Establishing a Language by Annotating a Corpus: the Case of Naija, a Post-creole Spoken in Nigeria Marine Courtin1, Bernard Caron2, Kim Gerdes3, Sylvain Kahane1 1 Modyco, Université Paris Nanterre & CNRS 2 Llacan, CNRS / IFRA Ibadan, CNRS 3 LPP, Université Sorbonne Nouvelle & CNRS marine.courtin@sorbonne-nouvelle.fr, bernard.caron@cnrs.fr, kim@gerdes.fr, skahane@parisnanterre.fr Abstract In this paper, we show that building a treebank can be used as a way to establish a language. Annotated corpus can be used as tools when arguing that some linguistic data belongs to a separate language (rather than a dialect or variety of another established language). We provide here a case study on a treebank of Naija, a Post-creole spoken in Nigeria which presents us with significant differences from treebanks of English in terms of existing constructions and frequency of several syntactic units. Keywords: Naija, Nigerian Pidgin, Treebank, Quantitative Linguistics, Typology 1. The Situation of Naija Naija also borrowed lexical items from other local Spoken by educated Nigerians, the Nigerian post-creole languages, in particular ideophones such as kpatakpata has been shown by Deuber (2005) to develop in Lagos as ‘completely’. a discrete language, separate from Nigerian English. This language, that we propose to call Naija, is now spoken as a second language by over 100 million speakers, all over Nigeria, a country of 180 millions people, where about 450 native languages are spoken with three dominating languages (Igbo, Yoruba, and Hausa). This new language has taken a considerable economical and cultural importance in Nigeria. Nevertheless, for its speakers, this language is often considered as an inferior version of English (they call it “Broken”) with a negative influence on Nigerian education. Most speakers are not conscious that, as a separate language with its own grammar and (1) sotay di rain sef kuku fall some house lexicon, it has a outstanding potential in favor of national dem down kpatakpata cohesion, since it is perceived as ethnically neutral, and so_that the rain EMPH commonly fall some house for regional integration, due to its intercomprehension PL down completely with Ghanaian and Cameroonian pidgins. ‘So that, often, the rain completely destroys houses.’ Considering the particular situation of this language, We use the Arborator (Gerdes 2013) as the online building a syntactic treebank takes a particular annotation tool for POS and dependency annotation. The significance. Of course, as for any language, a treebank Arborator’s exercise mode allows to present pre-annotated can be useful for many applications, such as the training sentences as exercises to newly recruited annotators. The of a syntactic parser. But here the treebank helps us to Arborator integrates the Mate parser (Bohnet 2010) that establish the existence of Naija as a language separate can be trained at any time which allows for quick and easy from (Nigerian) English, by showing constructions that bootstrapping of the annotation process. are specific to Naija (qualitative analysis) and constructions that are over-represented in Naija In order to allow for typological comparison and distance (quantitative analysis). measures on Naija, we use a surface-syntactic dependency annotation scheme that is compliant with standard 2. Tools and Workflow dependency annotation (e.g. prepositions as governors) and thus easy to learn and to apply, but which allows for a The study is based on a 750,000 word corpus collected all lossless transformation into Universal Dependencies (UD) around Nigeria. The transcription is a scientific and by means of a graph rewriting process (Guillaume 2012). political challenge by itself because most words stem Each treebank for the 75 languages of the UD database from English, but some of them have grammaticalized and must conform to the universal tagset for POS and are pronounced differently. We follow what is done in the dependency relation names. Language idiosyncrasies have (mostly informal) writing of Naija: keep the English to be encoded as additional features next to the POS or as spelling for lexical words, with exceptions for very subtypes of dependency relation names, e.g. in English the frequent words such as broda ‘brother’; and a more noun modifier (nmod) receives a subtype to describe the Saxon genitive: “John[’s] <-nmod:poss- book”. phonetic spelling for grammatical terms (dem ‘them’, im ‘him’, sey complementizer lit. say). Currently the treebank has 12,000 tokens and is available on the UD webpage. We intend to manually annotate 7 100,000 tokens and then to automatically parse the whole One of our hypothesis concerning these differences is that corpus. information packaging (or communicative structure) plays