Mapping Migration Discourse: The FR-MIGR-TWIT Corpus 2.0

(1.1) Corpus rationale

The rationale underlying FR-MIGR-TWIT is therefore twofold. On the one hand, the corpus is designed from a discourse-analytic perspective, in that it documents migration-related political discourse in a contemporary CMC environment marked by rapid circulation and platform-specific forms of textual organization (hashtagged word sequences, emojis, multimedia-linked textuality, relation between actors, etc.). (see Paveau, 2013; Longhi, 2013, 2016, among others). In this respect, the FR-MIGR-TWIT corpus contributes to updating the empirical basis of the field by integrating a new type of linguistic material drawn from CMC.

On the other hand, the annotated corpus structure makes it possible to examine how recurrent migr- lexical forms may display different frequency patterns, syntactic distributions, semantic role assignments, modification patterns, as well as their involvement in list and parallelism structures across political orientations and over time, thereby capturing both ideological differentiation and short-term diachronic evolution.

(1.2) Corpus design

The corpus contains 17,397 tweets posted or reposted by 39 Twitter/X accounts associated with French political actors and parties. It focuses on their (re)posts between 1 January 2011 and 30 June 2022 that contain at least one occurrence of a lexical form derived from the root migr-, including both neologisms such as immigrationniste (‘immigrationist’) and non-spaced word sequences containing migr- derivatives. These linguistic forms, yielding approximately 20,000 occurrences, constitute the core annotation units and are referred to as the migr-lexicon. Table 1 presents the 10 most frequent items in the migr-lexicon across the corpus.

The selection of accounts was based on explicit criteria aimed at ensuring political representativeness, breadth across the political spectrum, and sufficient discursive productivity, both on the CMC platform itself through (re)posting activity and in terms of media visibility: (i) participation in presidential elections; (ii) representation at the European level, that is, in the European Parliament; (iii) sufficient discursive productivity on Twitter/X, both in general and with respect to (re)posts containing migr-lexicon (henceforth migr-tweets), during at least part of the period between 2011 and 2022; and (iv) affiliation with structured political parties. The first two criteria were intended in particular to ensure political representativeness and media visibility. Regarding the discursive productivity of migr-tweets, their proportion among all (re)tweets posted by a given account was taken into consideration. For the selected accounts, the total number of all (re)tweets was retrieved using the Twitter API v2 Full-archive counts endpoint (Jeon, 2023, 2025a, 2025b).

Based on the traditional French left–right political spectrum, as well as on the labels under which the affiliated parties of political actors are themselves positioned, two subcorpora were constructed, which together enable systematic comparison across political orientations:

• FR-R-MIGR-TWIT groups together actors and parties situated on the center-right to far-right continuum (n = 16) and comprises 11,761 tweets.

• FR-L-MIGR-TWIT groups together actors and parties situated on the left to far-left continuum (n = 23) and comprises 5,636 tweets.

Table 2 presents the Twitter/X accounts associated with the actors and parties included in FR-R-MIGR-TWIT, while Table 3 presents those included in FR-L-MIGR-TWIT. The representative type of each account and the proportion of migr-tweets are also provided. The FR-R/FR-L distinction should be understood as a deliberately coarse-grained analytical partition intended to capture broad tendencies across the French left–right political divide. The aim is to identify and quantify systematic tendencies between these two macro-orientations. The methodological implications and limitations of this partition are discussed further in section (4).

(1.3) Annotation structure and theoretical grounding

The annotation structure of FR-MIGR-TWIT is built around the migr-lexicon, which serves as the central annotation unit in the corpus. The annotation layers were established to operationalize theoretically motivated dimensions of variation in migration-related political discourse. The main analytical dimensions of the annotation, conducted manually in ANALEC (see section (3.2)), are summarized in Table 4.

The model combines (i) syntactic annotation adapted from the Rhapsodie project (Lacheret et al., 2014; Lacheret et al., 2019), (ii) semantic role annotation based on a reconfigured inventory adapted from Jurafsky and Martin (2019), Tellier (2004), and Saeed (2016), and (iii) annotation of modification, as well as list and parallelism structures, which are relevant for analysing categorization and discursive framing in political discourse.

