DAHRS

Divergence-Aware Hallucination-Remediated SRL Projection

A Python framework for improving cross-lingual Semantic Role Labeling (SRL) projection using divergence-aware alignment and hallucination mitigation techniques.

---

Overview

DAHRS is a framework for projecting semantic role labels (SRL) across languages while addressing a key challenge:

• Hallucinated role alignments — incorrect or spurious mappings that arise from cross-lingual divergences (naturally occurring structural and lexical differences between languages)

The system builds on alignment analysis techniques and introduces strategies to improve projection quality through:

• Divergence-aware alignment handling
• Function word filtering
• Structured alignment resolution (FCFA: First-Come-First-Assign)

---

Paper

This repo goes along with the paper:

"DAHRS: Divergence-Aware Hallucination-Remediated SRL Projection"

It covers the ideas, methods, and experiments behind what's behind implemented here.

https://arxiv.org/abs/2407.09283

---

Key Ideas

Divergence-Aware Alignment

Captures and analyzes:

• One-to-many alignments
• Many-to-one alignments
• Structural mismatches between languages

---

Hallucination Remediation

Reduces incorrect alignments by:

• Filtering function words
• Removing inconsistent mappings
• Prioritizing stable alignment structures

---

FCFA Alignment Strategy

Implements a First-Come-First-Assign (FCFA) approach:

• Sequential alignment resolution
• Conflict reduction in many-to-one mappings
• Improved consistency in projected labels

---

Quick Start

from fcfa_framework import FCFAFramework

framework = FCFAFramework()

src_tokens = {0: "I", 1: "do", 2: "n't", 3: "know"}
tgt_tokens = {0: "Je", 1: "ne", 2: "sais", 3: "pas"}

framework.load_alignment_output("0-0 1-1 2-3 3-2", (src_tokens, tgt_tokens))

print(framework.get_alignment_statistics())