a larger role in Naija than in English. To explore this hypothesis it is necessary that we dispose of an annotated 3. Qualitative Analysis corpus, as we need to measure the frequency of some A good number of morphosyntactic specificities of Naija structures (for example dislocations and cleft sentences), have called for an ongoing review of the annotating rather than their strict presence or absence in the scheme that was initially adopted for the language. language. For this purpose, we use all available treebanks of English in UD v2.1: UD_English-ParTUT (Bosco and Some of these specificities are linked to the influence of Sanguinetti, 2014), UD_English-LinES (Ahrenberg 2007), adstrate vernacular languages belonging mainly to the UD_English-EWT (Silveira et al., 2014), and v2.2 version Niger-Congo family. This is the case of emphatic of UD_Naija-NSC. We also parsed the Santa Barbara adverbial particles (e.g. sha, o) tagged with the ADV POS Corpus of Spoken American English (Du Bois et al. 2000- label, but whose function is characterized by the 2005) to get a reference of what spoken English might mod:emph dependency link. The influence of adstrate look like in terms of syntactic relations’ distribution. vernacular languages is observed in the use of Serial Verb Constructions, that is “monoclausal construction[s] The table below presents some of the interesting consisting of multiple independent verbs with no element differences between (1) written English, (2) spoken linking them and with no predicate-argument relation English and (3) spoken Naija : between the verbs.” (Haspelmath 2016) Such constructions appear in languages of Nigeria, such as det case obl dislo- ccomp aux cc Yoruba (Stahlke 1970) (see (2)), and it has already been cated shown that they are present in creoles languages. 9.4 % 10.6 % 5.8 % 0.0 % 1.1% 4.2 % 3.7 % (1) (2) mo mu iwwe wa ilwe (Yoruba, Aubry 2010) (2) 6.7 % 6.6 % 4.2 % ? 2.0 % 4.5% 4.3 % 1SG take book come home’ (3) 5.7 % 4.2 % 3.7 % 1.7 % 2.1 % 9.3 % 1.4 % ‘I brought a book to my home’ We used the subtyped relation compound:svc for these constructions, which do not exist in English (see (3)). To test the significance of the observed differences in frequency counts, we applied a Fisher's Exact Test for Other specificities are linked to the emergence of up to Count Data with simulated p-value (based on 2000 here undescribed structures which the corpus has enabled replicates), giving us an overall p-value of 0.0004998. us to identify. One of them is a focus structure where the focus particle na (which identifies the clefted constituent) Some differences such as the lower frequency of is doubled by the morpheme naim (which introduces the determiners are easily explained. A Naija sentence such as cleft clause). This morpheme originates in the no dey stay for middle of road would not require definite grammaticalization of the colocation na + im, lit. ‘it is’ + determiners in front of middle or road, while its English ‘him/it/her’. This discovery of a new structure is the result counterpart, don’t stay in the middle of the road, would. of a collaborative analysis done by the team of annotators during the production of the corpus. Another variation concerns the frequency of auxiliaries, which are more than twice as frequent in Naija than in The same ongoing grammaticalization process is observed English, regardless of the distinction written/spoken. We in the formation of TAM auxiliaries where full lexical then looked at the ratio of verb on auxiliaries to see which verbs (e.g. go ‘go’; come ‘come’ ; dey ‘exist’) coexist language had more complex verbal constructions and with their grammaticalized equivalents (go, future; come, found that Naija had the highest score (which means less realis; dey, imperfective). Likewise, the verb make, which auxiliaries per verb on average). already appears in Serial Verb Constructions to express the equivalent of the comitative case, is used as an auxiliary for converb forms (e.g. dem want make e go church ‘they want him to go to church’). This flourishing Verb / Auxiliaries ratio multifunctionality, typical of creole languages, creates (1) 1.9 challenges for the recognition of government. (2) 1.8 (3) 2.0 4. Quantitative analysis In creoles, it is usually assumed that there is a division of labor between the lexifier language which provides the Taking into account the fact that Naija also has the highest majority of the lexicon (in our case English) and substrate frequency of auxiliaries (9.3% against 4.2% for written languages in areal contact with the creole (in the case of English and 4.6% for spoken English) we observe that Naija these might be Yoruba, Igbo and Hausa for Naija must compensate by having a high frequency of example). We attempt to show quantitative evidence of verbs which can be accounted for by the compound:svc, structural similarities and differences between Naija and ccomp, acl:relcl and root relations. If we look more English. precisely at the distribution of these auxiliaries, it appears that it is the auxiliaries which are not shared with English 8 (dey, come, go, don, fit, for and neva) which are more between written and spoken French, which seems to frequent, while there is only one occurrence of the shared suggest that this might very well be a product of the genre auxiliary will. rather than a characteristic of the language. 