Linked Metaphors Aldo Gangemi1,2 , Mehwish Alam1 , and Valentina Presutti1 1. Semantic Technology Lab, ISTC-CNR, Rome, Italy 2. FICLIT, University of Bologna, Italy Abstract. The poster summarizes Amnestic Forgery, an ontology for metaphor semantics, based on MetaNet and Framester factual-linguistic linked data. An example of metaphor generation based on linked metaphors is shown. 1 Introduction A metaphor is a cognitive operation involving usage of natural language and cross- domain conceptual mapping. Its ontological interest goes beyond purely cognitive and linguistic aspects, since ontology-based extraction and representation of knowledge need to make the semantics of natural language explicit even in presence of such phe- nomena. More generally, metaphor detection can help in improving Natural Language Understanding which can in turn allow the machines to understand and interpret pres- ence of Metaphors. We have designed an OWL ontology for metaphors, compliant with the frame-based semantics of Framester1 , and populated it with data from Berkeley’s MetaNet [2]. It is the reference repository of conceptual metaphors, developed as a Semantic Wiki2 , main- tained through collaborative editing by multiple conceptual metaphor experts. In Con- ceptual Metaphor Theory (CMT) [5], a metaphor is supposed to map aspects of a target domain (e.g. love) to those of another domain (e.g. travel). MetaNet used FrameNet [1] frames as units of those domains, and frame elements as their aspects. The new ontology is called Amnestic Forgery3 which is deployed as an extension of Framester [4]. It is a knowledge graph represented as Linked Open Data, which integrates heterogeneous linguistic resources (OntoWordNet, VerbNet, FrameNet-OWL, BabelNet, etc.), factual datasets (DBpedia, YAGO, etc.), and foundational ontologies, by providing them a uni- fied formal semantics. 2 Amnestic Forgery: an ontology for MetaNet and beyond Framester is a large knowledge graph containing more than 50 million triples link- ing millions of linguistic, conceptual, or real world entities. This ontology is based on Descriptions and Situations (D&S) [3]. The D&S formal framework is appropriate to 1 http://etna.istc.cnr.it/framester2/sparql 2 https://metaphor.icsi.berkeley.edu/pub/en/ 3 This is a recursive name, since FORGERY IS AMNESIA is a new metaphor generated by means of the ontology itself, cf. Sect. 2. formalise the (informal) explanation of the FrameNet [1] core schema, which distin- guished between frames and frame occurrences, or situations: “For example, the Apply heat frame describes a common situation involving a Cook, some Food, and a Heating Instrument, and is evoked by words such as bake, blanch, boil, broil, brown, simmer, steam, etc. We call these roles frame elements and the frame-evoking words are lexical units in the Apply heat frame”. Once formalised in D&S, the FrameNet implicit schema becomes a semiotic passep- artout: a frame f , as a description, can be the reification of any relation ρ with ar- bitrarily variable arity, a frame element f e is a binary projection of ρ, and a lexical unit lu of f is a symbol, for which ρ (and its reified counterpart f ), and its projec- tions f e1...n act as intensional interpretations. A “common situation” s described by f is the extensional interpretation (aka denotation) of lu, whose intension is f . The FrameProjection class allows to integrate any predicate defined either intensionally or extensionally in ontologies, lexical resources, or other vocabularies or web formats, For example, the FrameNet frame Activity start as well as the VerbNet verb class verbclass-begin-55.1-1 are linked as intensionally equivalent to the Framester frame ActivityStart, while the synset synset-newcomer-noun-1 from WordNet, which is intensionally mapped to FrameNet Activity start, is extensionally repre- sented as a class of newcoming entities, and linked as a unary projection of a Framester class of newcoming situations, Newcomer.n.1, which on its turn is represented as a subclass of ActivityStart. In D&S, higher-level descriptions can also be defined, e.g. for meta-norms that de- scribe priority between other norms. This expressive capability can be used to repre- sent a metaphor as a mapping operation between frames, as informally represented by MetaNet and CMT. A metaphor itself is a kind of description, which incorporates roles (frame elements) for two frames (the source and target frames), as