Implicit Relation Linking for Question Answering over Knowledge Graph

May combine two claims.

The Paraphrase dictionary maps natural language patterns to DBpedia relations and currently covers 2,284 relations and 31,130 paraphrase pairs.

“Paraphrase maps NL patterns to DBpedia relations, which currently covers 2,284 relations and 31,130 paraphrase pairs.”

• cites Xue et al., 2020 sourceSupported

May combine two claims.

Existing relation linking methods perform well on several benchmark datasets but are still challenged by implicit relations prevalent in the real world.

“However, according to (Sakor et al., 2019), while existing RL methods perform well on several benchmark datasets, they are still challenged by implicit relations prevalent in the real world.”

• cites Sakor et al., 2019 sourceCould not verify

  The supplied passages are bibliographic metadata, with no section or page discussing relation-linking performance. Searches for benchmark performance and implicit relations returned no substantive full-text passage; the only topical text returned was an abstract, which cannot serve as evidence under the stated rubric. (LLM, conf=0.70)

  Experimental — accuracy not yet measured

May combine two claims.

Falcon and SLING perform poorly for other kinds of implicit relations and cannot handle relation linking through relation paths.

“However, they perform poorly for other kinds of implicit relations, and cannot handle RL through relation paths either.”

• cites Sakor et al., 2019 sourceCould not verify

  The available indexed passages are bibliographic metadata (no section or page), not discussion of Falcon or SLING. Searches for implicit relations, poor performance, and relation paths found no passage addressing either claimed limitation. (LLM, conf=0.72)

  Experimental — accuracy not yet measured
• cites Mihindukulasooriya et al., 2020 sourcePartly supported

  The Qualitative Analysis (p. 10) supports a limitation in SLING: for a grandchildren question requiring two child relations, it did not decompose the question into two triples. But the Results (p. 9) say SLING identified a relation implicit in the question, and the cited passages do not establish that Falcon and SLING generally perform poorly on other implicit relations or cannot handle relation paths as a class. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

EERL discovers implicit relations by using domains and ranges of properties in the knowledge graph, but it cannot handle relation linking through relation paths.

“For instance, EERL discovers the implicit relations by using the domains and ranges of properties in KG, but it cannot handle the issue of RL through relation paths.”

• cites Pan et al., 2019 sourceCould not verify

  The available Abstract describes EERL as using entities extracted from questions and DBpedia entities to compute relation candidates. Searches for property domains and ranges, implicit relations, and relation paths found no passage addressing those specific claims. The available copy therefore does not establish either the proposed mechanism or the asserted limitation. (LLM, conf=0.72)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim.

AMR can effectively capture semantic information at the sentence level.

“according to (Mihindukulasooriya et al., 2020), AMR can effectively capture the semantic information at the sentence level”

• cites Mihindukulasooriya et al., 2020 sourceSupported

May combine two claims.

Many relations in natural language are implicitly expressed and may not link to a single relation in the knowledge graph because of ambiguity in natural language and incompleteness of the knowledge graph.

“Due to the ambiguity of NL and the incompleteness of KG, many relations in NL are implicitly expressed, and may not link to a single relation in KG, which challenges the current methods.”

• Also stated in: § Question: · plain

  “However, according to (Sakor et al., 2019), while existing RL methods perform well on several benchmark datasets, they are still challenged by implicit relations prevalent in the real world.”

• cites Sakor et al., 2019 sourceCould not verify

  The passages supplied from our copy contain bibliographic metadata, not the paper’s discussion. Searches for implicitly expressed relations, natural-language ambiguity, knowledge-graph incompleteness, and linking to a single relation returned no substantive passage addressing the claim. (LLM, conf=0.72)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Previous works pre-collect paraphrase pairs between relation phrases in natural language and KG relations through frequent item mining or crowdsourcing.

“The previous works (Nakashole et al., 2012;Xue et al., 2020) pre-collect some paraphrase pairs (e.g.,”

• cites Nakashole et al., 2012Not checked — no full text of the source
• cites Xue et al., 2020 sourceSupported

May combine two claims.

The paraphrase dictionary constructed by Xue et al. (2020) establishes a mapping relationship between natural language phrases and relations in the knowledge graph, which can be prior knowledge for implicit relation phrases with vague meanings.

“Considering the work (Xue et al., 2020) that constructs a dictionary of paraphrases and establishes a mapping relationship between NL phrases and relations in KG, which can be prior knowledge for implicit relation phrases with vague meanings”

• cites Xue et al., 2020 sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

93.8% of people in Freebase do not have place of birth, and 78.5% of people do not have nationality.

