<\/p>\n\n\n\n
\u672c\u6587\u4ecb\u7ecd\u4e86\u4e00\u4e2a\u6b63\u5728\u8fdb\u884c\u7684\u6570\u5b57\u4eba\u6587\u9879\u76ee\uff0c\u8be5\u9879\u76ee\u65e8\u5728\u5229\u7528\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u8bed\u6599\u5e93\u8bed\u8a00\u5b66\u3001\u673a\u5668\u5b66\u4e60\u548c\u7a7a\u95f4\u5206\u6790\u7684\u4f18\u52bf\u6765\u63a8\u8fdb\u5bf9\u5927\u91cf\u5386\u53f2\u8bed\u6599\u5e93\u7684\u8ba1\u7b97\u5206\u6790\u3002\u4f5c\u4e3a\u4e00\u4e2a\u6848\u4f8b\u7814\u7a76\uff0c\u8be5\u9879\u76ee\u96c6\u4e2d\u5728relacones Geogr\u00e1ficas de la Nueva Espa\u00f1a(1577-1585)\uff0c\u4e86\u89e3\u58a8\u897f\u54e5\u65e9\u671f\u6b96\u6c11\u65f6\u671f\u7684\u5173\u952e\u8bed\u6599\u4e4b\u4e00\u3002<\/p>\n\n\n\n
\u8be5\u9879\u76ee\u4f7f\u7528\u4e00\u79cd\u540d\u4e3a\u5730\u7406\u6587\u672c\u5206\u6790(GTA)\u7684\u8ba1\u7b97\u673a\u8f85\u52a9\u65b9\u6cd5\uff0c\u4e3a\u5386\u53f2\u6587\u672c\u7684\u89e3\u6790\u4ee5\u53ca\u6d89\u53ca\u5730\u540d\u548c\u5206\u6790\u6982\u5ff5\u7684\u8bcd\u6c47\u6807\u8bb0\u63d0\u4f9b\u4e86\u81ea\u52a8\u624b\u6bb5\uff0c\u7136\u540e\u5c06\u8fd9\u4e9b\u8bcd\u6c47\u4e0e\u5730\u7406\u4f4d\u7f6e\u8054\u7cfb\u8d77\u6765\u3002\u5728\u6587\u672c\u89e3\u6790\u4e2d\u52a0\u5165\u5730\u7406\u7a7a\u95f4\u667a\u80fd\uff0c\u53ef\u4ee5\u63a2\u7d22\u5386\u53f2\u8bed\u6599\u5e93\u4e2d\u7684\u9690\u85cf\u5730\u7406\u548c\u53d9\u4e8b\u3002\u672c\u6587\u63d0\u4f9b\u4e86\u8bed\u6599\u5e93\u7684\u4e00\u822c\u6982\u8ff0\uff0c\u4e00\u6b65\u4e00\u6b65\u5730\u63cf\u8ff0\u4e86GTA\u65b9\u6cd5\uff0c\u5e76\u62a5\u544a\u4e86\u8fc4\u4eca\u4e3a\u6b62\u53d6\u5f97\u7684\u8fdb\u5c55\u3002<\/p>\n\n\n\n
<\/p>\n\n\n\n
Abstract<\/p>\n\n\n\n
\u00a0This article introduces an ongoing Digital Humanities project aimed at leveraging the benefits of Natural Language Processing, Corpus Linguistics, Machine Learning, and Spatial Analysis for advancing the computational analysis of vast historical corpora.\u00a0As a case study, the project\u00a0concentrates on the\u00a0Relaciones Geogr\u00e1ficas de la Nueva Espa\u00f1a<\/em>\u00a0(1577\u20131585), one of the key corpora for understanding the early colonial period of Mexico.\u00a0<\/p>\n\n\n\n Using a computer\u2013assisted methodology called Geographical Text Analysis (GTA), the project offers automatic means for parsing historical texts and the markup of words referring both to place names (toponyms) and analytical concepts that are then linked to their geographic locations. Adding geospatial intelligence to the parsing of texts allows exploring hidden geographies and narratives in the historic corpus. The article provides a general overview of the corpus, describes the GTA methodology step by step, and reports on the progress achieved so far.<\/p>\n\n\n\n Keywords: Relaciones Geogr\u00e1ficas de la Nueva Espa\u00f1a; early colonial history; Mexico; geographical text analysis; named entity recognition; name entity disambiguation; named entity classification; geographic collocation; natural language processing; machine learning.<\/p>\n\n\n\n 1. Introduction<\/strong><\/p>\n\n\n\n This article introduces the project Digging into Early Colonial Mexico: A Large\u2013Scale Computational Analysis of Sixteenth\u2013Century Historical Sources<\/em>, and describes the progress achieved so far.[1]<\/sup><\/a> The project focuses on developing novel methods and tools for mining new knowledge from textual and pictorial historical records to facilitate the study of early Latin American history. It concentrates on one of the most important corpora related to the geography, culture, religion, economy, and history of native and Spanish peoples in Mexico and Guatemala, namely the Relaciones Geogr\u00e1ficas de la Nueva Espa\u00f1a<\/em> (Geographic Reports from New Spain), and specifically those compiled between 1577 and 1585, hereafter referred to as RGs.<\/p>\n\n\n\n The proposed methodology\u2014called Geographical Text Analysis (GTA)\u2014enables the automatic markup of words referring both to place names (toponyms) and analytical concepts relevant to historical research, and adds geospatial intelligence to the parsing of texts. This process is useful not only to find the most salient terms in textual corpora or to identify and disambiguate personal and geographical names, but also to facilitate the discovery of unsuspected data patterns and relationships, such as spatial correlation of events, interaction of people, links of persons and themes to places, and many others.<\/p>\n\n\n\n One of the main contributions of this project has been the development of an analytical model based on deep-learning and adapted to the specific challenges found in texts like the RGs. In order to achieve named entity recognition, disambiguation and classification, GTA employs methodologies from the fields of Natural Language Processing (NLP), Corpus Linguistics (CL) and Machine Learning (ML). Although NLP, CL and ML techniques are used frequently in textual analysis today, they are most commonly applied to contemporary texts composed in English. Little work has been done to adapt their use to historical corpora written in old language forms like the sixteenth\u2013century Spanish.