Latent semantic analysis — LSA via Sklearn

Slaps Lab

Nov 25, 2020·2 min read

Quick write up on using the CountVectorizer and TruncatedSVD from the Sklearn library, to compute Document-Term and Term-Topic matrices. After setting up our model, we try it out on simple, never before seen documents in order to label them.

from sklearn.feature_extraction.text import CountVectorizer
from sklearn.decomposition import TruncatedSVDdocuments = [
    'Basketball is my favorite sport.',
    'Football is fun to play.',
    'IBM and GE are companies.'
]cv = CountVectorizer()
bow = cv.fit_transform(documents)n_topics = 2
tsvd = TruncatedSVD(n_topics)

Helper Methods

• using these to simplify viewing a document-topic matrix

def set_topics(df, n_topics):
    topics = list(range(n_topics))
    df.columns = [ f'topic_{t}' for t in topics ]
    for t in range(n_topics):
        df[f'topic_{t}'] = df[f'topic_{t}'].abs().round(2)
    return dfdef svd_to_pandas(svd_results, n_topics):
    df = pd.DataFrame(svd_results)
    df = set_topics(df, n_topics)
    return df

Document-Topic

matrix = tsvd.fit_transform(bow)
doc_to_topic = to_pandas(matrix, n_topics)
doc_to_topic

Term-Topic

term_to_topic = pd.DataFrame(
    tsvd.components_,
    columns = cv.get_feature_names()
).Tterm_to_topic = set_topics(term_to_topic, n_topics)
term_to_topic

Hold-Out Documents

hold_outs = [
    'The basketball game is Friday.', ## expect Topic 0
    'IBM stock stinks lately.' ## expect Topic 1
]hold_out_bow = cv.transform(hold_outs)
to_pandas(tsvd.transform(hold_out_bow), n_topics)