Main

app.py

@@ -0,0 +1,41 @@
+import gradio as gr
+from sentiwordnet_calculator import SentimentPipeline
+
+pipe = SentimentPipeline("Tanor/SRGPTSENTPOS4", "Tanor/SRGPTSENTNEG4")
+
+
+def calculate(text):
+    result = pipe(text)
+    # Visual representation
+    visual = result
+    # Numerical representation
+    numerical = {key: round(value, 2) for key, value in result.items()}
+    # Create a formatted string
+    numerical_str = ", ".join(f"{key}: {value}" for key, value in numerical.items())
+    return visual, numerical_str
+
+iface = gr.Interface(
+    fn=calculate, 
+    inputs=gr.inputs.Textbox(lines=5, placeholder="Enter your text here..."), 
+    outputs=[gr.outputs.Label(num_top_classes=3), "text"],
+    title="Sentiment Analysis for Serbian",
+    description="""
+    This tool performs sentiment analysis on the input text using a model trained on Serbian dictionary definitions. 
+    The pretrained model [sr-gpt2-large model by Mihailo Škorić](https://huggingface.co/JeRTeh/sr-gpt2-large), 
+    was fine-tuned on selected definitions from the Serbian WordNet. Please limit the input to 300 tokens. 
+    The outputs represent the Positive (POS), Negative (NEG), and Objective (OBJ) sentiment scores.
+    """,
+    examples=[
+        ["osoba koja ne prihvata nove ideje"],
+        ["intenzivna ojađenost"],
+        ["uopštenih osećanja tuge"],
+        ["žalostan zbog gubitka ili uskraćenosti"],
+        ["činjenje dobra; osećaj dobrotvornosti"],
+        ["Jako pozitivno osećanje poštovanja i privrežen..."],
+        ["usrećiti ili zadovoljiti"],
+        ["Korisna ili vredna osobina"],
+    ]
+
+)
+
+iface.launch()

requirements.txt

@@ -0,0 +1,4 @@
+transformers==4.28.1
+#git+https://github.com/huggingface/transformers
+transformers[torch]
+xformers

sentiwordnet_calculator.py

@@ -0,0 +1,64 @@
+from transformers import pipeline
+
+class SentimentPipeline:
+    """
+    This class defines a custom sentiment analysis pipeline using Hugging Face's Transformers.
+    
+    The pipeline uses two separate models for predicting positive/non-positive and 
+    negative/non-negative sentiment respectively.
+
+    Inputs:
+        Single text string or a list of text strings for sentiment analysis.
+
+    Returns:
+        If a single text string is provided, a single dictionary is returned with POS, NEG, and OBJ scores.
+        If a list of text strings is provided, a list of dictionaries is returned with each dictionary 
+        representing POS, NEG, and OBJ scores for the corresponding text.
+
+    Usage:
+        sentiment_pipeline = SentimentPipeline(YOUR_POS_MODEL, YOUR_NEG_MODEL)
+        result = sentiment_pipeline("Your glossed text here")
+        results = sentiment_pipeline(["Your first glossed text here", "Your second glossed text here"])
+    """
+
+    def __init__(self, model_path_positive, model_path_negative):
+        """
+        Constructor for the SentimentPipeline class.
+        Initializes two pipelines using Hugging Face's Transformers, one for positive and one for negative sentiment.
+        """
+        self.pos_pipeline = pipeline('text-classification', model=model_path_positive)
+        self.neg_pipeline = pipeline('text-classification', model=model_path_negative)
+
+    def __call__(self, texts):
+        """
+        Callable method for the SentimentPipeline class. Processes the given text(s) and returns sentiment scores.
+        """
+        
+        # Check if input is a single string. If it is, convert it into a list.
+        if isinstance(texts, str):
+            texts = [texts]
+
+        results = []
+        for text in texts:
+            # Run the text through the pipelines
+            pos_result = self.pos_pipeline(text)[0]
+            neg_result = self.neg_pipeline(text)[0]
+
+            # Calculate probabilities for positive/non-positive and negative/non-negative.
+            # If the label is POSITIVE/NEGATIVE, the score for positive/negative is the score returned by the model, 
+            # and the score for non-positive/non-negative is 1 - the score returned by the model.
+            # If the label is NON-POSITIVE/NON-NEGATIVE, the score for non-positive/non-negative is the score returned by the model,
+            # and the score for positive/negative is 1 - the score returned by the model.
+            Pt, Pn = (pos_result['score'], 1 - pos_result['score']) if pos_result['label'] == 'POSITIVE' else (1 - pos_result['score'], pos_result['score'])
+            Nt, Nn = (neg_result['score'], 1 - neg_result['score']) if neg_result['label'] == 'NEGATIVE' else (1 - neg_result['score'], neg_result['score'])
+
+            # Calculate POS, NEG, OBJ scores using the formulas provided
+            POS = Pt * Nn
+            NEG = Nt * Pn
+            OBJ = 1 - POS - NEG
+
+            # Append the scores to the results
+            results.append({"POS": POS, "NEG": NEG, "OBJ": OBJ})
+
+        # If the input was a single string, return a single dictionary. Otherwise, return a list of dictionaries.
+        return results if len(results) > 1 else results[0]