Work in progress.
The goal of this GROBID module is to recognize in textual documents any expressions of measurements (e.g. pressure, temperature, etc.), to parse and normalization them, and finally to convert these measurements into SI units. We focus our work on technical and scientific articles (text, XML and PDF input) and patents (text and XML input).
As part of this task we support the recognition of the different value representation: numerical, alphabetical, exponential and date/time expressions.
Finally we support the identification of the "quantified" substance related to the measure, e.g. silicon nitride powder in
As the other GROBID models, the module relies only on machine learning and it uses linear CRF. The normalisation of quantities is handled by the java library Units of measurement.
The latest released version of grobid-quantities is 0.6.0. The current development version is 0.6.1.
You can find the latest documentation here.
The results (Precision, Recall, F-score) for all the models have been obtained using 10-fold cross-validation (average metrics over the 10 folds).
Evaluated on the 30/04/2020.
| Labels | Precision | Recall | F1-Score |
|---|---|---|---|
<unitLeft> |
96.45 | 95.06 | 95.74 |
<unitRight> |
88.96 | 68.65 | 75.43 |
<valueAtomic> |
85.75 | 85.35 | 85.49 |
<valueBase> |
73.06 | 66.43 | 68.92 |
<valueLeast> |
85.68 | 79.03 | 82.07 |
<valueList> |
68.38 | 53.31 | 58.94 |
<valueRange> |
90.25 | 88.58 | 88.86 |
| all (micro avg.) | 88.96 | 85.4 | 87.14 |
| Labels | Precision | Recall | F1-Score |
|---|---|---|---|
<base> |
98.95 | 99.02 | 98.98 |
<pow> |
97.2 | 98.49 | 97.83 |
<prefix> |
98.34 | 98.47 | 98.38 |
| all (micro avg.) | 98.7 | 98.89 | 98.8 |
| Labels | Precision | Recall | F1-Score |
|---|---|---|---|
<alpha> |
96.9 | 98.84 | 97.85 |
<base> |
85.14 | 74.48 | 79 |
<number> |
98.07 | 99.05 | 98.55 |
<pow> |
80.05 | 76.33 | 77.54 |
<time> |
73.07 | 86.82 | 79.26 |
| all (micro avg.) | 96.15 | 97.95 | 97.4 |
The current average results have been calculated using micro average which provides more realistic results by giving different weights to labels based on their frequency. The paper "Automatic Identification and Normalisation of Physical Measurements in Scientific Literature", published in September 2019 reported average evaluation based on macro average.
This project has been created and developed by science-miner since 2015, with additional support by Inria, in Paris (France) and the National Institute for Materials Science, in Tsukuba (Japan).
If you want to cite this work, please simply refer to the github project with optionally the Software Heritage project-level permanent identifier:
grobid-quantities (2015-2021) <https://github.com/kermitt2/grobid-quantities>, swh:1:dir:dbf9ee55889563779a09b16f9c451165ba62b6d7
Here's a BibTeX entry using the Software Heritage project-level permanent identifier:
@misc{grobid-quantities,
title = {grobid-quantities},
howpublished = {\url{https://github.com/kermitt2/grobid-quantities}},
publisher = {GitHub},
year = {2015--2021},
archivePrefix = {swh},
eprint = {1:dir:dbf9ee55889563779a09b16f9c451165ba62b6d7}
}GROBID and grobid-quantities are distributed under Apache 2.0 license.
The documentation is distributed under CC-0 license and the annotated data under CC-BY license.
If you contribute to grobid-quantities, you agree to share your contribution following these licenses.
Contact: Patrice Lopez (patrice.lopez@science-miner.com), Luca Foppiano (luca@foppiano.org)