The Latent Dirichlet Allocation (LDA) model has been recently used as a method of identifying latent aspects in customer reviews. In our previous work, we proposed Twofold-LDA to identify both aspects and positive or negative sentiment in review sentences. We incorporated domain knowledge (i.e. seed words) to produce more focused aspects and provided a user-friendly chart quantifying sentiment. Our previous work made an assumption that one sentence contains one aspect, but in this study we wish to extend our model to remove this assumption. Experimental results show that our extended model improves the performance for every aspect in the datasets. We also show the importance of seed words for identifying desired aspects.
|Title of host publication||Advances in Computational Intelligence Communications in Computer and Information Science|
|Place of Publication||Switzerland|
|Number of pages||11|
|Publication status||Published - 2012|