In this paper a set of information extraction experiments over historical texts are described. The experiments were done over the Spanish book Observaciones de Curvo written by Suarez de Ribera in 1735 based on the 1707 Curvo Semedo's work Observaciones medicas doutrinaes de cem casos gravissimos to evaluate which information can be extracted in a fully automatized way. Using publicly available NLP tools we extracted named entities (persons and places) and identi ed events. This information was used to populate a specialized ontology, allowing the application of powerful visualization and inference processes. A preliminary evaluation of the quality of the extracted information showed that, in spite of the use of generic NLP tools, this process is able to automatically identify relevant information and to help human experts in the creation of historical knowledge bases.
Text information extraction is an increasingly relevant NLP task, aiming to automatically structure unstructured text. On the other hand, historical documents have a huge potential amount of information, which is not easily accessible to researchers or citizens.
In this context, a project aiming to automatically populate a specialized ontology with information extracted from historical texts was created by researchers of the University of Evora, Portugal, and UFRGS { Universidade Federal do Rio Grande do Sul, RS, Brasil.
In this paper we describe the initial experiments done over the Spanish book Observaciones de Curvo written by Francisco Suarez de Ribera in 1735 based on the 1707 Curvo Semedo's work Observaciones medicas doutrinaes de cem casos gravissimos.
It is important to refer that all the processing was done by applying NLP computational tools without any human intervention: from the OCR until the ontology population. Our main goal with this option was to analyse and evaluate how a pipeline of computational processes was able to deal with historical texts in a fully automatized way.
In the next section we will brie y describe the used corpus; in section 3 we will present the applied methodology; in section 4 a more detailed discussion of the NLP architecture will be presented; in section 5 the used ontology is described; and in section 6 a preliminary evaluation is presented and discussed. 2
As already referred, as base corpus for the research work we have selected the book Observaciones de Curvo written by Francisco Suarez de Ribera in 1735 based on the Curvo Semedo's work Observaciones medicas doutrinaes de cem casos gravissimos. This book is available for download (in pdf and txt formats) from the Spanish National Library3 and is composed by 549.267 tokens. It is important to refer that the text version is the output of a OCR process but it was not revised. An example of the existent problems can be seen from the initial sentences:
OBSERVACION PRIMERA.
DE UNA COLICA NEPHRITICA, que a igio al Excekntifsimo fen~or Principe de Ligne, y Marques de Arronches.
N feis de E n e r o del an~o de 1 68<$. a igio al dicho Excelentifsimo fen~or la colica nephritU ca : t e n g o a Tentado, que la verdadera cien-: cia no con ne en lo r u i d o f o , o campanudo de las p a l a b r a s , ni en la apariencia, o pompa exterior de los v e n i d o s , mas si en las obras ordenadas con acier"< t o , y efectuadas con felicidad:
As it can be seen from this example there are many problems with the OCR quality of the document. We had two main options: a) manually or semi-manually revise the texts; b) use the text as is, without any revision. We decided to select option b) because one of our goals was to evaluate which information can be extracted from historical documents in a fully automatized way.
A distinct approach was followed in a distinct project over the original work
of Curvo Semedo[
1. Lexical analysis 2. Syntactical analysis As working methodology we followed a classical NLP pipeline of processes: 3 http://bdh-rd.bne.es/viewer.vm?id=0000081871&page=1 3. Semantical analysis 4. Ontology population
The lexical and syntactical analysis identi es lemmas, perform part-of-speech tagging and dependency parsing. With the semantic analysis we are able to do named entities recognition (persons, organizations, places, time) and semantic role labelling over the results of the previous modules. The last module { ontology population { receives as input the output of the previous module and creates instances of an ontology, allowing the formal representation of the extracted information.
As NLP tools we have used Freeling[
In the scope of this work we have focused on the extraction of entities and
events and we have used a specialized ontology developed in OWL in the context
of another research project[
The architecture is based on a pipeline of NLP modules and it is represented in gure 1.
Each sentence is processed by a series of modules - part of speech tagging, named entity recognition, dependency parsing, semantic role labelling, subjectverb-object identi cation, and the creation of ontology instances in OWL.
