Multilingual Clusters and Gender in Nordic Twitter

Steven Coats
English Philology, Faculty of Humanities
steven.coats@oulu.fi

CLARIN-PLUS Workshop “Creation and Use of Social Media Resources”
May 19th, 2017

Goals of research

Language use, bi- and multilingualism and gender: Small but significant differences in grammatical feature use according to gender have been found in large-scale corpus studies, including in online and social media (e.g. Argamon et al. 2007, Bamann et al. 2014, Coats 2016)

Are analogous gender-based differences evident at a higher level of discourse organization (language choice, bi- and multilingualism)?

Explore extent of language use by gender on Twitter for the Nordic countries

Investigate similarities or differences by gender among bi- and multilingual Nordic Twitter users

Data collection | Streaming API and gender disambiguation

Tweets with populated place attributes collected using Tweepy (Roesslein 2015) from the Twitter Streaming API 9 November 2016 – 19 April 2017 (~530m tweets)
Usernames of 162,035 unique authors of tweets with Nordic country_code fields (Greenland, Iceland, Faroe Islands, Norway, Denmark, Sweden, Åland, Finland) identified
Name frequency information for 15,284 names given to males and 17,763 names given to females obtained from Nordic statistical offices and used to induce a probability distribution for name gender

\[ P(X=x_{m}) = \frac{\sum_{\omega: X(\omega)=x_m}}{\sum_{\omega: X(\omega)=x_m}+\sum_{\omega: X(\omega)=x_f}} \]

Gender assigned to users whose author_name values either begin with or include as a discrete element Nordic names with \(P(x_{m},x_{f})>\) 0.8
Method assigned gender to 102,513 users

Data collection | REST API and source filtering

Timelines (up to 3250 tweets) downloaded from the Twitter REST API for each gendered user
Users with \(>\) 50% Nordic place values and at least 20 tweets retained in data
tweet_source filtered for sources more likely to be used by human agents to prevent (some) automated tweets sent by apps such that send English-language messages about music playlists, user location, workout details, weather, etc.

Data collection | Language detection

Languages of individual tweets were identified with the following procedure:
Remove sequences that can cause inaccurate language detection (usernames, URLs, hashtags, emojis)
Require three-way agreement: Retain only tweets whose language as reported by the native Twitter algorithm agrees with the language as reported by both the compact language detector 2 (Sites 2014) and langid (Lui and Baldwin 2012)
Working dataset: 25,509,920 tweets by 44,906 unique authors

Tweet density by language

Map polygons from Natural Earth, maps from Open Street Maps

Use of language by country and gender

For each user, the proportion of all of his/her tweets in each of the 87 languages in the data was calculated
For each of the five countries under consideration, the male-female balance was determined with a t-test of population means
To account for unequal variance in the population distributions for some languages, 95% confidence intervals were constructed by using stratified non-parametric bootstrapping based on 10,000 resamples
The 28 most frequently occurring languages overall are shown here

Use of language by country and gender: Iceland

Use of language by country and gender: Norway

Use of language by country and gender: Denmark

Use of language by country and gender: Sweden

Use of language by country and gender: Finland

Summary language use by gender

Males use more Arabic, Farsi, Turkish; females use more Thai, Tagalog, Japanese, Russian
These facts may represent gender differences in patterns of migration to the Nordics

Iceland, Norway and Denmark: Males use more national languages, females use more English
Sweden and Finland: Females use more national languages, males use more English
These facts may represent language shift

Labovian sociolinguistics: In changing linguistic situations, females use prestige markers more than males (1990)

Quantifying bi- and multilingualism

Users were defined as bi- or multilingual if for at least 2 languages each comprised \(\geq\) 10% of the total number of his/her tweets
In this data, males are more bi- or multilingual than females

Males: 41.3% bi- or multilingual; Females: 33.5% bi- or multilingual
Difference significant at p = 5.107e-16 (Fisher’s Exact Test, two-tailed)

Creating networks of bi- and multilinguals

To quantify the strength of connections between languages on Twitter in the Nordics, networks of bi- and multilingual users were created

