Chrono at SemEval-2018 Task 6: A System for Normalizing Temporal Expressions

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 97–101 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics Chrono at SemEval-2018 Task 6: A System for Normalizing Temporal Expressions Amy L. Olex, Luke G. Maffey, Nicholas Morton, Bridget T. McInnes Virginia Commonwealth University, Department of Computer Science Richmond, Virginia, USA {alolex, maffeyl, mortonn, btmcinnes}@vcu.edu Abstract Temporal information extraction is a challeng- ing task. Here we describe Chrono, a hybrid rule-based and machine learning system that identifies temporal expressions in text and nor- malizes them into the SCATE schema. Af- ter minor parsing logic adjustments, Chrono has emerged as the top performing system for SemEval 2018 Task 6: Parsing Time Normal- izations. 1 Introduction Understanding and processing temporal informa- tion is vital for navigating life. The human mind processes subtle temporal expressions in- stantly and effortlessly; however, it is difficult for computers to do the same. Identifying, process- ing, and utilizing this information requires knowl- edge and understanding of syntax, semantics, and context to link temporal information to related events and order them on a time-line. SemEval 2018 Task 6 (Laparra et al., 2018) aims to nor- malize fine-grained temporal information and re- lationships into the Semantically Compositional Annotations for Temporal Expressions (SCATE) schema developed by (Bethard and Parker, 2016). This scheme aims to improve upon the current TIMEX3/TimeML (Pustejovsky et al., 2003) stan- dard by representing a wide variety of temporal expressions, allowing for events to act as anchors, and using mathematical operations over a time- line to define the semantics of each annotation. To address this challenge, we developed Chrono1, a hybrid rule-based and machine learning (ML) Python package that normalizes temporal expres- sions into the SCATE schema. 1https://github.com/AmyOlex/Chrono 2 The Chrono System Our approach to building this hybrid system includes four processing phases: 1) text pre- processing, 2) flagging numeric and temporal to- kens, 3) temporal expression identification, and 4) SCATE normalization. 1) Text Pre-processing: Python’s Natural Lan- guage Toolkit (NLTK) WhitespaceTokenizer and part-of-speech (POS) tagger (Bird and Loper, 2004) process raw text files to identify individual tokens, token spans, and POS tags. Punctuation is not handled at this phase as it is important for identifying correct spans. 2) Flagging Numeric and Temporal Tokens: All numeric tokens are flagged regardless of con- text. Subsequent phases utilize contextual infor- mation to determine if a numeric token is part of a temporal expression. Depending on the task, a rule may remove all or some punctuation, and/or convert tokens to lowercase prior to parsing. In the following, RP and LC denote Removing all Punctuation and converting to LowerCase, re- spectively. Numeric Flagging: Tokens are flagged as nu- meric if either 1) the token has a POS tag of “CD” (Cardinal Number), or 2) the text can be converted to a numeric expression. Textual representations of numeric expressions are converted to numer- ics with the Word2Number2 Python module. A custom method recognizes ordinals from “first” to “thirty-first” and converts them into the associated numerics 1 to 31, …

Content truncated to 500 words