Fasttext and Gradient Boosted Trees at GermEval-2017 on Relevance Classification and Document-level Polarity
Abstract
This paper describes the submissions to the Shared Task on Aspect-based Sentiment in Social Media Customer Feedback for the GermEval 2017-workshop for the two subtasks Relevance Classification (task A) and Document-level Polarity (task B). For each subtask, the results of the same three systems were submitted: a fastText classifier, enhanced with pretrained vectors, gradient boosted trees (GBTs) trained on bag-of-words (BOWs), and an ensemble of GBTs, respectively trained on word embeddings and on BOWs. For the subtask Relevance Classification, the best system yields a micro-averaged F1-score of 0.895 on the dev set. For the subtask Document-level Polarity, the best system achieves 0.782 on the test set. The proposed system achieved the second place out of twelve systems submitted by seven teams for task A for both test sets. For task B, the proposed system achieved the first place for test set one and the second place for test set two out of 17 systems submitted by eight different teams.
Keywords
GermEval-2017