Repository location

https://doi.org/10.5281/zenodo.17828433

The dataset is also archived in Ortolang (https://www.ortolang.fr/market/corpora/fr-migr-twit-corpus-20).

Repository name

FR-MIGR-TWIT Corpus 2.0

Object name

FR-MIGR-TWIT_2.0.zip

Format names and versions

The corpus is distributed in two complementary versions. A fully annotated version, including the complete annotation layers and all associated metadata, is provided in CSV (UTF-8) and XML formats. These files are named FRMIGRTWIT_v2.csv and FRMIGRTWIT_v2.xml, respectively. A standalone corpus-text version containing essential tweet-level metadata, such as author username and identifier, tweet identifier, and tweet URL, is provided in TEI XML format within the tei/subfolder and is distributed as yearly files covering 2011–2022, following the filename pattern FRMIGRTWIT_YYYY_text.tei.xml, where YYYY denotes the year. The repository also includes a Python script, query_frmigrtwit.py, which takes FRMIGRTWIT_v2.csv as input and enables the generation of concordances from simple queries (Jeon & Pietrandrea, 2025).

Creation dates

The FR-R module was created between 2021-09-01 and 2022-11-01, including the completion of truncated retweets. The FR-L module was created between 2023-03-01 and 2023-07-31, likewise including the completion of truncated retweets. These two modules were subsequently compiled into FR-MIGR-TWIT 2.0 and underwent manual annotation until 2024-05-31, followed by an inter-annotator agreement phase completed on 2025-04-30.

Dataset creators

Sangwan Jeon, Paola Pietrandrea, Elena Battaglia.

Dataset annotator

Sangwan Jeon

Language

French-language data; variable names in English (e.g. data__text, data__created_at, data__author_id, etc.)

License

CC BY-NC-SA 4.0.

Publication date

2025-12-18.

(3.1) Data collection and corpus construction

Data collection was conducted in three successive steps. First, tweets containing at least one occurrence of migr- derivatives were collected from the 39 accounts selected according to the criteria described in subsection (1.2). This initial phase aimed to retrieve the textual content of relevant tweets together with their associated metadata. To this end, data were collected through the Full-archive search endpoint of the Twitter API v2 (Academic Research track). To facilitate collaborative querying and the reuse of request settings, API calls were executed through the Postman interface. Table 5 presents the lexical query terms used in the Full-archive search endpoint, and Table 6 summarizes the tweets.fields parameters retrieved for corpus construction and metadata archiving.

The second phase relied on the Full-archive Tweet counts endpoint and aimed to record the overall volume of tweets and retweets published by the selected accounts. This made it possible to operationalize the discursive productivity of (re)tweets containing migr- derivatives by measuring their proportion relative to the total posting activity of each account. The third phase consisted in manually completing truncated retweets by tracking the URL of the original post retweeted. Owing to technical limitations in the retrieved retweet text, some retweets exceeding 140 characters were returned in truncated form, restricted to the first 140 characters. These retweets were therefore manually completed in order to restore the full textual content of the post.

(3.2) Annotation model and implementation

Annotation was performed manually using ANALEC, a linguistic annotation and analysis tool developed at the CNRS research unit Lattice (Landragin, Poibeau & Victorri, 2012), which made it possible to implement a fine-grained multi-layer annotation scheme. The analytical dimensions introduced in subsection (1.3) are implemented in the annotation scheme through the field names summarized in Table 7. Figure 1 displays the annotation structures implemented in ANALEC, including the drop-down menus used to assign syntactic function and semantic role labels.

Figure 1

In this model, the central annotation unit is the migr-lexicon occurrence. Each annotated occurrence is annotated through a set of fields corresponding to these dimensions, including lexical identification, lemmatization, syntactic function, semantic role, modification, list and parallelism structures, and tweet-level metadata (i.e., rt, username, date). In addition to the intrinsic properties of the migr-lexicon occurrence, the annotation also includes information linked to structurally related units, such as list or parallelism structures (list_parallelism), conjuncts (conjunct), modifier lemmas (lemma_modif), or the lemma of a modified noun (lemma_noun-1), as well as tweet-level metadata manually integrated into the annotation files. As an illustration, Figure 2 presents the annotated occurrence migr-lexicon-9075 in ANALEC interface, and Figure 3 illustrates the corresponding XML encoding.

