RepGraph: Visualising and Analysing Meaning Representation Graphs
Jaron Cohen∗ Roy Cohen∗ Edan Toledo∗ Jan BuysDepartment of Computer Science
University of Cape Town{CHNJAR003, CHNROY002, TLDEDA001}@myuct.ac.za [email protected]
Abstract
We present RepGraph, an open source visu-alisation and analysis tool for meaning repre-sentation graphs. Graph-based meaning rep-resentations provide rich semantic annotations,but visualising them clearly is more challeng-ing than for fully lexicalized representations.Our application provides a seamless, unifyinginterface with which to visualise, manipulateand analyse semantically parsed graph datarepresented in a JSON-based serialisation for-mat. RepGraph visualises graphs in multi-ple formats, with an emphasis on showing therelation between nodes and their correspond-ing token spans, whilst keeping the represen-tation compact. Additionally, the web-basedtool provides NLP researchers with a clear, vi-sually intuitive way of interacting with thesegraphs, and includes a number of graph anal-ysis features. The tool currently supports theDMRS, EDS, PTG, UCCA, and AMR seman-tic frameworks. A live demo is available athttps://repgraph.vercel.app/.
1 Introduction
Broad-coverage semantic graphs provide richerrepresentations of sentence meaning than surface-level syntax or lexicalised semantic dependencies(Oepen et al., 2019). The breadth of meaning repre-sentation approaches now includes a large numberof semantic graph frameworks — each with theirown respective strengths and weaknesses at encod-ing the meaning of natural language (Koller et al.,2019). Recently, a growing body of work has fo-cused on parsing to or generating from graph-basedmeaning representations (Hershcovich et al., 2017;Buys and Blunsom, 2017; Zhang et al., 2019; Songet al., 2018). The outputs of many other syntacticand semantic analysis tasks can also be representedas graphs, where labelled nodes correspond to to-ken spans and edges to relations between thesespans (Jiang et al., 2020).
∗ These authors contributed equally to this work
Figure 1: Hierarchical visualisation of the EDS graphfor the sentence “Champagne and dessert followed.”
Visualisations of constituency trees, syntacticdependency trees and semantic dependency graphsare well established for teaching, analysing and pre-senting the outputs of those representations. How-ever there is no similar established standard forvisualising broad-coverage semantic graphs, dueto diverging approaches to representing semanticsin different frameworks, as well as the challengesinvolved in visualising both the graph and the cor-respondence to the sentence it represents clearly.
To stem the “Balkanisation” caused byframework-specific approaches, CoNLL 2019 and2020 hosted shared tasks on Cross-FrameworkMeaning Representation Parsing (MRP; Oepenet al., 2019, 2020). These tasks provided a uniformabstract graph representation with a JSON-basedserialization format and standardized datasets.However, current visualisation tools are eitherframework-specific or fail to represent graphsclearly and consistently across frameworks (§5).
RepGraph is an open source, web-based visu-alisation and analysis tool for meaning representa-tion graphs, with support for multiple frameworks.
The tool provides three novel visualisation formats,each designed to better elicit specific inherent qual-ities of the graphs (§3). In contrast to previousapproaches, two of our visualisations (hierarchicaland tree-like) represent the graph and sentence to-kens in an integrated manner, clearly showing therelationship between nodes and the tokens or tokenspans they correspond to. The third (flat) is akin toexisting dependency graph visualisations.
The tool provides a unified, intuitive and feature-rich platform for interacting with meaning repre-sentation graphs. Analysis functionality includes:displaying subgraphs, searching graphs by node la-bel or subgraph, comparing graphs visually, testinggraph properties such as planarity, and providingdataset-level statistics (§4).
RepGraph is targeted towards everyone work-ing with meaning representation graphs, including:researchers developing parsers or generators; com-putational linguists performing semantic analysis;NLP practitioners using graphs in downstream ap-plications; and students learning about these repre-sentations. Parsing is not currently integrated, butparser output can easily be processed. RepGraphis also not intended to be an annotation tool; suchfunctionality is orthogonal to what we provide.
Five semantic graph frameworks are currentlysupported: Dependency Minimal Recursion Se-mantics (DMRS; Copestake, 2009), ElementaryDependency Structures (EDS; Oepen and Lønning,2006), Prague Tectogrammatical Graphs (PTG;Hajic et al., 2012), Universal Conceptual CognitiveAnnotation (UCCA; Abend and Rappoport, 2013)and Abstract Meaning Representation (AMR; Ba-narescu et al., 2013). Some framework-specificnormalisations are performed to improve compat-ibility and enable a unified approach to visualisa-tion. While our development focused on Englishdatasets, most of the frameworks also support otherlanguages, and the tool can easily be extended tosupport additional frameworks.1
2 System Description
A semantic graph is a triple (T,N,E), where Nis a set of nodes, E ⊆ N ×N is a set of directededges, and T ⊂ N is a set of top nodes (Oepenet al., 2019).
