With the staggering growth of language models in the last few years, language technology is rapidly taking over some of the most influential procedures in modern society such as recruitment, teaching, business, legislation and legal systems. For example, instead of hiring a slow human worker to pore over hundreds of resumes in a job opening, an automatic resume analyzer can do it in a matter of minutes. Instead of wasting time and money in expensive lawsuits and trials, language models can analyze evidence and build adequate argumentation for defendants in court. 

The recent success of language models owes to two major factors: (i) their massive size reaching hundreds of billions of parameters such as GPT3 or ChatGPT, and (ii) the smart notion of pretraining them on colossal textual corpora with very little annotation and curation. Although pretraining on unlabeled datasets facilitated the adoption of human language by models, it also made it easy for them to absorb harmful subjective beliefs contained in those corpora. Indeed, a growing body of research is warning that language models inherited a large swath of human social biases and stereotypes from datasets. As a result, language models run the risk of siding with male applicants in job offers (because of the stereotype casting men as more competent and skillful than women); discriminating against people of color in court (because of the stereotype casting Blacks as supporters of crime and violence); not to mention the risk of propagating these stereotypes to kids when language models are used in teaching settings. In this thesis, we aim to characterize and measure social bias encoded in language models, and quantify the discrimination damage when these models are employed in downstream applications. Also, we propose three novel methods to reduce the amount of bias from language models: BiasMeter, ADV-Debias and AttenD operating on data, text embeddings and the attention mechanism respectively. 

In contrast to stereotypes, subjectivity can sometimes be beneficial to language models. For example, a task-oriented conversational agent can make use of subjective attributes in user utterances to enable subjective search. Also, subjectivity can enhance opinion and emotion mining from online reviews. Previous research shows that failing to explicitly model subjectivity in user-facing language technology such as chatbots and search ultimately results in user dissatisfaction. In this thesis, we focus on search and textual similarity, and propose methods to augment them with subjectivity. Be it for desired (subjective attributes) or undesired subjectivity (bias, stereotypes and prejudice), we provide extensive evaluation and validation of the proposed techniques.