GeoSense-AI: Fast Location Inference from Crisis Microblogs

This paper presents an applied AI pipeline for realtime geolocation from noisy microblog streams, unifying statistical hashtag segmentation, part-of-speech-driven proper-noun detection, dependency parsing around disaster lexicons, lightweight named-entity recognition, and gazetteer-grounded disambiguation to infer locations directly from text rather than sparse geotags. The approach operationalizes information extraction under streaming constraints, emphasizing low-latency NLP components and efficient validation against geographic knowledge bases to support situational awareness during emergencies. In head to head comparisons with widely used NER toolkits, the system attains strong F1 while being engineered for orders-of-magnitude faster throughput, enabling deployment in live crisis informatics settings. A production map interface demonstrates end-to-end AI functionality ingest, inference, and visualization–surfacing locational signals at scale for floods, outbreaks, and other fastmoving events. By prioritizing robustness to informal text and streaming efficiency, GeoSense-AI illustrates how domain-tuned NLP and knowledge grounding can elevate emergency response beyond conventional geo-tag reliance.

💡 Research Summary

GeoSense‑AI is an applied artificial‑intelligence pipeline designed to infer geographic locations from noisy, real‑time microblog streams during crises. The system addresses the fundamental challenge that only a tiny fraction of tweets contain explicit geo‑tags (≈0.36 % in developing regions), forcing responders to extract location references directly from text. To meet the dual demands of accuracy and low latency, the authors construct a modular streaming pipeline that combines several lightweight yet complementary techniques.

First, hashtags are segmented using a statistical word‑segmentation algorithm based on unigram probabilities and dynamic programming. This step recovers embedded location names (e.g., “#ChennaiFloods”) and improves recall, while subsequent gazetteer validation filters out spurious segments. Next, a normalization stage removes URLs, mentions, retweet markers, and inserts word boundaries at CamelCase transitions, deliberately preserving capitalization to protect proper‑noun cues.

The core linguistic analysis proceeds in three layers. (1) POS‑driven syntactic pattern matching extracts noun phrases that follow location‑indicative prepositions, directions, and suffixes (city, district, hospital, etc.). (2) Dependency parsing augments pattern matching by locating tokens within three to four edges of disaster‑related keywords (flood, earthquake, dengue) in the dependency tree, capturing locations that appear in non‑canonical constructions. (3) A lightweight NER component (spaCy’s pre‑trained model) tags GPE, FAC, and LOC entities as a safety net, adding minimal overhead.

All candidate strings from the above stages are then verified against a geographic knowledge base. Two gazetteer options are provided: GeoNames for broad coverage with moderate granularity, and OpenStreetMap for fine‑grained data at higher computational cost. Exact matching is followed by fuzzy matching to tolerate spelling variations; matched entries yield latitude/longitude and administrative hierarchy. The system prioritizes matches within the target region (India in the experiments) but retains global entries for future expansion.

The authors evaluate GeoSense‑AI on a curated dataset of 1,000 manually annotated crisis tweets (99 containing verifiable Indian locations) drawn from a larger collection of 239 k tweets collected during the 2017 dengue and flood outbreaks. Baselines include simple n‑gram gazetteer look‑ups (UniLoc, BiLoc), Stanford NER, the Twitter‑specific NER of Ritter et al., spaCy’s built‑in NER, and Google Cloud Natural Language API. All baselines use the same GeoNames gazetteer for fairness.

Results show that the GeoNames‑based variant (GeoLoc) achieves the highest F1 score of 0.814 (precision 0.799, recall 0.830), outperforming UniLoc (F1 0.517), BiLoc (F1 0.548), Stanford NER (F1 0.699), TwitterNLP (F1 0.588), spaCy NER (F1 0.711), and Google Cloud (F1 0.579). Importantly, GeoLoc processes the entire evaluation set in 1.19 seconds, roughly 150 × faster than Stanford NER (≈175 s) and dramatically faster than the OSM‑based variant (≈712 s). The OSM variant attains higher recall (0.889) but suffers from low precision (0.338) and prohibitive latency, illustrating the trade‑off between granularity and real‑time feasibility.

Error analysis identifies two main failure modes. False negatives arise from location names absent from the gazetteer, creative spellings beyond fuzzy‑matching thresholds, and indirect references via landmarks. False positives stem from common nouns that coincide with minor place names, over‑segmented hashtags, and person names that match gazetteer entries. The authors suggest expanding the gazetteer with locally used informal names and improving fuzzy‑matching thresholds as future work.

Beyond the core extraction engine, the paper describes a production‑grade web interface that visualizes extracted locations on an interactive map, enabling emergency managers to monitor evolving hotspots in near real time. The system’s design demonstrates that a carefully engineered combination of domain‑specific rule‑based methods, lightweight neural models, and efficient knowledge‑base grounding can deliver both the speed required for streaming crisis informatics and the accuracy needed for actionable situational awareness.

In conclusion, GeoSense‑AI offers a practical, scalable solution for real‑time location inference from crisis microblogs, outperforming traditional NER‑centric approaches in both speed and balanced precision‑recall performance. Future extensions include multilingual support, multimodal integration (images, user profiles), and automated gazetteer updates to further enhance coverage and robustness.