def runBaseline():
current_alg = "baseline"
print("Testing " + current_alg + " on " + train_path + " and " + dev_path + "...")
# prelim set-up
train = util_baseline.readExamples(train_path)
dev = util_baseline.readExamples(dev_path)
featureExtractor = baseline.extractWordFeatures
# calculate weights
weights = baseline.learnPredictor(train, dev, featureExtractor, eta=0.1, numIters=10)
# performance analysis
trainError = util_baseline.evaluatePredictor(train, lambda(x) : \
(1 if util_baseline.dotProduct(featureExtractor(x), weights) >= 0 else -1))
devError = util_baseline.evaluatePredictor(dev, lambda(x) : \
(1 if util_baseline.dotProduct(featureExtractor(x), weights) >= 0 else -1))
print "Error percentage on train examples: " + str(trainError)
print "Error percentage on dev examples: " + str(devError)
util_baseline.outputErrorAnalysis(train, featureExtractor, weights, testDir, 'train-error-analysis')
util_baseline.outputErrorAnalysis(dev, featureExtractor, weights, testDir, 'dev-error-analysis')
def windowPredictor(raw_in, features, weights):
#---Params:
use_punc = False
window_size = 4 #Best is 4
flip_thresh = 0.7 #Best is 0.4
up_thresh = 0.3
down_thresh = 0
upscale_amount = 2.5 #Best is 2.5
downscale_amount = 0 #Best is 0
flip_amount = -1 #Best is 1
#---
w = weights.copy()
negs = ["hardly", "neither", "no", "nobody","not","cannot","didn't","haven't","never","lacking","lack","wasn't","isn't","can't","won't","doesn't","hasn't","without"] #ISSUE: Should use stemming.
pivots = ["but", "however"]
amps = ['much', 'very']
puncs = ['.', ',']
raw_list = raw_in.split()
neg_positions = [raw_list.index(neg) for neg in raw_list if neg in negs]
pivot_positions = [raw_list.index(pivot) for pivot in raw_list if pivot in pivots]
amp_positions = [raw_list.index(amp) for amp in raw_list if amp in amps]
punc_positions = [raw_list.index(punc) for punc in raw_list if punc in puncs]
change_words = set()
flip_words = set()
upscale_words = set()
downscale_words = set()
for pos in neg_positions:
if use_punc:
for p in punc_positions:
if p > pos:
break
segment = raw_list[pos+1:pos+1+window_size]
else:
segment = raw_list[pos+1:pos+1+window_size]
for word in segment:
change_words.add(word)
flip_words.add(word)
for pos in amp_positions:
if use_punc:
for p in punc_positions:
if p > pos:
break
segment = raw_list[pos+1:p]
else:
segment = raw_list[pos+1:pos+1+window_size]
for word in segment:
change_words.add(word)
upscale_words.add(word)
for pos in pivot_positions:
if use_punc:
for p in punc_positions:
if p > pos:
break
up_segment = raw_list[pos+1:p] #How to treat down case?
else:
up_segment = raw_list[pos+1:pos+1+window_size]
down_segment = raw_list[pos-window_size:pos]
for word in down_segment:
#print(raw_in)
change_words.add(word)
downscale_words.add(word)
for word in up_segment:
change_words.add(word)
upscale_words.add(word)
for word in change_words:
try: #ISSUE: What should we do if there are unseen words?
if word in flip_words and abs(w[word]) >= flip_thresh:
w[word] = w[word]*(flip_amount)
if word in upscale_words and abs(w[word]) >= up_thresh:
w[word] = w[word]*(upscale_amount)
if word in downscale_words and abs(w[word]) >= down_thresh:
w[word] = w[word]*(downscale_amount)
except KeyError:
pass
return (1 if util_baseline.dotProduct(features, w) >= 0 else -1)