This work is motivated by the emergence of social sensing as a new paradigm of collecting observations about the physical environment from humans or devices on their behalf. These observations may be true or false, and hence are viewed as binary claims. A fundamental problem in social sensing applications lies in ascertaining the correctness of claims and the reliability of data sources without knowing either of them a priori. We refer to this problem as truth discovery. Prior works have made significant progress to addressing the truth discovery problem, but two significant limitations exist: (i) they ignored the fact that claims reported in social sensing applications can be either relevant or irrelevant to the topic of interests. (ii) They either assumed the data sources to be independent or the source dependency graphs can be represented as a set of disjoint trees. These limitations led to suboptimal truth discovery results. In contrast, this paper presents the first social sensing framework that explicitly incorporates the topic relevance feature of claims and arbitrary source dependency graphs into the solutions of truth discovery problem. The new framework solves a multidimensional maximum likelihood estimation problem to jointly estimate the truthfulness and topic relevance of claims as well as the reliability and topic awareness of sources. We compared our new scheme with the state-of-the-art truth discovery solutions using three real world data traces collected from Twitter in the aftermath of Paris Shooting event (2015), Hurricane Arthur (2014) and Boston Bombing event (2013) respectively. The evaluation results showed that our schemes significantly outperform the compared baselines by identifying more relevant and truthful claims in the truth discovery results.