Multi-Source Spatial Entity Linkage

Multi-Source Spatial Entity Linkage Suela Isaj Esteban Zimányi Torben Bach Pedersen

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

Multi-Source Spatial Entity Linkage 2019 SSDBM 10.1145/3340964.3340979 isaj2019-spatial-entity-linkage isaj2019-spatial-entity-linkage.txt, the project's staged copy 56cb61cd9b7512d58f69f636202e29627b1ae7c2 gold-oa computer-science

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them method Quotation withheld because this record is not cleared for redistribution. spatial entity linkage across location-based sources Quotation withheld because this record is not cleared for redistribution. system-or-tool-building system-or-tool-building Quotation withheld because this record is not cleared for redistribution. mechanisms the blocking structure's splitting criterion is changed from CAPACITY to density and diagonal, because a count of points is not a spatial quantity: a dense city block should split before its capacity is reached, and two farm points far apart may still be one entity Quotation withheld because this record is not cleared for redistribution. mechanisms a point is allowed to be assigned to MORE THAN ONE child of the split, so that two nearby entities are not separated merely because a dividing line happened to fall between them Quotation withheld because this record is not cleared for redistribution. mechanisms the address is deliberately NOT compared by edit distance, because syntactic and spatial closeness come apart in both directions - one changed digit is 650 metres, eight added characters is the same building - and is instead scored by equality, inclusion, or nothing Quotation withheld because this record is not cleared for redistribution. mechanisms category and keyword similarity is taken as the MAXIMUM over all cross-pairs of the two entities' keyword lists under a Wordnet-based measure, rather than by string matching, so that synonyms count as agreement Quotation withheld because this record is not cleared for redistribution. mechanisms the weighted scoring function is removed entirely and replaced by Pareto optimality: points on one frontier have equal utility, so the method says a pair is better than another WITHOUT quantifying by how much, and needs no attribute weights to do it Quotation withheld because this record is not cleared for redistribution. mechanisms the whole threshold problem is collapsed to ONE number - which skyline level separates the classes - in place of a threshold per attribute similarity plus another for the function that aggregates them Quotation withheld because this record is not cleared for redistribution. mechanisms THE UNSUPERVISED CUT-OFF WORKS BECAUSE THE POSITIVE CLASS IS NOT A CLUSTER: matching pairs are scattered, since several different patterns can make a match, so instead of intra-class compactness the method separates them by their distance to the NEGATIVE class, and sets the cut-off where that inter-class distance starts to increase Quotation withheld because this record is not cleared for redistribution. mechanisms the attributes used to BUILD the ground truth - phone, opening hours - are excluded from the similarity computation, and the stated reason is that they are used to construct the ground truth Quotation withheld because this record is not cleared for redistribution. Ground truth for the full dataset is mostly automatic, not manually checked, because manual labelling at that scale was not feasible. Quotation withheld because this record is not cleared for redistribution. Performance on the full dataset is lower than on the manually labelled sample, possibly an artefact of the automatic labelling rather than of the method. Quotation withheld because this record is not cleared for redistribution. The skyline-based algorithms could in principle degrade to cubic complexity. Quotation withheld because this record is not cleared for redistribution. The ranking step is computationally expensive because it requires repeated Pareto frontier computation. Quotation withheld because this record is not cleared for redistribution. The attributes that define a correct match are the ones excluded from the scoring, so identity is fixed by whichever fields were reserved. what this limits: any accuracy figure reported here: the attributes used to construct the ground truth are deliberately excluded from the similarity computation, so what counts as the same entity is defined by one set of attributes and predicted from a different set Quotation withheld because this record is not cleared for redistribution. String similarity is stated to be inapplicable to the attribute that looks most like a key, with a worked counterexample. what this limits: any use of string-similarity metrics on addresses: the paper states they cannot be used, giving two addresses with a Levenshtein similarity of 0.963 that are 650 meters apart Quotation withheld because this record is not cleared for redistribution. Coordinates fail as an identifier in both directions at once. what this limits: any linkage of these sources on coordinates: different spatial entities may share coordinates, as with shops in a shopping mall, and one entity may appear at coordinates 82 meters apart in two sources Quotation withheld because this record is not cleared for redistribution. The methods this work builds on were fitted to a different kind of entity, and the authors say the resemblance cannot be assumed. what this limits: carrying an entity-resolution method over to spatial entities: the authors state such solutions are learned on human entity datasets and that the resemblance of