Figure 2

Figure 3

(3.3) Annotation campaign and inter-annotator agreement

The annotation campaign was carried out by trained annotators (final-year undergraduate students in linguistics at Université de Lille) following a shared protocol, a training phase, and successive calibration steps. Inter-annotator agreement was evaluated on a subset of the corpus against a reference annotation, and Table 8 reports the comparison with annotator A for illustration. This reference-based approach was preferred to pairwise comparison between annotators because the aim was to assess the extent to which trainee annotations converged with an expert annotation already established and used in the research underlying the corpus (Jeon, 2025a). To ensure feasibility, the annotators each worked on a randomly selected sample of tweets, preceded by a training set. Agreement was calculated by aligning annotated migr-lexicon occurrences with the reference annotation and comparing the annotated properties using exact agreement and Cohen’s kappa. Overall, agreement reached satisfactory levels for lexical and morphosyntactic layers, while semantic role annotation yielded lower but still interpretable scores. Disagreements were analyzed and used to refine the annotation guidelines.

(4.1) Subcorpus design and keyword selection

The corpus design distinguishes two political subcorpora, FR-R and FR-L, in order to enable comparison between contrasting communities of practice. One methodological challenge, however, lies in the quantitative imbalance between these two subcorpora, with right-wing and far-right actors contributing to a substantially larger number of migration-related tweets and migr-lexicon occurrences than left-wing and far-left actors. This imbalance partly reflects a corpus design constraint, but it also corresponds to a meaningful empirical property of the dataset: migration is not equally salient across political communities, and the right-wing continuum shows a stronger and more sustained investment in migration-related references over the period considered. In this sense, the asymmetry should be interpreted as a characteristic of the political discourse under study. These differences are also supported by statistical testing conducted on the distribution of migr-lexicon occurrences, as well as on the proportion of tweets containing migr-lexicon occurrences (Jeon, 2025a, 2025b).

A further point which was raised during the review process was that the disparity between left- and right-wing discourse might stem from differences in lexical choice, and it was suggested that the feasibility of including migration-related terms not derived from the migr-lexicon, such as demandeur d’asile ‘asylum seekers’ or réfugiés ‘refugees’, be examined. Indeed, these terms were tested during the collection phase, during which preliminary observations suggest that this disparity cannot be explained solely by the use of alternative migration-related lexical items, notably because, across the 39 selected accounts, tweets containing these terms without any co-occurring migr- derivative remained rare in the French context. In other words, non-migr- migration-related terms frequently co-occurred with migr-lexicon occurrences rather than functioning as exclusive alternatives to them. These observations contributed to centering the corpus on the migr-lexicon, at least in the French context. Interestingly, this pattern differs from observations made in the UK context, where non-migr- migration-related terminology appears more autonomous.

Another remark concerned the representativeness of the selected political accounts. In particular, some major French center-right political parties are not represented through their official accounts, notably @lesRepublicains and @Renaissance, representing respectively Les Républicains (formerly UMP) and Renaissance (inherited from En Marche).

It should be noted that the construction of FR-R corpus, initially developed as a pilot corpus of political migr-tweets, followed an iterative process. During the initial account selection stage, migration-related discourse was found to be particularly concentrated among individual accounts associated with, or originating from, the Rassemblement National (formerly Front National), including @J_Bardella, @MLP_officiel, @MarionMarechal, @NicolasBay_, and @f_philippot. Frequent retweeting practices between these accounts and the organizational account @RNational_off were also observed, which lead to the inclusion of the latter in the corpus. This decision is further supported by the proportion rates reported in the subsequent analysis:² @RNational_off displayed a substantially higher proportion of migr-tweets (8.13%) (see Table 2) than @lesRepublicains (.66%; 505/76,800) and @Renaissance (.20%; 44/22,200).

Nevertheless, several accounts presenting lower proportion rates were also retained in order to capture variation across political affiliations. Given the macro-comparative scope of the present study, the absence of these accounts is unlikely to substantially affect the overall analytical tendencies observed in the corpus. Future research could nevertheless integrate migr-tweets from @lesRepublicains and @Renaissance, particularly in the context of account-level quantitative and qualitative analyses.

