Extracting Opinion Targets Using Attention-Based Neural Model
- PDF / 1,301,099 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 62 Downloads / 239 Views
ORIGINAL RESEARCH
Extracting Opinion Targets Using Attention‑Based Neural Model Saja Al‑Dabet1 · Sara Tedmori1 · Mohammad Al‑Smadi2 Received: 4 April 2020 / Accepted: 24 July 2020 © Springer Nature Singapore Pte Ltd 2020
Abstract Extracting opinion-target expression is a core subtask to perform aspect-based sentiment analysis which aims to identify the discussed aspects within a text associated with their opinion targets and classify the sentiment as positive, negative, or neutral. This paper proposes a deep learning model to tackle the opinion-target expression extraction task. The proposed model is composed of bidirectional long short-term memory as an encoder, long short-term memory as a decoder with an attention mechanism, and conditional random fields. This model, which operates at the sentence level, is designed to extract opinion targets for the Arabic language. The proposed model’s performance is evaluated using SemEval-2016 annotated dataset for the hotels’ domain. Experimental results show that the proposed model outperforms the baseline and the prior works, where it achieved an F1 measure of 72.83%. Keywords Aspect-based sentiment analysis · Deep learning · Opinion-target expression extraction
Introduction Nowadays, web services are used by many people around the world to share their experiences about different topics like travelling destinations, services, products, and events. Users’ generated content in social media websites (e.g. reviews, microblogs, posts, tweets, etc.) has an effective role in customers’ purchase decision, where it compromises a valuable resource for companies that aim to assess the satisfaction level of their customers and evolve better products and services for them. This can lead to a worthy insight for companies and customers. Sentiment analysis (SA) in general aims to analyze opinionated reviews as positive, negative, or neutral. Aspect-based sentiment analysis (ABSA) is a finegrained type of SA whose purpose is to identify the discussed aspects, opinion targets, and sentiment polarity of the opinionated text [1, 2]. Extracting opinion-targets expressions (OTEs) is one of the goals of ABSA as it extracts the explicitly mentioned opinion-expressions referring to specific aspects from within a sentence. These opinion expressions can be single words or phrases, each represented as an offset of clearly stated OTEs. For example, in a laptop * Saja Al‑Dabet [email protected] 1
Princess Sumaya University for Technology, Amman, Jordan
Jordan University of Science and Technology, Irbid, Jordan
2
review, “The screen resolution is amazing, but the price is very expensive”. The reviewer discusses two main OTEs which are “screen resolution” and “price”. This task can be helpful for different natural language processing (NLP) applications such as opinion summarization and question answering [3]. OTE extraction can be modeled as a sequence-labeling task where the purpose is to treat the entered text as a sequence of words x and then map them to the corresponding sequence
Data Loading...
