Frame-based Ontology Population from Text with PIKES

Transcription

1 Ontology Summit 2017 AI, Learning, Reasoning, and Ontologies April 5th, 2017 Frame-based Ontology Population from Text with PIKES Francesco Corcoglioniti Fondazione Bruno Kessler Trento, Italy joint work with Marco Rospocher and Alessio Palmero Aprosio

2 Knowledge Extraction Knowledge Extraction (KE) from text cross-disciplinary field (SW, NLP & KR) allowing bridging Web & Semantic Web a.k.a. Ontology Learning & Population [Cimiano, 2006] Ontology Learning: extraction of terminological knowledge (TBox) Ontology Population: extraction of assertional knowledge (Abox) knowledge items: instances, binary relations, semantic frames Semantic Frames n-ary relations, events & structured entities reified as ontological instances linked by role properties to participant instances schema defined by frame-based ontologies, e.g., FrameBase, ESO role property frame instance ( frame class) role property role property participant instance 2 participant instance... participant instance

3 Frame-based Ontology Population Why? fixed Tbox favors using extracted knowledge (compare with Open IE) expressivity of frames better than binary relations Semantic Role Labeling (SRL) fits well (and also other NLP tasks) predicate: supporter.01 Example: G. W. Bush and Bono are very strong supporters of the fight of HIV in Africa. argument: argm-mnr argument: arg0 frb:fe-taking_sides-cognizer arg0 arg0 :support_event_1 support.01 event frb:fe-taking_sides dbpedia:bush dbpedia:bono G. W. Bush Bono G. W. Bush G. W. supporters Bush Bono and Bono dbyago:person dbyago:person PER PER frb:fe-taking_sides-issue arg1 :fight_event_1 of the fight.01 fight of HIV event in Africa frb:fe-hostile_encounter frb:fe-taking_ argm-mnr frb:fe-hostile_ sides-degree encounter-side_2 arg1 attr:very-1r_strong-1a very strong ks:attribute 3 argument: arg1 dbpedia:hiv HIV of HIV owl:thing MISC argm-loc frb:fe-hostile_ encounter-place dbpedia:africa Africa in Africa dbyago:location GPE/LOC

4 Introducing PIKES* a tool for Frame-based Ontology Population from English text extracting semantic frames typed w.r.t. FrameBase instances typed w.r.t. YAGO and disambiguated w.r.t. DBpedia representing all contents in RDF + named graph based on a 2-phase approach open source (*) PIKES Is a Knownedge Extraction Suite [Corcoglioniti, Rospocher, Palmero Aprosio, TKDE 2016] 4

5 Data Model Resource layer ks:resource text...metadata... Mention layer ks:mentionof ks:mention offset in text...attributes... Instance layer ks:denotes / ks:implies ks:expresses a linguistically annotated piece of text about something of interest ks:instance describes Assertion (graph)...triples on instances... M3 M2 MentionSubclass...attributes... :resource1 a ks:resource; :mention1 a ks:framemention; dct:created " "; ks:mentionof :resource1; nif:isstring "G. W. Bush nif:beginindex 36; nif:endindex 46; and Bono are very nif:anchorof "supporters"; strong supporters of ks:synset wn30:n the fight of HIV in ks:predicate pmo:nb10_support.01; Africa.". ks:role pmo:nb10_support.01_arg01; ks:expresses :graph1; ks:denotes :supporters_entity; ks:implies :support_event_1. an entity (person...), semantic frame (event ) or attribute :graph1 { :supporters_entity a dbyago:supporter :support_event_1 a a frb:frame-taking_sides; frb:fe-taking_sides-cognizer :supporters_entity. } based on: Corcoglioniti et al. KnowledgeStore: a storage framework for interlinking unstructured and structured knowledge. IJSWIS

