RTE Knowledge Resources
Knowledge resources have shown their relevance for applied semantic inference, and are extensively used by applied inference systems, such as those developed within the Textual Entailment framework.
This page presents a list of the knowledge resources used by systems that have participated in the last RTE challenges. The first table lists the publicly available resources, the second one lists unpublished resources. Both tables are sortable by Resource name, type, author and number of users.
RTE Participants are encouraged to add information about all kind of knowledge resources used, from standard existing resources (e.g. WordNet) to knowledge collections created for specific purposes, which can be made available to the community.
Call for Resources
In order to help the research, all the participants are invited to contribute, sharing their own resources with the RTE community.
Making the resources available to be used by other systems has several advantages. On the one hand, it helps improve the TE technology; on the other hand, it offers an opportunity to further test and evaluate the resource.
RTE6 - Call for Resources
An ablation test consists of removing one module at a time from a system, and rerunning the system on the test set with the other modules, except the one tested.
Ablation test are meant to help better understand the relevance of the knowledge resources used by RTE systems, and evaluate the contribution of each of them to the systems' performances. In fact, comparing the results achieved in the ablation tests to those obtained by the systems as a whole allows assessing the contribution given by each single resource.
RTE5 - Ablation Tests
Publicly available Resources
|Resource||Type||Author||Brief description||PAST Users ||RTE4 Users||RTE5 Users||RTE6 Users||Usage info|
|WordNet||Lexical DB||Princeton University||Lexical database of English nouns, verbs, adjectives and adverbs||3||21||18||Users|
|eXtended Wordnet||Lexical DB||Human Language Technology Research Institute, University of Texas at Dallas||Extension of WordNet based on the exploitation of the information contained in WordNet definitional glosses: the glosses are syntactically parsed, transformed into logic forms and content words are semantically disambiguated. The Extended Wordnet is an ongoing project.||0||0||2||Users|
|Augmented Wordnet||Lexical DB||Stanford University||The resource is the result of the application of a learning algorithm for inducing semantic taxonomies from parsed text. The algorithm automatically acquires items of world knowledge, and uses these to produce significantly enhanced versions of WordNet (up to 40,000 synsets more).||0||0||1||Users|
|Verbnet||Lexical DB||University of Colorado Boulder||Lexicon for English verbs organized into classes extending Levin (1993) classes through refinement and addition of subclasses to achieve syntactic and semantic coherence among members of a class||2||2||1||Users|
|VerbOcean||Lexical DB||Information Sciences Institute, University of Southern California||Broad-coverage semantic network of verbs||2||3||6||Users|
|FrameNet||Lexical DB||ICSI (International Computer Science Institute) - Berkley University||Lexical resource for English words, based on frame semantics (valences) and supported by corpus evidence||1||1||2||Users|
|NomBank||Lexical DB||New York University||Lexical resource containing syntactic frames for nouns, extracted from annotated corpora||2||1||0||Users|
|PropBank||Lexical DB||University of Colorado Boulder||Lexical resource containing syntactic frames for verbs, extracted from annotated corpora||2||1||1||Users|
|Nomlex Plus||Lexical DB||New York University||Dictionary of English nominalizations: it describes the allowed complements for a nominalization and relates the nominal complements to the arguments of the corresponding verb||0||1||0||Users|
|Dekang Lin’s Thesaurus||Thesaurus||University of Alberta||Thesaurus automatically constructed using a parsed corpus, based on distributional similarity scores||0||1||1||Users|
|Grady Ward's Moby Thesaurus||Thesaurus||University of Sheffield||Thesaurus containing 30,260 root words, with 2,520,264 synonyms and related terms. Grady Ward placed this thesaurus in the public domain in 1996.||0||0||1||Users|
|Roget's Thesaurus||Thesaurus||Peter Mark Roget (Electronic version distributed by University of Chicago)||Roget's Thesaurus is a widely-used English thesaurus, created by Dr. Peter Mark Roget in 1805. The original edition had 15,000 words, and each new edition has been larger. The electronic edition (version 1.02) is made available by University of Chicago.||1||0||1||Users|
|Wikipedia||Encyclopedia||Free encyclopedia. Used for extraction of lexical-semantic rules (from its more structured parts), named entity recognition, geographical information etc.||0||3||6||Users|
|Umbel||Ontology||Structured Dynamics LLC, Coralville, IA||UMBEL stands for Upper Mapping and Binding Exchange Layer and is a lightweight ontology structure for relating Web content and data to a standard set of subject concepts||0||0||1||Users|
