A Comparison of String Similarity Measures for Toponym Matching

Abstract

The diversity of ways in which toponyms are specified often results in mismatches between queries and the place names contained in gazetteers. Search terms that include unofficial variants of official place names, unanticipated transliterations, and typos are frequently similar but not identical to the place names contained in the gazetteer. String similarity measures can mitigate this problem, but given their task-dependent performance, the optimal choice of measure is unclear. We constructed a task in which place names had to be matched to variants of those names listed in the GEOnet Names Server, comparing 21 different measures on datasets containing romanized toponyms from 11 different countries. Best-performing measures varied widely across datasets, but were highly consistent within-country and within-language. We discuss which measures worked best for particular languages and provide recommendations for selecting appropriate string similarity measures.

Learning Advanced Similarities and Training Features for Toponym Interlinking

“…The metasimilarity proposed in [5] takes into account accentuation and other languagespecific aspects of toponym names, in a four-stages process. The set of algorithms evaluated in [2] were assessed in the toponym interlinking problem by [12]. The authors experimented on place names listed in the GEOnet Names Server, that contains romanized toponyms from 11 different countries.…”

“…We note that the baselines we compare with cover a large part of the presented literature, presented in the following papers:[2,3,5,12,14].…”

Stable assessment of the quality of similarity algorithms of character strings and their normalizations

“…Comparison of models and algorithms used for highlighting requires arrays of similar strings of various origins [1], which are usually comes from either unpublished personal data arrays [2][3][4][5], or from hand-marked linguistic corps or thesauri, as in [6], or from artificially generated data [7]. The public unavailability of some excludes the reproducibility of experiments and an independent assessment of the quality of the initial data, while the high labor-consuming nature of others also limits their volume and availability.…”

“…Values of errors of metrics in the group(4 ) ofthe languages pairs ({de, hu}, {en, hu}, {eo, hu}, {fr, hu}) 8% ± 2.4% 29.5% ± 1.0% 30.4% ± 1.4% 31.2% ± 2.3% lcs 33.2% ± 2.4% 30.8% ± 1.3% 32.2% ± 1.8% 32.3% ± 2.2% qgram3 33.4% ± 3.3% 28.9% ± 0.6% 33.0% ± 1.5% 32.3% ± 2.9% dl,lv,osa 34.3% ± 4.3% 32.6% ± 1.0% 33.5% ± 1.1% 33.6% ± 2.9% cosine3 35.1% ± 3.1% 31.9% ± 0.6% 35.6% ± 1.4% 34.7% ± 2.6% AVERAGE 36.3% ± 2.2% 33.6% ± 1.0% 35.6% ± 0.9% 35.5% ± 1.9% qgram2 37.0% ± 3.2% 34.7% ± 0.8% 37.8% ± 1.9% 36.8% ± 2.6% qgram1 39.6% ± 3.4% 37.4% ± 1.0% 39.9% ± 1.3% 39.3% ± 2.5% jwp 41.0% ± 2.3% 36.3% ± 1.3% 39.6% ± 1.4% 39.5% ± 2.5% jw 41.1% ± 2.4% 36.9% ± 1.2% 40.0% ± 1.4% 39.8% ± 2.4% cosine2 40.8% ± 3.5% 38.5% ± 0.8% 41.2% ± 1.0% 40.5% ± 2.5% cosine1 43.1% ± 3.4% 43.4% ± 1.1% 46.0% ± 0.8% 44.3% ± 2.7%…”

Stable assessment of the quality of similarity algorithms of character strings and their normalizations

“…Сравнение используемых для выделения моделей и алгоритмов нуждается в массивах похожих строк различного происхождения [1], в качестве которых обычно используют либо публично недоступные массивы персональных данных [2][3][4][5], либо вручную размеченные лингвистические корпусы или тезаурусы, как в [6], а иногда и искусственно сгенерированные данные [7]. Закрытость одних исключает воспроизводимость экспериментов и независимую оценку качества исходных данных, а высокая трудозатратность других тоже ограничивает их объём и доступность.…”