DescriptionUsers rely increasingly on online reviews, forums and blogs to exchange information, practical tips, and stories. Such social interaction has become central to their daily decision-making processes. However, user-authored content is in a free-text format, usually with very scant structured metadata information. Users often face the daunting task of reading a large quantity of text to discover potentially useful information. This thesis addresses the need to automatically leverage information from user-authored text to improve search and to provide personalized recommendations matching user preferences. We first focus on developing accurate text-based recommendations. We the rich information present in user reviews by identifying the review parts pertaining to different product features and the sentiment expressed towards each feature. We derive text-based ratings which serve as alternate indicators of user assessment of the product. We then cluster similar users based on the topics and sentiments in their reviews. Our results show that using text yields better user preference predictions than those from the coarse star ratings. We also make fine-grained predictions of user sentiments towards the individual product features. In the interactive and social forum sites users frequently make connections with other users, enabling them to find the right person to answer their questions. A challenge then, is to score and rank the short snippets of forum posts while taking into account these personal connections. In this thesis, we learn user similarities via multiple indicators like shared information needs, profiles, or topics of interest. We develop a novel multidimensional model that uniformly incorporates the heterogeneous user relations in finding similar participants, to predict future social interactions and enhance keyword search. Search over user-authored data like forums requires providing results that are as complete as possible and yet are focused on the relevant information. We address this problem by developing a new search paradigm that allows for search results to be retrieved at varying granularity levels. We implement a novel hierarchical representation and scoring technique for objects at multiple granularities. We also present a score optimization algorithm that efficiently chooses the best $k$-sized non-overlapping result set. We conduct extensive user studies and show that a mixed granularity set of results is more relevant to users than standard post-only approaches. In summary, this thesis studies the problems in understanding user behavior from textual content in online reviews and forums. We present efficient techniques to learn user preferences and similarities to enhance search and recommendations.