6. Topic ExtractionΒΆ
As described in section dsm-label, word-embeddings provide an efficient way to model semantic simliarity between words. A DSM-based word-embedding maps words, which are semantically, related to similar vectors. Now, we are interested in semantically related documents. We like to find out which documents are semantically related in the sense that they refer to similar topics. We already know, that documents can be modelled by BoW-vectors and that we can determine similarity between documents by just calculating the similarity (or distance) between the corresponding BoW-vectors. However, the problem with this approach is, that we can describe the same topic with different words. For example the sentences The president visited Europe means essentially the same as the sentence The head of state had a trip to EU capitals. Applying BoW both sentences would have totally different vectors, even though they describe the same topic.
In NLP, the goals of topic extraction are
given a corpus of documents, find out which topics are discussed in these documents
find out which documents describe which topics
topics may be better features than words. Therefore topic extraction may be applied as preprocessing for other tasks, such as text-classification.
Algorithms for topic extraction are e.g. Latent Semantic Indexing (LSI), Non-Negative Matrix Factorisation (NNMF) or Latent Dirichlet Allocation. In this notebook LSI is described.