The reconciled knowledge graphs are typically used for multidocument summarization, or to detect knowledge evolution across document series. This paper focuses on reconciling knowledge graphs generated from two text documents about similar events described di erently. Our approach employs and extends MERGILO, a tool for reconciling knowledge graphs extracted from text, using word similarity and graph alignment. Complete semantic representation of events are generated using FRED, a semantic web machine reader, jointly with Framester, a linguistic linked data hub represented using a novel formal semantics for frames. Event-reconciliation is mainly performed via similarities based on the graph structure of frames using RDF2Vec graph embeddings, and the subsumption hierarchy of semantic roles as de ned in Framester. Our approach is evaluated over a coreference resolution task.
This study targets the problem of knowledge reconciliation (KR) [18] from the perspective of events. KR is useful in providing a combination of multiple graphs generated by multiple texts describing the same event. This merged graph provides a graph based summary of multiple texts which is more easily comprehensible by users and machines and usable by the algorithms providing interactive exploration of graphs/text analytics through visualization methods.
MERGILO [18] is a tool for reconciling knowledge graphs extracted from
text, it rst computes the word similarity between the node labels and then
performs graph alignment over the complete graphs. When di erent verbs
denote similar events and di erent agents play slightly di erent roles, the string
matching techniques as introduced in MERGILO might not be appropriate in
the KR process. For overcoming this limitation we use Frame Semantics which
describes a situation in the text with the help of frames and roles. For
identifying frames and semantic roles of entities in a text we use FRED [12], a machine
reader which generates event-centered knowledge graphs from two di erent texts.
Then, the similarity between these events is computed by calculating the
similarity between the corresponding FrameNet [
The rest of this paper is structured as follows. Section 2 lists the data sources, resources and tools we have adopted in our methodology. Section 3 includes state of the art work. Then, Section 4 gives some details of MERGILO and its functionalities and explains how frame semantics have been employed for improving MERGILO. Section 5 shows a precision-recall analysis for the presented approach on the EECB dataset. Finally, Section 6 concludes the paper with discussions, remarks and highlights some future directions. 2
FrameNet [
Event initial state
Event
Event end state prec.
prec.
Objective influence
Motion Transitive action
Control Intentionally affect
Mass motion
Motion Noise Invasion Scenario
Invading prec. prec.
Attack Conquering
Repel
Besieging resources to important ontological and linked data resources such as DBpedia, YAGO, DOLCE-Zero [20], schema.org etc.
Framester keeps the original FrameNet graph where the nodes represent the FrameNet frames and the edges represent di erent semantic relations between the frames i.e., Inheritance, SubFrame, CausativeOf etc. Figure 1 shows a part of FrameNet graph. Framester also contains a new subsumption hierarchy of semantic roles (i.e., frame elements) and added generic roles on top of the frame speci c roles. Figure 2 shows a part of the Framester role hierarchy associated with the framester role agent.2
and http://www.ontologydesignpatterns.org/ont/framester/data/framesterrole.ttl# are gfe: and framesterrole: respectively. FRED [12] 3 is a machine reader which generates ontological structure from natural language text using Discourse Representation Theory (DRT), frame semantics and Ontology Design Patterns. FRED uses Boxer,4 an open source tool for deep parsing of natural language using Combinatory Categorial Grammar (CCG) and produces event-based, semantic representations of natural language. The Discourse Representation Structures (DRS) produced by Boxer use VerbNet thematic roles. These functionalities implemented in FRED help in the event detection task for our method. 3
Approaches for integrating knowledge include cross-document coreference
resolution (when knowledge is represented as text documents) and ontology matching
(when knowledge is in a machine-readable form). Cross-document coreference
resolution aims at associating mentions about a same entity (object, person,
concept, etc.) across di erent texts [
instance matching. Similarly, [15] shows a greedy iterative algorithm for aligning knowledge bases with millions of entities and facts. These approaches are characterised by the preferred large size of the ontologies/datasets treated (for best performance), which is rarely (probably never) derived from text sources. MERGILO, as other knowledge integration tools [15], employs graph alignment, a more general and widely studied problem [26]. Note that all these approaches are connected and related to the classical graph matching problem [22]. We address this problem from the perspective of events, by taking advantage of frame embeddings i.e., the vector representations of linguistic frames and semantic roles.
