Aidan Phillips
commited on
Commit
·
f5893dd
1
Parent(s):
d2375b8
clean up fluency code
Browse files- categories/fluency.py +100 -46
- scorer.ipynb +11 -12
categories/fluency.py
CHANGED
@@ -5,86 +5,126 @@ import numpy as np
|
|
5 |
import spacy
|
6 |
import wordfreq
|
7 |
|
|
|
|
|
|
|
|
|
8 |
tool = language_tool_python.LanguageTool('en-US')
|
|
|
|
|
9 |
model_name="distilbert-base-multilingual-cased"
|
10 |
-
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
11 |
model = AutoModelForMaskedLM.from_pretrained(model_name)
|
12 |
model.eval()
|
|
|
13 |
|
|
|
14 |
nlp = spacy.load("en_core_web_sm")
|
15 |
|
16 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
return -np.log(wordfreq.word_frequency(word, lang) + 1e-12)
|
18 |
|
19 |
-
def
|
20 |
"""
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
}
|
30 |
-
]
|
31 |
-
}
|
32 |
"""
|
33 |
-
|
34 |
-
|
35 |
-
# print(input_ids)
|
36 |
-
offset_mapping = encoding["offset_mapping"][0]
|
37 |
-
# print(offset_mapping)
|
38 |
-
tokens = tokenizer.convert_ids_to_tokens(input_ids)
|
39 |
-
|
40 |
-
# Group token indices by word based on offset mapping
|
41 |
-
word_groups = []
|
42 |
current_group = []
|
43 |
-
|
44 |
prev_end = None
|
45 |
-
|
46 |
for i, (start, end) in enumerate(offset_mapping):
|
47 |
if input_ids[i] in tokenizer.all_special_ids:
|
48 |
continue # skip special tokens like [CLS] and [SEP]
|
49 |
-
|
50 |
if prev_end is not None and start > prev_end:
|
51 |
# Word boundary detected → start new group
|
52 |
-
|
53 |
current_group = [i]
|
54 |
else:
|
55 |
current_group.append(i)
|
56 |
-
|
57 |
prev_end = end
|
58 |
-
|
59 |
# Append final group
|
60 |
if current_group:
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
loss_values = []
|
64 |
for group in word_groups:
|
|
|
65 |
if group[0] == 0 or group[-1] == len(input_ids) - 1:
|
66 |
-
continue
|
67 |
|
|
|
68 |
masked = input_ids.clone()
|
69 |
for i in group:
|
70 |
masked[i] = tokenizer.mask_token_id
|
71 |
|
|
|
72 |
with torch.no_grad():
|
73 |
outputs = model(masked.unsqueeze(0))
|
74 |
logits = outputs.logits[0]
|
75 |
|
76 |
log_probs = []
|
77 |
for i in group:
|
|
|
78 |
probs = torch.softmax(logits[i], dim=-1)
|
79 |
true_token_id = input_ids[i].item()
|
80 |
prob = probs[true_token_id].item()
|
|
|
81 |
log_probs.append(np.log(prob + 1e-12))
|
82 |
|
|
|
83 |
word_loss = -np.sum(log_probs) / len(log_probs)
|
|
|
84 |
word = tokenizer.decode(input_ids[group[0]])
|
85 |
-
word_loss -= 0.6 *
|
86 |
loss_values.append(word_loss)
|
87 |
|
|
|
|
|
|
|
88 |
errors = []
|
89 |
for i, l in enumerate(loss_values):
|
90 |
if l < threshold:
|
@@ -92,36 +132,43 @@ def pseudo_perplexity(text, threshold=20, max_len=128):
|
|
92 |
errors.append({
|
93 |
"start": i,
|
94 |
"end": i,
|
95 |
-
"message": f"
|
96 |
})
|
97 |
|
98 |
-
error_rate = len(errors) / len(loss_values)
|
99 |
-
|
100 |
res = {
|
101 |
-
"score":
|
102 |
"errors": errors
|
103 |
}
|
104 |
|
105 |
return res
|
106 |
|
107 |
-
def
|
108 |
"""
|
109 |
-
|
110 |
-
|
111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
"""
|
113 |
-
score = 100 / (1 + np.exp(steepness * (
|
114 |
return round(score, 2)
|
115 |
|
116 |
def grammar_errors(text) -> tuple[int, list[str]]:
|
117 |
"""
|
|
|
118 |
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
|
|
123 |
"""
|
