In the last decade, online social networks have enabled people to interact in many ways with each other and with content. The digital traces of such actions reveal people's preferences towards online content such as news or products. These traces often result from interactions such as sharing or liking, but also from interactions in natural language. The continuous growth of the amount of content and of digital traces has led to information overload: surrounded by large volumes of information, people are facing difficulties when searching for information relevant to their interests. To improve user experience, information systems must be able to assist users in achieving their search goals, effectively and efficiently. This thesis is concerned with two important challenges that information systems need to address in order to significantly improve search experience and overcome information overload. First, these systems need to model accurately the variety of user traces, and second, they need to meaningfully explain search results and recommendations to users. To address these challenges, this thesis proposes novel methods based on machine learning to model user sentiment and preferences for information filtering systems, which are effective, scalable, and easily interpretable by humans. We focus on two prominent types of user traces in social networks: on the one hand, user comments accompanied by unary preferences such as likes, and on the other hand, user reviews accompanied by numerical preferences such as star ratings. In both cases, we advocate that by better understanding user text through mining its semantics and modeling its structure, we can not only improve information filtering, but also explain predictions to users. Within this context, we aim to answer three main research questions, namely: (i)~how do item semantics help to predict unary preferences; (ii)~how do sentiments of free-form user texts help to predict unary preferences; and (iii)~how to model fine-grained numerical preferences from user review texts. Our goal is to model and extract from user text the knowledge required to answer these questions, and to obtain insights on how to design better information filtering systems that are more effective and improve user experience. To answer the first question, we formulate the recommendation problem based on unary preferences as a top-N retrieval task and we define an appropriate dataset and metrics for measuring performance. Then, we propose and evaluate several content-based methods based on semantic similarities under presence or absence of preferences. To answer the second question, we propose a sentiment-aware neighborhood model which integrates the sentiment of user comments with unary preferences, either through fixed or through learned mapping functions. For the latter type, we propose a learning algorithm which adapts the sentiment of user comments to unary preferences at collective or individual levels. To answer the third question, we cast the problem of modeling user attitude toward aspects of items as a weakly supervised problem, and we propose a weighted multiple-instance learning method for solving it. Lastly, we show that the learned saliency weights, apart from being easily interpretable, are useful indicators for review segmentation and summarization.