Navigation – Plan du site

AccueilNuméros17DossierHorizon discursifPlace and Sentiment-based Life st...

Dossier
Horizon discursif

Place and Sentiment-based Life story Analysis

From the Spanish Republican Army to the French Resistance
Catherine Dominguès, Laurence Jolivet, Carmen Brando et Marion Cargill

Résumés

Le Réseau des acteurs de l’histoire et la mémoire de l’immigration (RAHMI) a lancé en 2008 un programme expérimental de collecte pour recueillir la mémoire oubliée de populations immigrées engagées dans la vie de la région. Différents groupes de personnes ont été visés par cette collecte qui a permis d’enregistrer des récits de vie dont ceux de Républicains espagnols s’étant exilés en France entre 1936 et 1939, et ayant participé à la résistance française. Le projet MATRICIEL (PEPS CNRS UPE 2016) s’est intéressé à la migration de ces Républicains espagnols sous l’angle des lieux mentionnés dans leurs récits, et des sentiments associés à ces lieux. Le projet visait à placer la parole des migrants au premier plan de l’analyse ; les objets d’étude choisis : les lieux, désignés par un nom propre : Barcelone, ou un nom commun : camp d’internement, et les sentiments associés distingués par leur polarité : positive ou négative, contribuent à valoriser les archives orales pour la construction d’une mémoire de l’immigration. Il s’agit dans cet article de présenter la démarche mise en place pour une analyse pluridisciplinaire de ce corpus de récits de vie, qui combine des méthodes et outils du traitement automatique des langues et de la cartographie. L’identification des lieux noms communs mentionnés dans les récits a été menée grâce à un modèle d’apprentissage supervisé. L’identification puis la représentation cartographique des lieux noms propres ont mis en évidence la répartition spatiale des parcours de vie des témoins, déterminés par le contexte historique et les choix personnels. L’annotation semi-automatique en sentiments ajoute une polarité aux récits. En perspective, l’analyse des types de lieux noms communs permettra d’évaluer la granularité utilisée par les témoins pour décrire les espaces vécus, et leur localisation contribuera à préciser la spatialité des récits.

Haut de page

Texte intégral

Introduction

  • 1 https://ich.unesco.org/fr/convention

1Migration is an old and multifaceted phenomenon. It has been ever renewed and new types of migration have emerged such as climate migration. It is a strong and often painful topic of interest, as shown by recent events and their media coverage, and academic productions on the subject. For its part, UNESCO has wished to promote the value of intangible cultural heritage and adopted in 2003 a convention for its safeguarding1; oral memory is thus acknowledged as an integral source of archives and particular emphasis is being placed on the contributions of immigration to national collective history and memory (Musée de l’histoire de l’immigration, 2014). In this context, the Réseau des acteurs de l’histoire et la mémoire de l’immigration (RAHMI) launched in 2008 an experimental program to collect forgotten stories of immigrant populations involved in local initiatives of the French region Nouvelle-Aquitaine. Several groups of people were targeted for collecting records of their life stories, one of them concerns the Spanish Republicans who went into exile in France between 1936 and 1939 and then committed to the French Resistance. 

2Life stories are traces of the migration history and the collection program has been an opportunity for associations, academia and archivists to work together. Several fields may be involved in the analysis of the life stories. In this paper, we use methodological approaches and tools combining natural language processing (NLP) and geomatics to provide new materials and reflect on other perspectives about migration.

3Places are structuring elements in all kinds of texts such as novels, newspaper articles, tourist guides, life stories (Caquard & Fiset, 2014). They represent relevant information and they are integrated into the description of an event or a fact (Nouvel et al., 2015). A place corresponds to a location, but not only that, as it is also defined by a category, usages and perceptions associated by those who live in, pass through, and speak about it. So, specifying the notion of place involves looking for additional features in terminology resources or ontologies (for example, geographic coordinates or administrative categories) , and extracting information as perception, sentiment, emotion or opinion which then can be associated with places from the stories. All these pieces of information correspond to features of a place.

4Our challenge is to specify the notion of place in the context of life stories and to define a spatial representation of these places. The notion of place in life stories should also take into account that life stories mean each individual speaks in perspective of other life stories. Unique individual stories can be articulated with respect to common experiences shared by a group of people and mentioned in several life stories. In particular, places can be experienced by one person or a group and there may exist regularities in places crossed during the journeys. Spatial representation should combine those features and propose a graphical reading of the life stories.

5The MATRICIEL project (PEPS CNRS UPE 2016-2017) focuses on the migration of the Spanish Republicans who participated to the RAHMI collecting program. The project aims at highlighting the voice of migrants by studying places and sentiment in order to explore memories and traces of immigration as a collective heritage. Our main research topic is textual data processing and analysis; it concentrates on the places mentioned in life stories and sentiment associated to these places. Expected outputs are extracted information to be then interpreted and turned into knowledge about historical facts. An automatic processing pipeline is defined from text analysis to graphic visualization. It includes NLP methods as well as descriptive and spatial data analysis tools.

6A state of the art enables to identify existing relevant approaches according to our objectives (section 2). The different parts of our defined global approach are integrated in section 3: first the life stories have been collected by the RAHMI and we then have formatted this material with metadata; the transcription of the audio records were corrected and enriched with indications about oral speech (repetition, hesitation); annotation guidelines were defined with schemes in order to automatically annotate and extract information. Evaluation of the annotations and results are described for places and sentiments (section 4). Discussion and perspectives are presented in section 5.

State of the art

7Life stories are narratives structured around places and the properties which can be attached to, sentiment in particular, and can be analyzed in order to be mapped. This section provides a state of the art of the main research topics of this work, i.e. place, place name, sentiment and story mapping.

  • 2 « la matière est généralement divisée selon la géographie »

8The definition of a place has its roots in the field of geography. Geographers such as Brunet et al. (1993, p. 485), and linguists like Dubois et al. (1994, p. 485), agree that “matter is generally divided according to geography” (our translation2) into mountains, rivers and water bodies, inhabited places, etc. Defining a place is a socially anchored process that makes it possible to distinguish, in a way that aims to be shared and permanent, a portion of territory and to locate it on the earth (Brunet et al., 1992, p. 485). A place is built by experience and its layout (coordinates and area), more or less precise, varies according to the scale of the representation (Tuan, 1977).