well as mapping rules between the respective roles. The design approach taken to formalise MetaNet is to use D&S in order to extract and formalise a metaphor, then to extract data and formally represent them according to that schema, and finally to align the schema and data to elements in the Framester knowledge graph. Fig. 1: The subgraph for the metaphor CRIME IS A DISEASE. We have firstly scraped tabular data from the MetaNet wiki4 , and we have designed a preliminary MetaNet schema that catches the intended meaning of the interface used to 4 It is a SemanticMediaWiki instance, but its data querying facility is not accessible. populate the MetaNet wiki. Secondly, we have refactored the extracted data according to this preliminary schema, and fine-tuned it against features deriving from the data entry variety in the wiki. The result is a refined schema, the Amnestic Forgery ontology, and its MetaNet data. Fig. 1 depicts a subgraph of MetaNet for the metaphor CRIME IS A DISEASE. The subgraph contains examples of the core relations in MetaNet, link- ing metaphors to their source and target frames, their role mappings, entailments, and possibly other more vague relations contributed by the users of the wiki. A summary diagram of the axiomatisation for Amnestic Forgery is shown in Fig. 2. The diagram uses a UML-class-diagram-oriented profile to sketch the core axioms for the Metaphor class, shown either as “attributes” within class boxes, or as either “as- sociations” (restrictions), or “generalisations” (subsumptions) across class boxes. The diagram summarizes the reuse of the Description class from D&S, which subsumes the Metaphor, Frame, and MetaphoricRoleMapping classes. A hierarchy of frame and role notions exemplify Framester schema alignements, and the treatment of seman- tic roles as both binary projection of frames, and OWL properties (binary relations). Association-like edges derive from either domain or range restrictions in the OWL en- coding of Amnestic Forgery, or from existential restrictions. Please refer to the OWL file for the full axiomatization, including imports, alignments, disjointness, and docu- mentation axioms5 . Generating new metaphors with MetaNet and Framester: In order to prove the advan- tages of having a large and formally rigorous knowledge base, we report here a query to Framester extended with Amnestic Forgery and MetaNet data. Given a MetaNet metaphor, the query is able to generate hundreds of novel intensional metaphors: an API is available online6 . Example results include e.g. the eponymous Amnestic Forgery as a linguistic rendering of the FORGERY IS AMNESIA metaphor, which appears to be ac- tually novel: no realisations can be found e.g. on the Web (based on Google searching). The query running behind the API Amnestic Forgery (see below), as announced earlier only uses adjective-noun phrase constructions, and their related senses and frames in Framester. prefix metanet: prefix framedata: prefix metaphordata: SELECT DISTINCT ?ssyn ?tsyn WHERE { metaphordata:CRIME_IS_A_DISEASE metanet:hasSourceFrame ?s ; metanet:hasTargetFrame ?t . ?s skos:closeMatch ?fns . ?fns a fn15schema:Frame . ?t skos:closeMatch ?fnt . ?fnt a fn15schema:Frame . ?fns skos:closeMatch ?ssyn . ?fnt skos:closeMatch ?tsyn . {?ssyn a wn30schema:AdjectiveSynset} UNION {?ssyn a wn30schema:AdjectiveSatelliteSynset} 5 https://github.com/alammehwish/AmnesticForgery 6 https://lipn.univ-paris13.fr/framester/en/metanet/ ?tsyn a wn30schema:NounSynset } Ongoing and future work bears multiple directions: completing the alignment of MetaNet frames to FrameNet frames (currently covering only about 25% of Metanet), automatically detecting metaphorical sentences in text; automatically enriching metaphoric knowledge graphs. Fig. 2: A class-diagram profile for the OWL axiomatisation of Amnestic Forgery. References 1. C. F. Baker, C. J. Fillmore, and J. B. Lowe. The Berkeley FrameNet Project. In Proc. of COLING 1998. 2. E. Dodge, J. Hong, and E. Stickles. Metanet: Deep semantic automatic metaphor analysis. In Third Workshop on Metaphor in NLP. ACL, 2015. 3. A. Gangemi. Norms and plans as unification criteria for social collectives. Autonomous Agents and Multi-Agent Systems, 17(1):70–112, 2008. 4. A. Gangemi, M. Alam, L. Asprino, V. Presutti, and D. R. Recupero. Framester: A wide coverage linguistic linked data hub. EKAW 2016. 5. G. Lakoff and M. Johnson. Metaphors we Live by. University of Chicago Press, Chicago, 1980.