Implement of L-LDA Model(Labeled Latent Dirichlet Allocation Model) with python
References:
- Labeled LDA: A supervised topic model for credit attribution in multi-labeled corpora, Daniel Ramage...
- Parameter estimation for text analysis, Gregor Heinrich.
- Latent Dirichlet Allocation, David M. Blei, Andrew Y. Ng...
An efficient implementation based on Gibbs sampling
The following descriptions come from Labeled LDA: A supervised topic model for credit attribution in multi-labeled corpora, Daniel Ramage...
Introduction:
Labeled LDA is a topic model that constrains Latent Dirichlet Allocation by defining a one-to-one correspondence between LDAโs latent topics and user tags. Labeled LDA can directly learn topics(tags) correspondences.
Gibbs sampling:
- Graphical model of Labeled LDA:
- Generative process for Labeled LDA:
- Gibbs sampling equation:
Usage
- new llda model
- training
- ?is_convergence
- update
- inference
- save model to disk
- load model from disk
- get top-k terms of target topic
Example
# @source code: example/exapmle.py
import sys
sys.path.append('../')
import model.labeled_lda as llda
# initialize data
labeled_documents = [("example example example example example"*10, ["example"]),
("test llda model test llda model test llda model"*10, ["test", "llda_model"]),
("example test example test example test example test"*10, ["example", "test"]),
("good perfect good good perfect good good perfect good "*10, ["positive"]),
("bad bad down down bad bad down"*10, ["negative"])]
# new a Labeled LDA model
# llda_model = llda.LldaModel(labeled_documents=labeled_documents, alpha_vector="50_div_K", eta_vector=0.001)
# llda_model = llda.LldaModel(labeled_documents=labeled_documents, alpha_vector=0.02, eta_vector=0.002)
llda_model = llda.LldaModel(labeled_documents=labeled_documents, alpha_vector=0.01)
print(llda_model)
# training
# llda_model.training(iteration=10, log=True)
while True:
print("iteration %s sampling..." % (llda_model.iteration + 1))
llda_model.training(1)
print("after iteration: %s, perplexity: %s" % (llda_model.iteration, llda_model.perplexity()))
print("delta beta: %s" % llda_model.delta_beta)
if llda_model.is_convergent(method="beta", delta=0.01):
break
# update
print("before updating: ", llda_model)
update_labeled_documents = [("new example test example test example test example test", ["example", "test"])]
llda_model.update(labeled_documents=update_labeled_documents)
print("after updating: ", llda_model)
# train again
# llda_model.training(iteration=10, log=True)
while True:
print("iteration %s sampling..." % (llda_model.iteration + 1))
llda_model.training(1)
print("after iteration: %s, perplexity: %s" % (llda_model.iteration, llda_model.perplexity()))
print("delta beta: %s" % llda_model.delta_beta)
if llda_model.is_convergent(method="beta", delta=0.01):
break
# inference
# note: the result topics may be different for difference training, because gibbs sampling is a random algorithm
document = "example llda model example example good perfect good perfect good perfect" * 100
topics = llda_model.inference(document=document, iteration=100, times=10)
print(topics)
# perplexity
# calculate perplexity on test data
perplexity = llda_model.perplexity(documents=["example example example example example",
"test llda model test llda model test llda model",
"example test example test example test example test",
"good perfect good good perfect good good perfect good",
"bad bad down down bad bad down"],
iteration=30,
times=10)
print("perplexity on test data: %s" % perplexity)
# calculate perplexity on training data
print("perplexity on training data: %s" % llda_model.perplexity())
# save to disk
save_model_dir = "../data/model"
# llda_model.save_model_to_dir(save_model_dir, save_derivative_properties=True)
llda_model.save_model_to_dir(save_model_dir)
# load from disk
llda_model_new = llda.LldaModel()
llda_model_new.load_model_from_dir(save_model_dir, load_derivative_properties=False)
print("llda_model_new", llda_model_new)
print("llda_model", llda_model)
print("Top-5 terms of topic 'negative': ", llda_model.top_terms_of_topic("negative", 5, False))
print("Doc-Topic Matrix: \n", llda_model.theta)
print("Topic-Term Matrix: \n", llda_model.beta)