9A place can be designated by its name. In this manner, the United Nations Group of Experts on Geographical Names (UNGEGN, GENUNG in French) addresses issues related to the standardization of geographical names. It provides assistance to countries that are creating their national toponymic structure to limit an uncontrolled proliferation of different names for the same places. Through their definitions, properties and uses, proper nouns (now Npr) are well suited to name places. This makes it possible to distinguish and designate, by a form defined in the language and in a conventional and stable way (Calabrese-Stienberg, 2009, p. 7), an element that emerges from its environment by geographical and/or social considerations. The landscape elements noticed, distinguished and designated by an Npr contribute to the knowledge and structuring of the physical world (Jonasson, 1994, p 18). In an international context, several places may share the same name: for example, in the collaborative gazetteer GeoNames3 Bordeaux can designate a town in France, Canada, Gabon, etc., a bay and a hill in Virgin Islands, a lake in California, etc. Reciprocally, several forms (used in respective languages) may be used to designate the same place: for example, the French city Bordeaux may be named Burdeos in Spanish, Bordeus in Catalan or Bordele in Basque, but the Gabon’s one has no other names (according to GeoNames). Nevertheless, these Npr do not make it possible to designate all the places and located objects, useful for human activity (Dominguès, 2018): common nouns (Nc) must be added to Npr. The places mentioned in life stories can be classified into two different linguistic categories: Npr which, according to Jonasson morphological description (Jonasson, 1994), can be pure: Barcelona, Ebro, or descriptive: Jardin des Plantes, or mixed: fort du Hâ, and Nc: zone occupée (occupied zone), usine de textile (textile factory), camp de concentration (concentration camp).

10Sentiment analysis has become a very popular task in NLP since the development of social networks and twitter. The notion exists in many domains but the emphasis is on interpreting a situation rather than the situation itself as triggering an emotion, as reflected in the definition given by the cognitive appraisal theory by Scherer et al. (2004) cited by Chatar-Moumni (2013): « Episodes of massive, synchronized recruitment of mental and somatic resources allowing to adapt or cope with a stimulus event subjectively appraised as being highly pertinent to the needs, goals and values of the individuals ».

11Typologies have been developed, particularly in psychology with those of Plutchik (1991) and Ekman (1992), where primary or secondary fundamental emotions combine to form others. Six emotions are common to these typologies: anger, disgust, fear, sadness, surprise, joy. In NLP, sentiment analysis addresses issues under sentiment, opinion, emotion and perception. Chatar-Moumni (2013) notices that the works on the linguistic expression of emotions address mainly the properties of the emotion lexicon. Indeed, several lexicons have been offered, most often translated and adapted from English. They gather words to which sentiment may be associated. The words are qualified by a valence which spreads between two extreme values: one characterizing an extremely positive sentiment, the other an extremely negative sentiment, or a polarity: positive or negative or neutral. The polarity may be balanced by an intensifier.

12SentiWordNet (Esuli & Sebastiani, 2006), (Baccianella et al., 2010) aims to identify and classify sentiments and opinions. It has been created from the semantic database WordNet (Miller, 1995) which represents concepts by a synset i.e. a group of synonymous words. These synsets have been automatically annotated with a polarity that can be neutral, positive or negative. aFinn (Nielsen, 2011) is an another example of lexicon; it contains 2,477 English terms to which whole valences between -5 and 5 have been assigned, the valence 0 is seen as neutral. The FEEL lexicon (a French expanded emotion lexicon) was proposed by Abdaoui et al. It was built from the English lexicon NRC-EmoLex (Word Emotion Association Lexicon), automatically translated and then manually validated, and extended using synonyms. Finally, the glossary contains more than 14,000 terms to which an emotion (one of the 6 basic emotions in Ekman’s typology: joy, surprise, anger, fear, sadness and disgust) is associated, and a polarity. The EMOTAIX emotional and affective lexicon (Piolat & Bannour, 2009) also aims to identify, categorize and count words that literally or figuratively describe psychological states in texts. Unlike SentiWordNet, aFinn and FEEL, it is directly made for French and hierarchically organized into referents (each one brings together other words sharing the same radical and a similar meaning) that describe the domain of emotions, feelings, moods, affects, emotional personality, temperaments with positive or negative polarities. At the top of the hierarchy, three supra-categories: malevolence, malaise, anxiety (with negative valences) versus benevolence, well-being, cold blood (with positive valences) organize nine super-categories. Each super-category contains 2 x 28 basic and opposite semantic categories: for example, disgust versus desire; disorientation versus calming; terror versus daring) which 3 other basic categories are added to: unspecified emotions, surprise and impassibility.

13Nevertheless, many studies also show that the emotional value is not only conveyed by the vocabulary but by syntactic, morphological, and discursive items in the texts. Consequently, in addition to the lexicons, other approaches have been studied, such as verb modality (which can express likelihood, ability, permission, request, capacity, advice, obligation, etc.) and negation, which makes it possible to specify and qualify, or even reverse, the polarities of the lexicon’s terms (Benamara et al., 2012), (Andreevskaia & Bergler, 2008), (Özdemir & Bergler, 2015).

14In the context of NLP, the notion of place should be brought closer to the definition of named entity (NE) and especially spatial named entity (SNE) which are well defined in the literature (Grouin et al., 2011). There are tools to automatically identify, extract, disambiguate and locate these SNE (Nouvel et al., 2015), (Moreno et al., 2017), (Melo & Martins, 2017), which operate gazetteers: BD Nyme, the French national mapping agency’s official database toponyms in France, GeoNames; therefore they are well adapted to the search for places designated by Npr. However, these tools and methods are not effective in identifying places designated by Nc whose relevance as place name varies according to the context and for which other methods must be implemented.

15In this context, we proposed (Brando et al., 2016) a method based on a machine learning approach which uses the Stanford-NER (Named Entity Recognition) tool. It implements linear chain Conditional Random Field (CRF) sequence models. The identification of the NE in texts takes form of an annotation associated to the NE mention affected. The NER tool needs learning corpora labeled using the BIO annotation: the head of each mention is marked as B-LOC and the rest of the mention as I-LOC; the O tag is used to indicate that the word does not belong to a NE; for example, in the sentence le camp de concentration surpeuplé était terrifiant (the overcrowded concentration camp was horrible), the manual and automatic annotations are shown in Table 1.

Table 1 – Manual and automatic annotations of the Nc place names according to the BIO format

Manual annotation

Automatic annotation

le

O

O

camp

B

B

de

I

I

concentration

I

I

surpeuplé

I

O

était

O

O

terrifiant

O

B

  • 4 For illustrated definitions, see https://en.wikipedia.org/wiki/F1_score

16In order to assess the performance of automatic tools, the results must be compared to results obtained manually. In this example, camp de concentration is recognized by both the automatic model and the manual annotation, it is therefore considered as a true positive; terrifiant has not been manually annotated and it has been recognized automatically, so it is a false positive; surpeuplé has been manually annotated and it has not been recognized automatically, so it is a false negative. The comparison is used to being based on the precision, recall and F-measure4 (or F-score) ratios:

precision = number of true positives / (number of true positives + number of false positives)

recall = number of true positives / (number of true positives + number of false negatives)

F - measure = 2 * precision * recall / (precision + recall)

17The identification of a place name and associated information in a text often aims to map this located information; it means defining the cartographic objects which graphically translate the place and its properties in the text, each cartographic object being characterized by its layout (point, line or area), symbol and location. These representation issues are at the heart of the sensitive mapping approach (Griffin & McQuoid, 2012) which, in a multidisciplinary context, aims to give a representation of a lived and perceived space: mapping of life stories (Caquard & Cartwright, 2014), (Olmedo, 2016), travel itineraries (Santos et al., 2017), novels (Rosemberg & Trouin, 2017), migration stories (Mekdjian, 2016), movies (Caquard & Fiset, 2014), indigenous mapping (Hirt, 2009, about maps by the Mapuche), collaborative construction of a city narrative (Rabie, 2017, about Le Grand Paris), etc. New representation solutions are then created to report these perceptual experiences of places, for example those based on textiles (Olmedo, 2016) or clay (Mekdjian & Amilhat Szary, 2015). These representations emphasize more the relationships between places and their features (danger, hunger, medical care, fear, death, etc.) than their location.