(4.2) Political categorization

Another methodological challenge concerns the political positioning of certain political actors whose placement on the left–right axis is not always stable or self-evident. This is particularly the case for actors such as Emmanuel Macron, who has long claimed to stand outside the traditional left–right divide. It should therefore be emphasized that, in the present study, the FR-R/FR-L distinction is not intended as a definitive ideological classification of political actors. Rather, it functions as a deliberately coarse-grained analytical partition designed to capture broad tendencies over time in the discursive treatment of migration-related entities and issues across the French political field. The annotation design nonetheless makes it possible to conduct finer-grained analyses at the account level, which remain beyond the scope of the study but constitute an important direction for future research.

(4.3) Inclusion of retweets

Another methodological decision concerns the inclusion of retweets in the corpus. From a purely text-based perspective, retweets may appear as repeated material. However, in the context of CMC, retweeting constitutes a platform-specific socio-technical discursive practice, embedded in the communicative ecology of Twitter/X (Crystal, 2011: 40–41; Paveau, 2013; Longhi, 2016: 112–113). From the perspective of political discourse analysis, retweets are therefore treated as observable traces of recirculation, reinforcement, and public uptake. They provide evidence of lexical recurrence, ideological alignment, and selective endorsement. Their inclusion is thus methodologically justified insofar as the present study is concerned not only with linguistic forms within (re)tweets as discourse units in their own right, but also with their circulation as decontextualized discourse segments, as well as with their persistence, visibility, and repeated activation in digital political communication.

(5.1) Referential reconfiguration and syntactic positioning

A first result concerns the semantic and syntactic reconfiguration of migrant and immigration over time. For immigration, the right-wing subcorpus shows a significant restructuring of semantic-role distributions (χ² (6) ≈ 43.13; p < .001) (Jeon & Pietrandrea, in press): although Topic remains dominant overall, it declines after 2015, while more dynamic and negatively oriented roles such as Force and Theme.Beneficiary.neg become more prominent. Rather than functioning primarily as a discursive topic label, immigration increasingly appears in configurations that construe it as a source of pressure, disturbance, or harmful consequences, as illustrated in (tw13):

• (tw13)   [Force] RT @MLP_officiel: «Notre pays est confronté à une explosion de [l’immigration], de la délinquance et l’on voit s’imposer partout un climat de grande violence criminelle, terroriste ou sociale. La délinquance silencieuse mais épouvantable à vivre gagne nos campagnes!» #HauteMarne #OnArrive (R_@RNational_off, 2019-02-23).

•            English translation: Our country is facing an explosion of [immigration] and delinquency, and we can see a climate of serious criminal, terrorist, and social violence imposing itself everywhere. A silent form of delinquency, yet dreadful to live with, is spreading into our rural areas.

In the left-wing subcorpus, distributional change is also significant (χ²(6) ≈ 24.7; p < .01), but much weaker: Topic remains predominant throughout the period, and the other roles, including Agent, Force, and Theme.Beneficiary.pos, remain rare (<5%) and do not alter the overall configuration. This suggests that immigration remains predominantly framed as a topic of political debate rather than as an axiologically polarized entity.

For migrant, the right-wing subcorpus shows a shift away from relatively descriptive uses, associated with Theme attribution, toward more impersonal, causative, and axiologically charged construals, associated with Force attribution, as illustrated in (tw14) and (tw15).

• (tw14)   L’UE va donc regarder à la jumelle les bateaux de migrants…Pour l’action, ce sera sans elle ni l’RPS, sponsors historiques de l’immigration (R_@f_philippot, 2015-06-22)

•            English translation: The EU will therefore watch the boats of migrants through binoculars…As for action, it will be without the EU or the RPS, historical sponsors of immigration.

• (tw15)   “Le port de #Calais est entrain de s’effondrer, notamment à cause de l’afflux de #migrants.” @F3nord @F3Picardie; (R_@MLP_officiel, 2015-12-02)

•            English translation: The port of #Calais is collapsing, notably because of the influx of #migrants.

