Institutional Repository
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Scaling text processing pipelines using Apache Spark

Katsani Merieme

Full record


URI: http://purl.tuc.gr/dl/dias/6D2F87E6-508B-4687-97B1-9A5CA530ACB6
Year 2017
Type of Item Diploma Work
License
Details
Bibliographic Citation Merieme Katsani, "Scaling text processing pipelines using Apache Spark", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2017 https://doi.org/10.26233/heallink.tuc.70651
Appears in Collections

Summary

Big data, which is derived from humans or machines, starting with social media and extending to smartphones or sensors, in forms of texts, images or transactions, is a continuously evolving field.Thus, the ongoing increase of data generated creates a need for knowledge extraction from it, through data analysis. Several areas are engaged in data mining, and in particular the area of machine learning which has been well established over the past years. Various techniques and methods of machine learning are trying to solve big data problems and these two areas consistnow an integral part. This particular combination is the main subject of this study, which aims to implement a large-scale text processing architecture. More specifically, this architecture focuses on processing streaming texts derived from Reddit in real-time and the classification thereof as sarcastic or non-sarcastic through a machine learning model. The architecture uses the latesttechnologies in the field of information processing through distributed platforms such as Apache Kafka and Spark as well as state-of-the-art but also simple and powerful ML algorithms, i.e Random Forests, Naive Bayes and Logistic Regression. After comparing the methodology and design of each individual piece forming the final layout, a selection of the most appropriate model is made followed by the implementation of the framework. Success rates exported were quite closeto the relevant literature and sometimes higher, depending on each technique examined. Finally, results are indexed in the distributed search engine Elasticsearch and are evaluated through the Kibana plugin.

Available Files

Services

Statistics