def retrieval_demo():
"""Function intended to illustrate retrieval in the experimental framework.
Intended as a basis for new experiments for those not intimately
familiar with the code.
"""
print 'Evaluation type: Retrieval'
print 'Graph type: Dependency'
print 'Centrality: PageRank'
print
print '> Reading data..'
desc_path = '../data/air/problem_descriptions_dependencies'
sol_path = '../data/air/solutions_preprocessed'
problems, _ = data.read_files(desc_path)
solutions, _ = data.read_files(sol_path)
print '> Creating solution representations..'
metric = freq_representation.FrequencyMetrics.TF_IDF
sol_vectors = freq_representation.text_to_vector(solutions, metric)
print '> Creating problem description representations..'
dicts = []
for i, doc in enumerate(problems):
print ' ',str(i)+'/'+str(len(problems))
g = graph_representation.construct_dependency_network(doc)
d = graph_representation.graph_to_dict(g, graph.GraphMetrics.PAGERANK)
dicts.append(d)
desc_vectors = graph_representation.dicts_to_vectors(dicts)
print '> Evaluating..'
score = evaluation.evaluate_retrieval(desc_vectors, sol_vectors)
print ' score:', score
print
def classification_comparison_freq(dataset='reuters'):
print '> Reading data..', dataset
training_path = '../data/' + dataset + '/training_preprocessed'
training_docs, training_labels = data.read_files(training_path)
test_path = '../data/' + dataset + '/test_preprocessed'
test_docs, test_labels = data.read_files(test_path)
results = {}
for metric in freq_representation.get_metrics():
print ' ', metric,
training_dicts = freq_representation.text_to_dict(
training_docs, metric)
test_dicts = freq_representation.text_to_dict(test_docs, metric)
print ' dicst -> vectors'
keys = set()
for d in training_dicts + test_dicts:
keys = keys.union(d.keys())
print ' vocabulary size:', len(keys)
training_rep = graph_representation.dicts_to_vectors(
training_dicts, keys)
test_rep = graph_representation.dicts_to_vectors(test_dicts, keys)
reps = {'training': training_rep, 'test': test_rep}
labels = {'training': training_labels, 'test': test_labels}
score = evaluation.evaluate_classification(reps, labels, mode='split')
results[metric] = score
print score
pp.pprint(results)
s = 'classification comparison \nrepresentation: frequency\nresult:\n' + str(
results) + '\n\n\n'
data.write_to_file(s, 'output/comparison/classification')
return results
def test_retrieval(orders=[1,2,3],order_weights=[1.0,1.53,1.51]):
"""
Test retrieval using different combinations of higher orders and weightings of these.
The list *orders* define which higher order relations to include.
The relative importance of the orders are defined by *order_weights*.
"""
print '> Reading cases..'
descriptions_path = '../data/air/problem_descriptions_preprocessed'
description_texts, labels = data.read_files(descriptions_path)
filenames = data.get_file_names(descriptions_path)
solutions_path = '../data/air/solutions_preprocessed'
solution_texts, labels = data.read_files(solutions_path)
solution_vectors = freq_representation.text_to_vector(solution_texts, freq_representation.FrequencyMetrics.TF_IDF)
print '> Creating representations..'
rep = []
for i, text in enumerate(description_texts):
print ' '+str(i)+"/"+str(len(description_texts))
g = graph_representation.construct_cooccurrence_network(text, orders=orders, order_weights=order_weights, doc_id='output/higher_order/air/'+labels[i]+'/'+filenames[i])
d = graph_representation.graph_to_dict(g, graph.GraphMetrics.WEIGHTED_DEGREE)
rep.append(d)
rep = graph_representation.dicts_to_vectors(rep)
print '> Evaluating..'
score = evaluation.evaluate_retrieval(rep, solution_vectors)
print 'orders:', orders
print 'score:', score
fname = 'output/higher_order/results/retr'
with open(fname, 'a+') as f:
s = reduce(lambda x,y:str(x)+str(y), orders)
f.write(str(s)+' '+str(score)+'\n')
return score
def retrieval_demo():
"""Function intended to illustrate retrieval in the experimental framework.
Intended as a basis for new experiments for those not intimately
familiar with the code.
"""
print 'Evaluation type: Retrieval'
print 'Graph type: Dependency'
print 'Centrality: PageRank'
print
print '> Reading data..'
desc_path = '../data/air/problem_descriptions_dependencies'
sol_path = '../data/air/solutions_preprocessed'
problems, _ = data.read_files(desc_path)
solutions, _ = data.read_files(sol_path)
print '> Creating solution representations..'
metric = freq_representation.FrequencyMetrics.TF_IDF
sol_vectors = freq_representation.text_to_vector(solutions, metric)
print '> Creating problem description representations..'
dicts = []
for i, doc in enumerate(problems):
print ' ', str(i) + '/' + str(len(problems))
g = graph_representation.construct_dependency_network(doc)
d = graph_representation.graph_to_dict(g, graph.GraphMetrics.PAGERANK)
dicts.append(d)
desc_vectors = graph_representation.dicts_to_vectors(dicts)
print '> Evaluating..'
score = evaluation.evaluate_retrieval(desc_vectors, sol_vectors)
print ' score:', score
print
def evaluate_tc_icc_retrieval():
graph_metrics = graph_representation.get_metrics(True, exclude_flow=True)
print '> Reading cases..'
corpus = 'air/problem_descriptions'
context = 'window'
solutions_path = '../data/air/solutions_preprocessed'
path = '../data/air/problem_descriptions_preprocessed'
description_texts, labels = data.read_files(path)
rep = {}
icc = {}
print '> Calculating ICCs..'
for metric in graph_metrics:
print ' ', metric
rep[metric] = []
centralities = retrieve_centralities(corpus, context, metric)
if centralities:
icc[metric] = graph_representation.calculate_icc_dict(centralities)
else:
icc[metric] = None
print '> Creating solution representations..'
solutions_texts, labels = data.read_files(solutions_path)
solutions_rep = freq_representation.text_to_vector(
solutions_texts, freq_representation.FrequencyMetrics.TF_IDF)
print '> Creating problem description representations..'
for i, text in enumerate(description_texts):
if i % 1 == 0:
print ' document', str(i) + '/' + str(len(description_texts))
g = graph_representation.construct_cooccurrence_network(
text, already_preprocessed=True, context='window')
for metric in graph_metrics:
if not icc[metric]: continue
#~ print ' ',metric
d = graph_representation.graph_to_dict(g, metric, icc[metric])
rep[metric].append(d)
g = None # just to make sure..
print '> Creating vector representations..'
for metric in graph_metrics:
if not icc[metric]: continue
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
results = {}
for metric in graph_metrics:
if not icc[metric]:
results[metric] = None
continue
vectors = rep[metric]
score = evaluation.evaluate_retrieval(vectors, solutions_rep)
print ' ', metric, score
results[metric] = score
pp.pprint(results)
data.pickle_to_file(
results, 'output/tc_icc/cooccurrence/' + corpus + '/retrieval.res')
return results
def classification_comparison_freq(dataset='reuters'):
print '> Reading data..', dataset
training_path = '../data/'+dataset+'/training_preprocessed'
training_docs, training_labels = data.read_files(training_path)
test_path = '../data/'+dataset+'/test_preprocessed'
test_docs, test_labels = data.read_files(test_path)
results = {}
for metric in freq_representation.get_metrics():
print ' ', metric,
training_dicts = freq_representation.text_to_dict(training_docs, metric)
test_dicts = freq_representation.text_to_dict(test_docs, metric)
print ' dicst -> vectors'
keys = set()
for d in training_dicts + test_dicts:
keys = keys.union(d.keys())
print ' vocabulary size:', len(keys)
training_rep = graph_representation.dicts_to_vectors(training_dicts, keys)
test_rep = graph_representation.dicts_to_vectors(test_dicts, keys)
reps = {'training':training_rep, 'test':test_rep}
labels = {'training':training_labels, 'test':test_labels}
score = evaluation.evaluate_classification(reps, labels, mode='split')
results[metric] = score
print score
pp.pprint(results)
s = 'classification comparison \nrepresentation: frequency\nresult:\n'+str(results)+'\n\n\n'
data.write_to_file(s, 'output/comparison/classification')
return results
def retrieval_comparison_graph(dataset='air', graph_type='co-occurrence', use_icc=False):
"""
Experiment used for comparative evaluation of different network
representations on retrieval.
graph_type = 'co-occurrence' | 'dependency'
`icc` determines whether to use _inverse corpus centrality_ in the vector representations.