The main goal is to identify events in the text, which are used to populate a prede ned ontology.
As already referred we used the Freeling framework [
Below is an example of the obtained output for a simple sentence of the corpus: ... y como por caufa de ella murieron en una cafa cinco hijos ...
The NER output is: y y CC 0.999989 como como CS 0.967153 por por SP 1 caufa caufa NCFS000 0.919869 de de SP 0.999961 ella el PP3FS00 1 murieron morir VMIS3P0 1 en en SP 1 una uno DI0FS0 0.951973 cafa cafa NCFS000 1 cinco 5 Z 0.999454 hijos hijo NCMP000 1
And the nal Freeling output is: Pred100.4: die.01|expire.01|perish.01 t100.20 (murieron) [ AM-LOC t100.21 (en una cafa) [t100.21 .. t100.23]
A1 t100.25 (cinco hijos) [t100.24 .. t100.25]
According to [
In the context of this work we used as basis the Simple Event Model (SEM) as a baseline model. A graphical representation of this ontology is given in Fig. 2 and its detailed design is presented in [8].
After applying the pipeline of NLP processes we were able to obtain the following output: { NER (named entities) 410 places (282 distinct) 2011 persons (1294 distinct) { Events 14005 events 2896 with subject { A0 8271 with direct objects { A1 1685 with indirect objects { A2 901 with place information { AM-LOC 2747 with adverbials { AM-ADV
We have done a preliminary evaluation of the quality of the extraction processes and we were able to calculate the precision of the extraction, i.e. the percentage of extracted concepts that are correct.
For each kind of information we present bellow this value and some examples: { Places
Precision: 21%
Sevilla Lisboa Salamanca
Rua de la Paz { Persons
Precision: 22%
Curvo Cardenal Rey Hypocrates
Galeno { Events
Precision: 5% (precision was calculated over a sample of 10% of the extracted events)
Verbo: murieron; A1: cinco hijos; AM-LOC: en una cafa Verbo: caera; A1: la enferma; AM-ADV: en una hydrope a Verbo: dar; A1: la Uncion; A2: a el paciente Verbo: tomava; A1: una taza com caldo, o agua; AM-LOC: en las manos; AM-ADV: hirviendo
As it can be seen from these examples, it was possible to extract relevant information and, in some cases, with an associated high level of complexity (see, for instance, the last example of the events). Nevertheless, the precision of extraction is still quite low.
We have made an analysis of the main error situations and the main sources of errors can be characterized in the following way: { Places: Most of the errors are related with incorrect OCR and with misclassi ed entities (e.g. persons classi ed as places). { Persons: The main source of errors for this class of entities is clearly the bad quality of the initial OCR, which created many incorrect words that, being unknown to the NLP tools, tend to be classi ed as proper nouns. { Events: As expected, the precision for this kind of information is very low.
This can be explained, again, by the poor quality of the initial OCR system and the nonexistence of a lexical and syntactical module adapted to the Spanish language of the XVIII century. 7
As main conclusion we believe we were able to show that it is possible to use standard NLP computational tools to automatically extract information from historical texts.
It is important to emphasize that we presented an initial evaluation with a not revised corpus directly generated from a OCR system. Thus, the obtained results are far from perfect and they show the relevance of having good quality texts as input to the processing pipeline. Nevertheless, the proposed approach can be used as a baseline and a basis for additional research work.
As future work, a deeper evaluation of the results should be done and NLP tools adapted to the lexicon and syntax of the historical corpora need to be developed.
We will also foresee the creation of a web based application for the visualization and access to the created ontology. 7. Quaresma, P., Nogueira, V.B., Raiyani, K., Bayot, R.: Event extraction and representation: A case study for the portuguese language. Information 10(6) (2019), https://www.mdpi.com/2078-2489/10/6/205 8. Van Hage, W.R., Malaise, V., Segers, R., Hollink, L., Schreiber, G.: Design and use of the simple event model (sem). Web Semantics: Science, Services and Agents on the World Wide Web 9(2), 128{136 (2011)