6 Data Model (2) Complete model: nif: < ks: < foaf: < ks:expresses contains triples about ks:resource ks:instance dct:title dct:creator dct:created ks:mentionof ks:mention ks:denotes / ks:implies ks:relationmention ks:participationmention ks:role ks:frame ks:framemention ks:predicate 6 Assertion (graph) ks:instancemention ks:coreferentialconjunct nif:beginindex ks:coreferential nif:endindex nif:anchorof ks:coreferencemention ks:synset ks:linkedto ks:argument ks:namemention ks:nerctype ks:timemention ks:norm.value rdf:type rdfs:label foaf:name owl:sameas rdfs:seealso ks:include frame/arg rel. ks:entity ks:frame ks:attribute ks:time ks:target ks:attributemention ks:normalizedvalue OWL time props. Instance layer Mention layer Resource layer

7 2-Phase Approach Resource layer G. W. Bush and Bono are very strong supporters of the fight of HIV in Africa. Phase 1 Linguistic Feature Extraction Mention layer G. W. Bush Bono ks:coref.conjunct very strong ks:arg. supporters ks:frame very strong supporters G. W. Bush and Bono [...] supporters ks:pred. fight ks:arg. HIV ks:frame fight of HIV supporters of [...] fight ks:frame Africa ks:arg. ks:arg. fight [...] in Africa ks:coreferential Phase 2 Knowledge Distillation Instance layer dbpedia:bush frb:fe-taking _sides-cognizer 7 dbpedia:bono attr:very-1r_strong-1a frb:fe-taking_sides-degree :support frb:fe-hostile :fight _encounter dbpedia:hiv dbpedia:africa -side_2 frb:fe-taking_sides-issue frb:fe-hostile_encounter-place

8 part-of-speech tagging POS NERC temporal expression recognition & norm. TERN word sense disambiguation semantic role labeling EL WSD SRL coreference resolution COREF dependency parsing 8 Coreference Frame named entity recognition & classification entity linking Attribute Time Type of mention Name NLP Task Instance ① apply several NLP tasks to input text ② map their outputs to mentions Participation Linguistic Feature Extraction DP

9 Linguistic Feature Extraction (2) Example: fight of HIV Extracted RDF mention graph via NERC, EL <..#char=54,59> a :FrameMention ; nif:anchorof fight :predicate pm:nb10-fight.01. via SRL via SRL, DP 9 <..#char=63,66> a :NameMention ; nif:anchorof HIV ; :nerctype :MISC ; :linkedto dbpedia:hiv. <..#char=54,66> a :ParticipationMention; nif:anchorof fight [ ] HIV :frame <..#char=54,59> ; :argument <..#char=63,66> ; :role pmo:nb10-fight.01-arg1.

10 Knowledge Distillation ① Rule-based conversion Mention layer mention graph mention attributes Rules Background knowledge (BK) mappings (e.g., PreMOn) further metadata (e.g., argument nominalization) Instance layer instances (entities, semantic frames) triples (participation, owl:sameas, ) named graphs links to mentions Example: body: mention with predicate annotation + predicate triggers argument nominalization (bk) head: create semantic frame & argument instances rules can be expressed in SPARQL ② Post-processing inference and owl:sameas smushing (i.e., merge owl:sameas instances) named graph merging in presence of same metadata / linked mentions 10

11 Knowledge Distillation (2) Mapping rules example: Mention layer Instance layer :mention1 a ks:framemention; nif:anchorof "supporters"; ks:synset wn30:n ; ks:predicate pmo:nb10_support.01; ks:role pmo:nb10_support.01_arg01; :g1 { :e1 a dbyago:supporter :ev1 a frb:frame-taking_sides; frb:fe-taking_sides-cognizer :e1. } Background knowledge pmo:nb10_support.01 a ks:argumentnominalization. 11 :mention1 ks:expresses :g1; ks:denotes :e1; ks:implies :ev1. INSERT {?m ks:denotes?i; ks:implies?if; ks:expresses?g. GRAPH?g {?i a ks:instance.?if a ks:frame } } WHERE {?m a ks:framemention; nif:anchorof?a, ks:predicate?s.?s a ks:argumentnominalization. BIND (ks:mint(?m) AS?g) BIND (ks:mint(?a,?m) AS?i) BIND (ks:mint(concat(?a, _pred ),?m) AS?if)

