Improving Accuracy in Language Analysis, the case of Multilingual Sentiment Analysis - Symbolic and Machine Learning Approaches
Accuracy is probably the main challenge that language technology is facing in Social Media and Big Data Analysis, due mainly to the poor quality and heterogeneity of texts.
Machine Learning approaches are getting established as mainstream technology in language analysis, as they offer many advantages. However, there are disadvantages too. For example, Machine Learning works as a black box, making it difficult for the user to identify and correct specific errors, i.e. making it difficult to increase accuracy on specific problems, hence for incremental improvement.
Symbolic approaches, based on dictionaries and grammars, tackle the problem of incremental improvement in an efficient way. Symbolic approaches work as a white box where users can find out why errors happen and correct them, providing a suitable environment for high accuracy and continuous improvement.
During the talk, a symbolic system will be presented. This system has been developed by Bitext and is in use in the industry in companies like Salesforce.
We will raise open questions for the audience, like "how do we achieve the quality required by markets like Social CRM?", "Can machine learning and symbolic approaches complement each other?", etc.