The MATRICIEL corpus and textual analysis pipeline

18Considering the objectives of the project MATRICIEL, our best alternative was to rely on an existing collection of life stories collected in the form of oral interviews as it was done by the RAHMI. This collection can be considered as a corpus because it presents a similar theme and structure. The interviews include life stories; witnesses are Spanish who sided in the Republican army during the Spanish civil war, went into exile in France and involved themselves in WW2 and/or in the French Resistance. The interviews are structurally consistent: the stories share a content structure; at the beginning, the interviewers talk about their childhood and parents; then they develop how they had involved in the civil war then in the French Resistance; they often conclude by their links with the associative and social life of Nouvelle-Aquitaine as well as the present ties with Spain.

Description of the corpus

19The initial RAHMI corpus is composed of oral recordings lasting a total of 18 hours and 30 minutes in which 18 persons tell excerpts of life stories. 11 are direct witnesses and 7 are indirect witnesses (i.e. spouse or child of a direct witness); 11 are male and 7 are female. The names of the witnesses, interviewers and any other people mentioned in the stories were anonymized, with the exception of historical well-known characters such as Francisco Franco or Charles de Gaulle. The interviews were conducted in French and are semi-directed, in other words, open-ended questions are asked by the members of RAHMI and the witnesses are free to tell their story. The interviewer systematically asks the same question at the beginning of the interview (translated here into English): Could you tell us who were your parents? What did they do for living? What was their degree of education? How did you arrive to France? In the remainder of the interview, the interviewer takes part only to ask for clarification of a fact or to prevent the speaker from wandering away from the story.

Transcription of the corpus and annotation of oral phenomena

20Initially, the set of 18 life stories were partially and manually transcribed by RAHMI members. Nevertheless we considered that an exhaustive and normalized transcription was necessary as many facts mentioned in the recordings were absent from transcriptions, also proper nouns were often misspelled and both prosody and multilingual phenomena were ignored. About this last manner, as most of the witnesses were Spanish native speakers, there was a frequent mixing of French and Spanish languages. Considering all these difficulties, transcriptions were revised and completed, independently by two human annotators.

21For doing so, transcription guidelines were defined according to the needs of our analysis and constituted the most challenging step of our work because of the inherent subjectivity of the task. We took as work base existing guidelines designed by the linguistics research community, in particular ESLO5 and ESTER26 conventions. RAHMI partial transcriptions were imported into the transcription software CLAN7 then the human annotators corrected, completed and aligned these texts to the sound signal. The speaking turns, foreign words, use of the mother tongue and translation, noise and sound context, short and long pauses, were transcribed and the disfluencies (repetitions, revisions) and markers of oral speech (pause, laugh, etc.)8 were marked up. The corpus MATRICIEL is made up of the 17 transcribed and aligned life stories.

22Following is a text excerpt successfully transcribed and annotated in oral phenomena of the life story of JV17 witness (translated here into English):we went to Barcelona there <revision> there was </revision> they had requisitioned because <repetition> it is </repetition> it was the French government that paid for the trip <mdd> huh </mdd> there they requisitioned hotels [pause] and we went I don’t remember very well if it was to the Ritz or <repetition> to the Regina </repetition> to the Regina <mdd> I think </mdd> next to the Plaza de Catalunya and then we stayed a few months [pause]”.

23Table 2 shows the numerical description of the whole corpus MATRICIEL and of three interviews that were used to illustrate the results obtained by the MATRICIEL pipeline (witnesses JV17, PP13 and VP12).

Table 2 – Numerical description of the corpus, with details about three life stories.

Interview

Duration

# tokens

VP12

01:52:46

17 142

PP13

00:10:54

1 500

JV17

00:52:24

10 808

Total

18:30:18

179 729

Annotation of place names and sentiment words

24Concerning place names, two main types are distinguished according to linguistic and geographical features, those designated as Npr, e.g. Argèles-sur-Mer and those based on Nc, e.g. camp de concentration (concentration camp) (see the State of the art section for detailed definitions).

25For single and multi-word expressions connoted with sentiment, we associated them with a polarity. Two modalities are used: positive or negative.

  • 9 https://gate.ac.uk/
  • 10 In the context of NLP, a pipeline is a sequence processing resources aiming at solving a given task

26Based on these criteria, transcribed stories were marked up in place names and sentiment words in a semi-automatic manner relying on two main approaches which are described in the following. We implemented them into the NLP software toolkit, GATE9 (Cunningham et al., 2013) as the MATRICIEL pipeline10. The entire pipeline is reusable to any other textual corpus in French, describing life experiences.

Relevance of automatic annotation methods

27As our approach is to annotate place names and sentiments with automatic methods, we computed inter-annotator agreement (IAA) on an excerpt of the corpus in order to state the relevance of automation for annotating the targeted pieces of information. A weak inter-annotator agreement would mean that human annotators cannot establish a consensual reference annotation; in that case, it would not be relevant to make an automatic annotation process since it would be impossible to legitimately assess it. 

28The agreement was calculated using the GATE tool IAA (Table 3), based on the annotations produced by two persons who separately annotated manually two entire life stories for the Nc (on VP12) and for the sentiment (JV17).

29Concerning the Nc, they were annotated using B-LOC and I-LOC tags for places, and the B-OBJ and I-OBJ tags for located objects relevant for migrants (e.g. means of transportation). The life story VP12 came up with the following precision, recall and F-measure, respectively, 0.89, 0.88 and 0.89, which is proof of strong agreement. We were able to identify and correct a few disagreements such as different choices of the annotation scope (e.g. café vs. big café); an annotator missed several annotations (e.g. in ils sont venus sulfater la vigne (they come to sulfate the vine), vigne (vine) referred to the Nc of place or the plant in this context); also the presence of a negation or hypothesis (designated by a modality) can also generate annotation difficulties.

30Sentiment annotation is based on specific words and embeds a polarity. Concerning the life story JV17, values of recall and precision for sentiment are respectively, 0.72, 0.73 and 0.73, which are smaller than the values for the Nc on VP12. This is mainly due to interpretation that may be different, especially concerning the importance of some words e.g. hiver (winter), relâché (released), or of markers in oral speech indicating a polarity, e.g. enfin.

