Enrichment of Renaissance Texts with Proper Names 
Summary: 
The aim of the Renom project was to enrich Renaissance texts with proper names. 
These texts present two challenges: they exhibit great diversity due to various spellings of words and are overladen with numerous XML-TEI tags introduced to save the exact format of the original edition. 
The task consisted of adding Named Entity tags to this format by tagging names, that had not been already tagged, and their left, and sometimes right, context when appropriate. 
In order to achieve this, we have improved free, open source program CasSys to parse texts with Unitex graph cascades and we have built specific dictionaries and cascades. 
The evaluation showed that the slot error rate of name tagging was 6.1%. 
Renaissance texts enriched in this way are used in a website that unites Humanities and tourism by allowing visitors to navigate maps with names.  
Keywords: Named entities, Renaissance texts, Graph cascades, CasSys, Humanities and tourism 
1 Motivation 
The Center of Higher Education of the Renaissance (CESR) offers the Humanist Electronic Libraries (BVH) on the Web for more than ten years: a great number of Renaissance books, written by Rabelais, Ronsard and other authors are represented as scanned and transcribed books. 
Their transcription, defined in the TEI format, strictly follows the presentation of the scanned original: paragraphs, line breaks, abbreviations, hyphens, lettering, etc. are all preserved. 
Figure 1 presents an extract of the novel Gargantua from Rabelais as presented on the BHV website2: one paragraph as transcribed and scanned text. 
The same paragraph tagged in TEI-format is given bellow (<p>---</p> denotes paragraph and <lb/> denotes line break). 
Rabelais wrote about Tours Region in France where the giants Grandgousier, Gargantua and Pantagruel move into. 
With the Renom project we planned to develop the so-called ‘literary tourism’ by introducing links between names and books: the website enables visitors to navigate through the novel using the proper names and to see where the imaginary or antique places were ‘located’. 
Figure 2 shows the imaginary Theleme Abbacy between the towns Chinon and Azay-le-Rideau, near Tours. 
Tourists are encouraged to visit the castles of these two towns and the Rabelais Museum near this ‘place’. 
The aim of this scientific project was to enrich the texts with tags indicating names of persons, personages and locations with supervised Natural Language Processing (NLP) techniques. 
This task, known as Named Entity Recognition (NER), has been well defined in NLP and is being tackled for two decades. 
In our project, the challenge was the existing TEI format, the extensive use of hyphens and the great variability in the spelling of proper names (and in the orthography of words on the whole). 
For instance, in the extract presented in Figure 1, the name Grandgousier is written Grand-<lb rend=”hyphen”/>gousier and the name Africa has two orthographies, Africque and Afrique. 
As introduced by MUC conferences (Message Understanding Conferences), the named entities include person names, location names, organization names, dates, percentages, currency (Chinchor, 1997) and sometimes titles, hours, occupations, etc. 
A state of the art of NER can be found in (Nadeau, Sekine, 2009); the main idea is to use internal and external evidence (MacDonald, 1996), i.e. the local context. 
For instance, in the sequence Hugues Thierry Salel l’ainé, seigneur de Seuille, the first name Hugues Thierry indicates that Hugues Thierry Salel is a person name (internal evidence) and the title seigneur (lord) proves that Seuille is a toponym (external evidence). 
For NER task three approaches are possible: machine learning, symbolic rules and hybrid approaches. 
Machine learning techniques need training corpus, not available for this task. 
For this reason we used an approach based on symbolic rules (Ait-Mokhtar et Chanod, 1997; Hobbs et al., 1997). 
To facilitate the cooperation between computer scientists, linguists and Renaissance experts, we chose Unitex platform4 (Paumier, 2003) because of its friendly interface (with use of graphs) and its free license. 
With Unitex, we can define cascade rule systems (CasSys menu) that benefit from all numerous functionalities of Unitex graphs. 
Cascades (Abney, 1991) are used in many NLP applications, such as chunking (Abney, 1996), syntactic analysis (Kokkinakis, Kokkinakis, 1999), morphological analysis (Alegria et al., 2001) and so on. 
Our system is inspired by (Friburger, Maurel, 2004). 
The NLP community is interested in ancient languages and ancient states of modern languages (see for instance (Denooz, Rosmordus, 2009)). 
In old French, orthography was not fixed and a lot of name variants existed. 
