CRI
A Competent Reader Imitator for detecting binomial names in an historical corpus
La Nature (1873–1962) is a French popular science magazine that spanned a large time period and a
large range of topics. It is available via ocerized archives so that it forms a corpus that is simultaneously diachronous,
heterogeneous, and noisy. Although these characteristics make it complex to analyze, La Nature is of great
interest for digital humanities studies on the evolution of thoughts in science, technology, and even politics.
The work presented in this article is part of research on the semantic annotation of these archives, which is discovering clues
for exploring them. One type of clue that has not been explored in a complex corpus such as La Nature is
binomial names, or more specifically, the named entities that refer to the Linnean
classification of life, e.g., Escherichia coli. To overcome this complexity, the concept of a Competent
Reader, who can detect binomial names even when obsolete, non-standard or defaced by OCR, is introduced. By imitating
a Competent Reader, our approach, which we call the Competent Reader Imitator (CRI), involves combining a
rule-based approach with a frequency argument. We show that this innovative method is robust to numerous variations and
consistently achieves an F-measure of about 70% despite diachronicity, heterogeneity, and noise, which are all known to impede
named entity recognition. Our method has many potential applications, such as in the study of chemical names and names of
scientific and technical artifacts, which could benefit from the Competent Reader imitation approach. Beyond our work on
La Nature, we hope this paper provides a set of tools and methods that are easily understandable, frugal, and
usable for a general public interested in exploring similar historical corpus.
Article outline
- 1.Introduction
- 2.Context information
- 2.1The Linnean classification
- 2.2The corpus of La Nature
- 2.2.1Structure of the corpus
- 2.2.2Volumetry of the corpus
- 3.Creation of a gold standard
- 3.1The Competent Reader Hypothesis — CRH
- 3.2The task
- 3.3The manual annotation
- 4.State of the art and baseline
- 4.1State of the art
- 4.2LINNAEUS
- 4.3TAXREF classifier
- 4.4QUAESITOR
- 5.A Competent Reader Imitator
- 5.1Latin
- 5.2A Rare-Frequent Threshold
- 5.3Handling deviations from the code
- 6.Evaluation
- 6.1Comparison with the state of the art
- 6.2Ex-post evaluation
- 6.3Sensitivity to accepted patterns
- 6.4Sensitivity to thematic variability
- 7.Conclusion & future work
- 7.1Carbon impact
- 7.2Conclusion / discussion
- 7.3Future work
- Author queries
-
References
This content is being prepared for publication; it may be subject to changes.