behaviours between the two cannot merely be assumed Quotation withheld because this record is not cleared for redistribution. manual checking of the labels on a sample of the pairs - 1,500 in the sample dataset and about 10,000 overall - against the automatic labels inferred from shared website or phone number eliminates: that the automatic labelling is wrong in the two ways the paper names: the same entity holding a different phone number or website, and different entities sharing a phone number · outcome (stated by the tagger, no quotation): partial Quotation withheld because this record is not cleared for redistribution. spatial entity linkage spatial entity linkage Quotation withheld because this record is not cleared for redistribution. spatial entity spatial entity Quotation withheld because this record is not cleared for redistribution. spatial object spatial object Quotation withheld because this record is not cleared for redistribution. ground truth ground truth Quotation withheld because this record is not cleared for redistribution. spatial blocking spatial blocking Quotation withheld because this record is not cleared for redistribution. QuadFlex QuadFlex Quotation withheld because this record is not cleared for redistribution. pairwise comparisons pairwise comparisons Quotation withheld because this record is not cleared for redistribution. SkyRank SkyRank Quotation withheld because this record is not cleared for redistribution. QuadSky QuadSky Quotation withheld because this record is not cleared for redistribution. spatial entity vs spatial object - THE DISTINCTION THE WHOLE PAPER RESTS ON: an object is identified by its coordinates alone, an entity is not, so the two words name things that need different identity conditions even though both sit at a point on a map Quotation withheld because this record is not cleared for redistribution. identity - AS A COMBINATION OF ATTRIBUTES, NOT A KEY: the paper states outright that what makes a spatial entity itself is several attributes together, and that inconsistency between sources makes that combination hard to infer Quotation withheld because this record is not cleared for redistribution. similarity - SYNTACTIC AND SPATIAL COME APART: a one-character change in an address is a 0.963 string similarity and 650 metres on the ground, which is why the paper refuses ordinary string metrics for exactly the field that looks most like an identifier Quotation withheld because this record is not cleared for redistribution. `QuadSky` - A PIPELINE WHOSE FOUR PARTS ARE NAMED TWICE OVER, ONCE BY FUNCTION AND ONCE BY ALGORITHM. `Quotation withheld` Three of the four carry a coined proper name in a parenthesis and one does not - `pairwise comparisons` is the stage the authors did not invent anything for, and the naming records that asymmetry exactly. A reader tracking `SkyRank` and a reader tracking `ranking the pairs` are following the same stage under a name that belongs to a different level. Quotation withheld because this record is not cleared for redistribution. extends → the authors' earlier conference paper [38] (QuadFlex + SkyEx) Quotation withheld because this record is not cleared for redistribution. builds-on → the skyline computation method of [38] for spatial blocking and pairwise comparison Quotation withheld because this record is not cleared for redistribution. builds-on → the lattice-based skyline algorithm of [42] Quotation withheld because this record is not cleared for redistribution. builds-on → WordNet [20] and the Wu & Palmer semantic similarity measure [21] Quotation withheld because this record is not cleared for redistribution. contradicts → the spatial entity matching tools of Karam et al. [17], Berjawi et al. [18] and Morana et al. [19] Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · spatial entity —distinct-from→ spatial object spatial entity spatial object EVIDENCE CORRECTED: RE-GROUNDED. The old span named only 'spatial object' and began mid-contrast ('However, while ...'). This sentence names both AND CARRIES THE DIFFERENTIA: an object is fully determined by its coordinates, an entity is geo-located AND has more. A distinction with its criterion. — strength of the relation: 1 Quotation withheld because this record is not cleared for redistribution. spatial entity —distinct-from→ spatial object spatial entity spatial object differs in: THE AXIS IS WHETHER IDENTITY BEYOND POSITION IS PART OF THE THING. 'Quotation withheld' Both are located; only one carries an identity, and that is the whole of the distinction. · strength of the relation: 1 · note: RE-GROUNDS another row of this record: the old span named ONE endpoint. Quotation withheld because this record is not cleared for redistribution. spatial blocking —part-of→ QuadSky spatial blocking QuadSky the relation itself is separately evidenced · strength of the relation: 1 · note: 'consists of' is the containment cue and it is in the gate's list, so this edge is cue-tested. The whole is named first and the parts second - CONTAINER_FIRST order. · the relation stated: QuadSky consists of four main parts: spatial blocking (QuadFlex), pairwise comparisons, ranking the pairs (SkyRank), and labelling the pairs Quotation withheld because this record is not cleared for redistribution. QuadFlex —part-of→ QuadSky QuadFlex QuadSky the relation itself is separately evidenced · strength of the relation: 1 · note: 'consists of' is the containment cue and it is in the gate's list, so this edge is cue-tested. The whole is named first and the parts second - CONTAINER_FIRST order. · the relation stated: QuadSky consists of four main parts: spatial blocking (QuadFlex), pairwise comparisons, ranking the pairs (SkyRank), and labelling the pairs Quotation withheld because this record is not cleared for redistribution. pairwise comparisons —part-of→ QuadSky pairwise comparisons QuadSky the relation itself is separately evidenced · strength of the relation: 1 · note: 'consists of' is the containment cue and it is in the gate's list, so this edge is cue-tested. The whole is named first and the parts second - CONTAINER_FIRST order. · the relation stated: QuadSky consists of four main parts: spatial blocking (QuadFlex), pairwise comparisons, ranking the pairs (SkyRank), and labelling the pairs Quotation withheld because this record is not cleared for redistribution. SkyRank —part-of→ QuadSky SkyRank QuadSky the relation itself is separately evidenced · strength of the relation: 1 · note: 'consists of' is the containment cue and it is in the gate's list, so this edge is cue-tested. The whole is named first and the parts second - CONTAINER_FIRST order. · the relation stated: QuadSky consists of four main parts: spatial blocking (QuadFlex), pairwise comparisons, ranking the pairs (SkyRank), and labelling the pairs Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · Sources describing the same physical world disagree: each covers only part of it, with different attributes, sometimes contradicting each other. EVIDENCE CORRECTED: RE-GROUNDED. The published string was not verbatim in EITHER copy of the paper. A quote field promises the author's words; this held a copy with the line-break hyphen closed in 'at- tributes'. Re-cut from the paper as it stands; the closing was correct scholarship and silently unverifiable. Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · The SkyEx-* algorithms reach 0.85 precision and 0.85 recall on a manually labeled set of 1,500 pairs. EVIDENCE CORRECTED: RE-GROUNDED. The published string was not verbatim in EITHER copy of the paper. A quote field promises the author's words; this held 'QuadSky achieves 0.85 precision and 0.85 recall'. The paper attributes those figures to 'our novel SkyEx-* family of algorithms', which is the LABELLING STEP of QuadSky - one of its four parts, beside QuadFlex, pairwise comparison and SkyRank. Naming the pipeline where the paper names the classifier moves a measured result onto a larger object than the one measured. Quotation withheld because this record is not cleared for redistribution. spatial entity linkage spatial entity linkage about this quote: QUOTE RECOVERED FROM THE PAPER'S OWN CHARACTERS. The recorded quote was LETTERS-ONLY IDENTICAL to this stretch of the staged text but not an exact substring of it - spaces, hyphens, ligature glyphs or mathematical operators differed, so it could never anchor. This is the file's own text between the same two letters: not produced by a rule about the damage, copied out of the file. Round-trip checked (the recovered span is letters-only equal to the original) and length-checked. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · 0.85 precision / 0.85 recall — QuadSky on a manually labelled set of 1,500 pairs EVIDENCE CORRECTED: RE-GROUNDED. The published string was not verbatim in EITHER copy of the paper. A quote field promises the author's words; this held 'QuadSky achieves 0.85 precision and 0.85 recall'. The paper attributes those figures to 'our novel SkyEx-* family of algorithms', which is the LABELLING STEP of QuadSky - one of its four parts, beside QuadFlex, pairwise comparison and SkyRank. Naming the pipeline where the paper names the classifier moves a measured result onto a larger object than the one measured. Quotation withheld because this record is not cleared for redistribution. 0.87 precision / 0.6 recall — QuadSky on a semi-manually labelled set of 777,452 pairs about this quote: QUOTE RECOVERED FROM THE PAPER'S OWN CHARACTERS. The recorded quote was LETTERS-ONLY IDENTICAL to this stretch of the staged text but not an exact substring of it - spaces, hyphens, ligature glyphs or mathematical operators differed, so it could never anchor. This is the file's own text between the same two letters: not produced by a rule about the damage, copied out of the file. Round-trip checked (the recovered span is letters-only equal to the original) and length-checked. The text as originally recorded is kept in evidence as extracted. Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · precision precision EVIDENCE CORRECTED: RE-GROUNDED. The published string was not verbatim in EITHER copy of the paper. A quote field promises the author's words; this held 'QuadSky achieves 0.85 precision and 0.85 recall'. The paper attributes those figures to 'our novel SkyEx-* family of algorithms', which is the LABELLING STEP of QuadSky - one of its four parts, beside QuadFlex, pairwise comparison and SkyRank. Naming the pipeline where the paper names the classifier moves a measured result onto a larger object than the one measured. Quotation withheld because this record is not cleared for redistribution. EVIDENCE CORRECTED (RE-GROUNDED) · recall recall EVIDENCE CORRECTED: RE-GROUNDED. The published string was not verbatim in EITHER copy of the paper. A quote field promises the author's words; this held 'QuadSky achieves 0.85 precision and 0.85 recall'. The paper attributes those figures to 'our novel SkyEx-* family of algorithms', which is the LABELLING STEP of QuadSky - one of its four parts, beside QuadFlex, pairwise comparison and SkyRank. Naming the pipeline where the paper names the classifier moves a measured result onto a larger object than the one measured. Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1145/3340964.3340979.