NEUROSENT-PDI at SemEval-2018 Task 7: Discovering Textual Relations With a Neural Network Model

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 848–852 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics NEUROSENT-PDI at SemEval-2018 Task 7: Discovering Textual Relations With a Neural Network Model Mauro Dragoni Fondazione Bruno Kessler Via Sommarive 18 Povo, Trento, Italy dragoni@fbk.eu Abstract Discovering semantic relations within textual documents is a timely topic worthy of inves- tigation. Natural language processing strate- gies are generally used for linking chunks of text in order to extract information that can be exploited by semantic search engines for per- forming complex queries. The scientific do- main is an interesting area where these tech- niques can be applied. In this paper, we de- scribe a system based on neural networks ap- plied to the SemEval 2018 Task 7. The sys- tem relies on the use of word embeddings for composing the vectorial representation of text chunks. Such representations are used for feeding a neural network aims to learn the structure of paths connecting chunks associ- ated with a specific relation. Preliminary re- sults demonstrated the suitability of the pro- posed approach encouraging the investigation of this research direction. 1 Introduction One of the emerging trends of natural language technologies is their application to scientific litera- ture. There is a constant increase in the production of scientific papers and experts are faced with an explosion of information that makes it difficult to have an overview of the state of the art in a given domain. Recent works from the semantic web, sci- entometry, and natural language processing (NLP) communities aimed to improve the access to scien- tific literature, in particular to answer queries that are currently beyond the capabilities of standard search engines. Examples of such queries include finding all papers that address a given problem in a specific way, or to discover the roots of a certain idea. The NLP tasks that underlie intelligent process- ing of scientific documents are those of informa- tion extraction: identifying concepts and recog- nizing the semantic relation that holds between them. Information extraction from corpora includ- ing relation extraction and classification normally involves a complicated multiple-step process. In this paper, we present a neural network strat- egy for addressing the challenge of extracting in- formative entities from scientific papers and infer- ring their semantic relation among a set of six al- ternatives. This challenge is part of the SemEval 2018 Task 7 (G´abor et al., 2018). One of the pil- lar of the proposed approach is that the word em- beddings used for representing the text within the neural network have been generated from a cor- pus containing only scientific papers instead of a general purpose one, like news repositories or Wikipedia. 2 System Implementation NeuroSent has been entirely developed in Java with the support of the Deeplearning4j library 1 and it is composed by following two main phases: • Generation of Word vectors (Section 2.1): raw text, appropriately tokenized using the Stanford CoreNLP Toolkit, is provided as in- put to a 2-layers …

Content truncated to 500 words