Gensim is a free Python library designed to automatically extract semantic topics from documents, as efficiently (computer-wise) and painlessly (human-wise) as possible.
Gensim is designed to process raw, unstructured digital texts (“plain text”). The algorithms in gensim, such as Latent Semantic Analysis, Latent Dirichlet Allocation and Random Projections discover semantic structure of documents by examining statistical co-occurrence patterns of the words within a corpus of training documents. These algorithms are unsupervised, which means no human input is necessary – you only need a corpus of plain text documents.
Once these statistical patterns are found, any plain text documents can be succinctly expressed in the new, semantic representation and queried for topical similarity against other documents.
Note
If the previous paragraphs left you confused, you can read more about the Vector Space Model and unsupervised document analysis on Wikipedia.
The creation of gensim was motivated by a perceived lack of available, scalable software frameworks that realize topic modelling, and/or their overwhelming internal complexity (hail Java!). You can read more about the motivation in our LREC 2010 workshop paper. If you want to cite gensim in your own work, please refer to that article (BibTeX).
You’re welcome to share your results and experiments on the mailing list.
The principal design objectives behind gensim are:
See also
If you’re interested in document indexing/similarity retrieval, I also maintain a higher-level package of document similarity server. It uses gensim internally.
Gensim is licensed under the OSI-approved GNU LGPLv2.1 license and can be downloaded either from its github repository or from the Python Package Index.
See also
See the install page for more info on gensim deployment.
The whole gensim package revolves around the concepts of corpus, vector and model.
In the Vector Space Model (VSM), each document is represented by an array of features. For example, a single feature may be thought of as a question-answer pair:
The question is usually represented only by its integer id (such as 1, 2 and 3 here),
so that the
representation of this document becomes a series of pairs like (1, 0.0), (2, 2.0), (3, 5.0)
.
If we know all the questions in advance, we may leave them implicit
and simply write (0.0, 2.0, 5.0)
.
This sequence of answers can be thought of as a vector (in this case a 3-dimensional vector). For practical purposes, only questions to which the answer is (or
can be converted to) a single real number are allowed.
The questions are the same for each document, so that looking at two vectors (representing two documents), we will hopefully be able to make conclusions such as “The numbers in these two vectors are very similar, and therefore the original documents must be similar, too”. Of course, whether such conclusions correspond to reality depends on how well we picked our questions.
Typically, the answer to most questions will be 0.0
. To save space,
we omit them from the document’s representation, and write only (2, 2.0),
(3, 5.0)
(note the missing (1, 0.0)
).
Since the set of all questions is known in advance, all the missing features
in a sparse representation of a document can be unambiguously resolved to zero, 0.0
.
Gensim does not prescribe any specific corpus format; a corpus is anything that, when iterated over, successively yields these sparse vectors. For example, set([(2, 2.0), (3, 5.0)], ([0, -1.0], [3, -1.0])) is a trivial corpus of two documents, each with two non-zero feature-answer pairs.
We use model as an abstract term referring to a transformation from one document representation to another. In gensim documents are represented as vectors so a model can be thought of as a transformation between two vector spaces. The details of this transformation are learned from the training corpus.
For example, consider a transformation that takes a raw count of word occurrences and weights them so that common words are discounted and rare words are promoted. The exact amount that any particular word is weighted by is determined by the relative frequency of that word in the training corpus. When we apply this model we transform from one vector space (containing the raw word counts) to another (containing the weighted counts).
See also
For some examples on how this works out in code, go to tutorials.