import gradio as gr import torch from transformers import RobertaForSequenceClassification, RobertaTokenizer import numpy as np from transformers import pipeline # Load sentiment analysis model sentiment_analyzer = pipeline("sentiment-analysis") # Load model and tokenizer model_name = "SamanthaStorm/abuse-pattern-detector-v2" model = RobertaForSequenceClassification.from_pretrained(model_name, trust_remote_code=True) tokenizer = RobertaTokenizer.from_pretrained(model_name, trust_remote_code=True) # Define labels (18 total) LABELS = [ "gaslighting", "mockery", "dismissiveness", "control", "guilt_tripping", "apology_baiting", "blame_shifting", "projection", "contradictory_statements", "manipulation", "deflection", "insults", "obscure_formal", "recovery_phase", "non_abusive", "suicidal_threat", "physical_threat", "extreme_control" ] # Custom thresholds for each label THRESHOLDS = { "gaslighting": 0.25, "mockery": 0.15, "dismissiveness": 0.30, "control": 0.43, "guilt_tripping": 0.19, "apology_baiting": 0.45, "blame_shifting": 0.23, "projection": 0.50, "contradictory_statements": 0.25, "manipulation": 0.25, "deflection": 0.30, "insults": 0.34, "obscure_formal": 0.25, "recovery_phase": 0.25, "non_abusive": 0.99, "suicidal_threat": 0.45, "physical_threat": 0.20, "extreme_control": 0.36 } PATTERN_LABELS = LABELS[:15] DANGER_LABELS = LABELS[15:18] EXPLANATIONS = { "gaslighting": "Gaslighting involves making someone question their own reality or perceptions, often causing them to feel confused or insecure.", "blame_shifting": "Blame-shifting is when one person redirects the responsibility for an issue onto someone else, avoiding accountability.", "projection": "Projection involves accusing the victim of behaviors or characteristics that the abuser themselves exhibit.", "dismissiveness": "Dismissiveness is the act of belittling or disregarding another person's thoughts, feelings, or experiences.", "mockery": "Mockery involves ridiculing or making fun of someone in a hurtful way, often with the intent to humiliate them.", "recovery_phase": "Recovery phase refers to dismissing or invalidating someone’s process of emotional healing, or ignoring their need for support.", "insults": "Insults are derogatory remarks aimed at degrading or humiliating someone, often targeting their personal traits or character.", "apology_baiting": "Apology-baiting is when the abuser manipulates the victim into apologizing for something the abuser caused or did wrong.", "deflection": "Deflection is the act of avoiding responsibility or shifting focus away from one's own actions, often to avoid accountability.", "control": "Control tactics are behaviors that restrict or limit someone's autonomy, often involving domination, manipulation, or coercion.", "extreme_control": "Extreme control involves excessive manipulation or domination over someone’s actions, decisions, or behaviors.", "physical_threat": "Physical threats involve any indication or direct mention of harm to someone’s physical well-being, often used to intimidate or control.", "suicidal_threat": "Suicidal threats are statements made to manipulate or control someone by making them feel responsible for the abuser’s well-being.", "guilt_tripping": "Guilt-tripping involves making someone feel guilty or responsible for things they didn’t do, often to manipulate their behavior.", "emotional_manipulation": "Emotional manipulation is using guilt, fear, or emotional dependency to control another person’s thoughts, feelings, or actions.", "manipulation": "Manipulation refers to using deceptive tactics to control or influence someone’s emotions, decisions, or behavior to serve the manipulator’s own interests.", "non_abusive": "Non-abusive language is communication that is respectful, empathetic, and free of harmful behaviors or manipulation.", "obscure_formal": "Obscure or overly formal language used manipulatively to create confusion, avoid responsibility, or assert superiority." } def calculate_abuse_level(scores, thresholds): triggered_scores = [score for label, score in zip(LABELS, scores) if score > thresholds[label]] if not triggered_scores: return 0.0 return round(np.mean(triggered_scores) * 100, 2) def interpret_abuse_level(score): if score > 80: return "Extreme / High Risk" elif score > 60: return "Severe / Harmful Pattern Present" elif score > 40: return "Likely Abuse" elif score > 20: return "Mild Concern" else: return "Very Low / Likely Safe" def analyze_messages(input_text, risk_flags): input_text = input_text.strip() if not input_text: return "Please enter a message for analysis." sentiment = sentiment_analyzer(input_text)[0] sentiment_label = sentiment['label'] sentiment_score = sentiment['score'] adjusted_thresholds = THRESHOLDS.copy() if sentiment_label == "NEGATIVE": adjusted_thresholds = {key: val * 0.8 for key, val in THRESHOLDS.items()} inputs = tokenizer(input_text, return_tensors="pt", truncation=True, padding=True) with torch.no_grad(): outputs = model(**inputs) scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy() pattern_count = sum(score > adjusted_thresholds[label] for label, score in zip(PATTERN_LABELS, scores[:15])) danger_flag_count = sum(score > adjusted_thresholds[label] for label, score in zip(DANGER_LABELS, scores[15:18])) contextual_flags = risk_flags if risk_flags else [] contextual_risk_score = len(contextual_flags) if contextual_risk_score >= 2: danger_flag_count += 1 critical_flags = ["They've threatened harm", "They monitor/follow me", "I feel unsafe when alone with them"] high_risk_context = any(flag in contextual_flags for flag in critical_flags) non_abusive_score = scores[LABELS.index('non_abusive')] if non_abusive_score > adjusted_thresholds['non_abusive']: return "This message is classified as non-abusive." abuse_level = calculate_abuse_level(scores, THRESHOLDS) abuse_description = interpret_abuse_level(abuse_level) if danger_flag_count >= 2: resources = "Immediate assistance recommended. Please seek professional help or contact emergency services." else: resources = "For more information on abuse patterns, consider reaching out to support groups or professional counselors." scored_patterns = [(label, score) for label, score in zip(PATTERN_LABELS, scores[:15])] top_patterns = sorted(scored_patterns, key=lambda x: x[1], reverse=True)[:2] top_pattern_explanations = "\n".join([ f"\u2022 {label.replace('_', ' ').title()}: {EXPLANATIONS.get(label, 'No explanation available.')}" for label, _ in top_patterns ]) result = ( f"Abuse Risk Score: {abuse_level}% – {abuse_description}\n\n" f"Most Likely Patterns:\n{top_pattern_explanations}\n\n" f"⚠️ Critical Danger Flags Detected: {danger_flag_count} of 3\n" "The Danger Assessment is a validated tool that helps identify serious risk in intimate partner violence. " "It flags communication patterns associated with increased risk of severe harm. " "For more info, consider reaching out to support groups or professionals.\n\n" f"Resources: {resources} \n\n" f"Sentiment: {sentiment_label} (Confidence: {sentiment_score*100:.2f}%)" ) if contextual_flags: result += "\n\n⚠️ You indicated the following:\n" + "\n".join([f"• {flag}" for flag in contextual_flags]) if high_risk_context: result += "\n\n🚨 These responses suggest a high-risk situation. Consider seeking immediate help or safety planning resources." return result iface = gr.Interface( fn=analyze_messages, inputs=[ gr.Textbox(lines=10, placeholder="Enter message here..."), gr.CheckboxGroup(label="Do any of these apply to your situation?", choices=[ "They've threatened harm", "They isolate me", "I’ve changed my behavior out of fear", "They monitor/follow me", "I feel unsafe when alone with them" ]) ], outputs=[ gr.Textbox(label="Analysis Result"), ], title="Abuse Pattern Detector" ) if __name__ == "__main__": iface.launch()