“Besides, according to the statistics in (Stadelmaier and Padó, 2019), 93.8% of people in Freebase do not have place of birth, and 78.5% of people do not have nationality”

• cites Stadelmaier and Padó, 2019 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

LC-QuAD 1.0 contains 5,000 questions and corresponding SPARQL queries over DBpedia version 2016-04.

“LC-QuAD 1.0 (Trivedi et al., 2017) is a large complex questions dataset for KBQA, which contains 5,000 questions and corresponding SPARQL queries over DBpedia (version 2016-04).”

• cites Trivedi et al., 2017 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SimReL uses AMR for question representation and employs a Siamese network with BERT to link relations in an end-to-end manner.

“SimReL (Naseem et al., 2021) also uses AMR for question representation, and employs a Siamese network with BERT to link relations in an end-to-end manner.”

• cites Naseem et al., 2021 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

SimReL reports its F1-score improvement on LC-QuAD as 2.4%.

“As a reference, SimReL reports its F1-score improvement on LC-QuAD as 2.4% (Naseem et al., 2021).”

• cites Naseem et al., 2021 sourcePartly supported

  The Results and Discussion section reports that incorporating AMR into relation linking improved the NSQA system’s F1 on the LC-QuAD 1.0 test set by 2.4, reaching 44.5. It does not describe this as a 2.4% relative improvement, and the reported gain is for the integrated KBQA system rather than SemReL’s standalone relation-linking F1. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SLING leverages AMR for preprocessing, trains a Transformer-based model based on distant supervision, and integrates several complementary signals for relation linking.

“SLING (Mihindukulasooriya et al., 2020) leverages AMR for preprocessing, trains a Transformer-based model based on distant supervision, and integrates several complementary signals for RL.”

• cites Mihindukulasooriya et al., 2020 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Falcon enhances entity and relation linking through cross-KG entity and relation alignment and basic principles of English morphology.

“Falcon (Sakor et al., 2019) enhances entity and relation linking through cross-KG entity and relation alignment and basic principles of English morphology.”

• cites Sakor et al., 2019 sourceCould not verify

  The supplied passages are bibliographic metadata, and searches for cross-KG alignment, English morphology, and Falcon's method returned no passage from the paper's body. The indexed abstract addresses the claim, but the audit rules exclude abstracts as evidence, so the claimed mechanism cannot be verified from eligible passages. (LLM, conf=0.69)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

EARL models the linking problem as the generalized traveling salesman problem and uses connection density to link entities and relations jointly.

“EARL (Dubey et al., 2018) models the linking problem as the generalized traveling salesman problem, and uses connection density to link entities and relations jointly.”

• cites Dubey et al., 2018 sourceSupported

May combine two claims.

In Figure 3, gAnswer's relation linking module incorrectly links the relation to dbo:birthPlace, leading to a wrong SPARQL query and wrong answer; ImRL correctly identifies dbo:birthDate, modifies the SPARQL query and returns the correct answer.

“In Figure 3, the relation linking module of gAnswer incorrectly links the relation in the question to dbo:birthPlace, which leads to the wrong SPARQL query and the wrong answer.In contrast, ImRL can correctly identify the relation dbo:birthDate, modify the SPARQL query and return the correct answer.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In the first PathSQ case, the relation phrase 'involved in' needs to be expressed using the relation path dbo:commander → dbo:militaryBranch according to the gold standard.

“In the first case from PathSQ, the relation phrase”

May combine two claims.

Entity dbr:Rudi_Hedman does not have relation dbo:nationality in DBpedia, so the gold relation path dbo:nationality is missing in the candidate list.

“However, since entity dbr:Rudi_Hedman does not have relation dbo:nationality in DBpedia, the gold relation path dbo:nationality is missing in the candidate list.”

May combine two claims.

In the second case, there is no entity in the QALD-9 question, which causes SimReL to be unable to enumerate any candidate relations since it depends on entity linking.

“In the second case, there is no entity in this question of QALD-9”

• Also stated in: § Ablation study of embedding models. · plain

  “causes that SimReL cannot enumerate any candidate relations since it depends on the result of entity linking.”

The quote only partly states this claim.

ImRL correctly predicts 10 of the 46 QALD-9 testing questions with unseen relations.

