UIT-DANGNT-CLNLP at SemEval-2017 Task 9: Building Scientific Concept Fixing Patterns for Improving CAMR

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 909–913, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics UIT-DANGNT-CLNLP at SemEval-2017 Task 9: Building Scientific Concept Fixing Patterns for Improving CAMR Khoa Dang Nguyen Faculty of Computer Science University of Information Technology VNU-HCM ndkhoa@nlke-group.net Dang Tuan Nguyen Faculty of Computer Science University of Information Technology VNU-HCM dangnt@uit.edu.vn Abstract This paper describes the improvements that we have applied on CAMR baseline parser (Wang et al., 2016) at Task 8 of SemEval-2016. Our objective is to in- crease the performance of CAMR when parsing sentences from scientific articles, especially articles of biology domain more accurately. To achieve this goal, we built two wrapper layers for CAMR. The first layer, which covers the input data, will normalize, add necessary information to the input sentences to make the input de- pendency parser and the aligner better handle reference citations, scientific fig- ures, formulas, etc. The second layer, which covers the output data, will mod- ify and standardize output data based on a list of scientific concept fixing patterns. This will help CAMR better handle bio- logical concepts which are not in the train- ing dataset. Finally, after applying our ap- proach, CAMR has scored 0.65 F-score1 on the test set of Biomedical training data2 and 0.61 F-score on the official blind test dataset. 1 Introduction Since Abstract Meaning Representation (AMR) was published by Banarescu et al. (2013) for the first time in 2013, it has been considered by many researchers in Natural Language Processing do- main. In this trend, the task of AMR has been held continuously for two years in SemEval-2016 and SemEval-2017. There have been many parsers 1Currently, the datasource for constructing the list of con- cept fixing patterns are the training, develop and test set of Biomedical training data 2This data is all freely available to download at http://amr.isi.edu/download.html shown its outstanding performance for high F- score points like RIGA (Barzdins and Gosko, 2016), CAMR, CU-NLP (Foland and Martin, 2016), etc. Inspired by the performance of CAMR in SemEval-2016 Task 8, we selected it as our base- line parser for SemEval-2017 Task 9 - Subtask 1: Parsing Biomedical Data. The parsing task of 2017 has a particular domain but there are many scientific terms, formulas, reference quotations, numbers, etc. That makes the task of 2017 very challenging. According to Wang et al. (2015a,b), the accu- racy of CAMR depends greatly on the accuracy of input dependency parser. When we conduct train- ing and testing on Biomedical training data used for SemEval-2017, we have found that CAMR is not good in handling the reference citations, sci- entific figures, formulas, etc which are commonly used in scientific papers. This is partly due to the dependency parser not correctly handling this kind of information. And also, the aligner can not ful- fill its mission. We have built the first wrapper to support solving the problem related to these infor- mation in input data. At the same time, we …

Content truncated to 500 words