[2]<\/sup><\/a> The RG corpus additionally contains other features that complicate the use of such tools, which makes very difficult the automatic \u201creading\u201d of the documents. These include the presence of multi-lingual expressions in Nahuatl, Zapotec, Otomi, and many other indigenous languages (Harvey, 1972), as well vocabulary only used in certain regions at the time. <\/p>\n\n\n\n A second contribution of our project has been to facilitate, using Geographical Information Systems (GIS), the automatic linking of the results of these processes of textual analysis with locational data. This combination of NLP, CL and ML techniques with spatial analysis methods creates powerful resources to discover unsuspected geographies in the textual reports. Our model will allow scholars of colonial Latin America to learn new information, extract it, cross\u2013link it to other document resources, and, most importantly, study significant historical topics, such as those we discuss in section 3.5 below. <\/p>\n\n\n\n Apart from the GTA model itself, we foresee several other products of this project. The first is a digital version of the RGs corpus, annotated in accordance with analytical categories that can be applied to other corpora through an interoperable ontology. The second is a digital gazetteer containing geographic coordinates as well as political, cultural, administrative and physical data, not only of the sixteenth-century sites mentioned in the RGs, but from other relevant sources and historical periods. The third is a set of software tools that link the spatial information contained in the gazetteer with the textual and pictorial contents of the RGs in order to facilitate the discovery and analysis of geographies not immediately apprehended with normal reading.<\/p>\n\n\n\n In writing this paper, our goal is to raise awareness about the possibilities of applying computer-assisted methods to the analysis of historical texts. Thus, we describe every step of GTA in the context of our specific study case, hoping that this will encourage more Humanities researchers to adopt this kind of approach.<\/p>\n\n\n\n The rest of the article is organized as follows: Section 2 provides an overview of the Relaciones Geogr\u00e1ficas<\/em>corpus; Section 3 describes our analytic methodology and reports on our progress; Section 4 describes future applications to the study of history; and the final section offers a general conclusion.<\/p>\n\n\n\n 2. The Sixteenth Century Relaciones Geogr\u00e1ficas de la Nueva Espa\u00f1a<\/em><\/strong><\/p>\n\n\n\n The compilation of the Relaciones Geogr\u00e1ficas de la Nueva Espa\u00f1a<\/em> was part of a large, long\u2013standing project, sponsored first by Emperor Charles V and then by King Philip II, who sought to collect geographic, economic, historic and cultural information from all the kingdoms and territories under their rule. The initiative emerged not only from the need to control the exploitation of resources and resolve important issues such as land possession, management of laborand the transfer of privileges to colonists, but also to compile scientific knowledge and fulfill intellectual curiosity (Delgado L\u00f3pez 2010; De la Garza 1983). This involved both collecting descriptions and making maps (Mundy 1996: 1\u201327). The whole project included Castile, Aragon, Italy, the Netherlands, Portugal (since its annexation in 1581), and the viceroyalties of Peru and New Spain.<\/p>\n\n\n\n For the New World enterprise, the Spanish king, through the Council of Indies, commissioned his leading cosmographers to gather information by means of questionnaires. These were conceived by Juan de Ovando and sent to the Indies in 1569 (37 questions), 1570 (200 questions) and 1573 (135 questions). The responses that came back to Madrid, however, were scarce, probably because the burden of collecting data was put on high-level ecclesiastical and government officials who lacked the necessary knowledge or were not willing to answer so many questions. <\/p>\n\n\n\n The\u00a0project was revived by Juan L\u00f3pez de Velasco, cosmographer\u2013chronicler of\u00a0Philip II, who in 1577 devised more feasible enquiries. The first one attempted to gather\u00a0observations about moon eclipses from various parts of the Indies (Cline 1964). This would\u00a0allow solving the elusive problem of determining longitude coordinates for the American\u00a0territories, which in turn would make it possible to establish correctly the position of the Indies with respect to Europe in a universal map (Edwards 1969; Mundy 1996: 17\u201319). A second independent enquiry sought to complete the scarce descriptions received in the previous years, sending only fifty carefully formulated questions to be answered not by high-level authorities, as in previous attempts, but by\u00a0corregidores\u00a0<\/em>and\u00a0alcaldes mayores,\u00a0<\/em>the local administrative officials with direct knowledge of the situation in towns, villages and hamlets, who also received printed instructions on how to answer the survey.\u00a0Figure 1\u00a0shows the first page (out of three) of the printed\u00a0Instrucci\u00f3n y Memoria<\/em>\u00a0sent as a guide to the respondents.<\/p>\n\n\n\n The questionnaire included very detailed information about the life in New Spain. One single query might cover topics such as demography, type of settlements, intercultural opinions, and ethnolinguistics.[3]<\/sup><\/a> There are also questions about local history, including the names and settlement of towns, and about agricultural and mineral resources. See Appendix 1 for a complete translation of the questionnaire. <\/p>\n\n\n\n Responses to the questionnaire were produced between 1579 and 1585 and are now held in Spain, Scotland, and the United States. In total there are 168 written reports corresponding to an equal number of main\u00a0cabeceras\u00a0<\/em>(\u201chead-towns\u201d or main districts). These also contain descriptions of 248 subordinate jurisdictions, 414 towns and many smaller villages and hamlets (Cline 1964). Together they cover more than half the political subdivisions existing in New Spain circa 1580 (Cline, 1972c).[4]<\/sup><\/a>\u00a0Figure 2 reproduces one of the manuscript pages of the Relaci\u00f3n de Atlatlauhca (present\u2013day Atlatlauhc\u00e1n, Morelos).