The study of long non-coding RNAs (lncRNAs), greater than 200 nucleotides, is central to understanding the development and progression of many complex diseases. Unlike proteins, the functionality of lncRNAs is only subtly encoded in their primary sequence. Current in-silico lncRNA annotation methods mostly rely on annotations inferred from interaction networks. But extensive experimental studies are required to build these networks. In this work, we present a graph-based machine learning method called FGGA-lnc for the automatic gene ontology (GO) annotation of lncRNAs across the three GO subdomains. We build upon FGGA (factor graph GO annotation), a computational method originally developed to annotate protein sequences from non-model organisms. In the FGGA-lnc version, a coding-based approach is introduced to fuse primary sequence and secondary structure information of lncRNA molecules. As a result, lncRNA sequences become sequences of a higher-order alphabet allowing supervised learning methods to assess individual GO-term annotations. Raw GO annotations obtained in this way are unaware of the GO structure and therefore likely to be inconsistent with it. The message-passing algorithm embodied by factor graph models overcomes this problem. Evaluations of the FGGA-lnc method on lncRNA data, from model and non-model organisms, showed promising results suggesting it as a candidate to satisfy the huge demand for functional annotations arising from high-throughput sequencing technologies.