Event-based knowledge reconciliation using frame embeddings and frame similarity

Abstract

This paper proposes an evolution over MERGILO, a tool for reconciling knowledge graphs extracted from text, using graph alignment and word similarity. The reconciled knowledge graphs are typically used for multi-document summarization, or to detect knowledge evolution across document series. The main point of improvement focuses on event reconciliation i.e., reconciling knowledge graphs generated by text about two similar events described differently. In order to gather a complete semantic representation of events, we use FRED semantic web machine reader, jointly with Framester, a linguistic linked data hub represented using a novel formal semantics for frames. Framester is used to enhance the extracted event knowledge with semantic frames. We extend MERGILO with similarities based on the graph structure of semantic frames and the subsumption hierarchy of semantic roles as defined in Framester. With an effective evaluation strategy similarly as used for MERGILO, we show the improvement of the new approach (MERGILO plus semantic frame/role similarities) over the baseline.

Introduction

Several approaches have been proposed for extracting knowledge graphs from text. These knowledge graphs are generated with the aim of making unstructured text machine-readable  [1]. In case of multiple texts explaining similar events, it is more efficient and usable to provide the machine with a combination of multiple graphs generated by multiple texts. Using this merged graph, a machine reader can obtain knowledge contained in multiple texts from a single consolidated graph instead of reading several graphs. This problem, termed as ”Knowledge Reconciliation” (KR), has recently been addressed by MERGILO [2], a tool for reconciling knowledge graphs using graph alignment and word similarity. These reconciled knowledge graphs can further be utilized by specific NLP applications, in particular by graph-based text summarization (which aims at summarizing knowledge represented in multiple closely related pieces of text), for assessing sentence or document similarity, etc.

The current study mainly targets the problem of knowledge reconciliation from the perspective of events. In a text, a complete description of an event is syntactically denoted by a verb, since it defines a relation between event participants. The first step in the event-based knowledge reconciliation is to extract event-oriented knowledge graphs. For doing so, we use FRED, a machine reader presented in [1], which generates an RDF/OWL graph of any open domain input text.

For dealing with different lexical units describing the same or similar events, we enhance the existing pipeline by enriching the knowledge graphs generated by FRED with semantic frames as defined in FrameNet¹. For this purpose, this study further makes use of mappings between VerbNet² (i.e., VerbNet verb classes and VerbNet roles) and FrameNet, as contained in Framester [3]. Framester is a linguistic linked data hub formulated using a novel formal semantics for frames for improving semantic interoperability between linguistic resources. Framester uses the RDF version of FrameNet [4]³, formalizes the FrameNet graph in OWL, and introduces a very rich subsumption hierarchy related to FrameNet frame elements (semantic roles).

We use Framester graph representations as a way to improve similarity between the nodes and the edges, where nodes represent the frames and edges represent the roles. When different verbs denote similar events, i.e. different verbs evoke different frames which are somehow connected in the FrameNet graph using the semantic relations already defined in FrameNet (such as Inheritance, SubFrame, ...), we can greatly improve simple string matching techniques introduced in MERGILO with frame as well as semantic role similarity measures. For doing so we considered the similarities based on the graph structure of the FrameNet frames as well as the subsumption hierarchy associated to the semantic roles defined in Framester. FrameNet graph organizes frames using semantic relations; to benefit from this graphical structure we adapt WordNet similarity measures [5] to FrameNet graph. We further exploit the vector representations of frames using the FrameNet graph and the subsumption hierarchy of roles as represented in Framester. We follow the approach RDF2Vec [6] to generate graph based frame embeddings referred to as Frame2Vec. These graph-based embeddings make use of graph mining algorithms such as graph walks and graph kernels to traverse over the graph, which is further used for generating its vector representations. In order to find the similarity between two frames and between two roles, this study uses WordNet similarities and cosine similarity for obtaining better consolidation between multiple graphs, which lead to an improvement over the results of a baseline algorithm for knowledge reconciliation, MERGILO  [2]. MERGILO already computes the similarity between the roles represented as edges in the FRED graphs but it merely performs string matching for finding if the roles are similar. These embeddings can further be used for any NLP application, however in the current scenario we use it for knowledge reconciliation purposes.

More in detail, the paper is organized as follows. Section 2 introduces state of art and related work. Section 3 lists the data sources, resources and tools we have adopted in our methodology. Then, Section 4 gives some details of MERGILO and its functionalities for use as basis for the Section 5, which explains how frame semantics have been employed for improving MERGILO. Section 6 shows a precision-recall analysis for the presented approach on the dataset introduced in [2]. Finally, Section 7 concludes the paper with discussions, remarks and highlights some future directions.