<\/p>\n\n\n\n Alongside the texts, the corpus contains 78 maps. Some were painted by indigenous artists following native representations of places and landscape\u00a0(Figure 3 and Figure 4), while others were clearly drawn by creoles or peninsular Spaniards\u00a0(Figure 5).\u00a0In the words of Robertson (1972: 243), they represent \u201c\u2026the largest single group of interrelated early colonial pictorial manuscripts which has come down to us.\u201d[5]<\/sup><\/a>\u00a0Of these maps, 23 came from Mexico; 3 from Michoac\u00e1n; 1 from Nueva Galicia; 21 from Oaxaca; 18 from Tlaxcala; 3 from Yucat\u00e1n and 2 from Guatemala. There are also accounts of 26 manuscripts and 16 maps lost centuries ago (Cline 1972c: tables 6 and 10; Robertson 1972), as well\u00a0as an\u00a0additional map that disappeared more recently (Mundy 1996: 226).\u00a0<\/p>\n\n\n\n Despite the fact that some authors failed to answer several questions, both the texts and the maps offer favorable conditions to test new approaches for the analysis of large historical corpora. The corpus of RGs offers highly structured information \u2014due to the standardized questionnaire\u2014 which gives the reports a degree of homogeneity rarely seen in other sixteenth\u2013century collections. The geographic descriptions also cover more than half the territory of New Spain and the level of detail is remarkable not only in terms of orography, landscape, and natural resources, but also the way of life of native peoples and colonists, as well as the interactions between different social groups. Such richness, consistency, and span facilitate comparative research across small, medium, and large areas on topics of social, historical, ethnographic,linguistic, economic, and political interest.<\/p>\n\n\n\n A further reason for selecting the RGs is a desire to improve the accessibility and valorization of the corpus as a main source for historical enquiry. Seminal research has been conducted by Cline (1972), Bravo-Garc\u00eda (2018), Gerhard (1972), Gruzinski (1991), Moreno Toscano (1968), Mundy (1996) and a few other scholars (C\u00e1ceres Lorenzo, 2013; Delgado L\u00f3pez, 2010). The publications by Cline (1972) and Gerhard (1972), for example, were ground-breaking works of Mexican historical geography. Moreno Toscano (1968), on the other hand, used the complete textual information of the RGs to perform a matrix analysis oriented to reconstruct the economic geography of New Spain, whilst Mundy (1996) analyzed all the surviving maps to identify continuities and changes in the indigenous cartographic tradition during the colonial period. For his part, Gruzinski (1991: 77-103) has approached the corpus by studying how the contrasting \u2014sometimes conflicting\u2014 worldview of conquistadors and native peoples shaped the responses to the questionnaire, pointing out interesting details about the colonization of the imaginary.<\/p>\n\n\n\n Single documents of the Relaciones<\/em> have also been the focus of historical research (Afanador-Pujol, 2015; Ballesteros 2005; Barlow, 1949; Fern\u00e1ndez Christlieb and Garza Merodio, 2006; Morato-Moreno, 2017; Mundy, 2013; Nutall 1926, among others), but in general the textual and cartographic documents have served as ancillary sources in regional studies, rather than as a principal mine of data (Dahlgren, 1954; Brand, 1960; Gibson, 1967). This is due perhaps to the difficulty of cross\u2013linking information among the many documents or to a lack of tools for uncovering non\u2013explicit knowledge hidden in the subtext of the answers to the questionnaire.<\/p>\n\n\n\n 3. Methodology and current progress <\/strong><\/p>\n\n\n\n As discussed in the introduction, we are developing advanced tools for unveiling new information from the RGs corpus by applying a methodology known as Geographical Text Analysis (GTA). The resulting digital resources and software tools will be released at the end of the project under an open-access license. The methodology was originally produced by some members of this project to reveal new, unsuspected patterns of information by focusing on the geographic components of the historical narrative (Cooper and Gregory 2011; Murrieta et al. 2015; Donaldson et al. 2015; Gregory et al. 2015; Porter et al. 2015; Martins and Murrieta\u2013Flores 2017; Murrieta\u2013Flores and Gregory 2017; Gregory et al. 2018). <\/p>\n\n\n\n We are now in the process of refining and extending this methodology with contributions from all members of this project and from researchers engaged in parallel initiatives (e.g., Santos et al. 2017; Santos et al. 2018; Won et al., 2018). In this particular project we followed four phases: obtaining a computer-readable corpus, the creation of a gazetteer, the semantic annotation and geoparsing of the corpus, and data mining and historical analysis. In the following sections, we discuss each stage in detail.<\/p>\n\n\n\n 3.1 <\/strong>First phase: Obtaining a computer-readable corpus<\/strong><\/p>\n\n\n\n The first phase of the methodology is obtaining computer\u2013readable files from the more comprehensive editions of the RGs. We began with Francisco del Paso y Troncoso\u2019s Papeles de Nueva Espa\u00f1a<\/em>, which contains a significant portion of the corpus. Transcriptions from the original manuscripts were already available in electronic format, but it was necessary to identify and correct significant errors caused by the Optical Character Recognition (OCR) software. This is a common problem with old documents, to which Alpert\u2013Abrams offers an approach in her 2016 article on the corpus known as Primeros Libros<\/em>.