Spaces:
Sleeping
Sleeping
Commit
·
e91e5d5
1
Parent(s):
7e9f859
Enhance subjectivity prediction with detailed output and update Gradio interface
Browse files
app.py
CHANGED
@@ -7,7 +7,8 @@ import torch.nn as nn
|
|
7 |
|
8 |
# Define the model and tokenizer
|
9 |
model_card = "microsoft/mdeberta-v3-base"
|
10 |
-
finetuned_model = "MatteoFasulo/mdeberta-v3-base-subjectivity-sentiment-multilingual"
|
|
|
11 |
|
12 |
# Custom model class for combining sentiment analysis with subjectivity detection
|
13 |
class CustomModel(PreTrainedModel):
|
@@ -22,7 +23,7 @@ class CustomModel(PreTrainedModel):
|
|
22 |
|
23 |
self.classifier = nn.Linear(output_dim + sentiment_dim, num_labels)
|
24 |
|
25 |
-
def forward(self, input_ids, positive, neutral, negative, attention_mask=None, labels=None):
|
26 |
outputs = self.deberta(input_ids=input_ids, attention_mask=attention_mask)
|
27 |
|
28 |
encoder_layer = outputs[0]
|
@@ -66,24 +67,48 @@ def get_sentiment_values(text: str):
|
|
66 |
sentiments = pipe(text)[0]
|
67 |
return {k:v for k,v in [(list(sentiment.values())[0], list(sentiment.values())[1]) for sentiment in sentiments]}
|
68 |
|
69 |
-
#
|
70 |
def predict_subjectivity(text):
|
71 |
sentiment_values = get_sentiment_values(text)
|
72 |
|
73 |
model = load_model(model_card, finetuned_model)
|
74 |
tokenizer = load_tokenizer(model_card)
|
75 |
|
|
|
|
|
|
|
|
|
76 |
inputs = tokenizer(text, padding=True, truncation=True, max_length=256, return_tensors='pt')
|
|
|
|
|
|
|
77 |
|
78 |
outputs = model(**inputs)
|
79 |
logits = outputs.get('logits')
|
80 |
|
81 |
-
|
|
|
|
|
|
|
|
|
|
|
82 |
predicted_class = model.config.id2label[predicted_class_idx]
|
83 |
|
84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
|
86 |
-
#
|
87 |
demo = gr.Interface(
|
88 |
fn=predict_subjectivity,
|
89 |
inputs=gr.Textbox(
|
@@ -91,14 +116,12 @@ demo = gr.Interface(
|
|
91 |
placeholder='Enter a sentence from a news article',
|
92 |
info='Paste a sentence from a news article to determine if it is subjective or objective.'
|
93 |
),
|
94 |
-
outputs=gr.
|
95 |
-
label="
|
96 |
-
info="
|
97 |
),
|
98 |
title='Subjectivity Detection',
|
99 |
-
description='Detect if a sentence is subjective or objective using a pre-trained model.'
|
100 |
-
theme='huggingface',
|
101 |
)
|
102 |
|
103 |
-
|
104 |
-
demo.launch(share=True)
|
|
|
7 |
|
8 |
# Define the model and tokenizer
|
9 |
model_card = "microsoft/mdeberta-v3-base"
|
10 |
+
finetuned_model = "MatteoFasulo/mdeberta-v3-base-subjectivity-sentiment-multilingual-no-arabic"
|
11 |
+
THRESHOLD = 0.65
|
12 |
|
13 |
# Custom model class for combining sentiment analysis with subjectivity detection
|
14 |
class CustomModel(PreTrainedModel):
|
|
|
23 |
|
24 |
self.classifier = nn.Linear(output_dim + sentiment_dim, num_labels)
|
25 |
|
26 |
+
def forward(self, input_ids, positive, neutral, negative, token_type_ids=None, attention_mask=None, labels=None):
|
27 |
outputs = self.deberta(input_ids=input_ids, attention_mask=attention_mask)
|
28 |
|
29 |
encoder_layer = outputs[0]
|
|
|
67 |
sentiments = pipe(text)[0]
|
68 |
return {k:v for k,v in [(list(sentiment.values())[0], list(sentiment.values())[1]) for sentiment in sentiments]}
|
69 |
|
70 |
+
# Modify the predict_subjectivity function to return additional information
|
71 |
def predict_subjectivity(text):
|
72 |
sentiment_values = get_sentiment_values(text)
|
73 |
|
74 |
model = load_model(model_card, finetuned_model)
|
75 |
tokenizer = load_tokenizer(model_card)
|
76 |
|
77 |
+
positive = sentiment_values['positive']
|
78 |
+
neutral = sentiment_values['neutral']
|
79 |
+
negative = sentiment_values['negative']
|
80 |
+
|
81 |
inputs = tokenizer(text, padding=True, truncation=True, max_length=256, return_tensors='pt')
|
82 |
+
inputs['positive'] = torch.tensor(positive).unsqueeze(0)
|
83 |
+
inputs['neutral'] = torch.tensor(neutral).unsqueeze(0)
|
84 |
+
inputs['negative'] = torch.tensor(negative).unsqueeze(0)
|
85 |
|
86 |
outputs = model(**inputs)
|
87 |
logits = outputs.get('logits')
|
88 |
|
89 |
+
# Calculate probabilities using softmax
|
90 |
+
probabilities = torch.nn.functional.softmax(logits, dim=1)
|
91 |
+
obj_prob, subj_prob = probabilities[0].tolist()
|
92 |
+
|
93 |
+
# Predict the class given the decision threshold
|
94 |
+
predicted_class_idx = 1 if subj_prob >= THRESHOLD else 0
|
95 |
predicted_class = model.config.id2label[predicted_class_idx]
|
96 |
|
97 |
+
# Format the output
|
98 |
+
result = f"""Prediction: {predicted_class}
|
99 |
+
|
100 |
+
Class Probabilities:
|
101 |
+
- Objective: {obj_prob:.2%}
|
102 |
+
- Subjective: {subj_prob:.2%}
|
103 |
+
|
104 |
+
Sentiment Scores:
|
105 |
+
- Positive: {positive:.2%}
|
106 |
+
- Neutral: {neutral:.2%}
|
107 |
+
- Negative: {negative:.2%}"""
|
108 |
+
|
109 |
+
return result
|
110 |
|
111 |
+
# Update the Gradio interface
|
112 |
demo = gr.Interface(
|
113 |
fn=predict_subjectivity,
|
114 |
inputs=gr.Textbox(
|
|
|
116 |
placeholder='Enter a sentence from a news article',
|
117 |
info='Paste a sentence from a news article to determine if it is subjective or objective.'
|
118 |
),
|
119 |
+
outputs=gr.Textbox(
|
120 |
+
label="Results",
|
121 |
+
info="Detailed analysis including subjectivity prediction, class probabilities, and sentiment scores."
|
122 |
),
|
123 |
title='Subjectivity Detection',
|
124 |
+
description='Detect if a sentence is subjective or objective using a pre-trained model.'
|
|
|
125 |
)
|
126 |
|
127 |
+
demo.launch()
|
|