The Impact of Semantic Linguistic Features in Relation Extraction: A Logical Relational Learning Approach

Proceedings of Recent Advances in Natural Language Processing, pages 648–654, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_076 648 The Impact of Semantic Linguistic Features in Relation Extraction: A Logical Relational Learning Approach Rinaldo Lima1, Bernard Espinasse2, Fred Freitas3 1 Departamento de Computação, Universidade Federal Rural de Pernambuco, Brazil 2 LIS UMR CNRS 7020, Aix-Marseille Université/Université de Toulon, France 3 Centro de Informática, Universidade Federal de Pernambuco, Brazil rinaldo.jose@ufrpe.br, bernard.espinasse@lis-lab.fr, fred@cin.ufpe.br Abstract Relation Extraction (RE) consists in detect- ing and classifying semantic relations be- tween entities in a sentence. The vast ma- jority of the state-of-the-art RE systems re- lies on morphosyntactic features and super- vised machine learning algorithms. This paper tries to answer important questions concerning both the impact of semantic- based features, and the integration of exter- nal linguistic knowledge resources on RE performance. For that, a RE system based on a logical and relational learning algo- rithm was used and evaluated on three ref- erence datasets from two distinct domains. The yielded results confirm that the classi- fiers induced using the proposed richer fea- ture set outperformed the classifiers built with morphosyntactic features in average 4% (F1-measure). 1 Introduction Relation Extraction (RE) consists in detecting and classifying binary semantic relations between enti- ties in a sentence. Many RE systems use statistical machine learn- ing techniques, such as feature-based and tree ker- nels-based methods (Choi et al., 2013) are based on a propositional hypothesis space for representing examples, i.e., they employ an attribute-value rep- resentation achieving robust results. However, this representation is not able to effectively capture structural information from parse trees without loss of information (Choi et al., 2013). More recently, other RE systems using deep learning techniques, such as Convolution Neural Networks (Nguyen and Grishman, 2015) and Recurrent Neural Net- works (Miwa and Bansal, 2016) have also been proposed. They are based on a dense vector repre- sentation of the input words retrieved from word embeddings (Mikolov et al., 2013). Some other RE systems rely on natural language processing (NLP) techniques for extracting rele- vant features from the text. They typically integrate shallow NLP tools for coping with lexical and syn- tactic aspects of the texts such as POS tagging, lem- matization, chunking, and syntactic parsing (Choi et al., 2013). However, according to Zouaq (2011), there are two main reasons to seriously consider deeper linguistic processing for RE: (i) it may pro- vide deeper semantic meaning; (ii) NLP tools are becoming sufficiently robust to be considered as re- liable tools for knowledge and model extraction. Contrarily to related work above, this work sub- scribes to the idea of performing RE employing the Logical and Relational Learning (LRL) (de Raedt, 2008) approach that can generate classification models from complex data structures such as graphs or multiple tables. More precisely, we rely on Inductive Logic Programming (ILP) (Muggle- ton, 1991) as one of the most successful relational learning techniques because it employs a symbolic and declarative representation for the examples and the extraction models are both understandable and interpretable by …

Content truncated to 500 words