<<

Maori¯ : A Corpus of English Tweets David Trye Andreea . Calude Computing and Mathematical School of General and Applied University of Waikato, New Zealand University of Waikato, New Zealand [email protected] [email protected]

Felipe Bravo-Marquez Te Taka Keegan Department of Computer Computing and Mathematical Sciences University of Chile & IMFD University of Waikato, New Zealand [email protected] [email protected]

Abstract this body of work is online presence of the loanwords - almost all studies come from Maori¯ loanwords are widely used in New (collaborative) written (highly edited, re- Zealand English for various social functions by within and outside of the vised and scrutinised newspaper language, Davies Maori¯ community. Motivated by the lack of and Maclagan 2006; Macalister 2009, 2006a,; linguistic resources for studying how Maori¯ Onysko and Calude 2013, and picture-books, Daly loanwords are used in social media, we present 2007, 2016), or from spoken language collected in a new corpus of tweets. the late 1990s (Kennedy and Yamazaki, 1999). We collected tweets containing selected Maori¯ In this paper, we build a corpus of New are likely to be known by New Zealand English tweets containing Maori¯ loan- Zealanders do not speak Maori.¯ Since over 30% of these words turned out to be ir- words. Building such a corpus has its challenges relevant (.., mana is a popular gaming term, (as discussed in Section 3.1). Before we discuss is a character from a Disney movie), these, is important to highlight the uniqueness we manually annotated a sample of our tweets of the situation between Maori¯ into relevant and irrelevant categories. This and (NZ) English. data was used to train machine learning mod- The language contact situation in New Zealand els to automatically filter out irrelevant tweets. provides a unique case-study for loanwords be- 1 Introduction cause of a number of factors. We list three partic- ularly relevant here. First, the direction of lexical of the most salient features of New Zealand transfer is highly unusual, namely, from an endan- English (NZE) is the widespread use of Maori¯ gered (Maori)¯ into a domi- words (loanwords), such as aroha (love), kai nant (English). The large-scale lex- (food) and (New Zealand). See ex. (1) ical transfer of this type has virtually never been specifically from (note the informal, con- documented elsewhere, to the best of our knowl- versational style and the Maori¯ loanwords empha- edge (see summary of current language contact sised in bold). situations in Stammers and Deuchar 2012, partic- ularly Table 1, p. 634). (1) Led the waiata for the manuhiri at the powhiri¯ for new staff for induction week. Secondly, because Maori¯ loanwords are “New Was told by the kaumatua¯ I did it with mana Zealand’s and New Zealand’s alone” (Deverson, and integrity. 1991, p. 18-19), and above speakers’ conscious- ness, their ardent study over the years provides a The use of Maori¯ words has been studied inten- fruitful of the use of loanwords across sively over the past thirty years, offering a com- genres, contexts and time. prehensive insight into the evolution of one of Finally, the aforementioned body of previous the youngest of English – New Zealand research on the topic is rich and detailed, and still English (Calude et al., 2017; Daly, 2007, 2016; rapidly changing, with use an in- Davies and Maclagan, 2006; De Bres, 2006; creasing trend (Macalister, 2006a; Kennedy and Degani and Onysko, 2010; Kennedy and Ya- Yamazaki, 1999). However, the jury is still out mazaki, 1999; Macalister, 2009, 2006a; Onysko regarding the for the loanword use (some and Calude, 2013). One aspect which is missing in hypotheses have been put forward), and the pat-