12 Implementation PIKES Java 1.8 on Linux / Mac OS X open source (GPL) Maven project on GitHub External dependencies Dbpedia Spotlight UKB need separate install 12 Stanford CoreNLP (tokenization, POStagging, lemmatization, NERC, TERN, DP, coref. resolution) Web UI Linguistic annotations Frontend Integrated dependencies Stanford CoreNLP Mate-tools Semafor RDFpro Input text ReST API Mention Extractor Mention graph Phase 1: Linguistic Feature Extraction DBpedia Spotlight (EL) Mate (SRL PB/NB) Semafor (SRL FrameNet) UKB (WSD) decoupling Mapping Rule Evaluator (rdfpro) Mapping rules Post processor (rdfpro) Mapping triples Knowledge graph Phase 2: Knowledge Distillation

13 Implementation (2) PIKES UI for: G.W. Bush and Bono are very strong supporters of the fight of HIV in Africa. Their March 2002 meeting resulted in a 5 billion dollar aid. 13

14 1 Evaluation: Throughput st Setup 109K Wikipedia-like pages (source: Simple English Wikipedia) 16 PIKES instances processing pages in parallel 24 CPU cores (dual Xeon E5-2430), 192GB ram, SSD Results 32 hours total processing time 358M triples (2M resource, 283M mention, 72M instance layers) item type Documents Sentences Tokens # items throughput [item/h] 24 cores 1 core* 109,242 3, ,584,406 50,000 3,125 23,877, ,000 47,100 (*) estimated based on time using all cores 14 time [s] # token per document

15 2nd Evaluation: FrameBase Population Gold standard 8 sentences (233 tokens) from [Gangemi, ESWC 2013] 166 triples (137 instances, 155 edges) manually annotated by 2 annotators Three SRL configurations Mate-tools PropBank, NomBank FrameBase via our mappings Semafor FrameNet FrameBase via existing mappings Mate-tools + Semafor leveraging outputs combination in Mention layer with Mate-tools: low recall & F1 due to mappings with both tools: F1 roles F1 all 15

16 3rd Evaluation: Comparison with FRED PIKES vs FRED [Presutti et al, EKAW 2012] on gold standard differently from PIKES, FRED is also able to extract TBox axioms comparison on VerbNet (VN) & FrameNet (FN) extracted triples PIKES nominal frames converted to FRED binary relation e.g., Iraqi official : { :rel1 :arg :per1, dbpedia:iraq } { :per1 :rel dbpedia:iraq } F1 16

17 Conclusions PIKES is a tool for Frame-based Ontology Population from English text extracting events and complex relations (semantic frames) representing all contents in RDF + named graph featuring 2-phases: linguistic feature extraction + knowledge distillation Benefits competitive with state of the art in terms of quality / throughput 2-phase decoupling allows tuning the two phases independenty Related/ongoing work 17 PreMOn lemon extension for predicate models KE4IR - knowledge extraction for IR KnowledgeStore - store for PIKES data KEM RDF/OWL model for knowledge extraction (ongoing work)

18 PreMOn = Predicate Model for Ontologies [Corcoglioniti, Palmero Aprosio, Rospocher, Tonelli, LREC 2016] Linguistic Linked Data resource (grounded in Lemon) representing predicate models and mapping resources: PB, NB, VN, FN, Semlink Homogeneously represents the semantic classes (e.g., rolesets in NB and PB, verb classes in VN, frames in FN) and semantic roles Benefits: ease of access and reuse of predicate model data abstract commonalities, keep peculiarities automated reasoning and SPARQL querying interlinking with third-party datasets Availability: download / SPARQL endpoint / URI dereferencing 18

