# Fileset

[presentation-jsap-2019.pdf](https://mdr.nims.go.jp/filesets/bda8619b-ea33-4328-b0bb-9b023b473a73/download)

## Creator

[ISHII, Masashi](https://orcid.org/0000-0003-0357-2832), [TANIFUJI, Mikiko](https://orcid.org/0000-0001-5284-6364), [FOPPIANO, Luca](https://orcid.org/0000-0002-6114-6164), [DIEB M. Thear](https://orcid.org/0000-0002-8111-2009), [SUZUKI, Akira](https://orcid.org/0000-0002-8167-0414)

## Rights

Creative Commons BY Attribution 4.0 International[Creative Commons BY Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)

## Other metadata

[Leveraging Segmentation of Physical Units through a Newly Open Source Corpus](https://mdr.nims.go.jp/datasets/a8459ee7-63bd-4722-8e30-7995d520ad00)

## Fulltext

presentation-jsap-2019Leveraging Segmentation of Physical Units through a Newly Open Source CorpusLUCA FOPPIANO, AKIRA SUZUKI, THAER M. DIEB, MASASHI ISHII AND MIKIKO TANIFUJI  MATERIAL DATA AND INTEGRATED SYSTEM (MADIS),  NATIONAL INSTITUTE FOR MATERIALS SCIENCE (NIMS) 1Content — Introduction and motivation  — Quantity extraction system   — Benchmark problem and our proposed solution  — Evaluation experiments   — Conclusions 2Introduction and motivation Text and Data Mining in scientific literature requires inevitably to deal with units of measurements and physical quantities  — The units recognition sub-task is an important step (measurement normalisation)  — Extraction of physical quantities is not a 
new subject  — different techniques have been already 
investigated   — there is no benchmark to evaluate 
different approaches (Reproducibility issue!) 3Quantity extraction system Use an open-source system called Grobid-quantities (developed in collaboration with P. Lopez)  Example data: 4Quantity  Extraction50(µ, g, 1) · (m, l, -1)0.05 kg/m3NormalisationValueUnit segmentation model  Grobid-quantities is hosted on GitHub https://github.com/kermitt2/grobid-quantities   https://github.com/kermitt2/grobid-quantitiesUnit segmentation model Segments raw text to product of triples (prefix, base, power), International System of Units   5µg/ml[(prefix, base, pow), (…)]Raw unit Tokenisation Labelling[(µ, g, 1), (m, l, 1)]µ  g  1  /  m  l  1 prefixprefixbasebasepowpowpow ml ︎  min − 1 ︎ kg −1Machine learning is important for dealing with variation having additional or missing characters from the original text.Problem  — No benchmark for evaluation   — The statistical distribution of units in specific subdomains creates biased evaluation results    602468K Gpa wt%00.751.52.25312.2 12.9 50 90 42.5 5.1 10 57.2Unit segmentation corpus We constructed a UNIt Segmentation CORpus [UNISCOR]   — “general dataset” covering broad area of Applied Physics   — open-source, available to be used (and improved) by anybody    7UNISCOR construction — Collected 3490 papers of Journal of Applied Physics  Suzuki Akira and Ishii Masashi, “Constructing a ”Unit dictionary” from scientific articles,” in Third International Work- shop on SCIentific DOCument Analysis (JSAI International Symposia on AI) (Springer, 2018).  — Automatic annotations using grobid-quantities   — Manually check the annotated data in collaboration with other researchers/engineers from NIMS 8Corpus statistics — extracted 1700 unique units:  — 400 simple units (e.g. m, l, etc..)  — 1300 complex units (e.g. m/s, etc..)  — Licence: Open source (CC-BY 4.0)   — Available at https://github.com/kermitt2/grobid-quantities/blob/master/resources/dataset/units/evaluation/unit-evaluation-corpus.tei.xml 9https://github.com/kermitt2/grobid-quantities/blob/master/resources/dataset/units/evaluation/unit-evaluation-corpus.tei.xmlhttps://github.com/kermitt2/grobid-quantities/blob/master/resources/dataset/units/evaluation/unit-evaluation-corpus.tei.xmlhttps://github.com/kermitt2/grobid-quantities/blob/master/resources/dataset/units/evaluation/unit-evaluation-corpus.tei.xmlhttps://github.com/kermitt2/grobid-quantities/blob/master/resources/dataset/units/evaluation/unit-evaluation-corpus.tei.xmlEvaluation experiments Experiment set-up:  — [GQ1] corpus created for training grobid-quantities (built with the application)   — [UNISCOR] evaluation corpus we are presenting (built independently)   10Evaluation experiment 1 We compare results on the same system:   — Training + evaluation using [GQ1]  — Training using [GQ1] and evaluation using [UNISCOR]   Results from evaluation on [GQ1] using standard approach  Results from evaluation using [UNISCOR]   11precision    recall       F1-score  82.27 81.12 81.64precision    recall       F1-score98.83 98.99 98.91Evaluation experiment 2 Comparison of two systems.   — Training and evaluation with [GQ1]:    — CRF: 98.86%   — BiLSTM + CRF: 98.38%  — Training with [GQ1] and evaluation with [UNISCOR]:  — CRF (lexicon + lexical features): F1 81.64%  — BiLSTM + CRF (embeddings): F1 74.09%  [UNISCOR] provides a benchmark that can be used to compare different systems  12Conclusions — We presented our Unit segmentation approach which relies on Machine Learning   — We release a UNIt Segmentation CORpus [UNISCOR] as Open-source (CC-BY).   — UNISCOR can be used as benchmark to provide evaluation measurement for unit recognition.   — In future:    — we extend the dataset to more units   — we add more evaluation datasets (quantities and value segmentation) 13Thank you 14