Recently, word embeddings have been used in variety of Information Retrieval
and Natural Language Processing applications. One recent application is used
for generating vector representations of word senses [13] and then these vector
representations are used for improving the results of word similarity and word
analogy tasks based on BabelNet word senses formally known as SensEmbed. [
Consider the two sentences: Sent1: \The Spaniards conquered the Incas." and Sent2: \The Incas were attacked by the Spaniards." They are describing the same event in the past using di erent words i.e., event of an attack or an invasion from Spaniards to Incas. Figure 3 shows the FRED graph of Sent1. Given two such knowledge graphs, MERGILO rst performs graph compression by merging the nodes in the same graph. The two compressed graphs are aligned by establishing a 1-1 correspondence between the nodes of the two graphs by maximizing a score function, which combines the similarity between aligned nodes and the similarity between aligned edges. In such a case, the similarity between \conquered" and \attacked" is not e ective since the word similarity is low, although in this context such words describe the same event.
For computing similarity between two nodes containing verb senses, the verb senses are rst mapped to frames using Framester mappings. For example, in Figure 3 s1 vn:data : Conquer 42030000 and for Sent2 we have s2 vn:data : Attack 33000000. According to Framester mappings, we obtain s1 Ñ tConqueringu and s2 Ñ tAttacku. These nodes are replaced by their corresponding frames. The edges containing the VN-roles are mapped to FNroles. For example, in Figure 3, the verb sense vn.data5:Conquer 42030000 evokes the roles vn.role:Agent and vn.role:Patient which are mapped to fe:Conqueror.conquering and fe:Theme.conquering respectively.
In the sentence in Figure 2, the roles evoked by the verb sense vndata:Attack 33000000 are vndata:Agent and vndata:Theme. The Framester mappings contains the following records for these roles: vndata:Agent.conquer_42030000 skos:closeMatch fe:Conqueror.conquering . vndata:Patient.conquer_42030000 skos:closeMatch fe:Theme.conquering . vndata:Agent.attack_33000000 skos:closeMatch fe:Assailant.attack . vndata:Theme.attack_33000000 skos:closeMatch fe:Victim.attack
Then the similarities are computed in two ways: (i) by considering the taxonomical structure imposed by the \inheritance" relation represented as fnschema6:inheritsFrom in Framester using Path Similarity, Wu-Palmers Similarity, Leacock-Chodorow Similarity; (ii) using Frame Embeddings.
Frame Embeddings using RDF2Vec: To learn latent numerical representation of the frames and roles in the FrameNet graph, we follow the RDF2Vec approach. First we transform the graph into a set of sequences of entities, which is then fed into a neural language models, resulting into vector representation of all the nodes in the graph in a latent feature space.
To convert the graph into a set of sequences of entities we use two approaches,
i.e., graph walks and Weisfeiler-Lehman Subtree RDF Graph Kernels. (i) Graph
Walks: given a graph G pV; Eq, for each vertex v P V , we generate all graph
walks Pv of depth d rooted in vertex v. To generate the walks, we use the
breadthrst algorithm. In the rst iteration, the algorithm generates paths by exploring
the direct outgoing edges of the root node vr. In the second iteration, for each of
the previously explored edges, the algorithm visits the connected vertices. The
nal set of sequences for the given graph G is the union of the sequences of
all the vertices PG vPV Pv. (ii) Graph Kernels: it computes the number of
sub-trees shared between two or more graphs by using the Weisfeiler-Lehman [
Once the set of sequences of entities is extracted, we build a word2vec model. Word2vec is a particularly computationally-e cient two-layer neural net model for learning word embeddings from raw text. There are two di erent algorithms, the Continuous Bag-of-Words model (CBOW) and the Skip-Gram model. The CBOW model predicts target words from context words within a given window, while the skip-gram model does the inverse. Once the training is nished, the cosine similarity is computed between two frames and roles. 5
The experiments were conducted for the task of Cross-document Coreference Resolution (CCR) on RDF graphs, which focuses on associating RDF nodes about a same entity (object, person, concept, etc.) across di erent RDF graphs generated from text. The data set used for the experimentation was obtained
by the EECB data set which speci es coreferent mentions (text fragment). Our
dataset was obtained by generating RDF graphs using FRED and associating
text mensions to graph nodes by manual annotations. The framework is built
on top of the original MERGILO code, which was released as a Python tool7.
