#compling

you’re in the midst of writing an email when suddenly, in grey text, google suggests the next sentence. mind reading? actually, it’s an application of natural language inference (nli) , the process by which computers determine if one sentence “entails” another. while google might seem to predict your email structure with uncanny accuracy, it can only handle the simple stuff; “let me know...”, “...does that work for you?”, and other phrases found in millions of emails. so at what point do these programs fail, and how can we fix it? that’s what computational linguists r. thomas mccoy, ellie pavlick, and tal linzen are trying to find out. 

Like, what is the distribution?  Is it the probability of the term appearing an in n-gram?  What size n-gram?  What else is in the n-gram?  Is it the probability of the term appearing in the document?  If so, what do n-grams have to do with this?  Like give me the formula so I know what we’re calculating here, mutual information is like the vaguest statistic ever without context

this is a long shot, but i was wondering if anyone knows of italian universities which offer computational linguistics courses?

i want to do an erasmus in italy but i can’t decide on a place because none seem to have everything i want OR specifically a comp ling program.

so any italian computational linguists here who can offer some insight? 

Gender Identity and Lexical Variation in Social Media

By: David Bamman, Jacob Eisenstein and Tyler Schnoebelen

Published by: Journal of Sociolinguistics Volume 18, Issue 2 April 2014 Pages 135–160

LL Abstract:

With a blend of quantitative and qualitative analysis, Bamman, Eisenstein and Schnoebelen explain their study of the relationship between gender, linguistic style, and social networks. They cluster Twitter users by lexical items instead of a male-female binary distinction, discovering that many clusters have strong gender orientations but conflict with the overall language stats found in the general population in previous research. The authors then train a machine-learning statistical classifier and measure the confidence with which it classifies individuals. The findings show that social network homophily is correlated with the use of same-gender markers, while individuals in general position themselves according to certain norms in order perform gender.

LL Summary:

The article begins with a brief overview of the study, a computational analysis of the impact of gender on lexical choices and social networks in a novel corpus of 14,000 individuals on Twitter. The authors continue with a background of the topic of gender in sociolinguistics, mentioning the theoretical outlook of gender as being constructed, maintained, and disrupted by linguistic practices. Computational analysis has looked at large data sets quantitatively to identify which words are the most accurate predictors of ‘attributes’ like age, gender, and regional origin. One Argamon et. al (2007) analyzed blogs and built a predicative model of gender that was more accurate than human judges. Computational literature distinguishes men and women in this way on the dimensions of ‘informativeness’ (resources that communicate propositions associated with men) and ‘involvement’ (resources that create interactions between speakers and environment associated with women). However, the complex role of gender in identity poses problems for quantitative analysis that only treats gender as an independent variable. The goal of this article is to bring the theory of resources having indexical fields of meaning used to create various stances and personae. In other words, social categories are performances by speakers, depending on the situational context. The next section of the article describes the data, collected from Twitter due to its public nature and ease of collection from the streaming API. After filtering for social network and gender and name distributions from census data, the data set contained over nine million tweets from 14,464 users. The first analysis was on the lexical markers of gender, a standard computational approach of dividing the tweets by gender and training a logistic regression classifier to identify gender (the dependent variable) by the independent variable (the 10,000 most frequent lexical items in the corpus). The classifier had 88 percent accuracy, and this analysis determined words most strongly associated with each gender. Pronouns, emotion terms, kinship terms, abbreviations, and some assent and negation terms were associated with female authors. Swears and taboo words were found to be more often associated with male authors. Overall these findings were similar to those of previous research, but this analysis determined eight categories of word classification. Each word was categorized in only one category depending on the salience of the ordered list of all eight categories. All differences between genders in these eight categories were statistically significant, but the direct association of word types with social categories is problematic. In the next section, the authors address this by using a quantitative method of clustering users according to similar lexical items or sets of words. Probabilistic clustering grouped linguistically similar authors together without regarding gender, resulting in seventeen clusters with fourteen which still had strong gender orientations. This organization showed the importance of intersectionality, or the impossibility of pulling different dimensions of social life into separate strands. These clusters are more like verbal repertoires, a complex interaction of social positions. Several differences between clusters reversed the earlier computational findings, suggesting that social categories like gender could not be separated from other aspects of identity. The data in this study exhibited gender homophily, the theory that people prefer similar people in their social networks. The authors also found a strong correlation between individuals with a greater proportion of same-gender ties and the use of more gendered language. They trained a classifier to identify gender based on the gender skew of social network, and it was found that the higher the gender skew, the more confident the classifier’s prediction would be. The article concludes with a discussion of building a data-driven model, how gender is constructed, and how machine learning in large data sets should be further investigated.

LL Recipe Comparison:

With its detailed convergence of fields of linguistics, this article is most like Linguine with Scallops Red Pepper and Broccolini:

In their article, the authors blend computational and sociolinguistic theories of gender and lexical use to examine how Twitter users construct gender through language. In this recipe, the fresh vegetables blend deliciously well with the fried scallops. I highly recommend both!