UPB at SemEval-2021 Task 8: Extracting Semantic Information on Measurements as Multi-Turn Question Answering

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 534–540 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 534 UPB at SemEval-2021 Task 8: Extracting Semantic Information on Measurements as Multi-Turn Question Answering Andrei-Marius Avram1,2, George-Eduard Zaharia1, Dumitru-Clementin Cercel1, Mihai Dascalu1 University Politehnica of Bucharest, Faculty of Automatic Control and Computers1 Research Institute for Artificial Intelligence, Romanian Academy2 {andrei marius.avram, george.zaharia0806}@stud.acs.upb.ro {dumitru.cercel, mihai.dascalu}@upb.ro Abstract Extracting semantic information on measure- ments and counts is an important topic in terms of analyzing scientific discourses. The 8th task of SemEval-2021: Counts and Mea- surements (MeasEval) aimed to boost research in this direction by providing a new dataset on which participants train their models to extract meaningful information on measure- ments from scientific texts. The competition is composed of five subtasks that build on top of each other: (1) quantity span identifi- cation, (2) unit extraction from the identified quantities and their value modifier classifica- tion, (3) span identification for measured enti- ties and measured properties, (4) qualifier span identification, and (5) relation extraction be- tween the identified quantities, measured en- tities, measured properties, and qualifiers. We approached these challenges by first identify- ing the quantities, extracting their units of mea- surement, classifying them with correspond- ing modifiers, and afterwards using them to jointly solve the last three subtasks in a multi- turn question answering manner. Our best per- forming model obtained an overlapping F1- score of 36.91% on the test set. 1 Introduction Our world revolves around quantities and units of measurement present in all texts, ranging from sci- entific texts to recipes. Nevertheless, the process of automatically extracting measurements is not triv- ial, considering that, in most situations, the quanti- tative structures are ambiguous and are not present in the same area within the text. Therefore, pars- ing the semantic relations becomes a ubiquitous task, since proper quantity identification leads to transformations towards an easy to follow quan- titative summary. Advantages of the previously mentioned process can be found in medical pre- scriptions (Adamo et al., 2015). As such, a system that can robustly and confidently identify medi- cation quantities, measurement units, as well as the medication itself has the potential to become a breakthrough for computer-based medicine and consultations. Another use case resides in ERP sys- tems where proper parsing of resource descriptions facilitates the identification of similar or duplicate items. The MeasEval - Counts and Measurements com- petition (Harper et al., 2021) organized by the 15th International Workshop on Semantic Evaluation (SemEval-2021) creates a new challenge in the area of Natural Language Processing, proposing five subtasks related to span identification, classi- fication, as well as relation extraction, that aim to improve the state of the art for the current field of measurement information extraction. We created a cascaded system to solve the stated problem that is composed of: (1) a subsystem that identifies quan- tities in the input text; (2) a subsystem that classi- fies their value modifiers; (3) a subsystem that ex- tracts their …

Content truncated to 500 words