As a fundamental task in natural language processing, language modeling aims to estimate the distribution of the word sequences. However, most existing algorithms have focused on the main text while often ignoring the vastly-accessible semantic attributes, e.g., titles, authors, sentiments and tags. To address this issue, we build three text datasets with a diversity of semantic attributes, and propose Semantic Attribute Modulated (SAM) language modeling, a novel language modeling framework that incorporates various attributes. Attributes are selected automatically with an attribute attention mechanism. We empirically examine the language model perplexities of several typical corpora, and demonstrate the superiority of our model with different combinations of the attributes. Lyric generation by taking both the title and author attributes into account further demonstrates the effectiveness of our method.
View on arXiv