ManTIME: Temporal expression identification and normalization in the TempEval-3 challenge

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 53–57, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics ManTIME: Temporal expression identification and normalization in the TempEval-3 challenge Michele Filannino, Gavin Brown, Goran Nenadic The University of Manchester School of Computer Science Manchester, M13 9PL, UK {m.filannino, g.brown, g.nenadic}@cs.man.ac.uk Abstract This paper describes a temporal expression identification and normalization system, Man- TIME, developed for the TempEval-3 chal- lenge. The identification phase combines the use of conditional random fields along with a post-processing identification pipeline, whereas the normalization phase is carried out using NorMA, an open-source rule-based tem- poral normalizer. We investigate the perfor- mance variation with respect to different fea- ture types. Specifically, we show that the use of WordNet-based features in the identifica- tion task negatively affects the overall perfor- mance, and that there is no statistically sig- nificant difference in using gazetteers, shal- low parsing and propositional noun phrases labels on top of the morphological features. On the test data, the best run achieved 0.95 (P), 0.85 (R) and 0.90 (F1) in the identifica- tion phase. Normalization accuracies are 0.84 (type attribute) and 0.77 (value attribute). Sur- prisingly, the use of the silver data (alone or in addition to the gold annotated ones) does not improve the performance. 1 Introduction Temporal information extraction (Verhagen et al., 2007; Verhagen et al., 2010) is pivotal for many Nat- ural Language Processing (NLP) applications such as question answering, text summarization and ma- chine translation. Recently the topic aroused in- creasing interest also in the medical domain (Sun et al., 2013; Kova´cevi´c et al., 2013). Following the work of Ahn et al. (2005), the temporal expression extraction task is now conven- tionally divided into two main steps: identification and normalization. In the former step, the effort is concentrated on how to detect the right bound- ary of temporal expressions in the text. In the nor- malization step, the aim is to interpret and repre- sent the temporal meaning of the expressions using TimeML (Pustejovsky et al., 2003) format. In the TempEval-3 challenge (UzZaman et al., 2012) the normalization task is focused only on two temporal attributes: type and value. 2 System architecture ManTIME mainly consists of two components, one for the identification and one for the normalization. 2.1 Identification We tackled the problem of identification as a se- quencing labeling task leading to the choice of Lin- ear Conditional Random Fields (CRF) (Lafferty et al., 2001). We trained the system using both human- annotated data (TimeBank and AQUAINT corpora) and silver data (TE3Silver corpus) provided by the organizers of the challenge in order to investigate the importance of the silver data. Because the silver data are far more numerous (660K tokens vs. 95K), our main goal was to rein- force the human-annotated data, under the assump- tion that they are more informative with respect to the training phase. Similarly to the approach pro- posed by Adafre and de Rijke …

Content truncated to 500 words