Turkish Address Research

Overview

Created in the Teknofest/Hepsiburada address-matching context, this collection preserves different approaches across 25 experiment directories and 19 notebooks. It contains rule baselines, TF-IDF, transformer embeddings, FAISS retrieval and GIS paths. Data contracts and environments vary between experiments; the collection is not a single production API or a validated algorithm leaderboard.

Problem and objectives

Free-form Turkish addresses contain abbreviations, letter variations, missing fields and place names repeated across provinces. These ambiguities complicate matching. Address-component normalization and competition-specific label prediction are also different objectives: a correct label does not establish that every address component was extracted correctly.

My role

Research and software work preserved in the repository. The sources do not document individual or team responsibility in detail.

Solution approach

Independent experiments investigate text cleaning, dictionaries, sparse features, embedding similarity and geographic matching. One GIS path combines geocoding, candidate retrieval and text scoring; another uses spatial joins with administrative polygons. A detailed research guide describes inputs and outputs, distinguishes the methods and exposes reproducibility gaps instead of implying uniform behavior.

Architecture flow

Conceptual flow derived from project descriptions; not an application screenshot.

Core components

• An index covering 25 historical experiment directories.
• Source-based model and artifact inventory for 19 notebooks.
• Text-similarity and FAISS candidate-retrieval experiments.
• Nominatim/KDTree hybrid and administrative-boundary spatial joins.

Technical challenges

• Ambiguity from Turkish letter distinctions and repeated administrative place names.
• Missing training data, variable dependencies and hard-coded file paths.
• Geographic leakage and near-duplicate addresses that can inflate evaluation scores.

Engineering decisions

• Normalization and label prediction are described as separate research objectives.
• Directory names are not treated as proof of trained models or superior methods.
• Experiments remain preserved without claiming a common production system.

Testing and validation

• Source and notebook structure were inspected statically; no common benchmark was rerun.
• A missing baseline-v2 import and syntax issues in two notebooks were documented.
• The PR expands the research guide; these documentation changes remain unmerged.

Outcomes

• A documented research archive with method families, data contracts and reproduction requirements.
• No newly verified Accuracy/F1, speed or production-deployment result is available.

Limitations and current status

• Training CSV data is absent; not every experiment runs from a clean setup.
• Redistribution rights and personal-data status of existing address data remain unresolved.
• Geocoding can transmit data to third parties; the portfolio uses no raw address examples.

Sources and documentation

• Repository source
• Open documentation and quality PR
• Reviewed README at PR head
• Research and reproducibility guide
• Notebook source inventory