For creation of digital textual corpora of preserved historical sources, automatic or semi-automatic extraction of speci c types of information is becoming a requested tool for many researchers active in the eld of digital humanities. With such tools, the e orts in digitization and semantic annotation will be greatly aided. For this reason, we propose a rule-based named-entity recognition system that can be used for location extraction from Latin text (so-call LOCALE). It is based on a set of computational linguistics rules for Latin language. Experimental results obtained on a set of 100 documents, which were further manually evaluated by human experts, showed that very promising results are achieved.
Historiography engages with a very wide variety of subjects, as long as they are connected to past human activity in some way. The traditional and still central source for the study of history has been and remains the written text. Historian's (textual) sources may vary widely in variety and quantity, but in general terms it is true the further back in time one reaches, the more scarce the sources.
Location names (settlements, castles, churches, fallows, waters, provinces, etc.) represent a speci c genus of information commonly contained in all (or almost all) written sources. They, rst and foremost, attest the existence of a given settled place at the time of creation of the text in which they appear. Furthermore, several social phenomenon may be inferred from location names such as the extent of political entities, medieval manors, but also enables insights into historical demography in the sense of the spreading of population over territories and so on.
Joined with personal names location names give the possibility to map the life paths of individuals in space and time, as well as to create models of interpersonal or other more abstract connections for a group of people mentioned in the sources. Such models may be built, depending on the type of sources utilized, for nobility or high clergy as well as for merchants or peasants. To sum, studies in social, economic, political, demographic, military, and many more aspects of history deal with topographic data in historical sources. Of course, other disciplines that are counted among the humanities may also make use of this sort of information.
Through creation of digital textual corpora of preserved historical sources automatic or semi-automatic extraction of speci c types of information is becoming a requested tool for many researchers active in the eld of digital humanities. With such tools, the e orts in digitization and semantic annotation will be greatly aided.
The long-term goal of the research will be to produce a thorough and reliable digital
(interactive) version of historical topography of the entire territory of the present-day Republic of
Slovenia. The rst step in this objective has already been successfully achieved in 2015 with the
project `Slovenian place-names in time and space', which covered the territory of the historical
province of Carniola [
The reminder of the paper is organized as follow: Section 2 provides the related work, Section 3 presents a rule-based system for location extraction from Latin text, Section 4 explains the data used for evaluation together with the experimental results, and nally Section 5 concludes the paper and provides directions for future work. 2
In the past few years, with a fast digitization of historical textual sources [
Named-entity recognition (NER) is a sub-task of IE, which aims to determine and identify
words or phrases in text into prede ned classes that describe the entities of a given domain [
Terminology-driven NER methods [
Despite the large amount of research done in NER for other domains, there are several studies
that are trying to address NER for historical text. In most cases, corpus-based NER methods
are applied, which also depend form the annotated historical corpus. Working with historical
corpus additional challenges should be considered, such as, language changes over time, spelling
variations, transliterations, and sources written in multiple languages among others. In [
From other side, the collection of historical Latin text is still relatively sparse compared to
English. In [
As we have already mentioned, working with historical corpora, especially Latin text, is challenging because language changes over time, spelling variations, and transliterations. Since the performance of corpus-based NERs is dependent from the quality of an annotated corpus, and having an annotated corpus is a time-consuming task, we propose a rule-base location namedentity recognition method, called LOCALE (ruLe-based lOCation nAmed-entity recognition method for Latin tExt), where the method does not using training annotated data, but a small set of manually created rules. The goal of this study is to nd if such a method can also provide promising results that can be used for location extraction from Latin text. 3.1
Since we are lack of annotated corpus, to extract location from Latin text, we propose a rulebased named-entity recognition method, called LOCALE. It is based on a set (i.e., grammar) of computational linguistics rules speci c for the Latin language. We consider that the Latin language has well de ned rules for presenting locations with a small number of exceptions.
The Latin language is an in ected language, which means that the type of the word can be determined from its su x. There are several examples when the su x is related with location. For example, the su x -ensis is an adjectival su x meaning `originating in' and it is used in modern Latin scienti c coinages, especially derivatives of place names. Another su x that can help in recognizing locations is the su x -ensem which is the accusative singular of -ensis, as well as the dative variant, -ensi and the ablative -ense. The plural form of the su x -ensis is -enses which is also used to indicate locations for example Aquilegiensis which derives from
Aquilegia. The genitive and the accusative forms of the plural form of the su x are -ensium and -enses, respectively. The su x -anus has a similar meaning and it indicates a relationship of position, possession, or origin. The accusative singular form -anum and plural masculine form -anarum are also very commonly used. The su x -ius and its dative -iis and accusative -iam/-ium forms are also used to form adjectives from nouns and indicates `belonging to'. The -ianus su x is frequently used in adjectives formed from proper names. Having a knowledge of these su xes and their meaning can bring to a conclusion that the lemma and the su x is crucial to work with NER.
Except for su xes there are also several characteristics by which we can determine a location. Some of these characteristics are:
Titles and functions: imperator, rex, episcopus, sancte.
Also, there are several issues that should be considered that can lead to a false positive location recognition. One such example are entities related to saints' names. This happens since their appearance may have di erent context, in some cases it signi es personal name and in some cases a location. For example in the phrase `. . . Ego Iohannes sancti Nicolai in Carcere Iulianus . . . ', sancti Nicolai is a location and in the case `. . . sancti Benedicti Aquilegensis . . . ', sancti Benedicti is a personal name.