<\/p>\n\n\n\n After completing six volumes of Papeles de Nueva Espa\u00f1a<\/em> (Del Paso y Troncoso 1905), the work continued with the RGs from Yucat\u00e1n, published in volume 11 of Colecci\u00f3n de documentos in\u00e9ditos de las posesiones de Espa\u00f1a en ultramar<\/em> (1898). Additionally, thanks to a generous permission from the National Autonomous University of Mexico (UNAM), we were also able to digitize a more recent version of the 54 RGs from Yucat\u00e1n transcribed and edited by De la Garza (1983) and to acquire in a digital format 114 RGs published in ten volumes by Acu\u00f1a (1982\u20131988). In the end, we obtained two versions of the whole corpus in machine\u2013readable format: the first one includes the notes, references and commentary from the scholars who transcribed and studied the documents, while the second contains only the transcription of the original RGs texts. A third version that includes all the NLP annotations was produced during the third phase of the project.<\/p>\n\n\n\n 3.2 Second phase: Creation of a gazetteer<\/strong><\/p>\n\n\n\n The second phase starts with a compilation of all the place names that exist in the corpus as well as many others from relevant sources, and then solving any linguistic and locational uncertainties that prevent their unequivocal identification.<\/p>\n\n\n\n 3.2.1 <\/em>Compilation of place names<\/strong>. <\/em>This is done by extracting toponym indices from published editions of the RGs, as well as from comprehensive studies on the historic geography of New Spain (Gerhard 1972a, 1972b, 1991; Cline 1972a, 1972b, 1972c; Moreno Toscano, 1968; Tanck de Estrada et al., 2005). Through this procedure, the UK and Mexican teams were able to compile a list of 14,654 place names (Liceras-Garrido et al. 2019).<\/em><\/strong><\/p>\n\n\n\n 3.2.2 Linguistic disambiguation<\/strong>. <\/strong><\/em>The resulting list was then subjected to a linguistic disambiguation process, during which the most difficult challenges were dealing with place names found in different native languages, as well as resolving name variations and spelling inconsistencies. Important names of cities, like Tlaxcala, have remained in original form since the sixteenth century, but many others have changed significantly over time for a variety of reasons, including: historic causes;[6]<\/sup><\/a> renaming due to changes in the boundaries of original settlements;[7]<\/sup><\/a> and addition of words such as the name of a patron saint to an indigenous toponym.[8]<\/sup><\/a> Another problem relates to spelling variations of the same toponym,[9]<\/sup><\/a> the application of Castilian transliterations to the names of ancient settlements,[10]<\/sup><\/a> or dealing with homonymies.[11]<\/sup><\/a><\/em><\/strong><\/p>\n\n\n\n State\u2013of\u2013the\u2013art methods to tackle these challenges are discussed in Nick and Tent (2017) and Santos et al. (2015, 2017, 2018). Additionally, there is a valuable body of research on the subject of Mexican toponomy, so we have benefited from publications by Anaya Monroy (1960), Guzm\u00e1n Betancourt (1987, 1989, 1998), Lefevre and Paredes Mart\u00ednez (2017), Le\u00f3n Portilla (2002), M\u00e1rquez and Ramos Navarro Wold (1998), Mundy (2014), Muntzel and Villegas Molina (2010), etc. This adds to the exhaustive research done by the editors of the RGs, especially Acu\u00f1a (1982\u20131988), De la Garza (1983) and Del Paso y Troncoso (1905). <\/p>\n\n\n\n 3.3.3. Geographic disambiguation.\u00a0<\/em><\/strong>Another challenge at this stage is finding latitude and longitude coordinates of towns and villages. The solution involves, first, detecting and merging duplicate entries by comparing the place name list with records found in online geographical databases, such as\u00a0the catalogue of indigenous localities, produced by Instituto Nacional de Estad\u00edstica Geograf\u00eda a Inform\u00e1tica (INEGI) and Comisi\u00f3n Nacional para el Conocimiento y Uso de la Biodiversidad (CONABIO);[12]<\/sup><\/a>\u00a0the geographic dataset created for the cartography of Mexico by INEGI; the general registrar of archaeological sites compiled by Instituto Nacional de Antropolog\u00eda e Historia (INAH);[13]<\/sup><\/a>\u00a0the GeoNames geographical database;[14]<\/sup><\/a>\u00a0and\u00a0the Getty Thesaurus of Geographical Names.[15]<\/sup><\/a>\u00a0Through both linguistic and locational disambiguation, we were able to obtain 12,748 correct names. Of these, we have so far been able to pin-point 3919physical locations\u00a0(Figure 6), which\u00a0are the entries used for the gazetteer.<\/p>\n\n\n\n The structure of the gazetteer complies with Linked Open Data standards to guarantee its interoperability between different systems and repositories and facilitate its application into other projects. We have relied on the experience of massive reference resources like the Getty Thesaurus of Geographical Names[16]<\/sup><\/a> and especially the Alexandria Digital Library Gazetteer (ADL), from which we adopted the relational database schema (Hill, 2000; Hill et al., 1999). For importing and exporting data, we adopted the file format JSON, created by the World Historical Gazetteer[17]<\/sup><\/a>, which is now an accepted standard in many textual annotation and analysis platforms, such as Recogito. The database of our gazetteer will be made available through the World Historical Gazetteer, the Pelagios Network, and through Recogito. In addition, the gazetteer will be also available in a standalone downloadable GIS format from the project\u2019s website and repositories.