136 Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 136–142 Florence, Italy, July 28 - August 2, 2019. 2019 Association for Computational Linguistics terns of use across different genres (it is unclear edited (mainly newspaper language, Macalister how language formality influences loanword use). 2006a; Davies and Maclagan 2006; Degani 2010), We find that Twitter data complements the so little is known about Maori¯ loanwords in recent growing body of work on Maori¯ loanwords in informal NZE interactions – a gap we hope to ad- NZE, by adding a combination of institutional and dress here. individual linguistic exchanges, in a non-editable With the availability of vast amounts of data, online platform. Social media language shares building Twitter corpora has been a fruitful en- properties with both spoken and , deavour in various , including Turkish but is not exactly either. More specifically, (S¸ims¸ek and Ozdemir¨ , 2012; C¸etinoglu, 2016), Twitter allows for creative expression and lexical Greek (Sifianou, 2015), German (Scheffler, 2014; innovation (Grieve et al., 2017). Cieliebak et al., 2017), and (American) English Our Twitter corpus was created by following (Huang et al., 2016) (though notably, not New three main steps: collecting tweets over a ten-year Zealand English, while a modest corpus of te period using “query words” (Section 3.1), man- reo Maori¯ tweets does exist, Keegan et al. 2015). ually labelling thousands of randomly-sampled Twitter corpora of mixed languages are tougher to tweets as “relevant” or “irrelevant” (Section 3.2), collect because it is not straightforward to detect and then training a classifier to obtain automatic data automatically. Geolocations predictions for the relevance of each tweet and de- can help to some extent, but they have limitations ploying this model on our target tweets, in a bid (most users do not use them to begin with). Recent to filter out all those which are “irrelevant” (Sec- work on has leveraged the presence of dis- tion 3.3). As will be discussed in Section2, our tinct scripts – the Roman and Arabic alphabet – to corpus is not the first of its kind but is the first cor- create a mixed language corpus (Voss et al., 2014), pus of New Zealand English tweets and the first but this option is not available to us. Maori¯ has collection of online discourse built specifically to traditionally been a spoken (only) language, and analyse the use of Maori¯ loanwords in NZE. Sec- was first written down in the early 1800s by Euro- tion4 outlines some preliminary findings from pean missionaries in conjunction with Maori¯ lan- our corpus and Section5 lays out plans for future guage scholars, using the Roman alphabet (Smyth, work. 1946). Our task is more similar to studies such as Das and Gamback¨ (2014) and C¸etinoglu(2016), 2 Related Work who aim to find a mix of two languages which share the same script (in their case, and En- It is uncontroversial that Maori¯ loanwords are both glish, and Turkish and German, respectively), but productively used in NZE and increasing in popu- our method for collecting tweets is not user-based; larity (Macalister, 2006a). The corpora analysed instead we use a set of target query words, as de- previously indicate that loanword use is highly tailed in Section 3.1. skewed, with some language users leading the way – specifically Maori¯ women (Calude et al., 2017; 3 The Corpus Kennedy and Yamazaki, 1999), and with certain topics of discourse drawing significantly In this section, we describe the process of build- of loanwords than others – specifically ing the Maori¯ Loanword Twitter Corpus (here- 1 those related to Maori¯ people and Maori¯ affairs, after, the MLT Corpus) . This process consists of Maoritanga¯ (Degani, 2010). The type of loan- three main steps, as depicted in Figure1. words being borrowed from Maori¯ is also chang- 3.1 Step 1: Collecting Tweets ing. During the first wave of borrowing, some two-hundred years ago, many flora and fauna In order to facilitate the collection of relevant data words were being borrowed; today, it is social for the MLT Corpus, we compiled a list of 116 tar- culture terms that are increasingly adopted, e.g., get loanwords, which we will call “query words”. aroha (love), whaea (woman, teacher), and tangi 1The corpus is available online at https: (Maori¯ funeral), see Macalister(2006a). However, //kiwiwords.cms.waikato.ac.nz/corpus/. the data available for loanword analysis is either Note that we have only released the tweet IDs, together with a download script, in accordance with Twitter’s terms and outdated (Calude et al., 2017; Kennedy and Ya- conditions. We have also released the list of query words mazaki, 1999), or exclusively formal and highly used.

137 1. Collect Tweets 2. Annotate Tweets 3. Deploy Model

Proud to be a Proud to be a kiwi tweet1 Love my crazy whanau Love my crazy whanau tweet2 Moana is my fav Princess Moana is my fav Princess ...

ne kuma fa haka ne kuma fa tweetm

Tweets Target

Figure 1: The corpus-building process.

