RNA-Seq based transcriptome assembly has been widely used to identify novel lncRNAs. However, the best-performing transcript reconstruction methods merely identified 21% of full-length protein-coding transcripts from H. sapiens. Those partial-length protein-coding transcripts are more likely to be classified as lncRNAs due to their incomplete CDS, leading to higher false positive rate for lncRNA identification. Furthermore, potential sequencing or assembly error that gain or abolish stop codons also complicates ORF-based prediction of lncRNAs. Therefore, it remains a challenge to identify lncRNAs from the assembled transcripts, particularly the partial-length ones. Here, we present a novel alignment-free tool, lncScore, which uses a logistic regression model with 11 carefully selected features. Compared to other state-of-the-art alignment-free tools (e.g. CPAT, CNCI, and PLEK), lncScore outperforms them on accurately distinguishing lncRNAs from mRNAs, especially partial-length mRNAs in the human and mouse datasets. In addition, lncScore also performed well on transcripts from five other species (Zebrafish, Fly, C. elegans, Rat, and Sheep). To speed up the prediction, multithreading is implemented within lncScore, and it only took 2 minute to classify 64,756 transcripts and 54 seconds to train a new model with 21,000 transcripts with 12 threads, which is much faster than other tools. lncScore is available at https://github.com/WGLab/lncScore.