The AI2 system at SemEval-2017 Task 10 (ScienceIE): semi-supervised end-to-end entity and relation extraction

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 592–596, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics The AI2 system at SemEval-2017 Task 10 (ScienceIE): semi-supervised end-to-end entity and relation extraction Waleed Ammar Matthew E. Peters Chandra Bhagavatula Russell Power Allen Institute for Artificial Intelligence, Seattle, WA, USA {waleeda,matthewp,chandrab,russellp}@allenai.org Abstract This paper describes our submission for the ScienceIE shared task (SemEval- 2017 Task 10) on entity and relation extraction from scientific papers. Our model is based on the end-to-end rela- tion extraction model of Miwa and Bansal (2016) with several enhancements such as semi-supervised learning via neural lan- guage models, character-level encoding, gazetteers extracted from existing knowl- edge bases, and model ensembles. Our of- ficial submission ranked first in end-to-end entity and relation extraction (scenario 1), and second in the relation-only extraction (scenario 3). 1 Task overview The ScienceIE shared task (Augenstein et al., 2017) focuses on information extraction from scientific papers. In the end-to-end evalua- tion scenario, participants were provided with a paragraph and asked to extract typed entities (Task, Material or Process) and relations (Hyponym-of or Synonym-of) in different scientific domains. Running example. Consider the following in- put sentence: “Here, we consider a radical pair in which the first electron spin is devoid of hyper- fine interactions, while the second electron spin in- teracts isotropically with one spin-1 nucleus, e.g. nitrogen.” The provided human labeled entity annotations for this sentence include: “electron spin” of type Process, “spin-1 nucleus” of type Material, and “nitrogen” of type Material. The only positive relation annotation labeled for this sentence is: Hyponym-of(“nitrogen”, “spin-1 nucleus”). We will use this example throughout the paper to illustrate various parts of our system (see Fig. 1). 2 System description In this section, we describe the components in our system. Text preprocessing. We process the input text using spaCy1 which provides sentence segmenta- tion, tokenization, part-of-speech (POS) tagging and labeled dependency parsing. Label encoding. We use the BILOU tagging scheme to encode the annotations for each of the three entity types. For a given entity type, each token in a sentence is labeled with: O if it does not belong to any entity, U if it belongs to a single- token entity, B if it is the beginning of a multi-word entity, L if it is the end of a multi-word entity, and I if it is inside a multi-word entity. In order to al- low the same token to participate in multiple enti- ties of different types, each token is assigned three labels (one for each entity type).2 The labeled data specify two kinds of rela- tions: a directional relation (Hyponym-of) and an undirectional relation (Synonym-of). In our system, we automatically convert Hyponym-of(e2,e1) into a new label Hypernym-of(e1,e2) when e2 follows e1 in the sentence. We generate the label No-relation(e1,e2) as needed in order to have a label for each pair (e1, e2) where e1 precedes e2 in a sentence. This encoding allows us to only consider entity pairs in …

Content truncated to 500 words