“Still, ImRL can correctly predict 10 of them”

Probably not a substantial claim: may not pass S3 (load-bearing).

PathSQ is based on DBpedia version 2016-04.

“It is based on DBpedia (version 2016-04)”

May combine two claims.

In Figure 4, ImRL does not recognize the same relation dbo:foundedBy as the gold SPARQL query, but it recognizes dbp:founder, which is actually equivalent to dbo:foundedBy.

“In Figure 4, although ImRL does not recognize the same relation dbo:foundedBy as the gold SPARQL query, it recognizes dbp:founder, which is actually equivalent to dbo:foundedBy.”

May combine two claims.

Although ImRL's candidate generation is more robust, the absence of entities has a great impact on it, making it give the wrong answer dbo:origin.

“Although the method of generating candidates of ImRL is more robust, the absence of entities also has a great impact on it, making it give the wrong answer dbo:origin.”

May combine two claims.

With the increase of candidate path number c, MRR, Hits@c and F1-score all increase.

“As shown in Table6, with the increase of c, MRR, Hits@c and F1-score all increase.”

May combine two claims.

ImRL achieves superior performance on LC-QuAD and QALD-9, where relations are mostly explicit, and PathSQ, where relations are all implicit.

“The results show that our method achieves superior performance on LC-QuAD and QALD-9, where the relations are mostly explicit, and PathSQ, where the relations are all implicit”

Classified as probably not a substantial claim.

In the QALD-9 dataset, 94% of the questions contain explicit single relations.

“Take QALD-9 for example, 94% of the questions are with explicit single relations.”

May combine two claims.

SimReL returns dbo:militaryBranch for the first case because it can only find the single relation whose meaning is most similar to the meaning of the whole question.

“The state-of-the-art method SimReL returns dbo:militaryBranch, as it can only find the single relation whose meaning is most similar to the meaning of the whole question as the answer.”

Classified as probably not a substantial claim.

Using rules to transform a dependency tree into a directed graph and using a heuristic search to find the shortest path between two nodes can effectively decompose a complex question into multiple simple triples, thereby discovering relations.

“Inspired by (Hu et al., 2018), which demonstrates that using rules to transform the dependency tree into a directed graph and using a heuristic search to find the shortest path between two nodes can effectively decompose the complex question into multiple simple triples, thereby discovering the relations.”

• cites Hu et al., 2018Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

Values provide both textual and structural knowledge of natural language and KG.

“Values provide both textual and structural knowledge of NL and KG.”

• AssumptionRests on an assumption: Paraphrases can be considered as prior knowledge about the diverse meanings of natural language.

May combine two claims.

Because knowledge graphs are incomplete, for some relation phrases it is not possible to find a corresponding single relation in the KG, and the phrase needs to be expressed by a relation path.

“which makes it fail to find the corresponding single relations in KG for some phrases and needs to be expressed by relation paths.”

Classified as probably not a substantial claim.

Conventional relation linking can be regarded as a simpler case of the extended relation linking task when the relation path is one hop.

“We extend the RL task to link phrases to relation paths, where conventional RL can be regarded as a simpler case when the path is one hop.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

QALD-9 provides 408 training questions and 150 test questions.

“It provides 408 training questions and 150 test questions”

• cites Usbeck et al., 2018Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

Keys provide textual knowledge from the perspective of natural language.

“Keys provide textual knowledge from the perspective of NL”

• AssumptionRests on an assumption: Paraphrases can be considered as prior knowledge about the diverse meanings of natural language.

Probably not a substantial claim: may not pass S3 (load-bearing).

6% of QALD-9 questions need implicit relation linking.

“6% of which need implicit RL.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Only 1% of LC-QuAD 1.0 questions involve implicit relation links.

“Only 1% of the questions involve implicit relation links.”

May combine two claims.

Existing benchmarks contain few implicit relations and cannot be used to evaluate implicit relation linking well.

“Existing benchmarks contain few implicit relations and cannot be used to evaluate implicit RL well.”

Probably not a substantial claim: may not pass S3 (load-bearing).

All six state-of-the-art methods compared support relation linking over DBpedia.

“We compare our ImRL with six state-of-the-art methods.All of them support RL over DBpedia.”

May combine two claims.

ImRL can capture similar relationships between dbo and dbp through the KG embedding module, showing that ImRL can leverage the structural information in the KG.

“Thus, ImRL can capture similar relationships between dbo and dbp through the KG embedding module, which also shows that ImRL can leverage the structural information in the KG.”

