We present a study of the relationship between gender, linguistic style, and social networks, using a novel corpus of 14,000 Twitter users. Prior quantitative work on gender often treats this social variable as a female/male binary; we argue for a more nuanced approach. By clustering Twitter users, we find a natural decomposition of the dataset into various styles and topical interests. Many clusters have strong gender orientations, but their use of linguistic resources sometimes directly conflicts with the population-level language statistics. We view these clusters as a more accurate reflection of the multifaceted nature of gendered language styles. Previous corpus-based work has also had little to say about individuals whose linguistic styles defy population-level gender patterns. To identify such individuals, we train a statistical classifier, and measure the classifier confidence for each individual in the dataset. Examining individuals whose language does not match the classifier's model for their gender, we find that they have social networks that include significantly fewer same-gender social connections and that, in general, social network homophily is correlated with the use of same-gender language markers. Pairing computational methods and social theory thus offers a new perspective on how gender emerges as individuals position themselves relative to audiences, topics, and mainstream gender norms.Comment: submission versio

With Twitter and Facebook blocked in China, the stream of information from Chinese domestic social media provides a case study of social media behavior under the influence of active censorship. While much work has looked at efforts to prevent access to information in China (including IP blocking of foreign websites or search engine filtering), we present here the first large-scale analysis of political content censorship in social media, i.e., the active deletion of messages published by individuals. In a statistical analysis of 56 million messages (212,583 of which have been deleted out of 1.3 million checked, more than 16%) from the domestic Chinese microblog site Sina Weibo, and 11 million Chinese-language messages from Twitter, we uncover a set a politically sensitive terms whose presence in a message leads to anomalously higher rates of deletion. We also note that the rate of message deletion is not uniform throughout the country, with messages originating in the outlying provinces of Tibet and Qinghai exhibiting much higher deletion rates than those from eastern areas like Beijing.

Adversarial training is a mean of regularizing classification algorithms by generating adversarial noise to the training data. We apply adversarial training in relation extraction within the multi-instance multi-label learning framework. We evaluate various neural network architectures on two different datasets. Experimental results demonstrate that adversarial training is generally effective for both CNN and RNN models and significantly improves the precision of predicted relations.

We consider the problem of automatically inferring latent character types in a collection of 15,099 English novels published between 1700 and 1899. Unlike prior work in which character types are assumed responsible for probabilistically generating all text associated with a character, we introduce a model that employs multiple effects to account for the influence of extra-linguistic information (such as author). In an empirical evaluation, we find that this method leads to improved agreement with the preregistered judgments of a literary scholar, complementing the results of alternative models.

We introduce a model for incorporating contextual information (such as geography) in learning vector-space representations of situated language. In contrast to approaches to multimodal representation learning that have used properties of the object being described (such as its color), our model includes information about the subject (i.e., the speaker), allowing us to learn the contours of a word's meaning that are shaped by the context in which it is uttered. In a quantitative evaluation on the task of judging geographically informed semantic similarity between representations learned from 1.1 billion words of geo-located tweets, our joint model outperforms comparable independent models that learn meaning in isolation.