<\/p>\n\n\n\n 3.4 Third phase: Semantic annotation of the corpus<\/strong><\/p>\n\n\n\n The third stage comprises four tasks: defining an annotation schema by selecting the most relevant analytic categories for the intended historical analysis; annotating manually, and in accordance with the previously defined annotation schema, two different samples of documents: one for training the NLP deep-learning model and another one to validate it; tuning the deep-learning model to automatically perform name entity recognition, disambiguation and classification, in order to obtain a fully annotated corpus; and assessing the performance of the model in accordance with the gold-standard validated by experts. <\/p>\n\n\n\n 3.4.1 Defining the annotation schema (ontology).<\/strong> <\/em><\/strong>The selection of the ontology categories can be done by consensus among expert linguists and historians, or with the help of automatic computer methods that extract a \u201cseed ontology\u201d directly from the corpus (Eynard et al., 2012). The first method is more time-consuming, but it is simpler and on many occasions yields better results, so this was the approach we followed.<\/p>\n\n\n\n The RGs ontology contains 21 entity classes and many more sub-categories relevant to social, political, territorial and economic terms appearing recurrently through the RGs (Table 2). For example, the general entity class \u201carchitecture\u201d includes sub-categories such as \u201cdomestic\u201d and religious\u201d, which offers the possibility to distinguish between house and convent buildings described in texts. Another class \u201ccosmogony\u201d includes sub-categories such as \u201critual\u201d, \u201cfestivity\u201d, \u201cdeity\u201d, etc. Each entity type follows the definitions and linked-data standards of DBpedia to ensure the interoperability of our dataset.<\/p>\n\n\n\n 3.4.2 Manual annotation of text samples for training the NLP model.\u00a0<\/em><\/strong>Using such entity classes we carried out a manual annotation of two samples of 18 out of 168 documents (10% of the corpus) from the RGs using the online tagging tools developed by Tagtog, a commercial company that generously provided a free license to our project.[18]<\/sup><\/a>\u00a0In order to ensure the quality of the annotated samples, five experts tagged the same documents independently. We then obtain statistics on the level of consensus among the experts by using the inter-annotator agreement tool from Tagtog.\u00a0Figure 7\u00a0illustrates an annotated document.<\/p>\n\n\n\n 3.4.3 Development of the NLP deep-learning model.<\/strong> <\/em>Once the training and gold-standard sample data are available, the actual development of the NLP model can begin. Our team in Portugal, led by Bruno Martins, combined two state-of-the-art approaches of deep-learning: neural networks and word embedding.<\/p>\n\n\n\n 3.4.3.1 Neural networks.<\/strong> A neural network is a computational model, loosely inspired on the neural structure of the brain, that is able to “learn by example” how to perform repetitive complex tasks. In the NLP and CL fields, it is applied, among other things, to perform word classification, or the assigning of a label to each word in an input text, indicating whether it belongs to a certain semantic category (e.g. toponym, person, animal, plant, institution, cultural concept), or to a \u201cpart-of-speech\u201d syntactic category (i.e. noun, pronoun, adjective, determiner, verb, adverb, preposition, conjunction, and interjection).<\/p>\n\n\n\n Modern NLP software libraries leverage such statistical models (Goldberg, 2017), making probabilistic decisions on the best annotations for previously unseen texts. One of the best is SpaCy. To run a deep-learning model in SpaCy we need to conduct a training phase during which the software is fed with the training data sample (in our case 10% of the corpus). This annotated training sample contains many known examples of the kind of words expected in the unannotated text (the remaining 90% of the corpus). <\/p>\n\n\n\n Because we know the correct answers \u2014thanks to the gold-standard data sample\u2014, we can give the model feedback on its predictions in the form of an error gradient or the loss function that calculates the difference between the training example and the expected output. If the accuracy of the annotations is not sufficient, then an iterative refining process can be carried out until accomplishing satisfactory results.<\/p>\n\n\n\n Then, to link the gazetteer with the corpus texts, it is necessary to assign correct longitude and latitude coordinates to each toponym recognized by the neural network. The Mordecai geoparsing library (Halterman, 2017) is a good tool for that job. It takes as input the place names recognized by SpaCy, and uses an index over the contents of the gazetteer to find the potential coordinates of extracted place names. In other words, the names recognized in the text are matched against similar names in the gazetteer. Through this process the gazetteer and the corpus are finally linked and ready for the implementation of query tools.<\/p>\n\n\n\n 3.4.3.2 Word-embedding.<\/strong> Complementing the above procedures, the word\u2013embedding algorithm makes it easier for the neural network to distinguish all of the different semantic contexts in which a word might be used. This process follows the idea that “a word is characterized by the company it keeps” (Firth, 1957). For example, in the phrase \u201c[t]he order of San Francisco was established in these lands in 1524\u201d, the term San Francisco<\/em> refers to an institution, but in \u201c[t]he mendicant friars helped the poor following the teachings of San Francisco\u201d, the same term refers to a person. In another context, it could mean a geographic location, as for instance in the sentence \u201cSan Francisco Tlalcilalcalpan is inhabited at the present by 18,721 people.