Nodes may optionally have zero or more prop-erties with associated values. The relationship be-
1A demo video is available at https://vimeo.com/user136369092/repgraph
tween the graph nodes and the input string is re-ferred to as anchoring or alignment. We assumethat the input is tokenized; a graph node may beanchored to a token, a token span, or a set of tokenspans.2 The alignment is not annotated in all frame-works; we assume that it can be obtained, using analigner, for (most) graph nodes.
The alignment between the nodes and input to-kens forms the basis of the design of our hierar-chical and tree-like visualisation formats. We alsodistinguish between surface nodes, which repre-sent the lexical items (tokens) they are aligned todirectly, and abstract nodes, which represent the se-mantic contribution of grammatical constructions.
This distinction is annotated explicitly in DMRSand EDS (which are based on the same under-lying annotations), but we extend it to the otherframeworks based on the alignments and frame-work properties.
2.1 System Architecture
The system consists of a web-based front-end user-interface through which users upload a file contain-ing a bank of MRP graphs in the Uniform GraphInterchange Format, which is serialised in JSONLines format (Oepen et al., 2019). The file isthen uploaded to and parsed by the back-end ofthe RepGraph application to create a transitorystructure of the dataset. The user can then pro-ceed to the main screen of the application (Fig.2). The main libraries used are React3, visx4, andmaterial-UI5 for the front end, and Spring-Boot6
for the backend. The source code can be found athttps://github.com/RepGraph/RepGraph.
3 Graph Visualisation
Our application provides three distinct visualisa-tions of meaning representation graphs in all thesupported frameworks, providing users with multi-ple perspectives of the same semantic information.
As can be seen in Figures 1–3, the visualisationsuse colour and annotations to represent the variouselements of the graphs — for example, surface andabstract nodes are differentiated and shown in dif-ferent colours. An expandable legend is provided
2The anchoring is annotated at character level in the MRPdata. We tokenise the input in a manner that is consistent withthe given annotations.
Figure 2: The RepGraph Main Page with an EDS graph visualised in the Tree-like format. The main elements onthe page are captioned.
to explain the different colours (see Fig. 2).The user can modify the default colours and spac-
ing of the visualisations from the settings in thepanel on the left of the main page. The placementof the edges and edge labels can also be manip-ulated easily by dragging them with the mouse.The examples in the figures, as well as the demodata provided in application, come from the sampleannotations of Wall Street Journal (WSJ) corpussentences provided for MRP 2020.7
We next discuss each of the visualisations, andin particular how node placement is calculated foreach of them (§3.1 - 3.3). Then we discuss howedge placement is determined across all the visu-alisations to minimise collisions (§3.4), and nor-malisations to make the anchoring more consistentacross frameworks (§3.5).
3.1 Hierarchical
The Hierarchical layout (Fig. 1) focuses on show-ing the natural hierarchy that occurs between theanchors of the semantic nodes in the graph. Nodeswith larger spans envelop nodes whose spans theysubsume in their range of tokens. Horizontal bars(brackets) are placed below the semantic nodes inthe graph to represent the alignment with their spanof tokens in the input sentence. Tokens are dis-played in the bottom row. For example, in Fig. 1,the top-most node with the label udef_q has abracket covering the token span “Champagne anddessert” below it.
7http://svn.nlpl.eu/mrp/2020/public/sample.tgz
3.2 Tree-like
The Tree-like visualisation (Fig. 2), places em-phasis on each node’s number of descendants. Atree-like structure is created by having nodes witha larger number of descendants positioned abovenodes with a lower number of descendants. The an-choring of nodes that are positioned at the bottomof the graph (directly above the tokens) is repre-sented with a dotted line to the tokens(s) they span.These nodes are positioned vertically in line withthe start of their token span. The alignment of othernodes is not indicated visually. The node placementis computed by running a topological sort on eachnode in order to get a list of its descendants, whichis used to determine its level in the tree.
3.3 Flat
The flat visualisation (Fig. 3), orders nodes hori-zontally based on the first token alignment. Ties be-tween nodes that are aligned to the same first tokenare broken by prioritising nodes with smaller spanlengths. Edges are curved above and below thelinear plane according to whether they are directedright or left. No tokens are shown in this format.The span(s) of each node and its corresponding textphrase(s) are present inside a tool-tip that appearswhen they are hovered upon. This visualisationis similar to the way that dependency graphs aretraditionally visualised (albeit using graph nodesrather than tokens as basic elements).
Figure 3: Example of the flat visualisation of the PTGgraph for the sentence “Champagne and dessert fol-lowed.”