Table 3 – Inter-annotator agreement for manual annotation, Nc in life story VP12 and sentiment in JV17.

IAA

Nc

Sentiment

R

P

F-measure

R

P

F-measure

VP12

0.88

0.89

0.89

JV17

0.72

0.73

0.73

31The results of IAA are good enough to allow us to proceed to automatic approaches for annotations.

32No IAA has been performed for the Npr. Indeed, disagreements about Npr place names is often based on spelling (uppercase, separators); since the corpus comes from oral interviews, the potential spelling inconsistencies have been processed by the transcription guide used in transcription. As expected, human annotators broadly (almost 1 when the rules of the annotation guide are strictly followed) recognized the same normalized forms of Npr.

Gazetteer- and lexicon-based annotation of Npr of places and sentiment words

33For finding Npr of places and sentiment words in the texts, we performed an exact string match automatic operation from lists of word forms we gathered in different ways.

34For Npr of places, these lists are referred to as gazetteers where each entry corresponds to a place name along with its geographical coordinates and geographical or administrative category (e.g. city, region, country, river, etc.). We relied basically on two geographical data sources to create the gazetteers, BD Nyme and GeoNames. We adapted this last one for limiting the geographical coverage to France, Spain and the World’s most populated cities. This limits place reference ambiguity issues, i.e. having too many entries to choose from, for a place name. We loaded these gazetteers into the ANNIE plugin within GATE.

35The constitution of a sentiment lexicon is a further complicated task. Indeed, there exist a few sentiment lexicons in French (see the State of the art section). We chose EMOTAIX (Piolat and Bannour 2009) as the basis for creating a sentiment lexicon adapted to the MATRICIEL corpus for the following reasons: SentiWordNet, aFinn and FEEL are translated from English and we noticed that even with human verification many possible translations were missing, EMOTAIX was instead specifically built for the French language; the context in which EMOTAIX is used, i.e. psychological studies based on interviews of people describing their life experiences, is somehow related to MATRICIEL.

  • 11 https://unitexgramlab.org/fr

36In order to further adapt the EMOTAIX sentiment lexicon to the MATRICIEL context, we firstly added inflected forms of words in EMOTAIX with the help of the UNITEX tool11 (Paumier, 2003) and then studied a sample of the life stories. At this point, we observed several ambiguity problems that we had to deal with, for instance, the polysemic nature of the form restaurant. In EMOTAIX, this form corresponds to the present participle of the verb restaurer (to restore) which has a positive polarity but in one of the texts of MATRICIEL, corresponds to the common noun restaurant and should not be considered as a sentiment word in this context. Another example of ambiguity is the adjective réservé (shy or booked, given the context). In the context of EMOTAIX, this form refers to someone shy and therefore associated to a negative polarity. However, in a more general context and in the MATRICIEL context, it corresponds more often to the past participle of the verb réserver (to book) and should not thus be considered as a sentiment word. For these reasons, we had to manually separate ambiguous and unambiguous terms in regards with their supposed meanings in the MATRICIEL context. Such task requires a good knowledge of the French language and the polysemy of terms, and resulted into the creation of a sentiment lexicon of unambiguous terms for MATRICIEL. We added up a small lexicon containing words and expressions suitable for the MATRICIEL corpus according to the studied sample, in particular a war related lexicon.

37The final sentiment lexicon contains around 34,773 single and multi-word entries including 23,851 negative polarity entries (e.g. translated in English feeling of injustice, tightness of heart, abandoned, slaughtered) and 10,922 positive polarity entries (e.g. inner peace, no problem, having soul, fascinating, entertaining). We then integrated this lexicon into the CLaC plugin (Özdemir & Bergler, 2015) in GATE. This plugin is being developed by the Computational Linguistics Laboratory in Concordia University in the context of the research project “Des récits à la carte” about methodologies, technologies and sensitive approaches for mapping the life stories of Rwandan refugees12. The CLaC plugin aims to identify specific words as negation, modality verbs and combine them to validate or reverse the emotion word polarity.

38Finally, the customized ANNIE plugin including the MATRICIEL gazetteers for identifying Npr of place, as well as the CLaC plugin with the MATRICIEL sentiment lexicon for recognizing sentiment words, are incorporated in the MATRICIEL pipeline. 

CRF-based annotation of Nc of places

39CRF provides good models for English which were trained on the CoNLL 2003 English training data for detecting persons, locations and organizations

  • 13 https://nlp.stanford.edu/software/CRF-NER.shtml

40For automatically identifying Nc of place in the MATRICIEL corpus, we opted for a supervised machine learning approach based on CRF which requires a trained model built upon a manually annotated corpus. Since very few models are available for the French language and none of them are well adapted to the context of MATRICIEL, we trained our own CRF using the Stanford NER tool13.

41After obtaining an acceptable inter-annotator agreement (see section 3.3.1), we were able to manually annotate a sample of the MATRICIEL corpus as training corpus. For creating the training corpus, it is necessary to choose a representative sample of corpora in which Nc place names look like the MATRICIEL corpus ones. The interviews tell peculiar stories based on varied situations: journey in many cities, life in the refugee camps, liberation of Bordeaux, etc., and the relevant Nc place names are varied because they correspond to the situations. So the training corpus is made up of excerpts of 15 life stories (out of the 17), and each excerpt only contains the two-thirds of a life story. During training, we experimented the influence of both size and relevancy of the training corpus in the place name annotation quality. Several place name annotated corpora were used (Brando et al., 2016). Lastly, a crowd-sourced Web corpus of people’s contributions related to urban planning issues in their city was added to the MATRICIEL excerpts. The place name annotation of this latter corpus was consistent with the guidelines we made to annotate the corpus MATRICIEL. The resulting training corpus counts 150,478 word forms.

42During training, there are several parameters which in combination can help to produce the best model; these are typically the right context (the number of words to the right of a given word), the left context (the number of words to the left of a given word), the consideration of morphosyntactic (part-of-speech) categories and lemmas, and the introduction of a specific lexicon. In this last manner, we integrated a small lexicon of head common nouns which names geographical and administrative categories, and can appear in the multi-word place names (e.g. ville in ville de Barcelone) built upon the GéOnto ontology (Mustière et al., 2009) - an ontology of the elements of the territory and its infrastructures - based on the French national mapping agency databases, with minor modifications. We experimented with many combinations of the parameters, and we obtained the best results (see Table 6), with the following parameters: no part-of-speech and lemma annotation, right and left contexts equal to 2, and the specific lexicon we produced.

43As a final step, we integrated the new model into the MATRICIEL pipeline in GATE through the Stanford NER plugin.

