Evaluating Industrial and Research Sentiment Analysis Engines on Multiple Sources

摘要

Sentiment Analysis has a fundamental role in analyzing users opinions in all kinds of textual sources. Computing accurately sentiment expressed in huge amount of textual data is a key task largely required by the market, and nowadays industrial engines make available ready-to-use APIs for sentiment analysis-related tasks. However, building sentiment engines showing high accuracy on structurally different textual sources (e.g. reviews, tweets, blogs, etc.) is not a trivial task. Papers about cross-source evaluation lack of a comparison with industrial engines, which are instead specifically designed for dealing with multiple sources. In this paper, we compare the results of research and industrial engines on an extensive experimental evaluation, considering the document-level polarity detection task performed on different textual sources: tweets, apps reviews and general products reviews, in both English and Italian. The experimental evaluation results help the reader to quantify the performance gap between industrial and research sentiment engines when both are tested on heterogeneous textual sources and on different languages (English/Italian). Finally, we present the results of our multi-source solution X2Check. Considering an overall cross-source average F-score on all of the results, X2Check shows a performance that is 9.1% and 5.1% higher than Google CNL, respectively on Italian and English benchmarks. Compared to the research engines, X2Check shows a F-score that is always higher than tools not specifically trained on the test set under evaluation; it is lower at most of 3.4% in Italian and 11.6% on English benchmarks, compared to the best research tools specifically trained on the target source.