In both of such cases, the lemma, occurring in the plural form, tends to appear in syntactically dependent positions, as in les bateaux de migrants ‘the boats carrying migrants’ (literally, ‘the boats of migrants’) or l’afflux de #migrants ‘the influx of #migrants’. Such configurations reduce the salience of its intrinsic [+human] features. This recurrent distribution contributes to a redistribution of agency and supports the representation of migration-related entities as affected, framed, or backgrounded rather than volitionally acting participants or social actors. In line with an anonymous reviewer’s suggestion, expressions such as explosion de l’immigration ‘explosion of immigration’, afflux de migrants ‘influx of migrants’, or submersion migratoire ‘migratory submersion’ reflect broader metaphorical framings of migration-related entities in anti-immigration discourse (Hart, 2013; Taylor, 2021), in which immigration and migrants are construed as overwhelming natural forces requiring containment.

In the left-wing corpus, by contrast, migrant remains predominantly associated with vulnerable human referents (Theme.beneficiary.pos, as illustrated earlier in example (tw6)). These patterns show that the combined annotation of semantic role, modification, and syntactic function makes it possible to quantify the referential reconfiguration of the migr-lexicon across political orientation (Jeon, 2025a; Jeon & Pietrandrea, in press).

(5.2) List structures as sites of diachronic ideological categorization

A second result concerns list and parallelism structures, which provide a privileged site for observing how migration-related lexemes are aligned with other semantic domains and integrated into broader ideological frames. As noted earlier (cf. (1.3.4)), these configurations correspond to coordinated or paratactic configurations in which several lexical items are aligned within the same syntactic level of clustering. Following previous work on list constructions (Blanche-Benveniste, 1990, 2011; Rabatel, 2011; Kahane & Pietrandrea, 2012; Masini et al., 2018), they are analyzed here as discursive devices that temporarily group heterogeneous items into a common interpretive frame, resulting in construction of ad hoc category, that is, context-dependent groupings whose coherence is not based on stable lexical taxonomy, but on pragmatic and discursive inference.

This functional dimension is especially relevant in political tweets, where the juxtaposition of migration-related lexemes with other terms may suggest shared properties, shared evaluative orientations, or common ideological relevance. In other words, the list does not merely enumerate, but contributes to contextual categorization by inviting the reader to infer a relation of association between otherwise heterogeneous referents. To compare these environments quantitatively, conjunct lexemes were assigned to a set of thematic categories designed to capture recurrent discursive domains in which migrant and immigration are embedded. Table 9 summarizes the category codes and provides lexical indicators for each category. Their proportional distributions across the two subcorpora are shown in Figure 4 and Figure 5.

Figure 4

Figure 5

In the right-wing subcorpus, the conjunct patterns associated with migrant show a clear short-term diachronic evolution. During the 2015–2016 period, corresponding to the so-called “migration crisis”, its most frequent conjuncts belong mainly to the security/judicial and geography domains. These environments frequently include emblematic place names such as Calais, Ceuta, and Aquarius, which foreground border pressure and crisis management, as illustrated in (tw16):

• (tw16)   N’attendons pas un nouveau drame! La loi doit s’appliquer, les clandestins doivent être reconduits. {#Calais | #Migrants | #Eurotunnel}; (R_@ECiotti, 2015-07-29)

•            English translation: Let us not wait for another tragedy! The law must be enforced, and clandestine migrants must be sent back. {#Calais | #Migrants | #Eurotunnel}.

From 2018 onward, and more clearly in 2019, the lexeme becomes increasingly embedded in list constructions involving economic and institutional terms, indicating a progressive extension from migration-specific reference to broader political and socio-economic framing. This shift is consistent with the wider political context, including debates surrounding the 2018 Marakech Pact and the 2019 European elections. In some later examples, list structures also operate through contrastive recategorization, as when réfugiés ‘refugees’ are explicitly opposed to migrants économiques ‘economic migrants’, as illustrated in (tw17):

• (tw17)   A #Melilla, il y a des centaines de jeunes hommes qui veulent passer la frontière de force : nous savons faire la différence entre {les réfugiés de guerre ukrainiens – avec qui il est normal que la France soit solidaire – | et ces #migrants économiques}. #Elysée2022 (@MLP_officiel, 2022-03-03)

•            English translation: In #Melilla, there are hundreds of young men trying to cross the border by force: we know the difference between {Ukrainian war refugees – with whom it is normal for France to show solidarity – | and these economic migrants}.

