Context-Aware Document-Level Location Selection for Finnish News Articles : a Hybrid Rule-Based and LLM Approach
Year: 2026
Type: dissertation
Abstract: Finnish morphological complexity and the absence of postal neighborhood names from standard NER training corpora make automatic location extraction from Finnish news articles particularly challenging. Superhood Oy, a Finnish neighborhood news platform, processes approximately 80 articles per day, each requiring manual location tagging. The present research was conducted in collaboration with Superhood Oy to address this problem by investigating whether a rule-based geoparsing and ranking pipeline can automatically identify the primary geographic location from Finnish news articles, reducing reliance on manual effort. A hybrid pipeline was designed and evaluated combining Stanza-based named entity recognition (NER), a postal-first hierarchical ranking strategy, a dynamic Geoapify geocoding resolver, and a Model Context Protocol (MCP) correction layer using the Llama 3.3 70B language model for document-level disambiguation. Four pipeline configurations were evaluated on a manually annotated dataset of 60 Finnish news articles. The pipeline configurations were tested in isolation so that the contribution of each component could be measured independently. The baseline configuration achieved an exact match accuracy of 66.67%. Introducing postal-first hierarchical ranking improved accuracy to 80.00%, a gain of 13.33 percentage points. Replacing the static resolver with a dynamic GeoAPI resolver reduced accuracy to 68.33% due to the introduction of noisy candidates. Adding the MCP correction layer recovered accuracy to 83.33%, exceeding both the baseline and the static resolver. The results show that LLM-based contextual reasoning via MCP can recover ranking errors introduced by dynamic geocoding, though at the cost of increased processing time and occasional regressions. The evaluation was conducted on 60 articles, and the results do not generalize beyond the evaluated dataset. The findings demonstrate that a hybrid rule-based and LLM-assisted approach is a viable direction for location extraction in morphologically complex, low-resource language settings.
Source: Theseus (Ammattikorkeakoulujen)
Author: Shehryar, Irum
Institution: Helsinki Metropolia University of Applied Sciences
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7161932661