For middle French (just before Renaissance French), specific tools and dictionaries (Souvay, 2004) or lemmatizers (Souvay, 2007) have been developed. 
In order to accomplish our task we also had to build specific dictionaries and cascades, which we have done in cooperation with Renaissance experts. 
The goal of Renom project is to help experts to tag texts precisely and to enhance dictionaries. 
These dictionaries contain proper names and their variants linked with unique keys (proposed by Renom and validated by experts) and locations linked with unique keys to the Geonames5 database. 
This pointer is used by the website to post the map. 
When some too tiny locations or ancient locations were not found in Geonames, experts added new entries to Geonames. 
2 corpus presentation  
The corpus contains 11 books: 
–Discours fantastiques (edition of 1566), Justin Tonnelier; 
–Courtisan (edition of 1538), Baldassare de Castiglione; 
–Voyage de Tours (edition of 1560) and 
Élégie sur les troubles d’Amboise (edition of 1563), Pierre de Ronsard; 
–Gargantua (edition of 1542),  
Pantagruel (edition of 1542),  
le Tiers Livre (editions of 1546 and 1552),  
le Quart Livre (editions of 1548 and 1552) and  
Brève déclaration (edition of 1552), François Rabelais. 
The format is very particular: as we explained in Section 1, it respects to the utmost degree the layout of the original edition, line breaks, footers, initial letters and so on. 
Sometimes, transcribers added some corrections as well. 
Some examples of this format are: 
• Initial letter (Pantagruel) 
• Transfer of the end of the first line at the end of the second one, after parenthesis, following the Latin citation on the same line – transcription is on three lines. 
• Footer (“Dace” truncated by the page number - 188) 
• Correction (insertion of an apostrophe) 
3 Typology used 
The CESR used TEI format for transcribed texts, so it was natural to adopt the same TEI typology. 
We used four types: geography (geogName), places (placeName), organizations (orgName) and persons (persName).  
3.1 Geography and places 
Geographic names were divided into two subtypes: geonyms (mountains, plains, plateaus, grottos…) and, hydronyms (oceans, seas, rivers, lakes, ponds…), for which a special value of the attribute type of the tag geogName was used. 
When possible, the geographic names were additionally specified by adding a specific internal tag geogFeat. 
Place names were sometimes also subtyped (cities, countries, estates and buildings) by using the same attribute type. 
3.2 Organizations 
Organizations were divided into three subtypes: people, estates and communities. 
The CESR chose not to link organizations with keys. 
Organizations could also be embedded. 
When it was difficult to decide whether a name refers to a place or an organization we inserted both tags, placeName and orgName. 
3.3 Persons 
In simplest cases just persName tags (with their keys) were used. 
We added internal tags for first names (foreName), surnames (surName) and particles (nameLink), where necessary. 
Finally, these tags were extended with titles or civilities (roleName) that were subtyped: nobility role, religious role, function or occupation, honor. 
When the title included a place name, it was also tagged: the lord of Essars is a person, but Essars is a place: 
We sometimes additionally specified personal names when nicknames or roles within the family were mentioned (‘the elder son’ below). 
Texts contained ambiguities. 
If possible, the expert would choose the right interpretation. 
For instance below, saint Martin de Candes may refer to a church or a person: 
4 Dictionaries 
As we mentioned earlier, it was often necessary that our NER recognizes the context of named entities. 
To that end we built a variant orthography dictionary, studying contexts and using an old first name list.  
For instance, in the Renaissance, the word “captain” was written in three different ways: capitaine, capiteine or cappitaine. 
We chose the synchronic entry as the lemma and we added features that NER could use (see sections 6.3 and 6.4): 
For instance, the second line contains five pieces of information: form (capiteine), lemma (capitaine), part of speech (N), feature (Military) and morphology (ms). 
We transformed these three CESR lists of names: persons, organizations, locations in the format of Unitex dictionaries (illustrated above), which were improved after each book was parsed. 
In produced dictionaries of names, a word has its key for lemma. 
These keys are used to link different orthographies and also to link locations to Geonames: 
In their previous work, CESR experts chose to use explicit keys (as loc_ancenis for the toponym Ancenis) so we were obliged to use the same approach. 
Table 1 presents the number of dictionary entries at the end of Renom project.  
5 Improvements of Unitex platform 
As we mentioned earlier, we chose Unitex platform to facilitate the cooperation between computer scientists, linguists and Renaissance experts. 