Most of these are individual words but some are this process with similar items). short phrasal units (tangata whenua, people of the After inspecting these tweets, it was clear that land; kapa haka, cultural performance). The list a large number of our query words were polyse- is largely derived from Hay(2018) but was modi- mous (or otherwise unrelated to NZE), and had fied to exclude function words (such as numerals) introduced a significant amount of noise into the and most proper , except five that have native data. The four main challenges we encountered English counterparts: Aotearoa (New Zealand), are described below. Kiwi(s) (New Zealander(s)), Maori¯ (indigenous First, because Twitter contains many different New Zealander), Pakeh¯ a¯ (European New Zealan- varieties of English, NZE being just one of these, der), non-Maori¯ (non-indigenous New Zealander). it is not always straightforward to disentangle the We also added three further loanwords which we of English spoken in New Zealand from deemed useful for increasing our data, namely other dialects of English. This could be a prob- haurangi (drunk), wairangi (drugged, confused), lem when, for instance, a Maori¯ like Moana and porangi¯ (crazy). (sea) is used in tweets to denote Using the Twitter Search API, we harvested the Disney movie (or its main character). 8 million tweets containing at least one query Secondly, Maori¯ words have forms word (after converting all characters to lower- with other Austronesian languages, such as case). The tweets were collected diachronically Hawaiian, Samoan and Tongan, and many speak- over an eleven year period, between 2007-2018. ers of these languages live and work (and tweet) We ensured that tweets were (mostly) written in in New Zealand. For instance, the word wahine lang:en English by using the parameter. (woman) has the same written form in Maori¯ A number of exclusions and further adjustments and in Hawaiian. But are not the were made. With the aim of avoiding redundancy only problematic words. Homographs with other, and uninformative data, retweets and tweets with genealogically-unrelated languages can also pose URLs were discarded. Tweets in which the query problems. For instance, the Maori¯ word hui (meet- word was used as part of a username or men- ing) is sometimes used as a proper in - tion (e.g., @happy kiwi) were also discarded. For nese, as can be seen in the following tweet: “Yo is those query words which contained macrons, we Tay Peng Hui okay with the tip of his finger?”. found that users were inconsistent in their macron Proper nouns constitute a third problematic as- use. Consequently, we consolidated the data by pect in our data. As is typical for many lan- adjusting our search to include both the macron guage contact situations where an indigenous lan- and the non-macron version (e.g., both Maori¯ and guage shares the same geographical space as an Maori). We also removed all tweets containing incoming language, Maori¯ has contributed many fewer than five tokens (words), due to insufficient place names and personal names to NZE, such as context of analysis. Timaru, Aoraki, Titirangi, Hemi¯ , Mere and so on. Owing to relaxed conventions on Twit- While these proper nouns theoretically as ter (and also the use of hashtags), certain query loanwords, we are less interested in them than in words comprising multiple lexical items were content words because the use of the former does stripped of spaces in order to harvest all variants of not constitute a choice, whereas the use of the lat- the phrasal units (e.g., kai moana and kaimoana). ter does (in many cases). The “choice” of whether As kai was itself a query word (in its own right), to use a loanword or whether to use a native En- we excluded tweets containing kai moana when glish word (or sometimes a native English phrase) searching for tweets containing kai (and repeated is interesting to study because it provides insights

138 into idiolectal lexical preferences (which words goal is to obtain automatic predictions for the rel- different speakers or writers prefer in given con- evance of each tweet in this corpus, according to texts) and relative borrowing success rates (Calude probabilities given by our model. et al., 2017; Zenner et al., 2012). We created (stratified) test and training sets that Finally, given the impromptu and spontaneous maintain the same proportion of relevant and irrel- of Twitter in general, we found that cer- evant tweets associated with each query word in tain Maori¯ words coincided with misspelled ver- the Labelled Corpus. We chose to include 80% of sions of intended native English words, e.g., whare these tweets in the training set and 20% in the test (house) instead of where. set (see Table1 for a break-down of relevant and The resulting collection of tweets, termed the irrelevant instances). Original Dataset, was used to create the Raw Cor- pus, as explained below. Train Test Total Relevant 1, 995 500 2, 495 3.2 Step 2: Manually Annotating Tweets Irrelevant 954 236 1, 190 We decided to the “noisy” tweets in our Total 2, 949 736 3, 685 data using supervised machine learning. Two coders manually inspected a random sample of Table 1: Dataset statistics for our labelled tweets. This 30 tweets for each query word, by checking the Table shows the relevant, irrelevant and total number of instances (i.e., tweets) in the independent training and word’s context of use, and labelled each tweet as test sets. “relevant” or “irrelevant”. For example, a tweet like that in example (1) would be coded as relevant Using the AffectiveTweets package (Bravo- and one like “ awesome!! Congrats to Tangi :)”, Marquez et al., 2019), our labelled tweets were would be coded as irrelevant (because the query transformed into feature vectors based on the word word tangi is used as a proper noun). Since 39 -grams they contain. We then trained various of the query words consistently yielded irrelevant classification models on this transformed data in tweets (at least 90% of the time), these (and the Weka (Hall et al., 2009). The models we tested tweets they occurred in) were removed altogether were 1) Multinomial Naive Bayes (McCallum from the data. Our annotators produced a total of et al., 1998) with unigram attributes and 2) L2- 3, 685 labelled tweets for the remaining 77 query regularised logistic regression models with differ- words, which comprise the Labelled Corpus (see ent word n-gram features, as implemented in LIB- Tables1 and4; note that irrelevant tweets have LINEAR2. We selected Multinomial Naive Bayes been removed from the latter for linguistic anal- as the best model because it produced the highest ysis). AUC, Kappa and weighted average -Score (see Assuming our coded samples are representative Table2 for a summary of results). Overall, logis- of the real distribution of relevant/irrelevant tweets tic regression with unigrams performed the worst, that occur with each query word, it makes sense to yielding (slightly) lower values for all three mea- also discard the 39 “noisy” query words from our sures. Original Dataset. In this way, we created the (un- After deploying the Multinomial Naive Bayes labelled) Raw Corpus, which is a fifth of the size model on the Raw Corpus, we found that (see Table4). 1,179,390 tweets were classified as relevant and We computed an inter-rater reliability score for 448,652 as irrelevant (with probability threshold = our two coders, based on a random sample of 200 0.5). tweets. Using Cohen’s Kappa, we calculated this Table3 shows examples from our corpus of value to be 0.87 (“strong”). In light of the strong each type of classification. Some tweets were between the initial coders, no further falsely classified as “irrelevant” and some were coders were enlisted for the task. falsely classified as “relevant”. A short explana- 3.3 Step 3: Automatically Extracting tion why the irrelevant tweets were coded as such Relevant Tweets is given in brackets at the end of each tweet. The next step was to train a classifier using the La- We removed all tweets classified as irrelevant, belled Corpus 2 as training data, so that the resulting https://www.csie.ntu.edu.tw/˜cjlin/ model could be deployed on the Raw Corpus. Our liblinear/

