Appendix for:

Amina Kadry, Laura Dietz.
Open Relation Extraction for Support Passage Retrieval: Merit and Open Issues. 
ACM SIGIR. 2017.

Paper

inside ./benchmark/wikidump/:

• query${qid}/
  • query${qid}_${entityid}/
    • query${qid}_${entityid}_${filename}

The query ids and entity ids are according to the REWQ Clueweb12 Dataset of Schuhmacher, Dietz, and Ponzetto available at http://mschuhma.github.io/rewq/

For example: ./query231/query231_5/... for query231_Sloth_(deadly_sin)

The filenames are specified below.

For every entity, entity's Wikipedia article is crawled and split it into candidate sentences.

${filename}=allSentences-SIDs ~ displays the sentences, and the associated sentence id (first column).

For example we refer to sentence number 7 as S\7 such as in query231_Sloth_(deadly_sin)\S\7 in the following.

${filename}=allClauses ~ lists extracted clauses from each sentence (first column is sentence number, e.g. 7)

From each sentence, multiple propositions are extracted with the ClausIE system. For example, we have four propositions extracted from the sentence above, for example P\7-0 in the following

• query231_Sloth_(deadly_sin)\P\7-0
• query231_Sloth_(deadly_sin)\P\7-11
• query231_Sloth_(deadly_sin)\P\7-2
• query231_Sloth_(deadly_sin)\P\7-3

${filename}=allSentences-SIDs-Extractions: ~ All sentences and extracted propositions

${filename}=allSentences-SIDs-IDs ~ Mapping of sentences and propositions to IDs (S for sentence and P for proposition) by line number.

${filename}=allSentences-SIDs-Extractions+IDs ~ Both of these files concatenated.

A source code example for how the sentences, propositions, and clauses are extracted can be found in ./src/ExampleNNFile.java

The following ${filename}'s provide the ground truth sentences for this query-entity pair. We denote

• sExplainingEntityRelevance: Ground Truth (AQ1)
• sContainingRelation: Rel (AQ2)
• sWithRelevantRelation: Rel rel (AQ3)
• sContainingClausieExtraction: ClausIE (AQ4)
• sContainingRelevantClausieExtraction: ClausIE rel (AQ5)
• sContainingQueryTerm: Sentences that contain at least one query term, stopwords do not count (Qterm)
• sEntityMentions: Sentences that mention the query entity (Name)
• sContainingRelevantEntity: Sentence that contain a relevant entity (not used in paper)

${filename}=allSentences_html.html ~ Displays the assessment interface used to collect the assessments for AQ1-5.

The source code for reading annotations and sentence can be found in ./src/New_AnnotationsReader.java

The ${filename} "allFeatures" lists all extracted features by our approach in SVM-light format. The query ids are renumbered to be unique per query-entity-pair. For reference the ${qid}, ${entityid}, and ${sentenceid} can be extracted from the comment at the end of the line. Example: # sentid = sent\query231_Sloth_(deadly_sin)\S\1

A full list of features (also see Table 1 of the paper).

Category "Text"

• 1 sentence length measured in number of words
• 2 sentence position measured as a fraction of the document
• 3 fraction words that are stop words
• 4 fraction of query terms covered by sentence
• 5 sum of ISF of query terms (ISF is inverse sentence frequency)
• 6 average of ISF of query terms
• 7 sum of TF-ISF of query terms
• 8 number of entities mentioned

Category "NLP"

• 9-12 for nouns/verbs/adjectives/adverbs: fraction of words with POS tag
• 13 whether sentence contains a named entity
• 14-16 for NER types PER/LOC/ORG: whether NER of type is contained

Category "Dependency Parse (DP)"

• 17 number of edges on the path between two entities in dependency tree
• 18 indicator whether path goes through root node
• 19 indicator whether path goes through query term

Category "ClausIE"

• 20 whether ClausIE generated an extraction from this sentence
• 21-27 for all seven clause types: whether clause of this type is extracted
• 28 proposition length measured in tokens
• 29 maximum constituent length (size of dependency tree) in proposition
• 30-32 for subject/object/both: if another entity is in subject and/or object position of the proposition
• 33-34 for subject/object position: if given entity is in position of proposition
• 35-36 for subject/object position: if any entity is in position of proposition
• 37-38 for subject/object position: if an entity link is in position of prop.
• 39-41 for subject/verb/object position: if a query term (ignoring stopwords) is in position of proposition
• 42-43 for subject/object position: if a named entity (NER) is in position of proposition

The source code for feature extraction can be found in ./src/FeatureExtraction.java

The empirical results for Experiment 1 "Relations and Relevance" (Table 2) are computed with the sheet ./spreadsheet/Statistics-Table2.xlsx.

The empirical results for Experiment 2 "Evaluation through LTR" (Figure 1 and Table 3) are computed with the sheet ./spreadsheet/L2R-Table3.xlsx.