124 |
-
|
125 |
matches = tool.check(text)
|
126 |
|
127 |
r = []
|
@@ -221,3 +268,10 @@ def __check_structural_grammar(text):
|
|
221 |
})
|
222 |
|
223 |
return issues
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
import spacy
|
6 |
import wordfreq
|
7 |
|
8 |
+
# setup global variables on import (bad practice, but whatever)
|
9 |
+
#--------------------------------------------------------------
|
10 |
+
|
11 |
+
# grammar checker
|
12 |
tool = language_tool_python.LanguageTool('en-US')
|
13 |
+
|
14 |
+
# masked language model and tokenizer from huggingface
|
15 |
model_name="distilbert-base-multilingual-cased"
|
|
|
16 |
model = AutoModelForMaskedLM.from_pretrained(model_name)
|
17 |
model.eval()
|
18 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name) # tokenizer
|
19 |
|
20 |
+
# spacy model for parsing
|
21 |
nlp = spacy.load("en_core_web_sm")
|
22 |
|
23 |
+
def __get_rarity(word, lang="en") -> float:
|
24 |
+
"""
|
25 |
+
Returns the rarity of a word in the given language. word_freq retuns a value
|
26 |
+
between 0 and 1, where 1 is the most common word. Therefore, taking the log results
|
27 |
+
in a value between 0 (log 1 = 0) and -27.63 (log 1e-12). We then negate it so super
|
28 |
+
rare words have a high score and common words have a low score.
|
29 |
+
|
30 |
+
Parameters:
|
31 |
+
word (str): The word to check.
|
32 |
+
lang (str): The language to check. Default is "en".
|
33 |
+
|
34 |
+
Returns:
|
35 |
+
float: The rarity of the word.
|
36 |
+
"""
|
37 |
return -np.log(wordfreq.word_frequency(word, lang) + 1e-12)
|
38 |
|
39 |
+
def __produce_groupings(offset_mapping, input_ids):
|
40 |
"""
|
41 |
+
Produce groupings of tokens that are part of the same word.
|
42 |
+
|
43 |
+
Parameters:
|
44 |
+
offset_mapping (list): The offset mapping of the tokens.
|
45 |
+
input_ids (list): The input ids of the tokens.
|
46 |
+
|
47 |
+
Returns:
|
48 |
+
list: A list of groupings of tokens.
|
|
|
|
|
|
|
49 |
"""
|
50 |
+
# Produce groupings of tokens that are part of the same word
|
51 |
+
res = []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
current_group = []
|
|
|
53 |
prev_end = None
|
|
|
54 |
for i, (start, end) in enumerate(offset_mapping):
|
55 |
if input_ids[i] in tokenizer.all_special_ids:
|
56 |
continue # skip special tokens like [CLS] and [SEP]
|
|
|
57 |
if prev_end is not None and start > prev_end:
|
58 |
# Word boundary detected → start new group
|
59 |
+
res.append(current_group)
|
60 |
current_group = [i]
|
61 |
else:
|
62 |
current_group.append(i)
|
|
|
63 |
prev_end = end
|
|
|
64 |
# Append final group
|
65 |
if current_group:
|
66 |
+
res.append(current_group)
|
67 |
+
|
68 |
+
return res
|
69 |
+
|
70 |
+
def pseudo_perplexity(text, threshold=4, max_len=128):
|
71 |
+
"""
|
72 |
+
Calculate the pseudo-perplexity of a text using a masked language model. Return all
|
73 |
+
words that exceed a threshold of "adjusted awkwardness". The threshold is a measure
|
74 |
+
in terms of log probability of the word.
|
75 |
+
|
76 |
+
Parameters:
|
77 |
+
text (str): The text to check.
|
78 |
+
threshold (float): The threshold for awkwardness. Default is 4.
|
79 |
+
max_len (int): The maximum length of the text. Default is 128.
|
80 |
+
|
81 |
+
Returns:
|
82 |
+
dict: A dictionary containing the score and errors.