"""
def make_dicts(docs, icc=None):
rep = []
for i, doc in enumerate(docs):
if i%100==0: print ' graph',str(i)+'/'+str(len(docs))
g = gfuns[graph_type](doc)
d = graph_representation.graph_to_dict(g, metrics[graph_type], icc)
rep.append(d)
return rep
postfix = {'co-occurrence':'_text', 'dependency':'_dependencies'}
gfuns = {'co-occurrence':graph_representation.construct_cooccurrence_network,
'dependency':graph_representation.construct_dependency_network}
metrics = {'co-occurrence':graph.GraphMetrics.WEIGHTED_DEGREE,
'dependency':graph.GraphMetrics.EIGENVECTOR}
print '--', graph_type
print '> Reading data..', dataset
path = '../data/'+dataset+'/problem_descriptions'+postfix[graph_type]
docs, labels = data.read_files(path)
print '> Creating solution representations..'
solutions_path = '../data/'+dataset+'/solutions_preprocessed'
solutions_texts, labels = data.read_files(solutions_path)
solutions_rep = freq_representation.text_to_vector(solutions_texts, freq_representation.FrequencyMetrics.TF_IDF)
icc = None
if use_icc:
print '> Calculating ICC..'
m = metrics[graph_type].split()[0]
print graph_type
if graph_type == 'co-occurrence':
p = 'output/centralities/co-occurrence/'+dataset+'/problem_descriptions/window/'+m+'.cent'
elif graph_type == 'dependency':
p = 'output/centralities/dependency/'+dataset+'/problem_descriptions/'+m+'.cent'
print ' fetching', p
icc = data.pickle_from_file(p)
print ' icc:', type(icc)
print '> Creating problem description representations..'
dicts = make_dicts(docs, icc)
descriptions_rep = graph_representation.dicts_to_vectors(dicts)#, remove_stop_words=True)
print '> Evaluating..'
results = evaluation.evaluate_retrieval(descriptions_rep, solutions_rep)
print results
s = 'retrieval comparison '
if use_icc: s += 'USING TC-ICC'
s += '\nrepresentation: '+graph_type+'\nresult: '+str(results)+'\n\n\n'
data.write_to_file(s, 'output/comparison/retrieval')
return results
def edge_direction_evaluation(direction):
"""
Evaluate impact of using different edge directions on dependency networks.
Values for *direction*: ``forward``, ``backward``, and ``undirected``.
"""
results = {'_edge-direction':direction}
print '------ CLASSIFICATION EVALUATION --------'
print '> Reading cases..'
descriptions_path = '../data/tasa/TASA900_dependencies'
texts, labels = data.read_files(descriptions_path)
print '> Creating representations..'
rep = []
for i, text in enumerate(texts):
if i%100==0: print ' ',str(i)+'/'+str(len(texts))
g = graph_representation.construct_dependency_network(text, direction=direction)
metric = graph.GraphMetrics.CLOSENESS
d = graph_representation.graph_to_dict(g, metric)
rep.append(d)
g = None # just to make sure..
rep = graph_representation.dicts_to_vectors(rep)
print '> Evaluating..'
score = evaluation.evaluate_classification(rep, labels)
print ' score:', score
results['classification'] = score
print '------ RETRIEVAL EVALUATION --------'
print '> Reading cases..'
descriptions_path = '../data/air/problem_descriptions_dependencies'
description_texts, labels = data.read_files(descriptions_path)
solutions_path = '../data/air/solutions_preprocessed'
solution_texts, labels = data.read_files(solutions_path)
solution_vectors = freq_representation.text_to_vector(solution_texts, freq_representation.FrequencyMetrics.TF_IDF)
print '> Creating representations..'
rep = []
for i, text in enumerate(description_texts):
if i%100==0: print ' ',str(i)+'/'+str(len(description_texts))
g = graph_representation.construct_dependency_network(text, direction=direction)
metric = graph.GraphMetrics.EIGENVECTOR
d = graph_representation.graph_to_dict(g, metric)
rep.append(d)
g = None # just to make sure..
rep = graph_representation.dicts_to_vectors(rep)
print '> Evaluating..'
score = evaluation.evaluate_retrieval(rep, solution_vectors)
print ' score:', score
results['retrieval'] = score
data.pickle_to_file(results, 'output/dependencies/stop_words_retr_'+direction)
pp.pprint(results)
return results
def centrality_weights_retrieval(weighted=True):
"""
Evaluate whether edge weights are beneficial to the depdendency
network represenation for the retrieval task.
"""
results = {'_is_weighted': weighted, '_evaluation': 'retrieval'}
graph_metrics = graph_representation.get_metrics(weighted)
print '> Reading cases..'
descriptions_path = '../data/air/problem_descriptions_dependencies'
description_texts, labels = data.read_files(descriptions_path)
solutions_path = '../data/air/solutions_preprocessed'
solution_texts, labels = data.read_files(solutions_path)
solution_vectors = freq_representation.text_to_vector(
solution_texts, freq_representation.FrequencyMetrics.TF_IDF)
rep = {}
for metric in graph_metrics:
rep[metric] = []
print '> Creating graph representations..'
for i, text in enumerate(description_texts):
if i % 10 == 0: print ' ', str(i) + '/' + str(len(description_texts))
g = graph_representation.construct_dependency_network(
text, weighted=weighted)
for metric in graph_metrics:
d = graph_representation.graph_to_dict(g, metric)
rep[metric].append(d)
g = None # just to make sure..
if i % 100 == 0:
if weighted:
postfix = '_weighted'
else:
postfix = '_unweighted'
data.pickle_to_file(
rep,
'output/dependencies/exp1_retr_tmp_' + str(i) + '_' + postfix)
print '> Creating vector representations..'
for metric in graph_metrics:
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
for metric in graph_metrics:
vectors = rep[metric]
score = evaluation.evaluate_retrieval(vectors, solution_vectors)
print ' ', metric, score
results[metric] = score
if weighted:
postfix = '_weighted'
else:
postfix = '_unweighted'
data.pickle_to_file(results, 'output/dependencies/exp1_retr' + postfix)
pp.pprint(results)
return results
def print_common_hub_words(rem_stop_words):
"""
Print a list of the most common hub words in the created networks.
Purpose of experiment was to show that hub words typically are stop words.
The *rem_stop_words* determine whether stop words are removed before creating
the networks.