Evaluation and results

44The annotation of places and sentiments employs automatic methods and must be assessed through comparison with manual annotations. It enables the production of maps. The evaluation process is presented according to two methods of text recognition: matching with gazetteers (for Npr of place and for sentiment) in 4.1 and using a learning model (for Nc of place) in 4.2. 

Automatic Npr and sentiment annotation with gazetteers

Npr

45The evaluation was made through two life stories PP13 and JV17, corresponding to 64 minutes, 6 % of the duration of the corpus record, and 10,848 words. The recall measure was calculated integrating the annotation scope (start and end of each annotated expression) and the identifier of the entry found in the gazetteer. This is equal to 0.96 and 0.92 respectively to PP13 and JV17 (see Table 4). The Npr that have not been found by the automatic pipeline are due to the following listed reasons (which are substantially the same as those concerning the entire MATRICIEL corpus). Examples are presented as with the Npr in the transcription not found with gazetteer and the corresponding Npr in the gazetteer:

  • referential ambiguity, e.g. San Juan can refer to several places of the same administrative category (here municipality) in Spain;

  • geographical category ambiguity, e.g. Lot can designate the river or an administrative area;

  • abbreviations with few ambiguity, e.g. Argelès for Argelès-sur-mer, Barcarès for Le Barcarès, Saint-Seurin for Saint-Seurin-sur-l’Isle;

  • abbreviations with ambiguity that can mean different places, e.g. Villeneuve can be Villeneuve-sur-Yonne or Villeneuve-sur-Lot; Sierra corresponds to Sierra Morena in context, but it could be others;

  • removal of the head of the noun which indicates the geographical type of the Npr, e.g. El Musel is port du Musel (in the city of Gijon), Niel means caserne de Niel in Bordeaux;

  • old official names, e.g. Castelnaud-Fayrac for Castelnaud-la-Chapelle;

  • old administrative entities, that have been fused e.g. Cosne-sur-Loire is now Cosne-Cours-sur-Loire;

  • too weak granularity, e.g. Amérique is less precise than Northern America or Southern America which correspond to gazetteer entries;

  • absence from gazetteer as vernacular names, e.g. Périgord;

  • uncommon spelling in entries of the gazetteer, e.g. Lac St André for Lac de Saint-André in Savoie;

  • Npr transformed in adjective (adjectivization), e.g. côte Cantabrique for côte de la province de Cantabrie;

  • spelling mistakes in the transcription process, e.g. Echourgnac for Échourgnac.

Sentiment

46Words of sentiments were recognized by the CLaC plugin using the dedicated gazetteer containing words and their polarity values. The manual annotation was guided by the characteristics of automatic annotation using gazetteers, meaning that the annotator focused on words and not on phrases or whole sentences that may also contain the context (which is explanatory and specific, and therefore difficult to reproduce). The evaluation was based on the stories PP13 and JV17 (see Table 4).

Table 4 – Results include assessment of sentiment automatic annotations through precision, recall, and F-measure.

Sentiment

#forms (reference)

#forms (automatic)

R

P

F-measure

PP13

64

8

0.13

1

0.22

JV17

228

86

0.30

0.80

0.44

47The numbers of annotated sentiment words are respectively 64 (16 with positive polarity and 48 with negative polarity) in comparison with 8 (2 and 6) automatic annotations, and 228 (60 and 168) in comparison with 86 (40 and 46) automatic annotations. Recall is 0.13 and 0.3. Precision is 1 and 0.8 (see Table 4). Those values were calculated by taking into account the scope of the annotations and the assigned polarity.

48The precision values are quite high for both life stories. We can conclude that words of sentiment contained in the gazetteer are relevant and associated with the right polarity. In contrast, the recall values are small. Several explanations exist and perspectives are formulated as follows to improve the approach:

  • integrating into the MATRICIEL gazetteer, existing thematic gazetteers or specific vocabulary, in particular the WW2 and the Spanish Civil War vocabulary;

  • taking into account the negation: negation markers can be automatically annotated by the CLaC plugin; the markers should be combined with the annotations about sentiment and interpreted in order to modify the polarity;

  • dealing with interrupted multi-word expressions that do not match with the gazetteer entries and generate false negatives, e.g. prendre bien soin (to take great carte) is not recognized as the common expression is prendre soin (to take carte);

  • identifying elements of context that may bring a polarity to factual elements: the context can be expressed nearby by an explicit sentence before or after the element to annotate, e.g. on avait faim (we were hungry) leads to associate a positive polarity to il connaissait beaucoup les herbes (he knew a lot about herbs); the context is also general, during the Spanish Civil War and WW2 e.g. une petite qui parle français (a little girl that speaks French) can be annotated as positive during the Retirada.

Maps

  • 14 https://qgis.org/en/site/

49One of our main objectives is to create cartographic representations from the MATRICIEL corpus annotated with our pipeline. For this reason, it was necessary, as previously explained, to annotate Npr of place relying on a gazetteer listing coordinates and categories of a given place. The maps aim to propose synthetic views of the locations in the life stories, number of occurrences, large-scale map of the places mentioned in a particular city. We also used sentiment words for creating thematic maps and presenting information about the occurrences of the polarities associated to Npr of place, the contrast between the polarities. These maps bring research questions about spatial distributions of features regarding the migration of Spanish Republicans. They were created using QGIS14 an open-source and very popular geographic information system (GIS) and are presented at the end of the next section.

50Annotating Npr with associated coordinates and features enables to produce maps which give synthetic views of the locations in the life stories and can bring questions about spatial distributions.

51A manual annotation was done in order to have a Npr reference annotation for Npr (the results of automatic annotation are not shown because the process is incomplete: referential disambiguation remains manual). Numbers of occurrences concerning the reference annotation are:

  • 2,787 occurrences for all annotations, minimum of 48 and maximum of 543 for one life story among the 17 ones ;

  • 351 occurrences for unique values.

52Regarding the life stories, reference annotation highlights some Npr. Logically, Espagne and France are mentioned in all the life stories and have the most occurrences (378 and 224 respectively). The five Npr that are mentioned in a large number of life stories are Allemagne in 14 life stories, Bordeaux in 13, Madrid in 13, Barcelone in 12, Argelès-sur-Mer in 11. Although Allemagne is mentioned in most of the life stories, the number of occurrences stays quite low, equal to 34.

53The type of the 351 different quoted places, regardless of the multiple quotes, are distributed as follows:

  • country and entity at regional level (administrative or not): 55

  • intermediate entity (administrative or not): 18

  • municipality: 255

  • site: 6

  • hydrography feature: 9

  • relief feature: 8

54Municipality names are widely predominant in the Npr mentions. They correspond to the lowest geographical level in the MATRICIEL gazetteer. We can also assume that municipalities represent a common located reference for most people, structuring their geographical space.