Unitex is open-source and free (LGPL license). 
With Unitex, one can parse texts with his own dictionaries (see Section 4) and write linguistic rules in a form of graphs with a very friendly interface; it is also possible to build cascades of graph with CasSys. 
A graph cascade is a succession of parsing graphs: the first graph parses the text, the second graph parses the text modified by the first graph and so on. 
For the Renom project we had to add three improvements of CasSys that will be explained in the following subsections: enabling the graph iteration until certain fixed point, and the use of Unitex morphological dictionaries and producing the more appropriate output files of the cascade. 
In Unitex system, graphs parse text with possibility to merge new sequences, replace others, use variables, move sequences, and insert information from dictionaries. 
5.1 Graph iteration 
We added to CasSys graph iteration until certain fixed point: the iterative graph parses the text and produces a modified text, then it parses this modified text and so on until the parsing does not modify the resulting text any more (this is the fixed point). 
We used iterative graphs primarily to build keys of embedded types (see section 6.5).  
5.2 unitex morphological dictionaries  
For a form occurring in a text a Unitex graph can extract information associated to it in dictionaries: its lemma, features or morphology codes. 
These dictionaries are named ‘morphological dictionaries’. 
We added the possibility to include this functionality into cascades. 
We defined our three dictionaries of names (see Section 4) as ‘morphological dictionaries’ which enabled us to link a name found in one of these dictionaries to its key. 
More information is given in Section 6.5.  
5.3 File output of a cascade 
The major idea when parsing a text with a cascade of graphs is to consider an already tagged piece of a text as a multiword expression (MWE) because the other graphs of the cascade than cannot parse inside it. 
A sequence of characters is recognized by the Unitex system as a MWE if it is enclosed with curly brackets. 
For instance, the XML tag <lb rend=”hyphen”/> will be interpreted as a MWE if curly brackets are added before and after the tag: 
That is done by the graph presented in Figure 4.  
The additional information inside curly brackets (after a comma and a dot) enables the subsequent graphs of the cascade to parse this MWE with Unitex expressions as <BaliseXML> (for a XML tag), <DFlb> or <hyphen>, depending on the necessary degree of precision. 
A disadvantage is that the resulting text is difficult to read. 
To overcome this we implemented a specific CasSys XML format that is used as an alternative cascade output: 
For that reason our cascades always come in a pair: the first one is used to parse a text while the second one transforms the specific CasSys XML format (the result of the first cascade) into the required format. 
The first graph of the cascade transforms all the original XML tags to MWEs. 
6 method 
Our corpus contained two groups of texts. 
The first group already contained tags for proper names (persName, geogName and placeName) that were added manually by experts; our work on these texts was to add internal tags (geoFeat, foreName, lastName…), to search for keys in the dictionaries and to extend names to named entities by adding tags such as roleName (Lord, Abbey…), genName (elder son…) and so on. 
In the second group of texts only formatting was annotated, which means that we had to recognize names in these texts, before doing the same work as for the first group of texts. 
We organized our work in four steps (see Figure 5). 
A few pre-tagged texts were parsed only from the third step. 
1. Preprocessing to rebuild truncated names at the end of a line or a page; 
2.Dictionary lookup and use of context rules to tag names; 
3.Consultation of dictionaries and application of internal and expanded rules, as presented before (firstName versus genName and so on). 
4. Extraction of names that are not in dictionaries. 
6.1 Hyphens 
As we said, a lot of hyphens occurred in texts. 
A first graph recognized XML tags as multiword expression (MWE) with the part of speech (POS) baliseXML. 
The other graphs of the first cascade rebuilt words. 
For instance, the graph given in Figure 6 recognizes a letter larger in size than others: the hi tag cuts the word and the graph rebuilds it. 
Two hi tags (start and end tags) become a new MWE with POS set to largerSup and its new attribute memorizes the scope of the highlighted sequence (here, value=”1”). 
6.2 Abbreviations 
The graph presented in Figure 7 recognizes the XML tags choice, abbr and expan, and builds a MWE8 with POS abbreviation. 
6.3 Names recognized from dictionaries  
The second step began with a dictionary lookup to tag names that were in dictionaries. 
Some names were ambiguous, so we used the context to disambiguate persons from locations or organizations. 