19 NomBank / 5576 / K triples 3474, 3474, 9841 (NB 1.0) X, Y, Z = # conceptualizations, # sem. classes, # sem. roles (total 36K, 33K, 77K 19.77M triples) 1081, 1664, 3410 (NB 1.0) 2754, 3899, 7950 (PB 1.7) 5248,, (VN 3.2) VerbNet , 4617, (SemLink 1.2.2c) 4402 / 484 / K triples 3617, 1713, 1904 (SemLink 1.2.2c) PropBank / 6181 / K triples 3473, 3473, 9905 (NB 1.0) 5416, 5563, (SemLink 1.2.2c) 4765, 6186, (PB 2.1.5) 596 concept. 596 FB classes 491 roles 252 FB props 1706, 1792, (PB 2.1.5) 9413 / 1018 / K triples 2246 concept FB classes 3083 roles 900 FB props OntoNotes 5 groupings FrameNet 1.6 FrameNet 1.5 PropBank / 8751 / K triples WordNet 3.1 synsets / 1204 / K triples, 891, 9132 (FN 1.6) concept FB classes 9632 roles 9632 FB props (provided by FB) FrameBase (FB) 8151 classes, 9632 props Produced with PreMOnitor FB classes 7575 synsets (provided by FB)

20 po we re d by PI KE S KE4IR KE4IR = Knowledge Extraction for Information Retrieval [Corcoglioniti, Dragoni, Rospocher, Palmero Aprosio ESWC 2016] PIKES analysis of query and documents to improve IR performances Semantics considered (e.g. astronomers influenced by Gauss ) URIs: dbpedia:carl_friedrich_gauss TYPE: dbyago:astronomer , dbyago:germanmathematicians FRAME: framebase:subjective_influence TIME: dbo:dateofbirth (1777), dbo:dateofdeath (1855) Prec@1 Prec@5 Prec@10 NDCG NDCG@10 MAP MAP@10 Textual KE4IR improvement 3.03% 1.71% 4.55% 2.64% 2.99% 3.50% 4.74% Approach 20

21 KnowledgeStore KnowledgeStore Scalable storage for text, NLP annotations and extracted knowledge [Corcoglioniti, Rospocher, Cattoni, et al - IJSWIS 2015] Any application HTTP access, possibly exploiting SPARQL client libs. create retrieve update delete CRUD endpoint KnowledgeStore Tools populator, exporter, statistics collector,... KnowledgeStore Client SPARQL endpoint Client-side Web UI Server-side KnowledgeStore Server (single node) Hadoop HDFS (name & data nodes) Representation 21 Any Java application HBase (mult. master & region nodes) Resource Mention Virtuoso (single node) Instance

22 Knowledge Extraction Model (KEM) Revision of PIKES data model grounded on semiotics and NIF (ongoing work) Resource layer Instance layer Mention layer Sign kem:compositeresource kem:hascomponent 1..* kem:fragmentof kem:resource 1 Meaning kem:compositefragment kem:hascomponent 2..* kem:fragment Referent semiotic triangle of meaning kem:isabout / kem:refersto kem:mention kem:instance kem:involvesreferentof kem:hasannotation kem:hasresourceannotation kem:annotation Indonesia Hit By Earthquake A United Nations assessment team was dispatched to the province after two quakes, measuring 7.6 and 7.4, struck west of Manokwari Jan. 4. At 22 kem:involves / kem:subject kem:conveys kem:graph contains RDF triples about kem:semanticannotation kem:substantiates kem:involvessubjectof Named Entity type: ORG dbpedia:united_nations dbpedia:united_nations rdf:type dbo:organisation; rdfs:label United

23 Ontology Summit 2017 AI, Learning, Reasoning, and Ontologies April 5th, 2017 Thank you! Francesco Corcoglioniti Fondazione Bruno Kessler Trento, Italy joint work with Marco Rospocher and Alessio Palmero Aprosio