|YAGO||Ontology||Max-Planck Institute for Informatics, Saarbrücken, Germany||Light-weight and extensible ontology. It contains more than 2 million entities and 20 million facts about these entities. The facts have been automatically extracted from Wikipedia and unified with WordNet.||0||0||1||Users|
|DBpedia||Ontology||Open community project||DBpedia is a community effort to extract structured information from Wikipedia and to make this information available on the Web. The DBpedia knowledge base currently describes more than 2.9 million things in 91 different languages and consists of 479 million pieces of information.||0||0||1||Users|
|DIRT Paraphrase Collection||Collection of paraphrases||University of Alberta||DIRT (Discovery of Inference Rules from Text) is both an algorithm and a resulting knowledge collection. The DIRT knowledge collection is the output of the DIRT algorithm over a 1GB set of newspaper text.||2||4||3||Users|
|TEASE Collection||Collection of Entailment Rules||Bar-Ilan University||Output of the TEASE algorithm||0||0||0||Users|
|BADC Acronym and Abbreviation List||Word List||BADC (British Atmospheric Data Centre)||Acronym and Abbreviation List||0||1||1||Users|
|Acronym Guide||Word List||Acronym-Guide.com||Acronym and Abbreviation Lists for English, branched in thematic directories||1||1||3||Users|
|Web1T 5-grams||Word list||Linguistic Data Consortium, University of Pennsylvania; Google Inc.||Data set containing English word n-grams and their observed frequency counts. The n-gram counts were generated from approximately 1 trillion word tokens of text from publicly accessible Web pages||0||1||0||Users|
|Normalized Google Distance (RTE3&RTE4)||Word Pair Co-occurrence||Saarland University||Co-occurrence of the word pairs in RTE3 and RTE4 using Normalized Google Distance (Cilibrasi and Vitanyi, 2004). The word pairs are all the possible combinations of content words in T and H. In practice, we used Yahoo! as the search engine.||0||0||1||Users|
|Normalized Google Distance (RTE5)||Word Pair Co-occurrence||Saarland University||Co-occurrence of the word pairs in RTE3 and RTE4 using Normalized Google Distance (Cilibrasi and Vitanyi, 2004). The word pairs are all the possible combinations of content words in T and H. In practice, we used Yahoo! as the search engine.||0||0||1||Users|
|GNIS - Geographic Names Information System||Gazetteer||USGS (United States Geological Survey)||Database containing the Federal and national standard toponyms for USA, associated areas and Antarctica||0||1||0||Users|
|Geonames||Gazetteer||Database containing eight million geographical names. It is integrating geographical data such as names of places in various languages, elevation, population and others from various sources.||0||1||0||Users|
|Sekine's Paraphrase Database||Collection of paraphrases||Department of Computer Science, New York University||Data-base created using Sekine's method, NOT cleaned up by human. It includes 19,975 sets of paraphrases with 191,572 phrases.||0||0||0||Users|
|Microsoft Research Paraphrase Corpus||Collection of paraphrases||Microsoft Research||Text file containing 5800 pairs of sentences which have been extracted from news sources on the web, along with human annotations indicating whether each pair captures a paraphrase/semantic equivalence relationship.||0||0||0||Users|
|Downward entailing operators||Collection of entailing operators||Department of Computer Science, Cornell University, Ithaca NY||System output of an unsupervised algorithm recovering many Downward Entailing operators, like 'doubt'.||0||0||1||Users|
|WikiRules!||Lexical Reference rule-base||Bar-Ilan University||Extraction of about 8 million lexical reference rules from the text body (first sentence) and from metadata (links, redirects, parentheses) of Wikipedia. Provides better performance than other automatically constructed resources and comparable performance to WordNet. Offers complementary knowledge to WordNet.||0||1||1||Users|
|DART||Collection of "world knowledge" propositions||Boeing Research and Technology||23 million tuples such as "airplanes can fly to airports", "rivers can flood" collected from abstracted parse trees.||0||0||0||Users|
|FRED||FrameNet-derived entailment rule-base||Bar-Ilan University||This package contains the outputs of the FRED algorithm, an algorithm which extracts entailment rules from FrameNet.||0||0||0||Users|
|DIRECT||Directional Distributional Term-Similarity Resource||Bar-Ilan University||This is a resource of directional distributional term-similarity rules (mostly lexical entailment rules) automatically extracted using the inclusion relation as described in (Kotlerman et.al., ACL-09).||0||0||0||Users|
|binaryDIRT||Entailment rules between binary templates using DIRT algorithm||Bar-Ilan University|| This resource contains entailment rules over binary templates learned over the Reuters corpus using
the DIRT algorithm of Lin and Pantel.