"""
results = {'_removing stop-words': rem_stop_words}
print '------ CLASSIFICATION EVALUATION --------'
print '> Reading cases..'
descriptions_path = '../data/tasa/TASA900_dependencies'
texts, labels = data.read_files(descriptions_path)
print '> Creating representations..'
fd = nltk.probability.FreqDist()
for i, text in enumerate(texts):
if i % 100 == 0: print ' ', str(i) + '/' + str(len(texts))
g = graph_representation.construct_dependency_network(
text, remove_stop_words=rem_stop_words)
hubs = graph.get_hubs(g, 10)
for h in hubs:
fd.inc(h[0])
g = None # just to make sure..
results['tasa'] = fd.keys()
print '------ RETRIEVAL EVALUATION --------'
print '> Reading cases..'
descriptions_path = '../data/air/problem_descriptions_dependencies'
description_texts, labels = data.read_files(descriptions_path)
print '> Creating representations..'
fd = nltk.probability.FreqDist()
for i, text in enumerate(description_texts):
if i % 100 == 0:
print ' ', str(i) + '/' + str(len(description_texts))
g = graph_representation.construct_dependency_network(
text, remove_stop_words=rem_stop_words)
hubs = graph.get_hubs(g, 10)
for h in hubs:
fd.inc(h[0])
g = None # just to make sure..
results['air'] = fd.keys()
if rem_stop_words:
modifier = 'without'
else:
modifier = 'with'
data.pickle_to_file(
results, 'output/dependencies/common_hubs_' + modifier + 'stop_words')
pp.pprint(results)
return results
def evaluate_tc_icc_retrieval():
graph_metrics = graph_representation.get_metrics(True, exclude_flow=True)
print '> Reading cases..'
corpus = 'air/problem_descriptions'
context = 'window'
solutions_path = '../data/air/solutions_preprocessed'
path = '../data/air/problem_descriptions_preprocessed'
description_texts, labels = data.read_files(path)
rep = {}
icc = {}
print '> Calculating ICCs..'
for metric in graph_metrics:
print ' ', metric
rep[metric] = []
centralities = retrieve_centralities(corpus, context, metric)
if centralities:
icc[metric] = graph_representation.calculate_icc_dict(centralities)
else:
icc[metric] = None
print '> Creating solution representations..'
solutions_texts, labels = data.read_files(solutions_path)
solutions_rep = freq_representation.text_to_vector(solutions_texts, freq_representation.FrequencyMetrics.TF_IDF)
print '> Creating problem description representations..'
for i, text in enumerate(description_texts):
if i%1==0: print ' document',str(i)+'/'+str(len(description_texts))
g = graph_representation.construct_cooccurrence_network(text, already_preprocessed=True, context='window')
for metric in graph_metrics:
if not icc[metric]: continue
#~ print ' ',metric
d = graph_representation.graph_to_dict(g, metric, icc[metric])
rep[metric].append(d)
g = None # just to make sure..
print '> Creating vector representations..'
for metric in graph_metrics:
if not icc[metric]: continue
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
results = {}
for metric in graph_metrics:
if not icc[metric]:
results[metric] = None
continue
vectors = rep[metric]
score = evaluation.evaluate_retrieval(vectors, solutions_rep)
print ' ', metric, score
results[metric] = score
pp.pprint(results)
data.pickle_to_file(results, 'output/tc_icc/cooccurrence/'+corpus+'/retrieval.res')
return results
def centrality_weights_retrieval(weighted=True):
"""
Evaluate whether edge weights are beneficial to the depdendency
network represenation for the retrieval task.
"""
results = {'_is_weighted':weighted, '_evaluation':'retrieval'}
graph_metrics = graph_representation.get_metrics(weighted)
print '> Reading cases..'
descriptions_path = '../data/air/problem_descriptions_dependencies'
description_texts, labels = data.read_files(descriptions_path)
solutions_path = '../data/air/solutions_preprocessed'
solution_texts, labels = data.read_files(solutions_path)
solution_vectors = freq_representation.text_to_vector(solution_texts, freq_representation.FrequencyMetrics.TF_IDF)
rep = {}
for metric in graph_metrics:
rep[metric] = []
print '> Creating graph representations..'
for i, text in enumerate(description_texts):
if i%10==0: print ' ',str(i)+'/'+str(len(description_texts))
g = graph_representation.construct_dependency_network(text, weighted=weighted)
for metric in graph_metrics:
d = graph_representation.graph_to_dict(g, metric)
rep[metric].append(d)
g = None # just to make sure..
if i%100==0:
if weighted:
postfix = '_weighted'
else:
postfix = '_unweighted'
data.pickle_to_file(rep, 'output/dependencies/exp1_retr_tmp_'+str(i)+'_'+postfix)
print '> Creating vector representations..'
for metric in graph_metrics:
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
for metric in graph_metrics:
vectors = rep[metric]
score = evaluation.evaluate_retrieval(vectors, solution_vectors)
print ' ', metric, score
results[metric] = score
if weighted:
postfix = '_weighted'
else:
postfix = '_unweighted'
data.pickle_to_file(results, 'output/dependencies/exp1_retr'+postfix)
pp.pprint(results)
return results
def do_context_size_evaluation_retrieval():
"""
Experiment evaluating performance of different context sizes for
co-occurrence networks in the retrieval task.
"""
results = {}
graph_metrics = graph_representation.get_metrics()
for metric in graph_metrics:
results[metric] = []
print '> Reading cases..'
descriptions_path = '../data/air/problem_descriptions_preprocessed'
description_texts, labels = data.read_files(descriptions_path)
solutions_path = '../data/air/solutions_preprocessed'
solution_texts, labels = data.read_files(solutions_path)
solution_vectors = freq_representation.text_to_vector(
solution_texts, freq_representation.FrequencyMetrics.TF_IDF)
for window_size in range(1, 11) + [20, 40, 80]:
print '-- window size:', window_size
rep = {}
for metric in graph_metrics:
rep[metric] = []
print '> Creating representations..'
# creating graphs and finding centralities
for i, text in enumerate(description_texts):
if i % 10 == 0: print i
g = graph_representation.construct_cooccurrence_network(
text, window_size=window_size, already_preprocessed=True)
for metric in graph_metrics:
d = graph_representation.graph_to_dict(g, metric)
rep[metric].append(d)
g = None # just to make sure..
# creating representation vectors
for metric in graph_metrics:
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
for metric in graph_metrics:
vectors = rep[metric]
score = evaluation.evaluate_retrieval(vectors, solution_vectors)
print ' ', metric, score
results[metric].append(score)
data.pickle_to_file(results, 'output/retr_context_' + str(window_size))
pp.pprint(results)
return results
def print_common_hub_words(rem_stop_words):
"""
Print a list of the most common hub words in the created networks.
Purpose of experiment was to show that hub words typically are stop words.
The *rem_stop_words* determine whether stop words are removed before creating
the networks.
"""
results = {'_removing stop-words':rem_stop_words}
print '------ CLASSIFICATION EVALUATION --------'
print '> Reading cases..'
descriptions_path = '../data/tasa/TASA900_dependencies'
texts, labels = data.read_files(descriptions_path)
print '> Creating representations..'
fd = nltk.probability.FreqDist()
for i, text in enumerate(texts):
if i%100==0: print ' ',str(i)+'/'+str(len(texts))
g = graph_representation.construct_dependency_network(text, remove_stop_words=rem_stop_words)
hubs = graph.get_hubs(g, 10)
for h in hubs:
fd.inc(h[0])
g = None # just to make sure..
results['tasa'] = fd.keys()
print '------ RETRIEVAL EVALUATION --------'
print '> Reading cases..'
descriptions_path = '../data/air/problem_descriptions_dependencies'
description_texts, labels = data.read_files(descriptions_path)
print '> Creating representations..'
fd = nltk.probability.FreqDist()
for i, text in enumerate(description_texts):
if i%100==0: print ' ',str(i)+'/'+str(len(description_texts))
g = graph_representation.construct_dependency_network(text, remove_stop_words=rem_stop_words)
hubs = graph.get_hubs(g, 10)
for h in hubs:
fd.inc(h[0])
g = None # just to make sure..
results['air'] = fd.keys()
if rem_stop_words:
modifier = 'without'
else:
modifier = 'with'
data.pickle_to_file(results, 'output/dependencies/common_hubs_'+modifier+'stop_words')
pp.pprint(results)
return results
def do_retrieval_experiments(
descriptions='air/problem_descriptions',
solutions='air/solutions',
graph_types=['co-occurrence', 'dependency', 'random'],
use_frequency=True):
"""
Experiment used for comparative evaluation of different network
representations on the retrieval task.