2 The lower frequencies for both oblique and case relations This over-representation seems to apply to cleft sentences are correlated: Naija seems to use less oblique as well. The subtype :cleft, which we used in the complement in favor of more direct objects. Locative annotation of both UD_Naija and UD_French_Spoken, complements can be expressed through Serial Verb can be found on 1.1 % of all relations in Naija, while it is Constructions with the place as direct object of the second considerably less frequent in spoken French (0.2%). verb as in (3). Another interesting findings is that Naija also shows three (3) government worker dem go dey enter go work times less coordinating conjunctions than English does government worker PL FUT PROG get_on go work (1.4% for Naija against 3.7% and 4.3% for written and ‘government workers will be getting on to go to work’ spoken English). This is interesting as we would expect a higher frequency of coordinations in spoken texts, to accommodate for lists and reformulations which are more common. In Naija it is not uncommon to have several coordinations without any coordinating conjunction as in (5) [conjuncts are underlined]. (5) Lagos don follow see dis kind rain o wey uproot tree take am block road spoil dose big billboard dem […] comot di roof of plenty house dem. ‘Lagos has experienced the kind of rain where trees were uprooted and blocked the road, destroyed those This role would be filled by an oblique complement big billboards […] and removed the roofing of lots of introduced by an adposition in English, as in the example houses.’ below: This suggests that Naija might favor other strategies such as juxtaposition rather than coordinated constituents linked with coordinating conjunctions. (4) We might also be interested in the differences in distribution of part-of-speech tags3 between English and Naija. Other differences do not show such clear-cut contrasts between English and Naija, but are still interesting as they indicate areas which might need to be investigated further. We measure that 1.7 % of all dependency relations1 in the Naija treebank are labeled dislocated. The mean length of sentences being around 10 tokens, this means that on average there is a dislocation in 1 sentence out of 6, which is very significant, even more so when compared to the 0.0004% frequency found in written English. Unfortunately our parser performs poorly on this relation (due to the lack of training data) and no reliable frequency Fig 1. Relative frequency of pos tags in English count of this relation type can be extracted from the spoken English corpus. We therefore look at spoken French (which has the reputation of being particularly prone to dislocations) to get a better sense of the significance of our findings, and find that 1.0 % of dependency links are dislocated (in the UD_French_Spoken, Lacheret and al., 2014). This indicates that dislocation is a major feature of spoken Naija. However, the variation in frequency of this 2 One reviewer also noted that some of the English corpora such dislocated link is not significantly more important as EWT were automatically converted from constituent between written English and spoken Naija than it is treebanks using rule-based systems which often fail to identify dislocated constructions. 1 3 punct links excepted We filtered tokens with PUNCT, X and SYM tags 9 5. Conclusion Annotators who were speakers of Naija reported that throughout the annotation process, their vision of Naija had changed. They noticed more readily that some syntactic phenomena were specific to Naija and that there were complex rules which governed the Naija grammar. We believe this to be an interesting pedagogical experiment where student annotators re-discover their language through the annotation of a corpus, and are confronted with regularities and patterns that sometimes went unnoticed in their day to day life (particularly so since speaking Naija is mostly depreciated). We think that claims of Naija being a separate language can better be supported using a treebank. Indeed, while Fig 2. Relative frequency of pos tags in Naija lexical differences are