|unaryBInc||Entailment rules between unary templates using BInc algorithm||Bar-Ilan University|| This resource contains entailment rules over unary templates learned over the Reuters corpus using
the BInc algorithm of Szpektor and Dagan (2008).
|New resource||Participants are encouraged to contribute||Users|
Not available Resources
The following table lists the unpublished resources used by RTE participants. Some of them have been developed by Users themselves specifically for RTE. Interested people may turn to authors to obtain further information.
|Resource||Type||Author||Brief description||PAST Users||RTE4 Users||RTE5 Users||RTE6 Users||Usage info|
|PARC Polarity Lexicon||Lexical DB||PARC - Palo Alto Research Center||Verbs classification with respect to semantic polarity||0||1||0||Users|
|Gazetteer from TREC||Gazetteer||NIST - National Institute of Standards and Technology||Cities and other geographical names||1||0||0||Users|
| DFKI Geographic Ontology
(to be released)
|Ontology||DFKI - German Research Center for Artificial Intelligence||Ontology containing geographic terms and two kinds of relations: the directional part-of relation, and the equal relation for synonyms and abbreviations of the same geographic area (e.g the United Kingdom, the UK, Great Britain, etc.)||0||1||0||Users|
|Geo||Collection of Entailment Rules||Bar-Ilan University; Tel-Aviv University||Meronymy entailment rules, based on TREC’s TIPSTER gazetteer.||0||0||1||Users|
|Regex||Collection of Entailment rules||Bar-Ilan University; Tel-Aviv University||Small set of entailment rules based on regular expressions, intended to address lexical variability involving temporal phrases||0||0||1||Users|
| Syntactic rule base
(to be released)
|Collection of Entailment Rules||Bar-Ilan University; Tel-Aviv University||A manually-composed collection of entailment rules which define parse tree transformations. The rules cover generic syntactic phenomena such as appositions, conjunctions, passive, relative clause, etc. (Bar-Haim et al., AAAI-07)||0||1||1||Users|
| Polarity rule base
(to be released)
|Collection of Entailment Rules||Bar-Ilan University; Tel-Aviv University||A manually-composed collection of entailment rules which detect predicates whose polarity is negative (e.g. didn't dance) or unknown (e.g. plans to dance). The rules capture diverse phenomena that affect polarity, e.g. verbal negation, modal verbs, conditionals, and certain verbs that induce negative or "unknown" polarity context. The latter were taken mainly from VerbNet. Extends a resource described in (Bar-Haim et al., AAAI-07)||0||1||0||Users|
|Lexical-Syntactic rule base||Collection of Entailment Rules||Bar-Ilan University; Tel-Aviv University||Extract lexical-syntactic entailment rules for predicates (verbal and nominal), including argument mapping. The resource is based on WordNet, Nomlex-Plus and Unary DIRT (Szpektor and Dagan, Coling 08)||0||1||0||Users|
|OPENU Collection||Collection of Entailment Rules and Patterns||Open University||Collections of rules, patterns etc. for RTE purpose, extracted from Reuter corpus parsed using Minipar.||1||0||0||Users|
|Abbr||Collection of rules for abbreviation||Bar-Ilan University; Tel-Aviv University||2000 Abbreviation rules, extracted from BADC and Acronym Guide||0||0||1||Users|
|UAIC Negation_list||Negation rules||„Al. I. Cuza“ University, Iasi, Romania||List of negative terms and words (verbs, adjectives, nouns) affecting modality or factuality of a infinitive verb preceded by the particle "to" (e.g. "believe","necessary", "attempt")||0||0||1||Users|
|DLSIUAES Negation_list||List of negative terms||University of Alicante||Basic list of negative terms.||0||0||1||Users|
|UAIC Quantifier_list||List of quantifiers||„Al. I. Cuza“ University, Iasi, Romania||List of quantifiers affecting entailment judgment. The quantifiers are taken from a list which contains expressions like “more than”, “less than”, or words such as “over”, “under”, etc.||0||0||1||Users|
|FBKirst StopWord list||List of frequent words|| FBK-Irst;
University of Trento - Italy
|A list of the 572 most frequent English words.||0||0||1||Users|
- ↑ RTE-3 data have been provided by participants by means of a questionnaire.
- ↑ RTE-4 data have been provided by participants and have been integrated with information extracted from the related proceedings.
- ↑ RTE-5 data have been provided by participants and have been integrated with information extracted from the related proceedings.
- ↑ RTE-6 data have been provided by participants and have been integrated with information extracted from the related proceedings.