Toggle comparison with frequency-based methods using *use_frequency*.
"""
results = {
'_solutions': solutions,
'_descriptions': descriptions,
'_evaluation': 'retrieval'
}
print '> Evaluation type: retrieval'
print '> Reading cases..'
descriptions_path = '../data/' + descriptions
descriptiondata = data.read_data(descriptions_path, graph_types)
solutions_path = '../data/' + solutions + '_preprocessed'
solution_texts, labels = data.read_files(solutions_path)
solution_vectors = freq_representation.text_to_vector(
solution_texts, freq_representation.FrequencyMetrics.TF_IDF)
print '> Evaluating..'
for gtype in graph_types:
print ' ', gtype
docs, labels = descriptiondata[gtype]
graphs = graph_representation.create_graphs(docs, gtype)
results[gtype] = {}
for metric in graph_representation.get_metrics():
print ' -', metric
vectors = graph_representation.graphs_to_vectors(graphs, metric)
results[gtype][metric] = evaluation.evaluate_retrieval(
vectors, solution_vectors)
if use_frequency:
print ' frequency'
results['freq'] = {}
for metric in freq_representation.get_metrics():
print ' -', metric
docs, labels = data.read_files(descriptions_path + '_preprocessed')
vectors = freq_representation.text_to_vector(docs, metric)
results['freq'][metric] = evaluation.evaluate_retrieval(
vectors, solution_vectors)
print
pp.pprint(results)
return results
def do_context_size_evaluation_retrieval():
"""
Experiment evaluating performance of different context sizes for
co-occurrence networks in the retrieval task.
"""
results = {}
graph_metrics = graph_representation.get_metrics()
for metric in graph_metrics:
results[metric] = []
print '> Reading cases..'
descriptions_path = '../data/air/problem_descriptions_preprocessed'
description_texts, labels = data.read_files(descriptions_path)
solutions_path = '../data/air/solutions_preprocessed'
solution_texts, labels = data.read_files(solutions_path)
solution_vectors = freq_representation.text_to_vector(solution_texts, freq_representation.FrequencyMetrics.TF_IDF)
for window_size in range(1,11)+[20,40,80]:
print '-- window size:',window_size
rep = {}
for metric in graph_metrics:
rep[metric] = []
print '> Creating representations..'
# creating graphs and finding centralities
for i, text in enumerate(description_texts):
if i%10==0: print i
g = graph_representation.construct_cooccurrence_network(text, window_size=window_size, already_preprocessed=True)
for metric in graph_metrics:
d = graph_representation.graph_to_dict(g, metric)
rep[metric].append(d)
g = None # just to make sure..
# creating representation vectors
for metric in graph_metrics:
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
for metric in graph_metrics:
vectors = rep[metric]
score = evaluation.evaluate_retrieval(vectors, solution_vectors)
print ' ', metric, score
results[metric].append(score)
data.pickle_to_file(results, 'output/retr_context_'+str(window_size))
pp.pprint(results)
return results
def corpus_dependency_properties(dataset = 'air/problem_descriptions'):
"""
Identify and pickle to file various properties of the given dataset.
These can alter be converted to pretty tables using
:func:`~experiments.print_network_props`.
"""
print '> Reading data..', dataset
corpus_path = '../data/'+dataset+'_dependencies'
(documents, labels) = data.read_files(corpus_path)
props = {}
giant = nx.DiGraph()
print '> Building networks..'
for i, text in enumerate(documents):
if i%10==0: print ' ',str(i)+'/'+str(len(documents))
g = graph_representation.construct_dependency_network(text,remove_stop_words=True)
giant.add_edges_from(g.edges())
p = graph.network_properties(g)
for k,v in p.iteritems():
if i==0: props[k] = []
props[k].append(v)
g = None # just to make sure..
print '> Calculating means and deviations..'
props_total = {}
for key in props:
props_total[key+'_mean'] = numpy.mean(props[key])
props_total[key+'_std'] = numpy.std(props[key])
data.pickle_to_file(giant, 'output/properties/dependency/corpus_network_air_all_no_stop_words')
data.pickle_to_file(props, 'output/properties/dependency/docs_air_all_no_stop_words')
data.pickle_to_file(props_total, 'output/properties/dependency/docs_air_all_no_stop_words_total')
def print_degree_distributions(dataset, context):
"""
Extracts degree distribution values from networks, and print them to
cvs-file.
**warning** overwrites if file exists.
"""
print '> Reading data..', dataset
corpus_path = '../data/'+dataset+'_text'
(documents, labels) = data.read_files(corpus_path)
degsfile = open('output/properties/cooccurrence/degrees_docs_'+dataset.replace('/','.'), 'w')
giant = nx.DiGraph()
print '> Building networks..'
for i, text in enumerate(documents):
if i%10==0: print ' ',str(i)+'/'+str(len(documents))
g = graph_representation.construct_cooccurrence_network(text,context=context)
giant.add_edges_from(g.edges())
degs = nx.degree(g).values()
degs = [str(d) for d in degs]
degsfile.write(','.join(degs)+'\n')
degsfile.close()
print '> Writing giant\'s distribution'
with open('output/properties/cooccurrence/degrees_giant_'+dataset.replace('/','.'), 'w') as f:
ds = nx.degree(giant).values()
ds = [str(d) for d in ds]
f.write(','.join(ds))
def corpus_properties(dataset, context):
"""
Identify and pickle to file various properties of the given dataset.
These can alter be converted to pretty tables using
:func:`~experiments.print_network_props`.
"""
print '> Reading data..', dataset
corpus_path = '../data/'+dataset+'_text'
(documents, labels) = data.read_files(corpus_path)
props = {}
#~ giant = nx.DiGraph()
print '> Building networks..'
for i, text in enumerate(documents):
if i%10==0: print ' ',str(i)+'/'+str(len(documents))
g = graph_representation.construct_cooccurrence_network(text,context=context)
#~ giant.add_edges_from(g.edges())
p = graph.network_properties(g)
for k,v in p.iteritems():
if i==0: props[k] = []
props[k].append(v)
g = None # just to make sure..
print '> Calculating means and deviations..'
props_total = {}
for key in props:
print ' ',key
props_total[key+'_mean'] = numpy.mean(props[key])
props_total[key+'_std'] = numpy.std(props[key])
data_name = dataset.replace('/','.')
#~ data.pickle_to_file(giant, 'output/properties/cooccurrence/giant_'+data_name)
data.pickle_to_file(props, 'output/properties/cooccurrence/stats_'+data_name)
data.pickle_to_file(props_total, 'output/properties/cooccurrence/stats_tot_'+data_name)
def do_context_sentence_evaluation_classification():
"""
Experiment evaluating performance of sentences as contexts for
co-occurrence networks in the classification task.
"""
print '> Reading cases..'
path = '../data/tasa/TASA900_text'
texts, labels = data.read_files(path)
print '> Evaluating..'
graphs = []
results = {}
for text in texts:
g = graph_representation.construct_cooccurrence_network(text, context='sentence')
graphs.append(g)
for metric in graph_representation.get_metrics():
print ' ', metric
vectors = graph_representation.graphs_to_vectors(graphs, metric, verbose=True)
score = evaluation.evaluate_classification(vectors, labels)
results[metric+' (sentence)'] = score
data.pickle_to_file(results, 'output/class_context_sentence')
pp.pprint(results)
return results
def word_vectors():
train = read_files('train')
test = read_files('test')
tokenizer = Tokenizer(num_words=20000)
tokenizer.fit_on_texts(train['text'])
train_seq = tokenizer.texts_to_sequences(train['text'])
test_seq = tokenizer.texts_to_sequences(test['text'])
x_train = sequence.pad_sequences(train_seq, maxlen=200) # shape (25000, 200)
y_train = train['label']
x_test = sequence.pad_sequences(test_seq, maxlen=200) # shape (25000, 200)
y_test = test['label']
return x_train, y_train.values, x_test, y_test.values
def solution_similarity_stats(dataset='air/solutions_preprocessed'):
"""
Plots histogram of solution-solution similarity distribution of a dataset.