3.4 Edge Handling
Edges are created as Quadratic Bézier Curves de-fined through 3 points (P0, P1 and P2):
B(t) = (1−t)2P0+2(1−t)tP1+t2P2, 0 ≤ t ≤ 1
where P0 is the edge’s source node position and P2
is the edge’s target node position.We designed a set of edge handling heuristics
that determine the value of P1 for each graph edge,with the aim to minimise overlap between edges,edge labels and nodes. The heuristics take into ac-count attributes of edges and their source and targetnodes that include their distance, whether they arein the same row or column, and the relative posi-tions of other nodes. Users can also drag edges andedge labels with their mouse pointer, manipulatingthe value of P1.
3.5 Framework Normalisation
In order to visualise and analyse graphs across allframeworks in a standardised way, we performedframework-specific normalisations that capture la-tent information whilst ensuring accuracy of thegraph visualisations.
We extract DMRS annotations from the Red-woods treebank8 using PyDelphin.9 Tokenisationand some additional normalisations were derivedfrom the Redwoods syntactic layer.
For the other frameworks we use the data pro-vided by the MRP 2020 shared task (Oepen et al.,2020). Input sentences were tokenised with Stan-ford CoreNLP (Manning et al., 2014).10
In EDS, nodes with the CARG property (namedentities) are treated as surface rather than abstractnodes, and displayed as such in the visualisation.
PTG graphs may contain nodes without tokenalignments. We align those nodes to their leftmostaligned children, or if that is not possible, theirleftmost aligned parents. These induced alignmentsare use only to determine the node layout, and arenot displayed visually.
UCCA nodes do not have explicit labels. Wetreat nodes with spans as surface nodes; their labelsare implicitly derived from their spans, using thecorresponding token strings as labels. Abstractnodes (nodes without annotated spans) implicitlyderive their spans as the union of the spans of theirdescendent nodes. We do not assign labels to them.
AMR annotations do not include node spans. Asour hierarchical and tree-like visualisations requirenode spans, an AMR aligner is used to obtain align-ments. The rule-based JAMR Aligner11 (Flaniganet al., 2014) is integrated into our tool to processonly AMR graphs that are uploaded without align-ments - this does not impact the language supportof the other frameworks. All aligned nodes aredesignated as surface nodes. Nodes that are leftunaligned by the aligner are aligned in the sameway as PTG to determine the layout.
Multiple Spans Graphs in the UCCA and PTGframeworks have nodes that are anchored to multi-ple, potentially discontiguous token spans. Thesenodes are handled as follows: If the multiple spansare contiguous, node are aligned to the union oftheir spans. Otherwise, dummy nodes are createdfor each additional non-adjacent span; the regu-lar node corresponds to the left-most span. Eachdummy node is placed above its aligned token spanand distinguished visually. It takes the label of itsoriginal node, with additional information to iden-tify which span it refers to and from which node itwas created. When the user hovers over the originalnode, the dummy nodes are highlighted.
4 Graph Analysis Features
4.1 Subgraph Display
The subgraph display tool allows users to examinespecific subgraphs of the currently visualised graph.Users can choose between two types of subgraphs:adjacent and descendent. Upon selection of a sub-graph type, the user selects a node on the graph
Figure 4: The Graph Comparison Tool showing which nodes and edges in the two UCCA graphs are similar (green)or dissimilar (red).
visualisation and clicks the display button to viewthe resulting subgraph.
Adjacent subgraphs are created from the nodesdirectly adjacent to the selected node. This allowsusers to focus in on the immediate neighbourhoodof a node.
Descendent subgraphs are created from the se-lected node’s descendants in the graph. This allowsusers to focus on a single node and the propagationof its descendants through the graph.
4.2 Subgraph Search
We provide two ways for users to find graphs match-ing specific patterns in large datasets:
Search by Node Label Set The user can enterone or more node labels as a query to search for allgraphs in the dataset which contain all of these nodelabels. This gives the user the ability to quickly findsentences and graphs containing nodes of interestand see how they are used in different graphs.
Search by Subgraph Pattern The user can alsovisually select a connected subgraph of the currentgraph by clicking to select the desired nodes andedges. After selection, the search will return allother graphs that contain this subgraph, and theuser can visualise any of these graphs.
4.3 Graph Comparison
The graph comparison interface, seen in Fig. 4,is broken up into two vertically-separated side-by-side panels that enable direct comparison betweentwo graphs. The user selects a graph from thedataset for each side, toggles the comparison set-
tings, and clicks compare. The similarities anddifferences between the graphs are displayed visu-ally through colouring the nodes and edges. If thesame node label appears in both graphs, all nodeswith that label are deemed as similar, even whenthe number of nodes with that label differ betweenthe two graphs.12 The tool also provides the optionto only compare surface or abstract nodes. Twomodes of graph comparison are provided:
Standard Comparison Nodes are compared us-ing only their labels, whereas edges are matchedif both the edge labels and the labels of the nodesthey are incident to are equal.