The graph given in Figure 8 adds a tag persName to personal names in the military context. 
6.4 Names recognized only from context 
After we recognized names from dictionaries, we used the same contexts to tag names that were not in the dictionaries, if the first letter was capitalized. 
The graph in Figure 9 tags unrecognized names with persName in the military context. 
6.5 Keys 
When names were identified, we searched their keys in dictionaries, if such entries existed. 
The graph presented in Figure 10 returns these keys in the form of the key attribute of the appropriate start tag. 
If the name was not found in dictionaries, other graphs built its possible key by concatenating keys of names it consists of. 
The attribute dic=”no” indicated to experts that the name (with this orthography) was not found in the dictionaries. 
The expert added it, with another key if it was a variant of an existing entry or with this key if it was really a new entry. 
To build a key was not trivial, because of embedded names. 
For instance, we had to add three keys for the name château du gué de Vede: one for the proper noun ‘Vede’ (the key extracted from dictionaries), one for ‘the ford of Vede’ and one for ‘the castle of the ford of Vede’: 
The graph that built keys of embedded names is applied iteratively. 
It calls four subgraphs, one for each type. 
Figure 11 presents this graph and one of its subgraphs that inserts a key in a placeName tag.  
6.6 Tags inside of names 
We inserted some tags inside name tags: tags for forenames and surnames inside persName tags and geogFeat tags inside geogName tags. 
For this task we used a Renaissance dictionary of forenames. 
The following instances posed some difficulties: a person with more than one forename (Hugues Thierry Salel), with a multiword surname (Jan Trivolse Guallo) or with a particle (Ulrich Thierry du Gallet). 
The graph given in Figure 12 tags one forename and one surname (tags are in subgraphs). 
After application of this graph both the names and their inside tags are considered as MWE. 
6.7 Extended tags 
We also extended persName to named entities that contained tags roleName (Lord, Abbey, teacher…), genName (elder son…) or addName (nicknames). 
We added new tags to the left or to the right of an already recognized personal name and we moved persName tags. 
The graph in Figure 13 adds tags to the left of a named entity when it is preceeded by a roleName. 
In this case the key does not change. 
6.8 Dictionary enhancement 
Finally, the two last cascades built a new file for dictionary enhancement: they erased the text, except the names that were not already in dictionaries or were in them but with different features. 
For instance, the last example: 
The entire list of names with the feature dic=”no” was transmitted to experts to improve the dictionaries.  
7 Evaluation 
To evaluate our work, we parsed two books of Pierre de Ronsard from our corpus (Voyage de Tours and Élégie sur les troubles d’Amboise). 
We did not use these books to construct our cascades – for developing them we used mainly the François Rabelais’ books. 
We computed a weighted variant of the slot error rate (SER) (Makhoul et al., 1999) used in French evaluation campaign. 
SER distinguished between three types of errors: 
Insertion (I - weight 1): we tagged words that are not names. 
Deletion (D - weight 1): we failed to tag a name.  
Tags with border errors: bad type (T - weight 0.5), tag outside or inside the proper name (E - weight 0.5) or both (TE - weight 1).  
If #R is the sum of the entities of the reference texts, the SER is computed by: 
With these counts, if #S is the sum of the detected entities, we can also compute the precision and the recall of our work: 
The tagged texts are entirely supervised, so border errors are less important. 
But we can also compute the type precision (correct recognition of types) and the limit precision (the boundaries of a named entity): 
Results are presented in Table 2. 
The experts of CESR wanted to read the whole corpus before its publication on the website. 
The SER of 6.1% indicates that their work was really improved. 
The significant number of deletion errors corresponds to a lot of names without context that were not represented in dictionaries. 
At the end, the experts enhanced them. 
8. Conclusion 
In this paper we presented the NER task applied to XML-TEI encoded Renaissance texts. 
The format of the corpus and the significant variations of vocabulary asked for the specific treatment as compared to contemporary texts. 
We used dictionaries and rule-based cascades and we obtained 6.1% of SER. 
Our system will enter the production line of transcribed texts. 
The most important texts of Rabelais are now online at the Renom website powered by a search engine for names. 
It uses keys to link name variants and to connect them to map locations.  
Acknowledgment 
This work is supported by Région Centre research program. 
Authors thank the CESR, particularly Sandrine Breuil, Marie-Luce Demonet, Jorge Fins and Marie Olivron. 