"""
print '> Reading data..', dataset
corpus_path = '../data/' + dataset
(documents, labels) = data.read_files(corpus_path)
print '> Creating vector representations..'
vectors = freq_representation.text_to_vector(
documents, freq_representation.FrequencyMetrics.TF_IDF)
print '> Calculating similarities..'
distances = scipy.spatial.distance.cdist(vectors.T, vectors.T, 'cosine')
diag = numpy.diag([2.0] * len(distances),
0) # move similarities of "self" to -1
distances = distances + diag
similarities = 1.0 - distances
similarities = similarities.ravel()
similarities = [s for s in similarities if s >= 0]
print plotter.histogram(similarities,
'similarity',
'# matches',
'',
bins=150)
print
print max(similarities)
print min(similarities)
print float(sum(similarities)) / len(similarities)
num = len([sim for sim in similarities if sim < 0.23])
print 'fraction sims < .23:', float(num) / len(similarities)
def test_scale_free():
import random
import data
import graph_representation
import plfit
import numpy
corpus_path = '../data/air/problem_descriptions_text'
(documents, labels) = data.read_files(corpus_path)
g = graph_representation.construct_cooccurrence_network(documents[0],context='sentence')
degree_sequence=sorted(nx.degree(g).values(),reverse=True) # degree sequence
dmax=max(degree_sequence)
degree_sequence = numpy.array(degree_sequence)
print degree_sequence
pl = plfit.plfit(degree_sequence)
p,ksv = pl.test_pl()
print
print
print
print
seq = [random.randrange(0,100) for i in range(len(degree_sequence))]
degree_sequence = numpy.array(seq)
print degree_sequence
pl = plfit.plfit(degree_sequence)
p,ksv = pl.test_pl()
print
print
print
print
def test_scale_free():
import random
import data
import graph_representation
import plfit
import numpy
corpus_path = '../data/air/problem_descriptions_text'
(documents, labels) = data.read_files(corpus_path)
g = graph_representation.construct_cooccurrence_network(documents[0],
context='sentence')
degree_sequence = sorted(nx.degree(g).values(),
reverse=True) # degree sequence
dmax = max(degree_sequence)
degree_sequence = numpy.array(degree_sequence)
print degree_sequence
pl = plfit.plfit(degree_sequence)
p, ksv = pl.test_pl()
print
print
print
print
seq = [random.randrange(0, 100) for i in range(len(degree_sequence))]
degree_sequence = numpy.array(seq)
print degree_sequence
pl = plfit.plfit(degree_sequence)
p, ksv = pl.test_pl()
print
print
print
print
def dataset_stats(dataset):
"""
Print and plot statistics for a given dataset.
A histogram is plotted with the document length distribution of the data.
"""
print '> Reading data..', dataset
corpus_path = '../data/' + dataset
(documents, labels) = data.read_files(corpus_path)
file_names = data.get_file_names(corpus_path)
lengths = []
empty = 0
for i, d in enumerate(documents):
d = preprocess.tokenize_tokens(d)
lengths.append(len(d))
if len(d) == 0:
print file_names[i], 'is empty'
empty += 1
lengths = numpy.array(lengths)
print '# documents:', len(documents)
print '# empty documents:', empty
print '# words:', sum(lengths)
print 'length avg:', lengths.mean()
print 'length stddev:', lengths.std()
print
print 'document lengths (sorted):', sorted(lengths)
plotter.histogram(lengths, '# tokens', '# documents', '', bins=80)
def dataset_stats(dataset):
"""
Print and plot statistics for a given dataset.
A histogram is plotted with the document length distribution of the data.
"""
print '> Reading data..', dataset
corpus_path = '../data/'+dataset
(documents, labels) = data.read_files(corpus_path)
file_names = data.get_file_names(corpus_path)
lengths = []
empty = 0
for i,d in enumerate(documents):
d = preprocess.tokenize_tokens(d)
lengths.append(len(d))
if len(d)==0:
print file_names[i],'is empty'
empty += 1
lengths = numpy.array(lengths)
print '# documents:',len(documents)
print '# empty documents:',empty
print '# words:',sum(lengths)
print 'length avg:',lengths.mean()
print 'length stddev:',lengths.std()
print
print 'document lengths (sorted):',sorted(lengths)
plotter.histogram(lengths,'# tokens','# documents','',bins=80)
def print_degree_distributions(dataset, context):
"""
Extracts degree distribution values from networks, and print them to
cvs-file.
**warning** overwrites if file exists.
"""
print '> Reading data..', dataset
corpus_path = '../data/' + dataset + '_text'
(documents, labels) = data.read_files(corpus_path)
degsfile = open(
'output/properties/cooccurrence/degrees_docs_' +
dataset.replace('/', '.'), 'w')
giant = nx.DiGraph()
print '> Building networks..'
for i, text in enumerate(documents):
if i % 10 == 0: print ' ', str(i) + '/' + str(len(documents))
g = graph_representation.construct_cooccurrence_network(
text, context=context)
giant.add_edges_from(g.edges())
degs = nx.degree(g).values()
degs = [str(d) for d in degs]
degsfile.write(','.join(degs) + '\n')
degsfile.close()
print '> Writing giant\'s distribution'
with open(
'output/properties/cooccurrence/degrees_giant_' +
dataset.replace('/', '.'), 'w') as f:
ds = nx.degree(giant).values()
ds = [str(d) for d in ds]
f.write(','.join(ds))
def test_document_lengths(dataset='mir'):
print '> Reading data..', dataset
path = '../data/' + dataset + '/problem_descriptions_preprocessed'
docs, _ = data.read_files(path)
names = data.get_file_names(path)
print "PROBLEM DESCRIPTIONS"
for i, d in enumerate(docs):
if not d:
print names[i], "is empty"
path = '../data/' + dataset + '/solutions_preprocessed'
docs, _ = data.read_files(path)
names = data.get_file_names(path)
print "SOLUTIONS"
for i, d in enumerate(docs):
if not d:
print names[i], "is empty"
def do_context_sentence_evaluation_classification():
"""
Experiment evaluating performance of sentences as contexts for
co-occurrence networks in the classification task.
"""
print '> Reading cases..'
path = '../data/tasa/TASA900_text'
texts, labels = data.read_files(path)
print '> Evaluating..'
graphs = []
results = {}
for text in texts:
g = graph_representation.construct_cooccurrence_network(
text, context='sentence')
graphs.append(g)
for metric in graph_representation.get_metrics():
print ' ', metric
vectors = graph_representation.graphs_to_vectors(graphs,
metric,
verbose=True)
score = evaluation.evaluate_classification(vectors, labels)
results[metric + ' (sentence)'] = score
data.pickle_to_file(results, 'output/class_context_sentence')
pp.pprint(results)
return results
def test_classification(orders=[1,2,3],order_weights=[1.0,1.53,1.51]):
"""
Test classification using different combinations of higher orders and weightings of these.
The list *orders* define which higher order relations to include.
The relative importance of the orders are defined by *order_weights*.