By contrast, the left-wing subcorpus displays a more stable thematic profile, especially for migrant, which remains predominantly embedded in list environments linked to humanitarian/asylum and social vulnerability referents. The lexeme immigration also shows a left/right contrast, although in a more gradual and heterogeneous way: in FR-R, its list environments progressively shift toward security/judicial and economic/institutional domains, whereas in FR-L they remain more strongly associated with humanitarian/asylum framing. When security/judicial conjuncts appear with immigration in the left-wing subcorpus, they often occur in explicitly contrastive or metadiscursive formulations that refer to the ideological positioning of political opponents, as illustrated in (tw18):

• (tw18)   Pour @mlp_officiel, “il y a un lien entre {l’immigration | et le terrorisme}”. Peuple métissé, nous vous méprisons, madame #f_inter #NoPasaran; (L_@JLMelenchon, 2012-04-19)

•            English translation: According to @mlp_officiel, “there is a link between {immigration | and terrorism}.” Mixed-race people, we despise you, madam. #f_inter #NoPasaran.

Taken together, these results suggest that list constructions do not merely reflect co-occurrence patterns, but participate in the short-term diachronic restructuring of ideological framing by stabilizing different ad hoc categories across political orientations.

Systematic differences thus emerge between the left- and right-wing subcorpora in lexical choices, syntactic constructions, and semantic-role distributions. These findings confirm that a fine-grained, multi-layer annotation scheme can support the study of ideological variation in political discourse on CMC platforms such as Twitter/X by linking lexical forms to linguistic constructions, discourse-level categorization, and micro-diachronic variation.

(6.1) Diachronic relevance

The diachronic structure of the corpus, spanning 2011–2022, makes it possible to investigate temporal dynamics on migration-related discourse over a time span that also captures the growing use of Twitter as a privileged tool of political communication, followed by its relative decline from 2019 onwards. The dataset supports the analysis of changes in the frequency, distribution, and combinatory patterns of migr-lexicon occurrences across time, including peaks associated with major political events. It therefore enables the study of gradual change, periods of acceleration, and stabilization phenomena in political language.

Beyond individual findings, the corpus provides a basis for modeling language change in digital environments, where discourse is persistent, searchable, and recirculated. This persistence differentiates social media corpora from traditional textual corpora and has implications for the temporality of recursivity of linguistic and discursive change.

(6.2) Reusability and scope

FR-MIGR-TWIT 2.0 is structured to support reuse in multiple research contexts, including:

• – micro-diachronic lexical and syntactic studies,

• – comparative political discourse analysis (including account-level comparisons),

• – computational modeling of language change and discursive variation,

• – studies of categorization, reference, and ideological framing in discourse

As detailed in section (2), FR-MIGR-TWIT 2.0 is entirely encoded in UTF-8 and available in CSV and XML formats. The full annotation layers and metadata were not distributed entirely in TEI/XML format. Nevertheless, the TEI/XML version of FR-MIGR-TWIT incorporates TEI-compatible structural principles, including unique textual identifiers (xml:id), user attribution (who), temporal metadata (when), and utterance-like segmentation (<u>, <seg>). In this respect, the corpus may be loaded into software environments such as TXM and IRaMuTeQ. The TEI/XML structure and metadata organization may facilitate interoperability with other French CMC corpora documenting political communication, such as Polititweet (CoMeRe repository) (Chanier et al., 2014).

More broadly, FR-MIGR-TWIT 2.0 is part of the multilingual MIGR-TWIT corpus family, which includes a UK module of migration-related political tweets, namely UK-R-MIGR-RA-2012-2022. The MIGR-TWIT corpus family was developed within the framework of the OLiNDiNUM Linguistic Observatory of Digital Discourse corpus infrastructure (Pietrandrea et al., 2026), alongside resources such as UK-EU-DEBATE-20-21, thereby creating opportunities for comparative analyses of the right-wing political discourse on migration and online public debate across languages and political contexts in Europe.

The combination of rich metadata, multi-layer annotation, and explicit corpus-construction criteria ensures transparency, reproducibility, and methodological transferability. As such, the corpus may be reused not only for qualitative and quantitative linguistic analysis, but also as a resource for annotation experiments.