For each language pair \((i,j)\) a contingency table of the number of bilinguals was set up:


	\(language_{i}\)	~\(language_{i}\)
\(language_{j}\)	\(O_{11}\)	\(O_{12}\)	\(= R_{1}\)
~\(language_{j}\)	\(O_{21}\)	\(O_{22}\)	\(= R_{2}\)
	\(= C_{1}\)	\(= C_{2}\)	\(= N\)

From this table a correlation coefficient \(\phi\) can be calculated:

\[ \phi_{ij}= \frac{(O_{11}O_{22}-O_{12}O_{21})}{\sqrt{R_{1}R_{2}C_{1}C_{2}}} \]

\(\phi\) ranges in value from \(-1\) to \(1\). A t-statistic and p-values were calculated for the \(\phi\) scores

Bilinguals’ use of language by country and gender: Iceland

Bilinguals’ use of language by country and gender: Norway

Bilinguals’ use of language by country and gender: Denmark

Bilinguals’ use of language by country and gender: Sweden

Bilinguals’ use of language by country and gender: Finland

Clusters for all Nordic males and females

Significance of gender differences

A z-score was calculated via a Fisher transform of \(\phi\) values according to the formula:

\[ z_{(i,j)} = \frac{z_{\phi_{(i,j)_{m}}}-z_{\phi_{(i,j)_{f}}}}{\sqrt{\frac{1}{n_{(i,j)_{m}}-3}+{\frac{1}{n_{(i,j)_{f}}-3}}}} \]

(Sheskin 2000: 792)

For languages \((i,j)\):

\(z_{\phi_{(i,j)_{m}}}\) is the Fisher transformed value of \(\phi\) for males

\(z_{\phi_{(i,j)_{f}}}\) is the Fisher transformed value of \(\phi\) for females

\(n_{(i,j)_{m}}\) and \(n_{(i,j)_{f}}\) are the number bilingual speakers in the male and female networks.

Summary of gender differences

Preliminary summary and conclusions

Aggregate language usage statistics may reflect both patterns of international mobility as well as cultural associations of languages/language learner preferences
English is the central language of bilingualism for both males and females
Males are more likely to be bilingual overall
Comparison of phi values somewhat supports sociolinguistic interpretation (more data needed)
There is some evidence for higher idea density (i.e. longer tweets) for bilinguals

Acknowledgements

Thanks to

Dorthe Larsen (Danmarks Statistik) for Danish name lists
Joar Skott (Statistika centralbyrån, Sweden) for Swedish name lists
Finland’s Center for Scientific Computing for data storage and computing resources

References

Argamon, S., M. Koppel, J. W. Pennebaker and J. Schler. 2007. Mining the blogosphere: Age, gender, and the varieties of self-expression. First Monday 12/9. http://pear.accc.uic.edu/ojs/index.php/fm/article/view/2003/1878

Bamann, D., J. Eisenstein and T. Schnoebelen. 2014. Gender identity and lexical variation in social media. Journal of Sociolinguistics 18(2), 135–160. http://onlinelibrary.wiley.com/doi/10.1111/josl.12080/full

Coats, S. 2016. Grammatical feature frequencies of English on Twitter in Finland. In L. Squires (ed.), English in computer-mediated communication: Variation, representation, and change. Berlin: De Gruyter. 179–210.

Labov, W. 1990. “The intersection of sex and social class in the course of linguistic change.” Language Variation and Change 2, 205–254.

Lui, M. and T. Baldwin. 2012. “Langid.py: An off-the-shelf language identification tool”. 50th Proceedings of the Association for Computational Linguistics, 25–30. Stroudsburg, PA: ACL. http://dl.acm.org/citation.cfm?id=2390475

Roesslein, J. 2015. Tweepy. Python package [Computer software]. http://www.tweepy.org

Sheskin, D. 2000. Handbook of parametric and non-parametric statistical procedures, 2nd ed. Boca Raton: Chapman and Hall.

Sites, D. 2013. Compact language detector 2. https://github.com/CLD2Owners/cld2