\u201d<\/p>\n\n\n\n We specifically pre-trained a word\u2013embedding algorithm in an unsupervised way to statistically predict the meaning of words in the context of real natural language utterances. Thus, we needed training data sufficiently big and varied to tackle the specific challenges of the RGs. One is that the spelling in the corpus varies considerably among the many authors. This is sometimes, in part, because words are spelled in accordance with their Latin origin, rather than their actual pronunciation. Another is that the corpus borrows heavily on vocabulary from indigenous languages, especially place names, and terms to identify flora, fauna, cultural concepts, etc.<\/p>\n\n\n\n Given the fact that the structure of modern Spanish still resembles to a high degree the language recorded in the RGs, we decided to combine modern and historical data sources to train our model. Contemporary data came from several Spanish corpora such as ANCora (Taul\u00e9 et al., 2008), IMPACT-es (S\u00e1nchez-Mart\u00ednez et al., 2013), or the Corpus del Espa\u00f1ol,<\/em>[19]<\/sup><\/a> while numerous examples of sixteenth\u2013century Spanish texts from different archives and ten dictionaries of indigenous languages (Garc\u00eda Cubas, 1888-1991; Pe\u00f1afiel, 1897; Robelo, 1897, 1900, 1902, 1902b; 1904; 1905; Starr, 1920) provided resources to deal with the language complexities particular to New Spain.<\/p>\n\n\n\n By thus leveraging modern and historical data in a combined training dataset, we can simultaneously benefit from having large amounts of contextual information from which the model can learn more easily, together with some data representing the specificities of historical Spanish. Furthermore, to avoid problems with out-of-vocabulary words, such as those brought forward by the inconsistent spelling or cultural expressions, we used a word\u2013embedding method that also creates representations for sub-words, and which is able to reconstruct entire word representations from such sub-words (Bojanowski et al., 2017).<\/p>\n\n\n\n 4. Future Work: Data mining and historical analysis<\/strong><\/p>\n\n\n\n At this stage, several data mining and analysis procedures can be carried out. The first is an adaptation of collocation analysis \u2014a technique from Corpus Linguistics (CL)\u2014 which allows for extracting the most statistically significant phrases that are found next to a term or keyword of interest (the \u201ccollocates,\u201d in CL jargon). Our variant of the technique is geographic<\/em> collocation analysis<\/em>, originally developed by the research group at Lancaster (Gregory et al., 2015, Murrieta-Flores et al., 2015, 2017a, 2017b). This facilitates the recognition and retrieval, across the corpus, of all place names that collocate or are close to a keyword of interest according to a predefined proximity threshold, usually defined in terms of the maximum number of words between the search\u2013term and the collocate. In this manner, if a researcher is interested in investigating economic aspects related to agriculture, they could retrieve, for instance, all phrases in the context of words like ma\u00edz<\/em>, siembra<\/em>, subsistencia<\/em>, mantenimientos, comida<\/em> (i.e. maize, sowing, subsistence, sustenance, food), which would appear linked to a map showing the geographic locations where such terms appear. <\/p>\n\n\n\n Furthermore, by putting words in context through collocation, a historian could extract in a few seconds \u2014as opposed to the weeks that it would take to access the information without GTA tools\u2014 interesting details about the first colonists in New Spain. Such a query, for instance, might involve running a collocate\u00a0search of the word \u201cconquistador\u201d\u00a0(Figure 8).\u00a0Another type of search can be conducted using general categories, such as employing the term\u00a0food<\/em>\u00a0to retrieve words like\u00a0beans, fish,\u00a0<\/em>or\u00a0bread<\/em>\u00a0(Figure 9).<\/p>\n\n\n\n In the following subsections, we highlight several topics of historic analysis that we plan to examine through this research.<\/p>\n\n\n\n 4.1 Historical phenomena changing the face of New Spain at the end of the sixteenth century<\/strong><\/p>\n\n\n\n One area of application for the techniques detailed in this study is understanding the coexistence of competing systems of land ownership and labor in New Spain during the colonial period. In New Spain during the short time of compilation of the RGs, there was a strong debate about the system of encomiendas <\/em>and the desire to replace it with the so-called repartimientos<\/em>. The encomienda<\/em> was the right to exploit the labor of indigenous populations, granted by the Spanish Crown to conquistadors, who at the beginning of the colonization held this privilege in perpetuity and could, therefore, pass it to their descendants with little intervention from the royal government in Madrid. In contrast, the repartimiento<\/em> only allocated the labor force of native workers for a fixed period of time in exchange for a meager remuneration. The latter gave the royal government more control over its subjects and resources. Neither the holders of encomiendas <\/em>nor repartimientos<\/em> were allowed to appropriate lands, as private property was reserved for haciendas<\/em>. These three systems coexisted in the 1570s despite the enactment of the so-called New Laws in 1542, which attempted to protect the rights of the indigenous populations and prevent the dispossession of their lands by abolishing the encomiendas<\/em>. Throughout the texts of the RGs, therefore, it is common to find claims by conquistadors\u2019 descendants of their encomienda<\/em> rights, and others defending the benefits of repartimientos<\/em>. A comprehensive study of the claims and contentions found in the RG corpus would improve our understanding of the period.