139 AUC Kappa F-Score Irrelevant tweets Relevant tweets f()<0.5 Haka ne! And i know even the son didnt get my chop ciggies Multinomial Naive Bayes Classified good guys get blood for body 2day so stopped talking 2 him. n = 1 0.872 0.570 0.817 irrelevant (0.282, ) just walked past and gave me the maori eyebrow lift and a Logistic Regression smile. were friends (0.337) n = 1 0.863 0.534 0.801 Whare has the year gone (0.36, Shorts and bare feet in this misspelling) whare (0.41) n = 1, 2 0.868 0.570 0.816 chegar na morena e falar can i Tangata whenua charged for n = 1, 2, 3 0.869 0.560 0.811 be your girlfriend can i (0.384, killing 6 #kereru¯ for Kai mean- foreign language) while forestry corps kill off n = 1, 2, 3, 4 0.869 0.563 0.813 widespread habitat for millions n = 1, 2, 3, 4, 5 0.869 0.556 0.810 #efficiency #doc (0.306) f(x)≥0.5 Te Wanganga Aotearoa’s Our whole worldview as Maori Classified moving to a new campus in is whanau based. Pakeha call Table 2: Classification results on the test set. The best relevant Palmy, but their media person it nepotism, tribalism, gangster- has refused to talk to us about it. ism, LinkedInism blah de blah. results for each column are shown in bold. The value #whatajoke (0.998, proper noun) It’s our way of doing stuff and of n corresponds to the type of word n-grams included it’s not going to change to suit another point of view. (0.995) in the feature space. I cant commit to anything but if Kia ora koutou - does anyone I were to commit to one song, know the te reo word for Corn- it would be kiwi - harry styles wall? (1.0) (0.791, proper noun) thereby producing the Processed Corpus. A sum- Why am I getting headaches out The New Zealand team do an- of no whero never get them :( other energetic haka though mary of all three corpora is given in Table4. I guess its all the (0.542, (0.956) spelling mistake)

4 Preliminary Findings Table 3: A selection of tweets and their classification types. The first three irrelevant tweets were classified As we are only just beginning to sift through the correctly (i.e. true negatives), as were the last three MLT Corpus, we note two particular sets of pre- relevant tweets (i.e. true positives). Function f(x) cor- liminary findings. responds to the posterior probability of the “relevant” First, even though our corpus was primarily class. The entries in brackets for the irrelevant exam- geared up to investigate loanword use, we are find- ples correspond to the values of f(x) and the ing that, unlike other NZE genres analysed, the why the target word was coded as irrelevant. Twitter data exhibits use of Maori¯ which is more in Raw Labelled Processed line with code-switching than with loanword use, Tokens (words) 28,804,640 49,477 21,810,637 see ex. (2-3). This is particularly interesting in Tweets 1,628,042 2,495 1,179,390 light of the reported increase in te reo Maori¯ lan- Tweeters (authors) 604,006 1,866 426,280 guage tweets (Keegan et al., 2015). Table 4: A description of the MLT Corpus’ three com- ponents (namely, the Raw Corpus, Labelled Corpus (2) Morena¯ e hoa! We must really meet IRL and Processed Corpus), which were harvested using when I get back to Tamaki¯ Makaurau! You the same 77 query words. have a fab day too!