• AssumptionRests on an assumption: Due to DBpedia's organizational structure, namespaces dbo and dbp have different representations of the same relations.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Rematch exploits WordNet and dependency parsing to model relation phrases with their underlying parts of speech, and uses various similarity metrics for ranking.

“Rematch (Mulang et al., 2017) exploits Word-Net and dependency parsing to model relation phrases with their underlying parts of speech, and uses various similarity metrics for ranking.”

• cites Mulang et al., 2017Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Classified as probably not a substantial claim.

Leveraging word embeddings and KG embeddings is common to both implicit and explicit relations.

“The reason is that leveraging word embeddings and KG embeddings is common to both implicit and explicit relations.”

May combine two claims.

All questions in PathSQ need to be linked through implicit relation paths, which is challenging for existing methods.

“This is because all the questions in PathSQ need to be linked through implicit relation paths, which is challenging for the existing methods.”

May combine two claims.

ImRL can achieve reasoning of relation paths according to the composition property of KG embeddings, thereby effectively improving the accuracy of implicit RL.

“By contrast, ImRL can achieve the reasoning of relation paths according to the composition property of KG embeddings, thereby effectively improving the accuracy of implicit RL.”

Classified as probably not a substantial claim. May combine two claims.

For explicit relations, because they usually contain a single relation, both R_Encoder and P_Encoder only need to encode one relation phrase or relation local name, which makes the matching of word embeddings easier.

“For explicit relations, since they usually contain a single relation, both R_Encoder and P _Encoder only need to encode one relation phrase or relation local name, which makes the matching of word embeddings easier.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

SIBKB uses PATTY as the underlying knowledge source and establishes a bipartite graph index from common phrases in natural language to KG relations for relation linking.

“SIBKB (Singh et al., 2017) uses PATTY as the underlying knowledge source, and establishes a bipartite graph index from common phrases in NL to KG relations for RL.”

• cites Singh et al., 2017 sourceNot checked — no full text of the source

May combine two claims.

ImRL can not only deal with relation phrases that are not explicitly expressed in natural language, but also perform relation path reasoning in the knowledge graph.

“demonstrating that ImRL can not only deal with relation phrases that are not explicitly expressed in NL, but also perform relation path reasoning in KG.”

The quote only partly states this claim.

The better results of all injection methods verify the effectiveness of knowledge injection.

“which verifies the effectiveness of knowledge injection.”

May combine two claims.

The knowledge in the dictionary may bring noises; combining part of the external knowledge and the original embeddings of the phrases can supplement more accurate external knowledge to enhance representations of implicit relation phrases.

“This shows that the knowledge in the dictionary may also bring noises, thus by combining part of the external knowledge and the original embeddings of the phrases, more accurate external knowledge can be supplemented to enhance the representations of implicit relation phrases.”

Classified as probably not a substantial claim.

The root cause of word embeddings being more important is that pre-trained language models (e.g., BERT) are trained on huge corpora with several downstream tasks, thus word embeddings can provide more prior knowledge.

“We believe that the root cause to the phenomenon is that pre-trained language models, e.g., BERT, are trained on huge corpora with several downstream tasks, thus the word embeddings can provide more prior knowledge.”

May combine two claims.

ImRL outperforms the latest method SimReL due to the use of KG embeddings and external knowledge of paraphrase dictionary as additional information.

“due to the use of KG embeddings and external knowledge of paraphrase dictionary as additional information, ImRL outperforms the latest method SimReL without the two information.”

May combine two claims.

The KG embedding model used by ImRL (RotatE) is trained on a small-scale corpus, and its embeddings focus on limited structural information, leading to KG embeddings providing relatively less information than word embeddings.

“While the KG embedding model used by ImRL, i.e., RotatE, is trained on a smallscale corpus, and the obtained embeddings more focus on limited structural information, which leads to that the KG embeddings provide relatively less information than the word embeddings.”

May combine two claims.

Path generation may cause correct relation paths to be missing in the candidate set, which causes the recall of ImRL to be lower than its precision.

“since the path generation may cause correct relation paths missing in the candidate set, the recall of ImRL is lower than precision.”

May combine two claims.

Even if lots of additional information is considered, it is still difficult to pick the most correct relation path from a set of candidates, which reveals the challenge of the relation linking task.

“Even if lots of additional information is considered, it is still difficult to pick the most correct one from a set of relation path candidates, which reveals the challenge of RL task.”