|
83 |
+
"""
|
84 |
|
85 |
+
# Tokenize the text and produce groupings
|
86 |
+
encoding = tokenizer(text, return_tensors="pt", return_offsets_mapping=True)
|
87 |
+
input_ids = encoding["input_ids"][0]
|
88 |
+
offset_mapping = encoding["offset_mapping"][0]
|
89 |
+
tokens = tokenizer.convert_ids_to_tokens(input_ids)
|
90 |
+
word_groups = __produce_groupings(offset_mapping, input_ids)
|
91 |
+
|
92 |
+
# Calculate the loss for each word group
|
93 |
loss_values = []
|
94 |
for group in word_groups:
|
95 |
+
# Skip special tokens (CLS and SEP)
|
96 |
if group[0] == 0 or group[-1] == len(input_ids) - 1:
|
97 |
+
continue
|
98 |
|
99 |
+
# Mask the word group
|
100 |
masked = input_ids.clone()
|
101 |
for i in group:
|
102 |
masked[i] = tokenizer.mask_token_id
|
103 |
|
104 |
+
# Get the model output distribution
|
105 |
with torch.no_grad():
|
106 |
outputs = model(masked.unsqueeze(0))
|
107 |
logits = outputs.logits[0]
|
108 |
|
109 |
log_probs = []
|
110 |
for i in group:
|
111 |
+
# Get the probability of the true token
|
112 |
probs = torch.softmax(logits[i], dim=-1)
|
113 |
true_token_id = input_ids[i].item()
|
114 |
prob = probs[true_token_id].item()
|
115 |
+
# Append the loss of the true token
|
116 |
log_probs.append(np.log(prob + 1e-12))
|
117 |
|
118 |
+
# Calculate the loss for the entire word group
|
119 |
word_loss = -np.sum(log_probs) / len(log_probs)
|
120 |
+
# Adjust the loss based on the rarity of the word
|
121 |
word = tokenizer.decode(input_ids[group[0]])
|
122 |
+
word_loss -= 0.6 * __get_rarity(word) # subtract rarity (rare words reduce loss)
|
123 |
loss_values.append(word_loss)
|
124 |
|
125 |
+
# Structure the results for output
|
126 |
+
average_loss = np.mean(loss_values)
|
127 |
+
|
128 |
errors = []
|
129 |
for i, l in enumerate(loss_values):
|
130 |
if l < threshold:
|
|
|
132 |
errors.append({
|
133 |
"start": i,
|
134 |
"end": i,
|
135 |
+
"message": f"Adjusted liklihood {l} over threshold {threshold}"
|
136 |
})
|
137 |
|
|
|
|
|
138 |
res = {
|
139 |
+
"score": __fluency_score(average_loss),
|
140 |
"errors": errors
|
141 |
}
|
142 |
|
143 |
return res
|
144 |
|
145 |
+
def __fluency_score(loss, midpoint=5, steepness=0.3):
|
146 |
"""
|
147 |
+
Transform the loss into a score from 0 to 100. Steepness controls how quickly the
|
148 |
+
score drops as loss increases and midpoint controls the loss at which the score is
|
149 |
+
50.
|
150 |
+
|
151 |
+
Parameters:
|
152 |
+
loss (float): The loss to transform.
|
153 |
+
midpoint (float): The loss at which the score is 50. Default is 5.
|
154 |
+
steepness (float): The steepness of the curve. Default is 0.3.
|
155 |
+
|
156 |
+
Returns:
|
157 |
+
float: The score from 0 to 100.
|
158 |
"""
|
159 |
+
score = 100 / (1 + np.exp(steepness * (loss - midpoint)))
|
160 |
return round(score, 2)
|
161 |
|
162 |
def grammar_errors(text) -> tuple[int, list[str]]:
|
163 |
"""
|
164 |
+
Check the grammar of a text using a grammar checker and a structural grammar check.
|
165 |
|
166 |
+
Parameters:
|
167 |
+
text (str): The text to check.
|
168 |
+
|
169 |
+
Returns:
|
170 |
+
dict: A dictionary containing the score and errors.
