This library provides functionality to find the labels of SKOS thesaurus concepts in (short) text. It is a reimplementation in Python of stwfsa combined with the concept scoring from [1]. A deterministic finite automaton is constructed from the labels of the thesaurus concepts to perform the matching. In addition, a classifier is trained to score the matched occurrences of the concepts.
The construction for the automaton requires a SKOS thesaurus represented as a rdflib
Graph
.
Concepts should be related to labels by skos:prefLabel
or skos:altLabel
.
Concepts have to be identifiable by rdf:type
.
The training of the predictor requires labeled short text.
Each training sample should be annotated with one or more concepts from the thesaurus.
Currently the algorithm does not yield satisfactory results for longer texts,
i.e., you should work with titles and keywords only, or possibly with abstracts but not with fulltexts.
First load your thesaurus.
from rdflib import Graph
g = Graph()
g.load('/path/to/your/thesaurus')
First, define the type URI for descriptors. If your thesaurus has a hierarchical structure that includes groups, you can optionally specify the type URI for sub-thesauri. In this case you should also specify the relationship that relates sub-thesauri to concepts. Furthermore you can indicate whether the thesaurus relation is a specialisation. For the STW this would be
descriptor_type_uri = 'http://zbw.eu/namespaces/zbw-extensions/Descriptor'
thsys_type_uri = 'http://zbw.eu/namespaces/zbw-extensions/Thsys'
thesaurus_relation_type_uri = 'http://www.w3.org/2004/02/skos/core#broader'
is_specialisation = False
Create the predictor
from stwfsapy.predictor import StwfsapyPredictor
p = StwfsapyPredictor(
g,
descriptor_type_uri,
thsys_type_uri,
thesaurus_relation_type_uri,
is_specialisation,
langs={'en'},
simple_english_plural_rules=True)
The next step assumes you have loaded your texts into a list X
and your labels in a list of lists y
,
such that for all indices 0 <= i < len(X)
. The list at y[i]
contains the URIs to the correct concepts for X[i]
.
The concepts should be given by their URI.
Then you can train the classifier:
p.fit(X, y)
Afterwards you can get the predicted concepts and scores:
p.suggest_proba(['one input text', 'A completely different input text.'])
Alternatively you can get a sparse matrix of scores by calling
p.predict_proba(['one input text', 'Another input text.'])
The indices of the concepts are stored in p.concept_map_
.
A trained predictor p
can be stored by calling p.store('/path/to/storage/location')
.
Afterwards it can be loaded as follows:
from stwfsapy.predictor import StwfsapyPredictor
StwfsapyPredictor.load('/path/to/storage/location')