<\/p>\n\n\n\n 4.2 The local government organization of New Spain toward the third quarter of the sixteenth century<\/strong><\/p>\n\n\n\n A second application of these methodologies has to do with disambiguating local structures of governance in New Spain. The early colonial period\u00a0was a time when the functions and spheres of influence of the four branches of colonial government were not well defined, and consequently the spatial boundaries of the judicial, administrative, military and ecclesiastical jurisdictions overlapped\u00a0(Figure 10).\u00a0This was especially true at the municipal level. Consequently, the corpus contains terms associated with several kinds of jurisdictions whose meanings were not consistent throughout the vast territory of New Spain (e.g.\u00a0partidos de cl\u00e9rigos<\/em>,\u00a0distrito de<\/em>\u00a0mision<\/em>,\u00a0custodia<\/em>,\u00a0guardian\u00eda<\/em>,\u00a0vicar\u00eda, presidencia,\u00a0<\/em>etc.). The subject has been a matter of ongoing debate among historians (Gerhard, 1972b: 66), but we believe that applying collocation analysis for extracting the contextual use of such words would shed light on how those institutions functioned in parallel to\u00a0alcald\u00edas mayores<\/em>\u00a0and\u00a0corregimientos<\/em>\u00a0and this would be an important contribution to the knowledge of this period.<\/p>\n\n\n\n 4.3 Settlement pattern analysis to elucidate the territorial organization of New Spain <\/strong><\/p>\n\n\n\n A third area for future study has to do with demarcating the boundaries of territories in New Spain. By reading the RGs, we could attempt to identify the names and boundaries (linderos<\/em>) of towns belonging to corregimientos<\/em> (territorial demarcations for administrative and judicial purposes) mentioned in the corpus. Circa 1570, there were 200 of such units whose frontiers have been outlined by Gerhard (1972a). However, the smaller demarcations (towns, villages) suffered adjustments in boundaries and jurisdictions. As Gerhard points out, at the time of the RGs, \u201c[e]arly civil and ecclesiastical divisions lost importance and were eliminated; others came into being as a result of conquest, proselytization, mining activity, and other causes\u201d (Gerhard, 1972b: 66-69). The application of a mathematical model that analyzes the maximum available territory around each settlement within a region, like the one proposed by Ducke and Kroefges (2008), could provide preliminary hypotheses on such demarcations. These would then be tested against the actual historical information extracted from the corpus through geographic collocation analysis. One of the many ways that such a study could begin is by extracting references on the routes that local administrators take for their visits to towns within their jurisdiction, the distances they traversed, the geographic features and topography of the landscape, etc. (Figure 10) This, and much other information, is available throughout the corpus, but more specifically in the answers to questions 7, 8, 11, and 12 (see Appendix 1).<\/p>\n\n\n\n 4.4 Analysis of historical diseases<\/strong><\/em><\/p>\n\n\n\n A fourth research question has to do with our understanding of the outbreak of disease (1576-1581) that took the lives of half the population of New Spain during the early colonial period (Acu\u00f1a-Soto et al., 2004). This was the second time that cocoliztli<\/em> (as the epidemic was called) had decimated the native population, causing the disappearance of whole communities, shortages in labor-force for the encomiendas<\/em> and repartimientos<\/em>, the displacement of people over different regions, or, on the contrary, the congregation of survivors (sometimes of different ethnic communities) in new Indian towns, all of which altered dramatically the traditional settlement pattern (Gerhard, 1972a: 24). How did this phenomenon vary over different regions? How many people died in each locality? How did this epidemic compare with others in the past? How did the new settlement change in comparison with Prehispanic times? These are all questions that GTA could help to answer by collocating words and phrases related to health (such as diseases, symptoms, and remedies) and analyzing them with spatial mathematical models of disease contagion. <\/strong><\/p>\n\n\n\n 4.5 The spatial distribution of natural resources (flora, fauna, minerals, etc.) and exploitation practices at the time of the RGs<\/strong><\/p>\n\n\n\n A fifth application of our methodology has to do with the study of natural resources. GTA will allow the extraction and analysis of regional historical information on natural resources such as plants. There are already interesting archaeological, historical and ethnographic research related to plants and their use in Mesoamerica, medicinal and otherwise (\u00c1lbores Z\u00e1rate, 2015; Ch\u00e1vez Mej\u00eda, 2020; Hersch, 1999). Some further research has been also dedicated to assessing the importance of the trees and edible plants introduced by Europeans. With the annotation of the RGs in this project and the development of GTA, it is now possible not only to extract information regarding native and European plants introduced in the different regions of New Spain, but also to contribute to further debates. It is known, for instance, that Mesoamerican cultures had an extended knowledge of plants for medicinal use and most of the research about this has been carried out using colonial primary sources. We could now create a map of all the plants mentioned in the RGs and compare this dataset with archaeological data and other primary sources that were dedicated to the recording of medicinal plants, such as the C\u00f3dice Badiano<\/em> or the Historia Natural de Nueva Espa\u00f1a. <\/em>This would allow researchers, for instance, to assess the continuation in the use of specific plants for remedies and consumption in specific regions of New Spain. Other interesting avenues of possible research with GTA analysis are the availability of food and animals both native and introduced, the use of natural resources, the existence of cultural artefacts, the presence of priests and religious orders in particular regions, as well as types of architecture.