55We mapped the Npr places in order to have an overview of the spatial content in single life stories and in the corpus. This kind of representation is more a visualization of the content than a map constructed with a message to spread. We tend to stay neutral and so exhaustive in the represented places. Figure 1 displays a heatmap of the mentioned Npr in the entire corpus. An heatmap calculates areas of graduated intensity (here the spatial density) by interpolation from punctual layout (here Npr). In order to have a coherent granularity of maps, only Npr corresponding to municipalities were selected (n = 255 for the entire corpus). The chosen extent for the map is the western part of Europe which corresponds to the a priori geographical space of the life stories. 

Figure 1 – Heatmap of the Npr corresponding to municipalities quoted in the corpus. Displayed toponyms are the Npr with more than 20 occurrences

Figure 1 – Heatmap of the Npr corresponding to municipalities quoted in the corpus. Displayed toponyms are the Npr with more than 20 occurrences

56Mapping the spatial density of Npr may support first interpretations about the geography of the life stories (which can then help supporting questions about this historical period). In Figure 1, several geographical aspects are highlighted. One is the spatial distribution of the Npr municipalities, with a broad majority in France and in Spain, a few in other countries (and only two outside the extent of the map: Stalingrad and Saint-Domingue). The second aspect deals with identifying high densities areas. For instance, these areas are: the surroundings of Bordeaux which is the main location of the interviews; area inside and nearby the Lot-et-Garonne department which contains living places of the interviewees after arriving in France and also Resistance places; the French-Spanish border on the Mediterranean and the Atlantic sides which were the main crossing points.

57A manual reference on sentiment propagated to Npr place names was built in order to have an overview of the polarities in the life stories. Characterizing polarity on places has the objective to potentially bring out similarity or difference on spatial configurations of polarities. A configuration can correspond to analyzing the polarities on one place. Another configuration can be studying polarities in spatially nearby places. The map in Figure 2 displays the absolute differences in polarities, negative or positive. Places such as Espagne, Ebre, Bordeaux and Allemagne are often mentioned, and correspond to great differences with a majority of negative polarities. Fewer differences have been observed for Npr with a majority of positive polarities, like Libourne, Châtellerault. Consequently, it would be interesting to investigate the reasons of the polarity spatial distribution by mapping every life story separately. Besides, the chronology of the life story events (childhood, exile, integration) may also influence the spatial distribution of the polarities.

Figure 2 – Highlighting contrasts in polarities related to Npr. Npr have been associated with the major polarity, negative or positive, and the count of differences between those two polarities

Figure 2 – Highlighting contrasts in polarities related to Npr. Npr have been associated with the major polarity, negative or positive, and the count of differences between those two polarities

Annotation with the NER-trained model for Nc

58This validation corpus is made up of the excerpts of the life stories which are not in the training corpus, i.e. the last third of the 15 life stories, and two full life stories JV17 and PP13, which are not used in the training corpus (150,478 words). The customized NER plugin annotated the validation corpus according to the BIO format (see Table 1) and produced four types of annotation: B-LOC, I-LOC for place names and B-OBJ, I-OBJ for located objects. To assess the performance of the Nc place name annotation model, the automatic annotations must be compared, for each type, with the manual ones which are the reference.

59In order to better explain the sources of differences, the assessment is firstly focused on the stories PP13 and JV17, relying on recall and precision measures.

Table 5 – Results (BIO format, stories PP13 and JV17) include assessment of automatic Nc place name annotation through precision, recall, and F-measure. The I-OBJ assessment is not included because only four were annotated in the JV17 reference (and none in PP13); none of them was automatically recognized.

Nc

B-LOC

I-LOC

B-OBJ

R

P

F-measure

R

P

F-measure

R

P

F-measure

PP13

0.69

1

0.81

0.44

1

0.61

NA

NA

NA

JV17

0.53

0.85

0.66

0.28

0.67

0.39

0.33

1

0.5

60Recall ratio is always worse than precision. It means that some Nc of place names are not found by the automatic pipeline. 

61Concerning multi-word place names, partial annotation generates both false positives and false negatives, e.g. usine (B-LOC true positive) d’armements (I-LOC false negative). The place name: Couvent de Breton de la rue Calvé is a more complicated example. The model recognizes two place names: Couvent de Breton and rue Calvé. So, it makes one B-LOC false positive rue and two I-LOC false negatives: de la and rue Calvé.

62Focusing on the B-LOC, these differences are:

  • The training corpus is made up of excerpts (two thirds) of every interview, but the interviews do not have all the same length so the contribution of each interview in the annotation model is not the same. In addition, the Nc place names vary according to the topics of the stories. So the model is not equally effective in Nc place name;

  • The manual annotation considers that the expressions made on the pattern chez <personal pronoun>, for example, chez moi (at home) are name places. Nevertheless, the model does not annotate them. A post-treatment of the automatically annotated corpus or the use of a locative preposition lexicon upon the model making is mandatory to automatically identify them (2 expressions in PP13 and 17 in JV17).

63The process must be assessed on the entire MATRICIEL corpus. We saw that the issues concerning the mistagging between B-LOC and I-LOC are predominant in the assessment. In order to identify other sources of errors, B-LOC and I-LOC were aggregated into a single LOC tag. Lastly, we decided not to include the results concerning located objects (annotated as B-OBJ and I-OBJ) as there were too few in the validation corpus. Table 6 shows the results concerning the entire validation corpus, relying on recall, precision and F-measure.

Table 6 – Results include assessment of automatic Nc place name (LOC) annotation through precision, recall, and F-measure

LOC

R

P

F-measure

Validation corpus

0.72

0.63

0.67

64The recall is fairly good; many examples have been well recognized by the model. Thus, single word place names such as pont (bridge), prison, village, école (school), gare (station) or frontière (border) are well recognized. As far as multi-word entities are concerned, place names such as base sous-marine (submarine base), rue de la Course, fort du Hâ, place Tourny (Tourny square), camp d’Hennebont (Hennebont camp) or gare de Saint-Médard (Saint-Médard station) are also well known. We also noticed that ligne de démarcation du côté de Langon (demarcation line entity on the Langon side) was entirely detected as a LOC.

65Besides issues about confusion between B-LOC and I-LOC tags, other sources of errors exist. There are false negatives, e.g. arènes de Madrid (Madrid arenas), bourse du travail (labour exchange), cours de l’Yser (Yser Avenue) or hôtels (hotels). The main explanation is that these place names are not or only slightly present in the training corpus. Another one is that several geographical and administrative words are missing in the geographical and administrative head noun lexicon: arènes, bourse

66There are false positives, e.g. équipe d’Agonac (Agonac’s team) or liaisons téléphoniques (telephone links). The first error implements the syntaxical pattern <Noun> de Npr which matches geographical (lac d’Annecy, Lake Annecy) or administrative (ville de Barcelone, Barcelona city) place names. The second one is due to the head noun liaisons which concerns both means of transportation and means of communication.

