Topic labelling using tomotopy.
These functions should be called by Aurum methods; end-users should not
need to invoke them directly.
Expand source code
"""
Topic labelling using `tomotopy`.
These functions should be called by `ignis.aurum.Aurum` methods; end-users should not
need to invoke them directly.
"""
import ignis.util
tp = ignis.util.LazyLoader("tomotopy")
class TomotopyLabeller:
"""
Provides an interface to the `tomotopy` First-order Relevance labeller.
Parameters
----------
model: tomotopy.LDAModel
A trained Tomotopy model
verbose: bool, optional
Whether or not to print verbose progress messages
workers: int, optional
Number of worker processes to use for the labeller -- Note that this *might*
affect other random processes (e.g., the topic modelling proper) because it
affects the order of concurrent/parallel operations
Other Parameters
----------------
The other keyword arguments are labeller-specific options.
(See the `tomotopy` docs for details.)
"""
def __init__(
self,
model,
min_cf=10,
min_df=5,
max_len=5,
max_cand=10000,
smoothing=0.01,
mu=0.25,
workers=8,
verbose=False,
):
self.labeller_type = "tomotopy"
self.model = model
# Bookkeeping: Save a record of the options used
self.options = {
"min_cf": min_cf,
"min_df": min_df,
"max_len": max_len,
"max_cand": max_cand,
"smoothing": smoothing,
"mu": mu,
"workers": workers,
"verbose": verbose,
}
if verbose:
print("Extracting label candidates from model...", flush=True)
extractor = tp.label.PMIExtractor(
min_cf=min_cf, min_df=min_df, max_len=max_len, max_cand=max_cand
)
candidates = extractor.extract(self.model)
if verbose:
print("Preparing First-order relevance labeller...", flush=True)
self.labeller = tp.label.FoRelevance(
self.model,
candidates,
min_df=min_df,
smoothing=smoothing,
mu=mu,
workers=workers,
)
if verbose:
print("Done.")
def get_topic_labels(self, topic_id, top_n=10):
"""
Gets the `top_n` labels from the saved `tomotopy.label.FoRelevance` labeller
for the given `topic_id`.
**NOTE**: `topic_id` is 1-indexed, not 0-indexed (i.e., it is in `range(1,
len(topics) + 1)`).
Parameters
----------
topic_id: int
ID of the topic to label.
top_n: int
Number of highest-scoring labels to generate.
Returns
-------
tuple of (str, float)
Returns (`label`, `score`) for each of the `top_n` labels for the topic.
"""
# Tomotopy topics are 0-indexed
tp_topic_id = topic_id - 1
return self.labeller.get_topic_labels(k=tp_topic_id, top_n=top_n)Classes
class TomotopyLabeller (model, min_cf=10, min_df=5, max_len=5, max_cand=10000, smoothing=0.01, mu=0.25, workers=8, verbose=False)-
Provides an interface to the
tomotopyFirst-order Relevance labeller.Parameters
model:tomotopy.LDAModel- A trained Tomotopy model
verbose:bool, optional- Whether or not to print verbose progress messages
workers:int, optional- Number of worker processes to use for the labeller – Note that this might affect other random processes (e.g., the topic modelling proper) because it affects the order of concurrent/parallel operations
Other Parameters
The other keyword arguments are labeller-specific options. (See the
tomotopydocs for details.)Expand source code
class TomotopyLabeller: """ Provides an interface to the `tomotopy` First-order Relevance labeller. Parameters ---------- model: tomotopy.LDAModel A trained Tomotopy model verbose: bool, optional Whether or not to print verbose progress messages workers: int, optional Number of worker processes to use for the labeller -- Note that this *might* affect other random processes (e.g., the topic modelling proper) because it affects the order of concurrent/parallel operations Other Parameters ---------------- The other keyword arguments are labeller-specific options. (See the `tomotopy` docs for details.) """ def __init__( self, model, min_cf=10, min_df=5, max_len=5, max_cand=10000, smoothing=0.01, mu=0.25, workers=8, verbose=False, ): self.labeller_type = "tomotopy" self.model = model # Bookkeeping: Save a record of the options used self.options = { "min_cf": min_cf, "min_df": min_df, "max_len": max_len, "max_cand": max_cand, "smoothing": smoothing, "mu": mu, "workers": workers, "verbose": verbose, } if verbose: print("Extracting label candidates from model...", flush=True) extractor = tp.label.PMIExtractor( min_cf=min_cf, min_df=min_df, max_len=max_len, max_cand=max_cand ) candidates = extractor.extract(self.model) if verbose: print("Preparing First-order relevance labeller...", flush=True) self.labeller = tp.label.FoRelevance( self.model, candidates, min_df=min_df, smoothing=smoothing, mu=mu, workers=workers, ) if verbose: print("Done.") def get_topic_labels(self, topic_id, top_n=10): """ Gets the `top_n` labels from the saved `tomotopy.label.FoRelevance` labeller for the given `topic_id`. **NOTE**: `topic_id` is 1-indexed, not 0-indexed (i.e., it is in `range(1, len(topics) + 1)`). Parameters ---------- topic_id: int ID of the topic to label. top_n: int Number of highest-scoring labels to generate. Returns ------- tuple of (str, float) Returns (`label`, `score`) for each of the `top_n` labels for the topic. """ # Tomotopy topics are 0-indexed tp_topic_id = topic_id - 1 return self.labeller.get_topic_labels(k=tp_topic_id, top_n=top_n)Methods
def get_topic_labels(self, topic_id, top_n=10)-
Gets the
top_nlabels from the savedtomotopy.label.FoRelevancelabeller for the giventopic_id.NOTE:
topic_idis 1-indexed, not 0-indexed (i.e., it is inrange(1, len(topics) + 1)).Parameters
topic_id:int- ID of the topic to label.
top_n:int- Number of highest-scoring labels to generate.
Returns
tupleof(str, float)- Returns (
label,score) for each of thetop_nlabels for the topic.
Expand source code
def get_topic_labels(self, topic_id, top_n=10): """ Gets the `top_n` labels from the saved `tomotopy.label.FoRelevance` labeller for the given `topic_id`. **NOTE**: `topic_id` is 1-indexed, not 0-indexed (i.e., it is in `range(1, len(topics) + 1)`). Parameters ---------- topic_id: int ID of the topic to label. top_n: int Number of highest-scoring labels to generate. Returns ------- tuple of (str, float) Returns (`label`, `score`) for each of the `top_n` labels for the topic. """ # Tomotopy topics are 0-indexed tp_topic_id = topic_id - 1 return self.labeller.get_topic_labels(k=tp_topic_id, top_n=top_n)