Following the application of data-processing procedures, it is possible to compare feature vectors, corpora and documents.
Comparing feature vectors
A vector can be thought of either as
- the co-ordinates of a point
- describing a line segment from the origin to a point
The view of a vector as a line segment starting at the origin is useful, as any two vectors will have an angle between them, corresponding to their similarity, as calculated by cosine similarity.
The cosine similarity of two vectors is the dot product of two vectors over the product of their magnitudes. It is a standard distance metric for comparing documents.
A quick way to compare corpora is to find words common to the whole dataset, but with a strong affinity to only one corpus. This is a function of how much higher their frequency is in that corpus than in the dataset.
Terms’ comparative affinities to two corpora
corpus2 are tables of lists of documents, returns a dictionary of terms and their affinity for
Enron CEO Jeff Skillings was a member of the Beta Theta Pi fraternity at Southern Methodist University (SMU). If we want to find secret fraternity code words used by the Betas, we can compare his fraternity emails (those containing SMU or Betas) to his other emails.
q)fraternity:jeffcorpus i:where (jeffcorpus[`text] like "*Betas*")|jeffcorpus[`text] like "*SMU*" q)remaining:jeffcorpus til[count jeffcorpus]except i q)summaries:key each 10#/:.nlp.compareCorpora[fraternity;remaining] q)summaries 0 / summary of the fraternity corpus `beta`homecoming`betas`smu`yahoo`groups`tent`reunion`forget`crowd q)summaries 1 / summary of the remaining corpus `enron`jeff`business`information`please`market`services`energy`management`company
This function allows you to calculate the similarity of two different documents. It finds the keywords that are present in both the corporas, and calculates the cosine similarity.
Cosine similarity of two documents
dict2 are dictionaries that consist of the document‘s keywords, returns the cosine similarity of two documents.
Given the queried email defined above, and a random email from the corpus, we can calculate the cosine similarity between them.
q)queryemail2:jeffcorpus[rand count jeffcorpus] q).nlp.compareDocs[queryemail`keywords;email2`keywords] 0.1163404
Comparing documents to corpus
Cosine similarity between a document and other documents in the corpus
keywordsis a list of dictionaries of keywords and coefficients
idxis the index of the feature vector to compare with the rest of the corpus
returns as a float the document’s significance to the rest of the corpus.
Comparing the first chapter with the rest of the book:
q).nlp.compareDocToCorpus[corpus`keywords;0] 0.03592943 0.04720108 0.03166343 0.02691693 0.03363885 0.02942622 0.03097797 0.0408502..