Beyond Linked Data - Exploiting Entity-Centric Knowledge on the Web

Editor's Notes

• #12: Definition 2.1. ith-Order Value Inconsistency (Dx, Dy, P) between the pair of datasets Dx, Dy with respect to the ith-Order single-value property P is the proportion of the equivalent entities in Dx and Dy having contradicting values in P Definition 2.2. ith-Order Value Incompleteness (Dx, Dy, P) between the pair of datasets Dx, Dy with respect to a ith-Order multi-value property P is the proportion of entities in Dx and Dy having dierent values in P. ISSUES: different can mean incorrectness as well as incompleteness
• #16: Filtering: identifying cluster of datasets which are similar to Ds (two metrics: LSA-based, Wordnet-based), threshold theta Ranking: cosine between profiles Experimentally better results than using the ranks from filtering step
• #17: Evalualtion: map for different similarity thresholds (theta) from filtering step when explaining: - why is MAP decreasing with higher similarity thresholds? "For the given intervals [0, 0.7], [0, 0.8] and [0, 0.9], with respect to the used measures, we have 100% of recall --> all datasets considered as true are present in the recommanda list.
• #23: Random Sampling: randomly selects resource instances from Ri 2 Di for further analysis in the proling pipeline. Weighted Sampling: weigh each resource as the ratio of the number of datatype properties used to dene a resource over the maximum number of datatype properties over all resources for a specic dataset. The weight for rk Fig. 1. Processing pipeline for generating structured proles of Linked Data graphs. is computed by wk = jf(rk)j=maxfjf(rj)jg (rj 2 Rijj = 1; ; n), where f(rk) represents the datatype properties of resource rk. An instance is included in a sample if, for a randomly generated number p from a uniform distribution, the weight wk such that wk > (1 􀀀 p). Such a strategy ensures that resources that carry more information (having more literal values) have higher chances of being included earlier at low cut-os of analysed samples. Resource Centrality Sampling: weighs each resource as the ratio of the number of resource types used to describe a particular resource (V 0 k Vk) divided by the total number of resource types in a dataset. The weight is dened by ck = jC0k j=jCj with C0k = C \ V 0 k. Similarly to `weighted sampling', for a randomly generated number p, rk is included in the sample if ck > (1 􀀀 p). The main motivation behind computing the centrality of a resource is that important concepts in a dataset tend to be more structured and linked to other concepts.
• #74: The underlying assumption is that very specific and targeted seed lists will require different crawling and relevance computation methods than very broad and unspecific seed lists. http://www.visualdataweb.org/relfinder/demo.swf?obj1=TWluaW9ucyAoZmlsbSl8aHR0cDovL2RicGVkaWEub3JnL3Jlc291cmNlL01pbmlvbnNfKGZpbG0p&obj2=U2FuZHJhIEJ1bGxvY2t8aHR0cDovL2RicGVkaWEub3JnL3Jlc291cmNlL1NhbmRyYV9CdWxsb2Nr&obj3=Sm9uIEhhbW18aHR0cDovL2RicGVkaWEub3JnL3Jlc291cmNlL0pvbl9IYW1t&obj4=TWljaGFlbCBLZWF0b258aHR0cDovL2RicGVkaWEub3JnL3Jlc291cmNlL01pY2hhZWxfS2VhdG9u&obj5=QWxsaXNvbiBKYW5uZXl8aHR0cDovL2RicGVkaWEub3JnL3Jlc291cmNlL0FsbGlzb25fSmFubmV5&obj6=RGVzcGljYWJsZSBNZSAyfGh0dHA6Ly9kYnBlZGlhLm9yZy9yZXNvdXJjZS9EZXNwaWNhYmxlX01lXzI=&obj7=U3RldmUgQ29vZ2FufGh0dHA6Ly9kYnBlZGlhLm9yZy9yZXNvdXJjZS9TdGV2ZV9Db29nYW4=&obj8=R2VvZmZyZXkgUnVzaHxodHRwOi8vZGJwZWRpYS5vcmcvcmVzb3VyY2UvR2VvZmZyZXlfUnVzaA==&name=REJwZWRpYSAobWlycm9yKQ==&abbreviation=ZGJw&description=TGlua2VkIERhdGEgdmVyc2lvbiBvZiBXaWtpcGVkaWEu&endpointURI=aHR0cDovL2RicGVkaWEuaW50ZXJhY3RpdmVzeXN0ZW1zLmluZm8=&dontAppendSPARQL=ZmFsc2U=&defaultGraphURI=aHR0cDovL2RicGVkaWEub3Jn&isVirtuoso=dHJ1ZQ==&useProxy=ZmFsc2U=&method=UE9TVA==&autocompleteLanguage=ZW4=&autocompleteURIs=aHR0cDovL3d3dy53My5vcmcvMjAwMC8wMS9yZGYtc2NoZW1hI2xhYmVs&ignoredProperties=aHR0cDovL3d3dy53My5vcmcvMTk5OS8wMi8yMi1yZGYtc3ludGF4LW5zI3R5cGUsaHR0cDovL3d3dy53My5vcmcvMjAwNC8wMi9za29zL2NvcmUjc3ViamVjdCxodHRwOi8vZGJwZWRpYS5vcmcvcHJvcGVydHkvd2lraVBhZ2VVc2VzVGVtcGxhdGUsaHR0cDovL2RicGVkaWEub3JnL3Byb3BlcnR5L3dvcmRuZXRfdHlwZSxodHRwOi8vZGJwZWRpYS5vcmcvcHJvcGVydHkvd2lraWxpbmssaHR0cDovL2RicGVkaWEub3JnL29udG9sb2d5L3dpa2lQYWdlV2lraUxpbmssaHR0cDovL3d3dy53My5vcmcvMjAwMi8wNy9vd2wjc2FtZUFzLGh0dHA6Ly9wdXJsLm9yZy9kYy90ZXJtcy9zdWJqZWN0&abstractURIs=aHR0cDovL2RicGVkaWEub3JnL29udG9sb2d5L2Fic3RyYWN0&imageURIs=aHR0cDovL2RicGVkaWEub3JnL29udG9sb2d5L3RodW1ibmFpbCxodHRwOi8veG1sbnMuY29tL2ZvYWYvMC4xL2RlcGljdGlvbg==&linkURIs=aHR0cDovL3B1cmwub3JnL29udG9sb2d5L21vL3dpa2lwZWRpYSxodHRwOi8veG1sbnMuY29tL2ZvYWYvMC4xL2hvbWVwYWdlLGh0dHA6Ly94bWxucy5jb20vZm9hZi8wLjEvcGFnZQ==&maxRelationLegth=Mg==
• #75: As shown in our experimental evaluation, specific entities within a seed list strongly reflect the crawl intent. {Pulp Fiction, Film, Entertainment}, the most specific entity \texttt{`Pulp Fiction'}, reflects the most specific crawl intent, whereas the entities \texttt{`Film'} and \texttt{`Entertainment'} provide contextual information, namely that \texttt{`Pulp Fiction'} is a movie. Motivated by this, we assume that the relevance of specific candidate entities is dependent on the seed entity they are related to. For example, candidate entities similar to entity \texttt{`Pulp Fiction'} will be ranked higher than entities that are similar to other seed entities.
• #79: The average improvement across different NDCG levels is 1.6% on depth 2 and 4.3% on depth 3, suggesting a positive effect of the attrition factor for the cases of our seed lists. On the other hand, the coherence of the seed list appears to have no significant impact on the suitability of particular configuration. Given the significantly increased runtime when crawling beyond hop 2, a crawl depth of 2 seems to provide optimal efficiency, and it is not advisable to crawl to a higher distance.
• #80: The average improvement across different NDCG levels is 1.6% on depth 2 and 4.3% on depth 3, suggesting a positive effect of the attrition factor for the cases of our seed lists. On the other hand, the coherence of the seed list appears to have no significant impact on the suitability of particular configuration. Given the significantly increased runtime when crawling beyond hop 2, a crawl depth of 2 seems to provide optimal efficiency, and it is not advisable to crawl to a higher distance.
• #81: This is due to the fact that high coherence seed lists have a more specific crawl intent, leading to narrow and often small result sets, and hence also a limited ground truth, while the low coherence lists have a much broader crawl intent as well as relevant entity set. This is reflected in our ground truth: the average number of entities labeled as related (score≥ 3 and beyond) is 208 for low coherence seed list, and 145 for high coherence seed lists. Meanwhile, the narrow search intent also causes more disagreement among crowdsourcing workers for generating the ground truth, which makes the results for high coherence seed lists less consensual. Another difficulty faced when evaluating the crawling task is the highly heteroge- neous and varied nature of the possible result sets, originating from a highly heteroge- neous Linked Data graph.