Masking and Transformer-based Models for Hyperpartisanship Detection in News

Proceedings of Recent Advances in Natural Language Processing, pages 111–120 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_014 111 Interactive Learning Approach for Arabic Target-Based Sentiment Analysis Husamelddin A.M.N Balla , Marisa Llorens Salvador and Sarah Jane Delany School of Computer Science Technological University Dublin Dublin, Ireland {husamelddin.balla,marisa.llorens,sarahjane.delany}@tudublin.ie Abstract Recently, the majority of sentiment analysis re- searchers focus on target-based sentiment anal- ysis because it delivers in-depth analysis with more accurate results as compared to tradi- tional sentiment analysis. In this paper, we propose an interactive learning approach to tackle a target-based sentiment analysis task for the Arabic language. The proposed IA- LSTM model uses an interactive attention- based mechanism to force the model to focus on different parts (targets) of a sentence. We investigate the ability to use targets, right and left contexts, and model them separately to learn their own representations via interactive modeling. We evaluated our model on two dif- ferent datasets: Arabic hotel review and Ara- bic book review datasets. The results demon- strate the effectiveness of using this interac- tive modeling technique for the Arabic target- based sentiment analysis task. The model ob- tained accuracy values of 83.10 compared to SOTA models such as AB-LSTM-PC which obtained 82.60 for the same dataset. 1 Introduction Sentiment analysis (SA) is one of the most prolific research areas in computer sciences, which aims to identify and extract user opinions from reviews. This technique has become an essential part of a wide range of applications in the areas of politics, business, advertising and marketing as it can help in identifying people’s opinions towards related targets (Tang et al., 2015). Arabic is considered among the top 4 languages in terms of internet us- age (Boudad et al., 2018). With the rapid growth of Arabic web content and low resources for analyz- ing Arabic opinion mining, the need for accurate Arabic sentiment analysis tools is very necessary. There are three main classification levels in senti- ment analysis: document-level, sentence-level, and target-level. Document-level SA aims to classify the sentiment expressed in the whole document. It considers the whole document as a basic informa- tion unit (talking about one topic). Sentence-level SA aims to classify the sentiment expressed in each sentence. In traditional sentiment analysis, the de- tailed opinions of all targets of the entity (which are required in many applications) are not provided. To acquire these details, we need to use the tar- get level. Target-based sentiment analysis (TBSA) aims to classify the sentiment with respect to the specific targets of entities. The opinion holders can give different opinions for different targets of the same entity, like in this sentence: “The hotel is clean with good services, but the room was too small”. Target-based sentiment analysis is a fine- grained task in sentiment analysis. This kind of fine-grained target-based analysis generally relies on machine learning techniques that require large domain-specific datasets with manual training data (Hu and Liu, 2004). Target-based sentiment analysis has become more popular in recent research as it delivers more accurate results compared …

Content truncated to 500 words