"""
print '> Reading cases..'
path = '../data/tasa/TASA900_text'
texts, labels = data.read_files(path)
filenames = data.get_file_names(path)
print '> Creating representations..'
rep = []
for i, text in enumerate(texts):
print ' '+str(i)+"/"+str(len(texts))
g = graph_representation.construct_cooccurrence_network(text, context='sentence', orders=orders, order_weights=order_weights, doc_id='output/higher_order/tasa/'+labels[i]+'/'+filenames[i])
d = graph_representation.graph_to_dict(g, graph.GraphMetrics.WEIGHTED_DEGREE)
rep.append(d)
rep = graph_representation.dicts_to_vectors(rep)
print '> Evaluating..'
score = evaluation.evaluate_classification(rep, labels)
print 'orders:', orders
print 'score:', score
fname = 'output/higher_order/results/class'
with open(fname, 'a+') as f:
s = reduce(lambda x,y:str(x)+str(y), orders)
f.write(str(s)+' '+str(score)+'\n')
return score
def classification_demo():
"""Function intended to illustrate classification in the experimental framework.
Intended as a basis for new experiments for those not intimately
familiar with the code.
"""
print 'Evaluation type: Classification'
print 'Graph type: Co-occurrence w/2-word window context'
print 'Centrality: Weighted degree'
print
print '> Reading data..'
corpus_path = '../data/tasa/TASA900_preprocessed'
docs, labels = data.read_files(corpus_path)
print '> Creating representations..'
dicts = []
for i, doc in enumerate(docs):
print ' ', str(i) + '/' + str(len(docs))
g = graph_representation.construct_cooccurrence_network(doc)
d = graph_representation.graph_to_dict(
g, graph.GraphMetrics.WEIGHTED_DEGREE)
dicts.append(d)
vectors = graph_representation.dicts_to_vectors(dicts)
print '> Evaluating..'
score = evaluation.evaluate_classification(vectors, labels)
print ' score:', score
print
def solution_similarity_stats(dataset='air/solutions_preprocessed'):
"""
Plots histogram of solution-solution similarity distribution of a dataset.
"""
print '> Reading data..', dataset
corpus_path = '../data/'+dataset
(documents, labels) = data.read_files(corpus_path)
print '> Creating vector representations..'
vectors = freq_representation.text_to_vector(documents, freq_representation.FrequencyMetrics.TF_IDF)
print '> Calculating similarities..'
distances = scipy.spatial.distance.cdist(vectors.T, vectors.T, 'cosine')
diag = numpy.diag([2.0]*len(distances),0) # move similarities of "self" to -1
distances = distances + diag
similarities = 1.0 - distances
similarities = similarities.ravel()
similarities = [s for s in similarities if s >= 0]
print plotter.histogram(similarities,'similarity','# matches','',bins=150)
print
print max(similarities)
print min(similarities)
print float(sum(similarities))/len(similarities)
num = len([sim for sim in similarities if sim < 0.23])
print 'fraction sims < .23:', float(num)/len(similarities)
def test_document_lengths(dataset='mir'):
print '> Reading data..', dataset
path = '../data/'+dataset+'/problem_descriptions_preprocessed'
docs, _ = data.read_files(path)
names = data.get_file_names(path)
print "PROBLEM DESCRIPTIONS"
for i, d in enumerate(docs):
if not d:
print names[i], "is empty"
path = '../data/'+dataset+'/solutions_preprocessed'
docs, _ = data.read_files(path)
names = data.get_file_names(path)
print "SOLUTIONS"
for i, d in enumerate(docs):
if not d:
print names[i], "is empty"
def classification_demo():
"""Function intended to illustrate classification in the experimental framework.
Intended as a basis for new experiments for those not intimately
familiar with the code.
"""
print 'Evaluation type: Classification'
print 'Graph type: Co-occurrence w/2-word window context'
print 'Centrality: Weighted degree'
print
print '> Reading data..'
corpus_path = '../data/tasa/TASA900_preprocessed'
docs, labels = data.read_files(corpus_path)
print '> Creating representations..'
dicts = []
for i, doc in enumerate(docs):
print ' ',str(i)+'/'+str(len(docs))
g = graph_representation.construct_cooccurrence_network(doc)
d = graph_representation.graph_to_dict(g, graph.GraphMetrics.WEIGHTED_DEGREE)
dicts.append(d)
vectors = graph_representation.dicts_to_vectors(dicts)
print '> Evaluating..'
score = evaluation.evaluate_classification(vectors, labels)
print ' score:', score
print
def do_retrieval_experiments(descriptions='air/problem_descriptions',
solutions='air/solutions',
graph_types=['co-occurrence','dependency','random'],
use_frequency=True):
"""
Experiment used for comparative evaluation of different network
representations on the retrieval task.
Toggle comparison with frequency-based methods using *use_frequency*.
"""
results = {'_solutions':solutions,
'_descriptions':descriptions,
'_evaluation':'retrieval'}
print '> Evaluation type: retrieval'
print '> Reading cases..'
descriptions_path = '../data/'+descriptions
descriptiondata = data.read_data(descriptions_path, graph_types)
solutions_path = '../data/'+solutions+'_preprocessed'
solution_texts, labels = data.read_files(solutions_path)
solution_vectors = freq_representation.text_to_vector(solution_texts, freq_representation.FrequencyMetrics.TF_IDF)
print '> Evaluating..'
for gtype in graph_types:
print ' ',gtype
docs, labels = descriptiondata[gtype]
graphs = graph_representation.create_graphs(docs, gtype)
results[gtype] = {}
for metric in graph_representation.get_metrics():
print ' -', metric
vectors = graph_representation.graphs_to_vectors(graphs, metric)
results[gtype][metric] = evaluation.evaluate_retrieval(vectors, solution_vectors)
if use_frequency:
print ' frequency'
results['freq'] = {}
for metric in freq_representation.get_metrics():
print ' -', metric
docs, labels = data.read_files(descriptions_path+'_preprocessed')
vectors = freq_representation.text_to_vector(docs, metric)
results['freq'][metric] = evaluation.evaluate_retrieval(vectors, solution_vectors)
print
pp.pprint(results)
return results
def evaluate_tc_icc_classification():
graph_metrics = graph_representation.get_metrics(True, exclude_flow=True)
print '> Reading cases..'
corpus = 'tasa/TASA900'
#~ corpus = 'tasa/TASATest2'
context = 'sentence'
path = '../data/' + corpus + '_text'
texts, labels = data.read_files(path)
rep = {}
icc = {}
print '> Calculating ICCs..'
for metric in graph_metrics:
print ' ', metric
rep[metric] = []
centralities = retrieve_centralities(corpus, context, metric)
if centralities:
icc[metric] = graph_representation.calculate_icc_dict(centralities)
else:
icc[metric] = None
print '> Creating graph representations..'
for i, text in enumerate(texts):
if i % 10 == 0: print ' ', str(i) + '/' + str(len(texts))
g = graph_representation.construct_cooccurrence_network(
text, context=context)
for metric in graph_metrics:
print ' ', metric
if not icc[metric]: continue
d = graph_representation.graph_to_dict(g, metric, icc[metric])
rep[metric].append(d)
g = None # just to make sure..
print '> Creating vector representations..'
for metric in graph_metrics:
if not icc[metric]: continue
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
results = {}
for metric in graph_metrics:
if not icc[metric]:
results[metric] = None
continue
vectors = rep[metric]
score = evaluation.evaluate_classification(vectors, labels)
print ' ', metric, score
results[metric] = score
pp.pprint(results)
data.pickle_to_file(
results,
'output/tc_icc/cooccurrence/' + corpus + '/classification.res')
return results
def test():
import data
dataset = 'test/freq'
corpus_path = '../data/'+dataset
(documents, labels) = data.read_files(corpus_path)
vectors = {}
print '> Building vector representations..'
for metric in get_metrics():
print ' ', metric
vectors[metric] = text_to_vector(documents, metric)
def test():
import data
dataset = 'test/freq'
corpus_path = '../data/' + dataset
(documents, labels) = data.read_files(corpus_path)
vectors = {}
print '> Building vector representations..'