<\/p>\n\n\n\n 4.6 Exchange networks, trading practices and communication routes<\/strong><\/p>\n\n\n\n An additional topic is the study of exchange networks and trading practices, as well as communication routes, both terrestrial and aquatic, among the populations described in the RGs.\u00a0One such study has already been conducted by Favila-V\u00e1zquez (2019), who successfully reconstructed aquatic communication networks in the Rio Balsas region by extracting information from the responses available to queries 38 to 47 of the RGs\u00a0(Figure 11).\u00a0These contain crucial data regarding navigation by sea, including location of ports and coastal landscape. Together with data from question 19 (focused on fluvial networks), the extraction of data from the RGs corpus allowed Favila-V\u00e1zquez to reconstruct a riverine and coastal system of communication that explains the spatial connectivity of the Valley of Mexico to the Pacific coast.\u00a0<\/p>\n\n\n\n 5. Conclusion<\/strong><\/p>\n\n\n\n Geographical Text Analysis can make an important contribution to the study of massive corpora, like the RGs, <\/em>by expediting the extraction and analysis of specific information that would be too slow or cumbersome to find in big datasets through traditional reading methods. It also can provide tools to uncover patterns behind the narrative of historical sources and link them to their geographic locations. Thus, GTA creates a way for historians to apply what in computational literary studies is called \u201cdistant reading.\u201d[20]<\/sup><\/a><\/p>\n\n\n\n At the same time, the use of geographic collocations and the concept of \u201ckeywords in context\u201d permit a more detailed, granular comprehension of the corpus by identifying those parts that deserve a \u201cclose reading.\u201d This flexibility has been demonstrated in previous research projects employing GTA (Donaldson et al., 2017; Murrieta\u2013Flores et al., 2017; Taylor et al., 2017). <\/strong>Having a multiplicity of \u201cplace names in context\u201d associated with keywords of interest opens an unprecedented range of possibilities in terms of identifying information patterns and hypothesis\u2013testing across large corpora, which in turn allows answering historical questions in a way that would be impossible otherwise. <\/p>\n\n\n\n In such a context, the proposal of GTA must not be seen as an attempt to replace the careful micro study of historical sources practiced by modern historians. More exactly, GTA is a methodology for parsing and crosslinking information at a macro scale throughout a whole corpus in a relatively short period of time, so historians can use their efforts and time more efficiently to focus on the more important documents or fragments of texts.<\/p>\n\n\n\n One of the main contributions of our project specifically has been adapting named entity recognition, disambiguation, and classification techniques to text collections that either belong to specific historical periods or might pose a further challenge due to their multi\u2013language nature. We have succeeded in developing a new deep-learning model to process the language of the RGs despite the orthographic variance across the corpus and the high frequency of indigenous vocabulary. Nevertheless, there is still much work to be done, especially for the indigenous languages of the Americas (Mager et al. 2018). <\/p>\n\n\n\n Finally, we have already produced a significant number of digital datasets of value, such as the gazetteer, which includes not only sixteenth\u2013century localities but place names from other Mexican historic periods. Hopefully, these will prove valuable resources for humanists engaged in studying the Latin American historical and archaeological past. The methods and tools being developed by the project have the advantage of being generic, and therefore, applicable to other corpora. <\/p>\n\n\n\n We are committed to releasing the digital products and software tools through several repositories, including GitHub and our own online platform. Our hope is that scholars from diverse fields will be able to identify the desired information, extract it, cross\u2013link it to other document resources, and most importantly, answer significant research questions: Where do certain phenomena happen more often? Who was responsible for certain events? What was happening in specific places at certain times? Such kind of enquiry will take historical analysis a step further, contributing significantly to advancing research on the colonial period in Latin America, and humanities research more generally. <\/p>\n\n\n\n![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/1-695x1024.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/2-739x1024.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/3-753x1024.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/4-1024x946.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/5-1024x747.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/6-1024x778.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/7-1024x475.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/8-1024x631.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/9-1024x635.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/10-1024x786.png\")
![\"\"](\"https:\/\/achieve.dhcn.cn\/wp-content\/uploads\/2022\/04\/11-1024x808.png\")