(3) Heh! He porangi toku ngeru - especially at 5 Conclusions and Future Work 5 in the morning!! Ata marie e hoa ma.I am well thank you. This paper introduced the first purpose-built cor- pus of Maori¯ loanwords on Twitter, as well as a Secondly, we also report the use of hybrid hash- methodology for automatically filtering out irrel- tags, that is, hashtags which contain a Maori¯ evant data via machine learning. The MLT Cor- part and an English part, for example #mycrazy- pus opens up a myriad of opportunities for future whanau, #reostories, #Matarikistar, #bringiton- work. mana, #growingupkiwi, #kaitoputinmyfridge. To Since our corpus is a diachronic one (i.e., all our knowledge, these hybrid hashtags have never tweets are time-stamped), we are planning to use been analysed in the current literature. Hybrid it for testing hypotheses about . hashtags parallel the phenomenon of hybrid com- This is especially desirable in the context of New pounds discussed by Degani and Onysko(2010). Zealand English, which has recently undergone Degani and Onysko report that hybrid compounds considerable change as it comes into the final stage are both productive and semantically , show- of dialect formation (Schneider, 2003). ing that the borrowed words take on reconceptu- Another avenue of future research is to automat- alised meanings in their adoptive language (2010, ically identify other Maori¯ loanwords that are not p.231). part of our initial list of query words. This could

140 be achieved by deploying a language detector tool Mark Cieliebak, Jan Milan Deriu, Dominic Egger, and on every unique word in the corpus (Martins and Fatih Uzdilli. 2017. A twitter corpus and benchmark Silva, 2005). The “discovered” words could be resources for German sentiment analysis. In 5th International Workshop on Natural Language Pro- used as new query words to further expand our cessing for Social Media, Boston, MA, USA, Decem- corpus. ber 11, 2017, pages 45–51. Association for Compu- In , we intend to explore the meaning of tational Linguistics. our Maori¯ loanwords using distributional semantic Nicola Daly. 2007. Kukupa,¯ koro, and kai: The use models. We will train popular word embeddings of Maori¯ items in New Zealand English algorithms on the MLT Corpus, such as Word2Vec children’s picture books. (Mikolov et al., 2013) and FastText (Bojanowski Nicola Daly. 2016. picturebooks in En- et al., 2017), and identify words that are close to glish and Maori.¯ Bookbird: A Journal of Interna- our loanwords in the semantic space. We predict tional Children’s Literature, 54(3):10–17. that these neighbouring words will enable us to un- Amitava Das and Bjorn¨ Gamback.¨ 2014. Identifying derstand the semantic make-up of our loanwords languages at the word level in code-mixed Indian so- according to their usage. cial media text. Finally, we hope to extrapolate these findings Carolyn Davies and Margaret Maclagan. 2006. Maori¯ by deploying our trained classifier on other online words–read all about it: Testing the presence of 13 discourse sources, such as Reddit posts. This has maori¯ words in four New Zealand newspapers from 1997 to 2004. Te Reo, 49. great potential for enriching our understanding of how Maori¯ loanwords are used in social media. Julia De Bres. 2006. Maori lexical items in the mainstream television in New Zealand. New 6 Acknowledgements Zealand English Journal, 20:17. Marta Degani. 2010. The Pakeha myth of one New The authors would like to thank former Hon- Zealand/Aotearoa: An exploration in the use of ours student Nicole Chan for a preliminary study Maori loanwords in New Zealand English. From in- on Maori¯ Loanwords in Twitter. Felipe Bravo- ternational to local English–and back again, pages Marquez was funded by Millennium Institute for 165–196. Foundational Research on Data. Andreea S. Marta Degani and Onysko. 2010. Hybrid Calude acknowledges the support of the NZ Royal compounding in New Zealand English. En- Society Marsden Grant. David Trye acknowl- glishes, 29(2):209–233. edges the generous support of the Computing and Tony Deverson. 1991. New Zealand English : the Mathematical Sciences group at the University of Maori dimension. English Today, 7(2):18–25. Waikato. Jack Grieve, Nini, and Diansheng Guo. 2017. Analyzing lexical emergence in Modern American English online. English Language & Linguistics, References 21(1):99–127. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Tomas Mikolov. 2017. Enriching word vectors with Pfahringer, Peter Reutemann, and Ian Witten. subword information. Transactions of the Associa- 2009. The WEKA data mining software: an update. tion for Computational Linguistics, 5:135–146. ACM SIGKDD explorations newsletter, 11(1):10– 18. Felipe Bravo-Marquez, Eibe Frank, Bernhard Jennifer Hay. 2018. What does it mean to “know a Pfahringer, and Saif . Mohammad. 2019. word?”. In Language and Society Conference of NZ AffectiveTweets: a Weka package for analyzing in November 2018 in , NZ, Wellington, affect in tweets. Journal of Machine Learning NZ. Research, 20:1–6. Yuan Huang, Diansheng Guo, Alice Kasakoff, and Jack Andreea Simona Calude, Steven , and Mark Grieve. 2016. Understanding US regional linguistic Pagel. 2017. Modelling loanword success–a soci- variation with twitter data analysis. Computers, En- olinguistic quantitative study of Maori¯ loanwords in vironment and Urban Systems, 59:244–255. New Zealand English. and Lin- guistic Theory. Te Taka Keegan, Paora Mato, and Stacey Ruru. 2015. Using Twitter in an indigenous language: An analy- Ozlem¨ C¸etinoglu. 2016. A Turkish-German code- sis of Te Reo Maori¯ tweets. AlterNative: An Inter- switching corpus. In International Conference on national Journal of , 11(1):59– Language Resources and Evaluation. 75.