IBM ILOG CPLEX 12.6.1 was used for solving the Integer Linear Program and
the experiments were conducted on a MacOS server with 6-Core Intel Xeon E5
3.50GHz and 64GB of RAM. We used the following metrics for evaluation: (i)
MUC [25]: Link-based metric that quanti es the number of merges necessary to
cover predicted and gold clusters; (ii) B3 [
For the current evaluation, MERGILO was considered as a baseline. Table 1 shows the results for the baseline method, the Wu-Palmer's similarity, the Path similarity, the Leacock-Chodorow similarity and the results for cosine similarity using (i) graph walks and (ii) graph kernels with FrameNet roles respectively. Here Frame2Vec refers to the vector representations generated for FrameNet frames and Role2Vec refers to the vector representations generated for frame elements i.e., semantic roles.
For the rst approach with graph walks, for each entity in the FrameNet graph 200 and 500 random walks were generated, each of depth 4 and 8. For each entity in the subsumption hierarchy of roles we generate 400 random walks with depth 4. For the Weisfeiler-Lehman algorithm, we use h 2 iterations and subgraph depth d 2, and after each iteration of the algorithm we extract all walks for each entity with the same depth. We use these sequences to build both CBOW and Skip-Gram models with the following parameters: window size = 5; number of iterations = 10; negative sampling for optimization; negative samples = 25; with average input vector for CBOW. We experiment with 200 and 500 dimensions for the entities' vectors.
The results clearly indicate that each model used for graph walks and graph kernels performs better than the MERGILO baseline for all the considered metrics, showing a clear advantage of using the proposed frame similarities for reconciling knowledge graphs. The Wu-Palmer, Path and Leacock Chodorow measures use the inheritance relations only whereas Frame2Vec employs either graph walks or graph kernels over the FrameNet frame graph as well as subsumption hierarchy of FrameNet roles using either only FrameNet roles or improved subsumption hierarchy of FrameNet roles as introduced in Framester. Based on these settings, vector representations are generated which are further used for computing the cosine similarity. In general, Frame2Vec, for its intrinsic construction, exploits
more semantics than the other similarity measures (Wu-Palmer, Path and Leacock Chodorow); for such a reason, Frame2Vec provides the highest results for almost each evaluation measure except for BLANC.
BLANC is more sensitive to wrong assignments when clusters of mentions are larger, since a wrong assignment lead to a higher number of wrong noncoreference links. Therefore, although BLANC is case-by-case coherent with the other measures (when BLANC is low, the other measures are low and viceversa), in the few cases when Frame2Vec is outperformed by other measures (WuPalmer, Path and Leacock Chodorow), the BLANC measure, and in particular the contribution given by non-coreference link, gives a much smaller score. These cases in uence the overall average and for this reason in Table 1 BLANC seems to have a di erent behaviour than the other measures.
The generated models i.e., vector representations of FrameNet frames generated using FrameNet graph and subsumption hierarchy of FrameNet roles using RDF2Vec are freely available on-line8. 8 http://lipn.univ-paris13.fr/~alam/Frame2Vec/
This paper presents a way to perform event-reconciliation for merging multiple event-oriented knowledge graphs originated from multiple texts. It uses existing tool MERGILO, a tool for reconciling knowledge graphs using word similarity and graph alignment. The current study exploits several path-based similarity measures for frames and semantic roles, i.e., following the approach RDF2Vec, graph-based frame embeddings were generated. The evaluation shows that the introduced approach is an e ective improvement over the baseline.