for metric in get_metrics():
print ' ', metric
vectors[metric] = text_to_vector(documents, metric)
def store_corpus_network(corpus, context):
print '> Constructing corpus network for', corpus
path = '../data/'+corpus+'_text'
store_path = 'output/giants/co-occurrence/'+corpus+'/'+context+'_graph.net'
if data.pickle_from_file(store_path, suppress_warning=True):
print ' already present, skipping'
return
texts, labels = data.read_files(path)
gdoc = ' '.join(texts)
giant = graph_representation.construct_cooccurrence_network(gdoc, context=context, already_preprocessed=False, verbose=True)
print '> Serializing and saving..'
data.pickle_to_file(giant, store_path)
def vectorization(vector_type):
train = read_files('train')
test = read_files('test')
if vector_type == 'CountVectorizer':
vectorizer = CountVectorizer(ngram_range=(1, 2), min_df=3, max_df=0.9, max_features=2000)
elif vector_type == 'TfidfVectorizer':
vectorizer = TfidfVectorizer()
else:
vectorizer = HashingVectorizer(n_features=200)
vectorizer.fit(train['text'])
x_train = vectorizer.transform(train['text'])
y_train = train['label'].values
x_test = vectorizer.transform(test['text'])
y_test = test['label'].values
return x_train, y_train, x_test, y_test
def evaluate_tc_icc_classification():
graph_metrics = graph_representation.get_metrics(True, exclude_flow=True)
print '> Reading cases..'
corpus = 'tasa/TASA900'
#~ corpus = 'tasa/TASATest2'
context = 'sentence'
path = '../data/'+corpus+'_text'
texts, labels = data.read_files(path)
rep = {}
icc = {}
print '> Calculating ICCs..'
for metric in graph_metrics:
print ' ', metric
rep[metric] = []
centralities = retrieve_centralities(corpus, context, metric)
if centralities:
icc[metric] = graph_representation.calculate_icc_dict(centralities)
else:
icc[metric] = None
print '> Creating graph representations..'
for i, text in enumerate(texts):
if i%10==0: print ' ',str(i)+'/'+str(len(texts))
g = graph_representation.construct_cooccurrence_network(text, context=context)
for metric in graph_metrics:
print ' ', metric
if not icc[metric]: continue
d = graph_representation.graph_to_dict(g, metric, icc[metric])
rep[metric].append(d)
g = None # just to make sure..
print '> Creating vector representations..'
for metric in graph_metrics:
if not icc[metric]: continue
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
results = {}
for metric in graph_metrics:
if not icc[metric]:
results[metric] = None
continue
vectors = rep[metric]
score = evaluation.evaluate_classification(vectors, labels)
print ' ', metric, score
results[metric] = score
pp.pprint(results)
data.pickle_to_file(results, 'output/tc_icc/cooccurrence/'+corpus+'/classification.res')
return results
def retrieval_comparison_freq(dataset='mir'):
print '> Reading data..', dataset
path = '../data/'+dataset+'/problem_descriptions_preprocessed'
docs, _ = data.read_files(path)
print '> Creating solution representations..'
solutions_path = '../data/'+dataset+'/solutions_preprocessed'
solutions_docs, _ = data.read_files(solutions_path)
solutions_rep = freq_representation.text_to_vector(solutions_docs, freq_representation.FrequencyMetrics.TF_IDF)
print '> Evaluating..'
results = {}
for metric in freq_representation.get_metrics():
print ' ', metric,
descriptions_rep = freq_representation.text_to_vector(docs, metric)
score = evaluation.evaluate_retrieval(descriptions_rep, solutions_rep)
results[metric] = score
print score
pp.pprint(results)
s = 'retrieval comparison \nrepresentation: frequency\ndataset:'+dataset+' \nresult:\n'+str(results)+'\n\n\n'
data.write_to_file(s, 'output/comparison/retrieval')
return results
def store_corpus_network(corpus, context):
print '> Constructing corpus network for', corpus
path = '../data/' + corpus + '_text'
store_path = 'output/giants/co-occurrence/' + corpus + '/' + context + '_graph.net'
if data.pickle_from_file(store_path, suppress_warning=True):
print ' already present, skipping'
return
texts, labels = data.read_files(path)
gdoc = ' '.join(texts)
giant = graph_representation.construct_cooccurrence_network(
gdoc, context=context, already_preprocessed=False, verbose=True)
print '> Serializing and saving..'
data.pickle_to_file(giant, store_path)
def retrieval_comparison_freq(dataset='mir'):
print '> Reading data..', dataset
path = '../data/' + dataset + '/problem_descriptions_preprocessed'
docs, _ = data.read_files(path)
print '> Creating solution representations..'
solutions_path = '../data/' + dataset + '/solutions_preprocessed'
solutions_docs, _ = data.read_files(solutions_path)
solutions_rep = freq_representation.text_to_vector(
solutions_docs, freq_representation.FrequencyMetrics.TF_IDF)
print '> Evaluating..'
results = {}
for metric in freq_representation.get_metrics():
print ' ', metric,
descriptions_rep = freq_representation.text_to_vector(docs, metric)
score = evaluation.evaluate_retrieval(descriptions_rep, solutions_rep)
results[metric] = score
print score
pp.pprint(results)
s = 'retrieval comparison \nrepresentation: frequency\ndataset:' + dataset + ' \nresult:\n' + str(
results) + '\n\n\n'
data.write_to_file(s, 'output/comparison/retrieval')
return results
def freq_classification(dataset='tasa/TASA900'):
results = {'_dataset':dataset,
'_evaluation':'classification'}
corpus_path = '../data/'+dataset
results['results'] = {}
for metric in freq_representation.get_metrics():
print metric
documents, labels = data.read_files(corpus_path+'_preprocessed')
vectors = freq_representation.text_to_vector(documents, metric)
r = evaluation.evaluate_classification(vectors, labels, mode='cross-validation')
results['results'][metric] = r
print ' ', r
print
pp.pprint(results)
return results
def test_dependency_graph():
(docs, labels) = data.read_files('../data/tasa/TASA900_dependencies')
graphs = []
for i, text in enumerate(docs):
print i
graphs.append(construct_dependency_network(text))
g = graphs[0]
print g.nodes()
print g.edges()
print '#graphs:', len(graphs)
pos = nx.spring_layout(g)
graph.draw_with_centrality(g, layout=pos)
def test_vocabulary_size(path = '../data/air/problem_descriptions_preprocessed'):
"""
Print vocabulary sizes for documents in dataset.
"""
texts, labels = data.read_files(path)
lengths = []
for text in texts:
text = text.split(' ')
l = len(list(set(text)))
lengths.append(l)
print ' ',l
lengths = numpy.array(lengths)
print 'avg', lengths.mean()
print 'max', lengths.max()
print 'min', lengths.min()
def test_dependency_graph():
(docs, labels) = data.read_files('../data/tasa/TASA900_dependencies')
graphs = []
for i, text in enumerate(docs):
print i
graphs.append(construct_dependency_network(text))
g = graphs[0]
print g.nodes()
print g.edges()
print '#graphs:', len(graphs)
pos = nx.spring_layout(g)
graph.draw_with_centrality(g, layout=pos)
def freq_classification(dataset='tasa/TASA900'):
results = {'_dataset': dataset, '_evaluation': 'classification'}
corpus_path = '../data/' + dataset
results['results'] = {}
for metric in freq_representation.get_metrics():
print metric
documents, labels = data.read_files(corpus_path + '_preprocessed')
vectors = freq_representation.text_to_vector(documents, metric)
r = evaluation.evaluate_classification(vectors,
labels,
mode='cross-validation')
results['results'][metric] = r
print ' ', r
print
pp.pprint(results)
return results
def do_context_size_evaluation_classification():
"""
Experiment evaluating performance of different context sizes for
co-occurrence networks in the classification task.