141 Graeme Kennedy and Shunji Yamazaki. 1999. The in- Clare Voss, Stephen Tratz, Jamal Laoudi, and Dou- fluence of Maori on the Nw Zealand English lexicon. glas M Briesch. 2014. Finding Romanized Arabic LANGUAGE AND COMPUTERS, 30:33–44. dialect in code-mixed tweets. In LREC, pages 2249– 2253. John Macalister. 2006a. The Maori lexical presence in New Zealand English: Constructing a corpus for Eline Zenner, Dirk Speelman, and Dirk Geeraerts. diachronic change. Corpora, 1(1):85–98. 2012. Cognitive Sociolinguistics meets loan- word research: Measuring variation in the suc- John Macalister. 2006b. the Maori presence in the cess of anglicisms in Dutch. Cognitive Linguistics, New Zealand English lexicon, 1850–2000: Evi- 23(4):749–792. dence from a corpus-based study. English World- Wide, 27(1):1–24.

John Macalister. 2009. Investigating the changing use of Te Reo. NZ Words, 13:3–4.

Bruno Martins and Mario´ Silva. 2005. Language identification in web pages. In Proceedings of the 2005 ACM symposium on Applied computing, pages 764–768. ACM.

Andrew McCallum, Kamal Nigam, et al. 1998. A com- parison of event models for naive Bayes text classi- fication. In AAAI-98 workshop on learning for text categorization, volume 752, pages 41–48. Citeseer.

Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Cor- rado, and Jeff Dean. 2013. Distributed representa- tions of words and phrases and their compositional- ity. In Advances in neural information processing systems, pages 3111–3119.

Alexander Onysko and Andreea Calude. 2013. Com- paring the usage of Maori¯ loans in spoken and - ten New Zealand English: A case study of Maori,¯ Pakeh¯ a,¯ and Kiwi. New perspectives on lexical borrowing: Onomasiological, methodological, and phraseological innovations, pages 143–170.

Tatjana Scheffler. 2014. A German twitter snapshot. In LREC, pages 2284–2289. Citeseer.

Edgar Schneider. 2003. The dynamics of New En- glishes: From identity construction to dialect birth. Language, 79(2):233–281.

Maria Sifianou. 2015. Conceptualizing politeness in Greek: Evidence from twitter corpora. Journal of Pragmatics, 86:25–30.

Mehmet Ulvi S¸ims¸ek and Suat Ozdemir.¨ 2012. Analy- sis of the relation between Turkish twitter messages and stock market index. In 2012 6th International Conference on Application of Information and Com- munication Technologies (AICT), pages 1–4. IEEE.

Patrick Smyth. 1946. Maori Pronunciation and the Evolution of Written Maori. Whitcombe & Tombs Limited.

Jonathan R Stammers and Margaret Deuchar. 2012. Testing the nonce borrowing hypothesis: Counter- evidence from English-origin in Welsh. Bilin- gualism: Language and Cognition, 15(3):630–643.

142