Another issue are some of the su xes that in the the majority of cases are an indication of a location. However, these su xes in some cases can be found as an ending of a personal name. For example the su x -ium indicates a location in the example `versus Rosacium' and a personal name in `prefatum Wernerium '.
LOCALE consists of two steps:
Each document (i.e., Latin text) is going under pre-processing by applying tokenization and generating list of entities that are usually followed by a location or come right after the location.
The list of tokenized words are checked by the set of computational linguistics rules, which de ne LOCALE grammar, in order to extract the location entities.
The rst step, that is the pre-processing, consists of a creation of a static list of entities by which we can determine a location like geographical entities (e.g., rivus, monte), titles and functions (e.g., imperator, episcopus) etc. This list was further used as a look-up table for checking the previous and following word of each individual token. The next pre-processing step is tokenization where we split the input document into individual tokens/words which are used as the basic input elements for the next step which is application of rules.
The LOCALE grammar consists of four rules:
The rst rule check each word if it stars with a capital letter. This is the rst condition to be ful lled for a further rule check to be applied on the speci c word. This comes from the fact that every location in Latin starts with a capital letter.
The second rule extracts words that are after the adverbs like de, in, iuxta, versus, ad, inter and so on. This comes from the fact that the adverbs in most of the cases are followed by a location. One problem that appears here is the listing of multiple locations after one of those adverbs, for example, `. . . iuxta Natissam, Stanowiza, Boriana, Potoch, Creda . . . '. For this reason, we extended this rule to take into account multiple locations that are after an adverb. An exception of this rule was also applied. This rule covered the cases of listing people that belong to a speci c location rather than locations, like `. . . domini Bertoldi senioris de Rosaciis, Henrici, Rantol . . . ' where Henrici and Rantol are personal names.
The third rule is related to words coming after and before the word. As mentioned above geographical locations, titles and functions are a good indicator that the following or previous word is a location. For this purpose we created a dictionary lookup list that consists of entities belonging to the title and function categories as well their derivations. For example, ecclesiam and ecclesias present the accusative case in plural and singular of the word ecclesia, respectively.
Finally, all other words that are not recognized by the three rules, are additionally checked for the abovementioned su xes that are related to locations.
The LOCALE source code is publicly available at https://github.com/Ivona221/LOCALE. 4
4.1
In this section, we explain the data set used for evaluation of our proposed method, together with the obtained results.
Historical topography deals with the identi cation of former location names that are found in various written historical records of the past. Since our group at the Milko Kos Historical Institute, Ljubljana, is highly involved in researching historical geographical locations (i.e., historical topography, which deals with the identi cation of former place names), to test the performance of the proposed method, we selected a data set that consists of medieval historical charters written and published at the territory of the Republic of Slovenia and its surrounding area.
Early documents produced in the present-day Slovenian territory were written mainly in
Latin [
The test data set is publicly available https://github.com/Ivona221/LOCALE. 4.2
After applying LOCALE on 100 documents, the result consists of location entities extracted from each document. Because we do not have an annotated corpus (i.e., the ground truth), the results were manually evaluated by human experts. In our case, we have one positive class, which is the location.
Going through the extracted entities, the human experts classi ed each of them as a true positive (TP), which means that the extracted entity from the method belongs to the positive class (i.e., a location entity is truly recognized), or a false positive (FP), which means that the extracted entity is incorrectly classi ed as location entity (i.e., the entity is not a location.) Additionally going through the Latin text, they also report the location entities that were not recognized by the method, or these entities are classi ed as false negatives (FNs) (i.e., the entity that is incorrectly classi ed that does not belong to the positive class.) We should point here, that the results provided by both human experts was the same. The number of TPs, FPs, and FNs is presented in Table 1.
Using the results reported in Table 1, the evaluation metrics for F1 score, precision, and recall, are presented in Table 2. Precision gives the percentage of the results that are relevant, while the recall provides the percentage of total relevant results correctly recognized by your method. The F1 score is a measure of accuracy, which is the harmonic mean of precision and recall. Higher values are preferable for all three evaluation metric.
In our evaluation, the last rule related to the su xes provides the most of the annotated locations. The most common su xes that were present in our corpus were -ensis, -anus and -us and their derivations in genitive, dative, accusative and ablative, as well as their plural forms.
Looking at the evaluation results, we can concluded that the rst attempt of extraction location entities from Latin text using a rule-based system provides very promising results. However, since our evaluation corpus consists of a 100 documents, we are planning to extend this work in future by:
The documents will be extracted from larger number of charters in order to have more
variability in the Latin text;
Comparing LOCALE with the corpus-based NER method presented in [
Sensitivity analysis of the LOCALE with regard to Mediaeval Latin variation over time and grammar.
Creating an annotated corpus that can be used for training corpus-based NERs. 5
To extract locations from Latin text, we proposed a rule-based named-entity recognition method, called LOCALE. The method is based on a small number of computational linguistics rules that are speci c for the Latin text. The evaluation was performed using early documents produced in the Slovenian territory from twelfth and thirteenth centuries. Experimental results obtained using a 100 documents, which were further evaluated by human experts, showed that promising results can be achieved resulting in 0.92 for precision, 0.91 for recall, and 0.91 for F1 score. 6
This work was partially supported by the project Toponomastical heritage of Primorska Region, which received funding from the Slovenian Research Agency (grant number L7-9424). J.S. and T.E. acknowledge the nancial support from the Slovenian Research Agency (research core funding number P2-0098).