90% accuracy on news articles while the sentiment analyzer achieves 80% accuracy. Future work is proposed to improve the stemmer and sentiment analysis.
Sentiment analysis for Serbian language
- 1. Sentiment analysis of sentences in Serbian language Nikola Milošević
- 2. Why to analyze sentiment in Serbian? ● Great industrial need – – Automated market research – ● Ads websites Customer satisfaction NLP tools for Serbian are not developed – Need for tools and resources – Almost no accessible tools through API
- 3. Serbian language ● Belongs to Indo-Europian language group ● Slavic language ● Highly inflectional ● 3 pronunciation types ● 3 dialect groups ● Write as you speak ● Latin and Cyrillic writing system
- 5. Tokenization and preprocessing ● Process of breaking a stream of text up into words ● Stop-word filtering ● Negation handling – – ● Adding NE_ prefix after negation All words before punctuation Irregular verbs
- 6. Stemming ● Process for reducing inflected words to their stem, base or root form ● Kešelj and Šipka (2008) ● Hand crafted rule based stemmer ● ~300 rules
- 7. Sentiment analysis ● Aim to build binary sentiment analysis ● General Serbian language ● No annotated corpus for Serbian ● Annotation work (~1000 small texts) ● Supervised machine learning
- 8. Naive Bayes ● Algorithm that learns fast ● Bag of words approach ● Assumption of conditional independence ● Laplace smoothing
- 9. Implementation ● Web API with presentation layer ● JSON communication ● Secured page for annotating ● Build using PHP and MySQL ● Web & Android
- 10. Results ● Stemmer – – 90% correct on news articles – ● Smallest and most precise stemmer Problems: small words, irregular inflections, voice changes Sentiment analyzer – 80% correct – Problems: Irony, ambiguity, small training data
- 11. Future work ● Stemmer – – ● Use snowball framework Build multi-step stemmer Sentiment analyzer – POS tagging – Complex negation handling – SVM algorithm
- 12. Thank you ● Available from http://inspiratron.org ● Contact: nikola.milosevic@postgrad.manchester.ac.uk