ULISBOA at SemEval-2017 Task 12: Extraction and classification of temporal expressions and events

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 1019–1023, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics ULISBOA at SemEval-2017 Task 12: Extraction and classification of temporal expressions and events Andre Lamurias∗1, Diana Sousa1, Sofia Pereira1, Luka A. Clarke2 and Francisco M. Couto1 1LaSIGE, Faculdade de Ciˆencias, Universidade de Lisboa, Portugal 2BioISI: Biosystems & Integrative Sciences Institute, Faculdade de Ciˆencias, Universidade de Lisboa, Portugal Abstract This paper presents our approach to partic- ipate in the SemEval 2017 Task 12: Clin- ical TempEval challenge, specifically in the event and time expressions span and attribute identification subtasks (ES, EA, TS, TA). Our approach consisted in train- ing Conditional Random Fields (CRF) classifiers using the provided annotations, and in creating manually curated rules to classify the attributes of each event and time expression. We used a set of com- mon features for the event and time CRF classifiers, and a set of features specific to each type of entity, based on domain knowledge. Training only on the source domain data, our best F-scores were 0.683 and 0.485 for event and time span iden- tification subtasks. When adding target domain annotations to the training data, the best F-scores obtained were 0.729 and 0.554, for the same subtasks. We ob- tained the second highest F-score of the challenge on the event polarity subtask (0.708). The source code of our system, Clinical Timeline Annotation (CiTA), is available at https://github.com/ lasigeBioTM/CiTA. 1 Introduction This paper presents our system CiTA (Clinical Timeline Annotation) to participate in the Se- mEval 2017 Task 12: Clinical TempEval chal- lenge. Our team participated in the subtasks cor- responding to the identification of the following ∗alamurias@lasige.di.fc.ul.pt properties: time expression spans, event expres- sion spans, time expression attributes, event ex- pression attributes. Time expressions had only one attribute, type, which could be DATE, TIME, DU- RATION, QUANTIFIER, PREPOSTEXP or SET. Event attribute identification consisted of four at- tributes: type (N/A, ASPECTUAL or EVIDEN- TIAL), polarity (POS or NEG), degree (N/A, most or little) and modality (ACTUAL, HEDGED, HY- POTHETICAL or GENERIC). For this challenge, we developed a sys- tem, named Clinical Timeline Annotation CiTA1, which uses IBEnt (Lamurias et al., 2015) to iden- tify the text spans of time and event entities based on machine learning and semantic techniques. CiTA also incorporates hand-crafted rules to as- sign the attributes to each entity. We trained one classifier for each entity type using Conditional Random Fields (CRF) and developed a set of rules for each attribute, based on the training data avail- able at each phase. This paper describes the fea- tures and resources used for each subtask, presents our results and discusses the main issues found. CiTA is publicly available in a GitHub reposi- tory 2. 2 Methods A corpus of clinical, pathology and radiology notes from the Mayo Clinic was available to the participants. This corpus contained notes for the source domain (colon cancer) and for the target domain (brain cancer). Each document was man- ually annotated …

Content truncated to 500 words