Discussion and perspectives

67The approach implemented – definition of the themes to be analyzed, construction of the corpus, linguistic analysis, formulation of thematic research questions into questions on the corpus or on the cartographic representation of the corpus – could be reproduced in order to contribute to the multidisciplinary study on other research themes. Consequently, our contribution is methodological and concerns the automation of information extraction about places, in a textual corpus. We have experimented NLP and geomatics methods for topics and resources about a specific migration event.

68Annotation and its evaluation depend on the resources used in the automatic process (suited to specific issues, reusable, shared, etc.). Very rich resources, such as BD Nyme and GeoNames for place names, EMOTAIX for sentiment analysis, are available. Nevertheless, they must be modified in two opposite directions. Resources must be adapted and cut in order to reduce ambiguity in recognition of place names and sentiments. They must be enriched to better target the theme and take into account the context. For example, the war vocabulary should be added to the sentiment lexicon in order to report the emotional situation of the witnesses. The polarity of some words and expressions (sabotage, revolt, arrest, pistol, etc.) would systematically be negative, whatever outcome, purpose, desire for justice of the witnesses.

69In order to recognize the Nc place names, we made a model based on machine learning. The results are satisfactory but they can be improved. The usual way to deal with this is to extend the learning corpus and, consequently, to improve the quality of the model. The difficulty now is to have access to annotated corpora which are comparable to the MATRICIEL one according to the objects of study, the research objectives and the form of the interviews. The most challenging perspective is to locate Nc place names; an idea would be to link Nc places to Npr ones using lexical and syntactic clues. The Npr place names were identified and located through gazetteers; results are good (if disambiguation is completed). 

70Although the sentiment lexicon was reduced by deleting ambiguous terms, several automatic annotations are wrong. For example, gâtée (spoiled) is negatively connoted in the lexicon; in the corpus, it qualifies a little girl “spoiled” by the attentions of an adult who looked after her; in the context, excessive care cannot be negative for a refugee child, and it may even be annotated positively.

71In addition, the annotation process must be adapted to the context: the term “refugee” is connoted to negative ideas, but used in the life story the “refugee status” expression seems to designate a more stable situation and may be annotated as a positive.

72Finally, sentiment may be linked to a series of sentences which makes it difficult to precisely annotate the relevant word(s). For instance, a witness describes a series of places where her family had successively to move. In this case, the negative sentiment is due to the multiplication of moves which were mandatory because of the war; this idea is not supported by one or several word(s).

73The cartographic representations, through their synoptic property, have raised research questions on migration of Spanish Republicans. Indeed, maps (for the entire corpus) show the places, designated in the life stories, which were crossed or of stay;they enable to study the spatial distribution of the corpus place names with those emblematic for the Spanish Republicans such as the border, battles, the Maquis in the south-west of France. In other words, are the most frequent places the most significant ones during that period for the French Resistance or Spanish migration? Other questions concern the spatial dimension of the place features: for example, do identical polarity modalities (positive or negative) correspond to nearby locations?

74The MATRICIEL project aims to extract information about places and sentiments from the life story corpus and to turn it into knowledge about the Spanish Republicans’ exile, commitment and installation in France. The maps enable to synthesize pieces of information and to demonstrate the spatial dimension of these phenomena. With this approach, new perspectives might open up as we are able to convert thematic research issues into questions for corpora suitable for those themes, and to propose relevant maps based on the extracted information.

Haut de page

Bibliographie

ABDAOUI Amine et al., 2017, “Feel: a french expanded emotion lexicon”, Language Resources and Evaluation, vol. 51, n°3, p. 833-855.

ANDREEVSKAIA Alina, BERGLER Sabine, 2008, “When specialists and generalists work together: Overcoming domain dependence in sentiment tagging”, Proceedings of the 46th Annual Meeting of the Association for Computational Linguistics, Human Language Technology Conference, p. 290-298.

BACCIANELLA Stefano, ESULI Andrea, SEBASTIANI Fabrizio, 2010, “SentiWordNet 3.0: an enhanced lexical resource for sentiment analysis and opinion mining”, LREC, vol. 10, p. 2200-2204.

BENAMARA Farah et al, 2012, “How do negation and modality impact on opinions?”, Proceedings of the Workshop on Extra-Propositional Aspects of Meaning in Computational Linguistics, Association for Computational Linguistics, p. 10-18.

BRANDO Carmen, DOMINGUÈS Catherine, CAPEYRON Magali, 2016, “Evaluation of NER systems for the recognition of place mentions in French thematic corpora”, Proceedings of the 10th Workshop on Geographic Information Retrieval, Association for Computing Machinery.

BRUNET Roger, FERRA Robert, THÉRY Hervé, 1992, Les mots de la géographie, dictionnaire critique, Paris, Reclus-Documentation française, 518 p.

CALABRESE-STEINBERG Laura, 2009, “Nom propre et dénomination évènementielle : quelles différences en langue et en discours ?”, Corela. Cognition, représentation, langage, vol. 7, n°1.

CAQUARD Sébastien, CARTWRIGHT William, 2014, “Narrative cartography: From mapping stories to the narrative of maps and mapping”, The Cartographic Journal, vol. 51, n° 2, p. 101-106.

CAQUARD Sébastien, FISET Jean-Pierre, 2014, “How can we map stories? A cybercartographic application for narrative cartography”, Journal of Maps, vol. 10, n° 1, p. 18-25.

CHATAR-MOUMNI Nizha, 2013, “L’expression verbale des émotions: présentation”, Langue française, n°4, p. 3-11.

CUNNINGHAM Hamish et al, 2013, “Getting more out of biomedical documents with GATE’s full lifecycle open source text analytics”, PLoS computational biology, vol. 9, n° 2, e1002854, https://doi.org/10.1371/journal.pcbi.1002854

DOMINGUÈS Catherine, 2018, “Lieu et nom de lieu, du texte vers sa représentation cartographique”, Proceedings of the 25e conférence sur le Traitement Automatique des Langues Naturelles, p. 519-528.

EKMAN Paul, 1992, “An argument for basic emotions”, Cognition & emotion, vol. 6, n°3-4, p. 169-200.

ESULI Andrea, SEBASTINI Fabrizio, 2006, “SentiWordNet: A publicly available lexical resource for opinion mining”, LREC, vol. 6, p. 417-422.

GRIFFIN Amy, McQUOID Julia, 2011,“At the intersection of maps and emotion: The challenge of spatially representing experience”, Kartographische Nachrichten, 2012, vol. 62, n°6, p. 291-299.GROUIN Cyril, et al., “Proposal for an extension of traditional named entities: From guidelines to evaluation, an overview”, Proceedings of the 5th  linguistic annotation workshop, Association for Computational Linguistics, p. 92-100.

HIRT Irène, 2009, “Cartographies autochtones. Éléments pour une analyse critique”, L’Espace géographique, vol. 38, n°2, p. 171-186.