Ongoing work concentrates on practical applications of frame embeddings in real systems, such as news series integration, knowledge graph evolution with robust event reconciliation (e.g. in streaming of texts where we expect relatedness or updates), or con ict detection across texts describing similar facts with di erent narratives or perspectives. 12. Aldo Gangemi, Valentina Presutti, Diego Reforgiato Recupero, Andrea Giovanni Nuzzolese, Francesco Draicchio, and Misael Mongiovi. Semantic Web Machine Reading with FRED. Semantic Web Journal, 2016. 13. Ignacio Iacobacci, Mohammad Taher Pilehvar, and Roberto Navigli. Sensembed: Learning sense embeddings for word and relational similarity. In ACL (1), pages 95{105, 2015. 14. Karin Kipper Schuler. Verbnet: A Broad-coverage, Comprehensive Verb Lexicon.
PhD thesis, Philadelphia, PA, USA, 2005. AAI3179808. 15. Simon Lacoste-Julien, Konstantina Palla, Alex Davies, Gjergji Kasneci, Thore Graepel, and Zoubin Ghahramani. Sigma: Simple greedy matching for aligning large knowledge bases. In KDD2013, pages 572{580, New York, USA, 2013. ACM. 16. Heeyoung Lee, Marta Recasens, Angel Chang, Mihai Surdeanu, and Dan Jurafsky.
Joint entity and event coreference resolution across documents. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 489{500, 2012. 17. Xiaoqiang Luo. On coreference resolution performance metrics. In Proc. of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pages 25{32. Association for Computational Linguistics, 2005. 18. Misael Mongiov , Diego Reforgiato Recupero, Aldo Gangemi, Valentina Presutti, and Sergio Consoli. Merging open knowledge extracted from text with MERGILO.
Knowl.-Based Syst., 108:155{167, 2016. 19. Roberto Navigli and Simone Paolo Ponzetto. BabelNet: The Automatic Construction, Evaluation and Application of a Wide-Coverage Multilingual Semantic Network. Arti cial Intelligence, 193:217{250, 2012. 20. A. G. Nuzzolese, A. Gangemi, V. Presutti, P. Ciancarini, and A. Musetti. Automatic Typing of DBpedia Entities. In Proc. of the International Semantic Web Conference (ISWC), Boston, MA, US, 2012. 21. Marta Recasens and Eduard Hovy. Blanc: Implementing the rand index for coreference evaluation. Natural Language Engineering, 17(04):485{510, 2011. 22. Diego Reforgiato Recupero. E cient graph matching. Encyclopedia of Data Warehousing and Mining, pages 736{743, 2009. 23. Petar Ristoski and Heiko Paulheim. Rdf2vec: Rdf graph embeddings for data mining. In ISWC, pages 498{514, 2016. 24. Fabian M. Suchanek, Serge Abiteboul, and Pierre Senellart. PARIS: Probabilistic alignment of relations, instances, and schema. In Proceedings of the VLDB Endowment, volume 5, pages 157{168. VLDB Endowment, 2011. 25. Marc Vilain, John Burger, John Aberdeen, Dennis Connolly, and Lynette Hirschman. A model-theoretic coreference scoring scheme. In Proceedings of the 6th conference on Message understanding, pages 45{52, 1995. 26. Joshua T Vogelstein, John M Conroy, Vince Lyzinski, Louis J Podrazik, Steven G Kratzer, Eric T Harley, Donniell E Fishkind, R Jacob Vogelstein, and Carey E Priebe. Fast approximate quadratic programming for graph matching. PLoS One, 10(4):e0121002, 2015. PMID: 25886624. 27. William Yang Wang and Diyi Yang. That's so annoying!!!: A lexical and framesemantic embedding based data augmentation approach to automatic categorization of annoying behaviors using# petpeeve tweets. In EMNLP, pages 2557{2563, 2015.