certainly noticeable between Naija and English, we believe that the identity of the language lies in its syntactic structure which is not as easily Naija has significantly more verbs while the English accessible from raw text or even tagged corpus. Having a corpus is a lot richer in nouns. Part of the over- treebank of Naija enables us to quantify the frequency of representation of verbs in Naija can be attributed to Serial some syntactic structures, which in turns helps us to Verb Constructions, with verbs in the second position evaluate the complexity and idiosyncracies of the Naija representing 1.48 % of all tokens, but this account does grammar, and to measure the distance the language has not suffice to explain such a gap. Investigating this taken from English. Comparisons between the two disparity, we also measured other relations involving languages could also yield interesting insights concerning verbal dependents such as ccomp. We find twice as many clausal complements with respectively 1.64 % and 0.82 % the ungoing creolization process of Naija. ccomp links in Naija and English. This indicates that looking at complex sentences in more details might Acknowledgments provide us with additional examples of differences between the two languages. We thank our reviewers for valuable remarks and We also expect that genre differences 4 between the corrections. This work is supported by the French treebanks play an important part in this repartition. Future National Research Agency (ANR) with the project work using a Nigerian English corpus of both spoken and NaijaSynCor written texts should allow us to better determine the extent of differences due to genre and the variety of English being considered. References Interestingly enough, even though Naija allows the dropping of pronouns they are still very frequent in our corpus. One possible explanation is that pronouns are highly susceptible to repetition and reformulation in Ahrenberg, L. (2007). "LinES: An English-Swedish spoken language. But it might also have to do with the Parallel Treebank". Proceedings of the 16th Nordic frequent topicalization of subjects through dislocation in Conference of Computational Linguistics Naija, as in (6), or with rhetorical devices which involve (NODALIDA, 2007). repeating the pronoun to emphasize parallelism as in (7). Aubry, N. (2010) Changements syntaxiques dans le Yorùbá de la presse (1930-2010) : traitement (6) dat man im pull over automatique d'un corpus diachronique et analyse des that man he pulls over résultats, PhD thesis, Inalco. ‘that man pulls over’ Bohnet, B. (2010). "Very high accuracy and fast (7) dem go bring am dem go seize am again. dependency parsing is not a contradiction." Proceedings of the 23rd international conference on computational they will bring it they will seize it again linguistics. Association for Computational Linguistics. ‘they will bring it and seize it again’ Bosco, C. and Sanguinetti, M. (2014). "Towards a Universal Stanford Dependencies parallel treebank". In Proceedings of the 13th Workshop on Treebanks and Linguistic Theories (TLT-13), Tubingen (Germany). 4 Deuber, D. (2005). Nigerian Pidgin in Lagos: Language There is a small portion of spoken English in UD_English- LinES, but apart from this the corpus we used is all written texts, contact, variation and change in an African urban with variations in terms of genres (news, wiki, nonfiction, blog, setting. Battlebridge Publications. emails, legal texts..). The Naija treebank is all spoken texts (conversations and interviews). 10 Du Bois, John W., Wallace L. Chafe, Charles Meyer, generalizations. Language and Linguistics, 17(3), 291- Sandra A. Thompson, Robert Englebretson, and Nii 319. Martey. (2000-2005). Santa Barbara corpus of spoken Jansen, B., Koopman, H., Muysken, P. (1978). Serial American English, Parts 1-4. Philadelphia: Linguistic verbs in the creole languages. Amsterdam Creole Data Consortium. Studies 2. 125–159. Gerdes, K. (2013). "Collaborative dependency Lacheret, A., Kahane, S., Beliao, J., Dister, A., Gerdes, annotation." Proceedings of the second international K., Goldman, J. P., Tchobanov, A. (2014). Rhapsodie: a conference on dependency linguistics (DepLing 2013). prosodic-syntactic treebank for spoken french. In Guillaume, B., Bonfante, G., Masson, P., Morey, M. and Language Resources and Evaluation Conference. Perrier, G. (2012). "Grew: un outil de réécriture de Silveira, N., Dozat, T., de Marneffe, M., Bowman, S., graphes pour le TAL (Grew: a Graph Rewriting Tool Connor, M., Bauer, J., and Manning, C. (2014)."A Gold for NLP)[in French]." Proceedings of JEP-TALN- Standard Dependency Corpus for English." LREC. RECITAL. Haspelmath, M. (2016). The serial verb construction: Comparative concept and cross-linguistic 11