Instance Selection Improves Cross-Lingual Model Training for Fine-Grained Sentiment Analysis

Proceedings of the 19th Conference on Computational Language Learning, pages 153–163, Beijing, China, July 30-31, 2015. c⃝2015 Association for Computational Linguistics Instance Selection Improves Cross-Lingual Model Training for Fine-Grained Sentiment Analysis Roman Klinger*‡ *Institute for Natural Language Processing University of Stuttgart 70569 Stuttgart, Germany Philipp Cimiano‡ ‡Semantic Computing Group, CIT-EC Bielefeld University 33615 Bielefeld, Germany roman.klinger@ims.uni-stuttgart.de cimiano@cit-ec.uni-bielefeld.de Abstract Scarcity of annotated corpora for many languages is a bottleneck for training fine- grained sentiment analysis models that can tag aspects and subjective phrases. We pro- pose to exploit statistical machine transla- tion to alleviate the need for training data by projecting annotated data in a source language to a target language such that a supervised fine-grained sentiment analysis system can be trained. To avoid a nega- tive influence of poor-quality translations, we propose a filtering approach based on machine translation quality estimation mea- sures to select only high-quality sentence pairs for projection. We evaluate on the language pair German/English on a corpus of product reviews annotated for both lan- guages and compare to in-target-language training. Projection without any filtering leads to 23 % F1 in the task of detecting aspect phrases, compared to 41 % F1 for in-target-language training. Our approach obtains up to 47 % F1. Further, we show that the detection of subjective phrases is competitive to in-target-language training without filtering. 1 Introduction An important task in fine-grained sentiment anal- ysis and opinion mining is the extraction of men- tioned aspects, evaluative subjective phrases and the relation between them. For instance, in the sentence “I really like the display but the battery seems weak to me.” the task is to detect evaluative (subjective) phrases (in this example “really like” and “seems weak”) and aspects (“display” and “battery”) as well as their relation (that “really like” refers to “display” and “seems weak” refers to “battery”). Annotating data for learning a model to extract such detailed information is a tedious and time- consuming task. Therefore, given the scarcity of such annotated corpora in most languages, it is in- teresting to generate models which can be applied on languages without manually created training data. In this paper, we perform annotation pro- jection, which is one of the two main categories for cross-language model induction (next to direct model transfer (Agi´c et al., 2014)). Figure 1 shows an example of a sentence to- gether with its automatically derived translation (source language on top, target language on bot- tom) and the alignment between both. Such an alignment can be used to project annotations across languages, e. g., from a source to target language, to produce data to train a system for the target lan- guage. As shown in the example, translation errors as well as alignment errors can occur. When using a projection-based approach, the performance of a system on the target language crucially depends on the quality of the translation and the alignment. In this paper we address two questions: • What is the performance on the task when training data for the …

Content truncated to 500 words