|
171 |
"""
|
|
|
172 |
matches = tool.check(text)
|
173 |
|
174 |
r = []
|
|
|
268 |
})
|
269 |
|
270 |
return issues
|
271 |
+
|
272 |
+
|
273 |
+
def main():
|
274 |
+
pass
|
275 |
+
|
276 |
+
if __name__ == "__main__":
|
277 |
+
main()
|
scorer.ipynb
CHANGED
@@ -11,14 +11,14 @@
|
|
11 |
},
|
12 |
{
|
13 |
"cell_type": "code",
|
14 |
-
"execution_count":
|
15 |
"metadata": {},
|
16 |
"outputs": [
|
17 |
{
|
18 |
"name": "stdout",
|
19 |
"output_type": "stream",
|
20 |
"text": [
|
21 |
-
"Sentence:
|
22 |
]
|
23 |
}
|
24 |
],
|
@@ -31,23 +31,22 @@
|
|
31 |
"print(\"Sentence:\", s) # Print the input sentence\n",
|
32 |
"\n",
|
33 |
"err = grammar_errors(s) # Call the function to execute the grammar error checking\n",
|
34 |
-
"flu = pseudo_perplexity(s, threshold=3.
|
35 |
]
|
36 |
},
|
37 |
{
|
38 |
"cell_type": "code",
|
39 |
-
"execution_count":
|
40 |
"metadata": {},
|
41 |
"outputs": [
|
42 |
{
|
43 |
"name": "stdout",
|
44 |
"output_type": "stream",
|
45 |
"text": [
|
46 |
-
"
|
47 |
-
"Perplexity 4.
|
48 |
-
"Perplexity
|
49 |
-
"Perplexity
|
50 |
-
"Perplexity 5.1115574262487735 over threshold 3.5: apples\n"
|
51 |
]
|
52 |
}
|
53 |
],
|
@@ -61,15 +60,15 @@
|
|
61 |
},
|
62 |
{
|
63 |
"cell_type": "code",
|
64 |
-
"execution_count":
|
65 |
"metadata": {},
|
66 |
"outputs": [
|
67 |
{
|
68 |
"name": "stdout",
|
69 |
"output_type": "stream",
|
70 |
"text": [
|
71 |
-
"
|
72 |
-
"Fluency Score:
|
73 |
]
|
74 |
}
|
75 |
],
|
|
|
11 |
},
|
12 |
{
|
13 |
"cell_type": "code",
|
14 |
+
"execution_count": 11,
|
15 |
"metadata": {},
|
16 |
"outputs": [
|
17 |
{
|
18 |
"name": "stdout",
|
19 |
"output_type": "stream",
|
20 |
"text": [
|
21 |
+
"Sentence: caveman speak weird few word good\n"
|
22 |
]
|
23 |
}
|
24 |
],
|
|
|
31 |
"print(\"Sentence:\", s) # Print the input sentence\n",
|
32 |
"\n",
|
33 |
"err = grammar_errors(s) # Call the function to execute the grammar error checking\n",
|
34 |
+
"flu = pseudo_perplexity(s, threshold=3.25) # Call the function to execute the fluency checking"
|
35 |
]
|
36 |
},
|
37 |
{
|
38 |
"cell_type": "code",
|
39 |
+
"execution_count": 12,
|
40 |
"metadata": {},
|
41 |
"outputs": [
|
42 |
{
|
43 |
"name": "stdout",
|
44 |
"output_type": "stream",
|
45 |
"text": [
|
46 |
+
"This sentence does not start with an uppercase letter.: caveman speak\n",
|
47 |
+
"Perplexity 4.2750282429106585 over threshold 3.25: caveman\n",
|
48 |
+
"Perplexity 5.191700905668536 over threshold 3.25: few\n",
|
49 |
+
"Perplexity 3.8370066187600944 over threshold 3.25: good\n"
|
|
|
50 |
]
|
51 |
}
|
52 |
],
|
|
|
60 |
},
|
61 |
{
|
62 |
"cell_type": "code",
|
63 |
+
"execution_count": 10,
|
64 |
"metadata": {},
|
65 |
"outputs": [
|
66 |
{
|
67 |
"name": "stdout",
|
68 |
"output_type": "stream",
|
69 |
"text": [
|
70 |
+
"100.0 80.14\n",
|
71 |
+
"Fluency Score: 90.07\n"
|
72 |
]
|
73 |
}
|
74 |
],
|