Open Information Extraction from Real Internet Texts in Spanish Using Constraints over Part-Of-Speech Sequences: Problems of the Method, Their Causes, and Ways for Improvement


Usually we do not know the domain of an arbitrary text from the Internet, or the semantics of the relations it conveys. While humans identify such information easily, for a computer this task is far from straightforward. The task of detecting relations of arbitrary semantic type in texts is known as Open Information Extraction (Open IE). The approach to this task based on heuristic constraints over part-of-speech sequences has been shown to achieve high performance with lower computational and implementation cost. Recently, this approach has gained spread and popularity. However, Open IE is prone to certain errors that have not yet been analyzed in the literature. Detailed analysis of the errors and their causes will allow for faster and more focused improvement of the methods for Open IE based on this approach. In this paper, we analyze and classify the main types of errors in relation extraction that are specific to Open IE based on heuristic constraints over part-of-speech sequences. We identify the causes of the errors of each type and suggest ways for preventing such errors, with corresponding analysis of their cost and scale of impact. The analysis is performed for extractions from two Spanish-language text datasets: the FactSpaCIC dataset of grammatically correct and verified sentences and the RawWeb dataset of unedited text fragments from the Internet. Extraction is performed by the ExtrHech system. Extracción abierta de información a partir de textos de internet en español utilizando reglas sobre categorías de palabras en secuencias: Problemas del método, sus causas y posibles mejoras

Extracted Key Phrases

5 Figures and Tables