JONASSON Kerstin, 1994, Le nom propre. Constructions et interprétations, Louvain-la-Neuve, Belgique, De Boeck Supérieur, « Champs linguistiques », 258 p.

MEKDJIAN Sarah, 2016, “Les récits migratoires sont-ils encore possibles dans le domaine des Refugee Studies ? Analyse critique et expérimentation de cartographies créatives”, ACME: An International E-Journal for Critical Geographies, vol. 15, n°1, p. 150-186, Retrieved from https://www.acme-journal.org/index.php/acme/article/view/1211/1168 [Accessed 2019.09.25]

MEKDJIAN Sarah, AMILHAT SZARY Anne-Laure, 2015, “Cartographies traverses, des espaces où l’on ne finit jamais d’arriver”, halshs-01242532.

MELO Fernando, MARTINS Bruno, 2017, “Automated geocoding of textual documents: A survey of current approaches”, Transactions in GIS, vol. 21, n°1, p. 3-38.

MILLER George A., 1995, “WordNet: a lexical database for English”, Communications of the ACM, vol. 38, n°11, p. 39-41.

MORENO Jose G. et al., 2017,“Combining word and entity embeddings for entity linking”, European Semantic Web Conference, p.337-352.

Le Guide de la Galerie des dons, 2014, Paris, Musée national de l’histoire de l’immigration éd., 256 p.

MUSTIÈRE Sébastien et al., 2009, “GéOnto: Enrichissement d’une taxonomie de concepts topographiques”, Spatial Analysis and GEOmatics SAGEO 2009, inria-00432628.NIELSEN Finn Årup, “Afinn”, Richard Petersens Plads, Building, 2011, vol. 321.

NOUVEL Damien, EHRMANN Maud, ROSSET Sophie, London, ISTE Group, 2015, « Sciences cognitives », 168 p.

OLMEDO Élise, 2016, “Pour une cartographie affective des récits des femmes de Sidi Youssef Ben Ali (Marrakech, Maroc)”, dans Fournier M (sous la dir. de), Cartographier les récits, Clermont-Ferrand, Presses universitaires Blaise Pascal, Collection CERAMAC, p. 179-196.

ÖZDEMIR Canberk, BERGLER Sabine, 2015, “Clac-sentipipe: Semeval2015 subtasks 10 b, e, and task 11”, Proceedings of the 9th International Workshop on Semantic Evaluation SemEval Association for Computational Linguistics, p. 479-485, Retrieved from https://www.aclweb.org/anthology/S15-2081 [Accessed 2019.09.25].

PAUMIER Sébastien, 2003, De la reconnaissance de formes linguistiques à l’analyse syntaxique. Volume 2, Manuel d’Unitex, Thèse de doctorat, IGM, Université de Marne-la-Vallée.

PIOLAT Annie, BANNOUR Rachid, 2009, “EMOTAIX: un scénario de Tropes pour l’identification automatisée du lexique émotionnel et affectif”, L’Année psychologique, vol. 109, n°4, p. 655-698.

PLUTCHIK Robert, 1991, The emotions, Lanham, Maryland, USA, University Press of America, 216 p.

RABIE Joseph, 2003, Ce qui fait lieu – Vers une éthique chorographique, Thèse de doctorat, École doctorale Ville, Transports et Territoires, Université de Marne-la-Vallée.

SANTOS Rui, MURRIETA-FLORES Patricia, MARTINS Bruno, 2017, “An Automated Approach for Geocoding Tabular Itineraries”, Proceedings of the 10th Workshop on Geographic Information Retrieval. Association for Computing Machinery.

SCHERER Klaus R. et al., 2004, “Emotions in everyday life: Probability of occurrence, risk factors, appraisal and reaction patterns”, Social Science Information, vol. 43, n°4, p. 499-570.

TUAN Yi-Fu, 1977, Space and place: The perspective of experience, Minneapolis, Minnesota, USA, University of Minnesota Press, 248 p.

Haut de page

Notes

1 https://ich.unesco.org/fr/convention

2 « la matière est généralement divisée selon la géographie »

3 www.geonames.org

4 For illustrated definitions, see https://en.wikipedia.org/wiki/F1_score

5 http://eslo.huma-num.fr/images/eslo/pdf/GUIDE_TRANSCRIPTEUR_V4_mai2013.pdf

6 http://www.afcp-parole.org/camp_eval_systemes_transcription/docs/Conventions_Transcription_ESTER2_v01.pdf

7 http://dali.talkbank.org/clan/

8 The tag is <mdd>…</mdd>.

9 https://gate.ac.uk/

10 In the context of NLP, a pipeline is a sequence processing resources aiming at solving a given task.

11 https://unitexgramlab.org/fr

12 http://storytelling.concordia.ca/content/newly-funded-projects-partnership-page-rwanda

13 https://nlp.stanford.edu/software/CRF-NER.shtml

14 https://qgis.org/en/site/

Haut de page

Table des illustrations

Titre Figure 1 – Heatmap of the Npr corresponding to municipalities quoted in the corpus. Displayed toponyms are the Npr with more than 20 occurrences
URL http://journals.openedition.org/rfsic/docannexe/image/7228/img-1.png
Fichier image/png, 307k
Titre Figure 2 – Highlighting contrasts in polarities related to Npr. Npr have been associated with the major polarity, negative or positive, and the count of differences between those two polarities
URL http://journals.openedition.org/rfsic/docannexe/image/7228/img-2.png
Fichier image/png, 120k
Haut de page

Pour citer cet article

Référence électronique

Catherine Dominguès, Laurence Jolivet, Carmen Brando et Marion Cargill, « Place and Sentiment-based Life story Analysis »Revue française des sciences de l’information et de la communication [En ligne], 17 | 2019, mis en ligne le 01 septembre 2019, consulté le 18 avril 2024. URL : http://journals.openedition.org/rfsic/7228 ; DOI : https://doi.org/10.4000/rfsic.7228

Haut de page

Auteurs

Catherine Dominguès

Institut national de l’information géographique et forestière, Laboratoire des sciences et technologies de l’information géographique (Saint-Mandé). Courriel : Catherine.Domingues@ign.fr

Laurence Jolivet

Institut national de l’information géographique et forestière, Laboratoire des sciences et technologies de l’information géographique (Saint-Mandé). Courriel : Laurence.Jolivet@ign.fr

Carmen Brando

École des hautes études en sciences sociales, Centre de recherches historiques (Paris). Courriel : carmen.brando@gmail.com

Marion Cargill

Institut national de l’information géographique et forestière, Laboratoire des sciences et technologies de l’information géographique (Saint-Mandé). Courriel : serge.weber@u-pem.fr

Haut de page

Droits d’auteur

CC-BY-NC-SA-4.0

Le texte seul est utilisable sous licence CC BY-NC-SA 4.0. Les autres éléments (illustrations, fichiers annexes importés) sont « Tous droits réservés », sauf mention contraire.

Haut de page
Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search