"""
results = {}
graph_metrics = graph_representation.get_metrics()
for metric in graph_metrics:
results[metric] = []
print '> Reading cases..'
path = '../data/tasa/TASA900_preprocessed'
texts, labels = data.read_files(path)
for window_size in range(1, 11) + [20, 40, 80]:
print '-- window size:', window_size
rep = {}
for metric in graph_metrics:
rep[metric] = []
print '> Creating representations..'
# creating graphs and finding centralities
for text in texts:
g = graph_representation.construct_cooccurrence_network(
text, window_size=window_size, already_preprocessed=True)
for metric in graph_metrics:
d = graph_representation.graph_to_dict(g, metric)
rep[metric].append(d)
g = None # just to make sure..
# creating representation vectors
for metric in graph_metrics:
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
for metric in graph_metrics:
vectors = rep[metric]
score = evaluation.evaluate_classification(vectors, labels)
print ' ', metric, score
results[metric].append(score)
data.pickle_to_file(results,
'output/class_context_' + str(window_size))
pp.pprint(results)
return results
def store_corpus_network(corpus):
print '> Constructing corpus network for', corpus
path = '../data/'+corpus+'_dependencies'
store_path = 'output/giants/dependency/'+corpus+'/graph.net'
if data.pickle_from_file(store_path, suppress_warning=True):
print ' already present, skipping'
return
texts, labels = data.read_files(path)
gdeps = {}
for i, text in enumerate(texts):
if i%1==0: print ' ',str(i)+'/'+str(len(texts))
d = pickle.loads(text)
for dep in d.keys():
gdeps[dep] = gdeps.get(dep, []) + d[dep]
giant = graph_representation.construct_dependency_network(gdeps,verbose=True,unpickle=False)
print '> Serializing and saving..'
data.pickle_to_file(giant, store_path)
def do_context_size_evaluation_classification():
"""
Experiment evaluating performance of different context sizes for
co-occurrence networks in the classification task.
"""
results = {}
graph_metrics = graph_representation.get_metrics()
for metric in graph_metrics:
results[metric] = []
print '> Reading cases..'
path = '../data/tasa/TASA900_preprocessed'
texts, labels = data.read_files(path)
for window_size in range(1,11)+[20,40,80]:
print '-- window size:',window_size
rep = {}
for metric in graph_metrics:
rep[metric] = []
print '> Creating representations..'
# creating graphs and finding centralities
for text in texts:
g = graph_representation.construct_cooccurrence_network(text, window_size=window_size, already_preprocessed=True)
for metric in graph_metrics:
d = graph_representation.graph_to_dict(g, metric)
rep[metric].append(d)
g = None # just to make sure..
# creating representation vectors
for metric in graph_metrics:
rep[metric] = graph_representation.dicts_to_vectors(rep[metric])
print '> Evaluating..'
for metric in graph_metrics:
vectors = rep[metric]
score = evaluation.evaluate_classification(vectors, labels)
print ' ', metric, score
results[metric].append(score)
data.pickle_to_file(results, 'output/class_context_'+str(window_size))
pp.pprint(results)
return results
def test_best_classification():
print '> Reading cases..'
path = '../data/tasa/TASA900_text'
texts, labels = data.read_files(path)
rep = []
print '> Creating representations..'
for i, text in enumerate(texts):
if i%100==0: print ' ',i
g = graph_representation.construct_cooccurrence_network(text, context='sentence')
d = graph_representation.graph_to_dict(g, graph.GraphMetrics.WEIGHTED_DEGREE)
rep.append(d)
g = None # just to make sure..
rep = graph_representation.dicts_to_vectors(rep)
print '> Evaluating..'
score = evaluation.evaluate_classification(rep, labels)
print ' ', score
def store_corpus_network(corpus):
print '> Constructing corpus network for', corpus
path = '../data/' + corpus + '_dependencies'
store_path = 'output/giants/dependency/' + corpus + '/graph.net'
if data.pickle_from_file(store_path, suppress_warning=True):
print ' already present, skipping'
return
texts, labels = data.read_files(path)
gdeps = {}
for i, text in enumerate(texts):
if i % 1 == 0: print ' ', str(i) + '/' + str(len(texts))
d = pickle.loads(text)
for dep in d.keys():
gdeps[dep] = gdeps.get(dep, []) + d[dep]
giant = graph_representation.construct_dependency_network(gdeps,
verbose=True,
unpickle=False)
print '> Serializing and saving..'
data.pickle_to_file(giant, store_path)
def test_best_classification():
print '> Reading cases..'
path = '../data/tasa/TASA900_text'
texts, labels = data.read_files(path)
rep = []
print '> Creating representations..'
for i, text in enumerate(texts):
if i % 100 == 0: print ' ', i
g = graph_representation.construct_cooccurrence_network(
text, context='sentence')
d = graph_representation.graph_to_dict(
g, graph.GraphMetrics.WEIGHTED_DEGREE)
rep.append(d)
g = None # just to make sure..
rep = graph_representation.dicts_to_vectors(rep)
print '> Evaluating..'
score = evaluation.evaluate_classification(rep, labels)
print ' ', score
def do_classification_experiments(
dataset='tasa/TASA900',
graph_types=['co-occurrence', 'dependency', 'random'],
use_frequency=True):
"""
Experiment used for comparative evaluation of different network
representations on classification.
Toggle comparison with frequency-based methods using *use_frequency*.
"""
results = {'_dataset': dataset, '_evaluation': 'classification'}
print '> Evaluation type: classification'
print '> Reading data..', dataset
corpus_path = '../data/' + dataset
docdata = data.read_data(corpus_path, graph_types)
print '> Evaluating..'
for gtype in graph_types:
print ' ', gtype
documents, labels = docdata[gtype]
graphs = graph_representation.create_graphs(documents, gtype)
results[gtype] = {}
for metric in graph_representation.get_metrics():
print ' -', metric
vectors = graph_representation.graphs_to_vectors(graphs, metric)
results[gtype][metric] = evaluation.evaluate_classification(
vectors, labels)
if use_frequency:
print ' frequency'
results['freq'] = {}
for metric in freq_representation.get_metrics():
print ' -', metric
documents, labels = data.read_files(corpus_path + '_preprocessed')
vectors = freq_representation.text_to_vector(documents, metric)
results['freq'][metric] = evaluation.evaluate_classification(
vectors, labels)
print
pp.pprint(results)
return results
def corpus_dependency_properties(dataset='air/problem_descriptions'):
"""
Identify and pickle to file various properties of the given dataset.
These can alter be converted to pretty tables using
:func:`~experiments.print_network_props`.
"""
print '> Reading data..', dataset
corpus_path = '../data/' + dataset + '_dependencies'
(documents, labels) = data.read_files(corpus_path)
props = {}
giant = nx.DiGraph()
print '> Building networks..'
for i, text in enumerate(documents):
if i % 10 == 0: print ' ', str(i) + '/' + str(len(documents))
g = graph_representation.construct_dependency_network(
text, remove_stop_words=True)
giant.add_edges_from(g.edges())
p = graph.network_properties(g)
for k, v in p.iteritems():
if i == 0: props[k] = []
props[k].append(v)
g = None # just to make sure..
print '> Calculating means and deviations..'
props_total = {}
for key in props:
props_total[key + '_mean'] = numpy.mean(props[key])
props_total[key + '_std'] = numpy.std(props[key])
data.pickle_to_file(
giant,
'output/properties/dependency/corpus_network_air_all_no_stop_words')
data.pickle_to_file(
props, 'output/properties/dependency/docs_air_all_no_stop_words')
data.pickle_to_file(
props_total,
'output/properties/dependency/docs_air_all_no_stop_words_total')