Strict Comparison This mode additionally re-quires that the concatenation of the token spanscorresponding to each node has to be the same (asstrings) in order for the nodes to match.
4.4 Graph PropertiesA number of graph properties can be evaluated onthe current graph.
Planarity An important property of a semanticgraph is whether it is planar, also referred to asnon-crossing (Kuhlmann and Oepen, 2016). To de-termine planarity, the nodes are ordered in a similarmanner as for the flat visualisation (§3.3), exceptthat nodes with the exact same span are placed be-low one another in the same horizontal position,and have their edges transferred to the node rep-resenting their position in the linear ordering. All
12This contrasts to Smatch (Cai and Knight, 2013) whichfinds a 1-to-1 alignment between the nodes to measure graphoverlap, while for visualisation purposes we rather show allpossible correspondences.
84
edges are represented as arcs above the nodes.Nodes with no token alignment are excluded
from the construction as it is assumed they can beplaced anywhere to avoid causing conflicts. Edgesbetween nodes in the same linear ordering posi-tion are also excluded as they have no impact onpotential crossing edges.
The check for planarity, once the linear order-ing has been established, follows the definition ofplanarity outlined in (Gómez-Rodríguez and Nivre,2010). After running the planarity test, the modi-fied graph can be visualised. Crossing edges arehighlighted in red.
Graph Connectivity A graph is connected if anundirected path exists between every pair of nodes(Kuhlmann and Oepen, 2016). AMR graphs arealways connected, but this is not guaranteed forthe other frameworks. Graph connectivity is deter-mined using a variation of Breadth First Search.
Longest Directed and Undirected Path(s) Ourtool also has functionality to find the longest di-rected or undirected paths in any graph. Thelongest path is highlighted on the graph. If there aremultiple longest paths, the user can cycle throughthem. The paths are found using derivations ofTopological Sort and Breadth First Search. Beforethe longest path is derived, the graph is checked forcycles using a derivation of Depth First Search.
4.5 Dataset Statistics
The dataset statistics tool can produce a numberof global statistics on the current dataset, whichmay be useful for comparing datasets. The statis-tics included are a subset of the ones used byKuhlmann and Oepen (2016) to compare semanticgraph frameworks, and include average number ofnodes, average span of node, and percentage ofcyclic graphs, amongst others.
5 Related Work
The plethora of tools and visualisation options cur-rently available are fragmented and often built with-out semantic graphs in mind. In the context ofthe MRP Shared Tasks, mtool13 was developedto provide various functionality including formatconversion, graph analysis, and evaluation. How-ever, mtool is a command-line only tool and doesnot provide a simple and intuitive way to interactwith large datasets of semantic graphs. In termsof graph visualisation, mtool supports the output
13https://github.com/cfmrp/mtool
(a) DMRS graph visualised with delphin-viz
(b) UCCA graph visualised with mtool
Figure 5: Visualisations provided by previous work
of graphs in DOT language that is compiled withGraphviz.14 Unfortunately, since Graphviz is ageneral-purpose visualization tool, the visualisa-tions produced are often complex and inadequateat capturing framework-specific assumptions. Asan example, the UCCA visualisation in Fig. 5blacks lexical information.
A number of other framework-specific visualisa-tion tools exist, including for DMRS15 and AMR(Saphra and Lopez, 2015), but are limited in termsof their functionality and framework support. Thedelphin-viz DMRS visualisation (Fig. 5a) is simi-lar to our flat visualisation, but lacks the structureconveyed by our other visualisations.
6 Conclusion
The proliferation of graph-based and span-basedmeaning representations has created the need fora general platform for analysing these represen-tations. RepGraph provides such a unified plat-form for the visualisation and analysis of semanticgraphs in multiple frameworks. The tool allowsusers to explore and analyse these representations
in an interactive and intuitive manner that is not pos-sible with existing tools. Our work aims to makemeaning representations accessible to a broaderaudience than researchers invested in a particularframework, while providing new insights into theserepresentations through novel visualisations.
For future work, we plan to include support foradditional semantic frameworks, including span-based representations such as semantic role la-belling. We also intend to integrate parsers forall supported frameworks. This will provide signif-icant utility by allowing users to upload text files orenter sentences directly, and use RepGraph straightaway rather than having to run a parser separately.
ReferencesOmri Abend and Ari Rappoport. 2013. Universal Con-
ceptual Cognitive Annotation (UCCA). In Proceed-ings of the 51st Annual Meeting of the Associationfor Computational Linguistics (Volume 1: Long Pa-pers), pages 228–238, Sofia, Bulgaria. Associationfor Computational Linguistics.
Laura Banarescu, Claire Bonial, Shu Cai, MadalinaGeorgescu, Kira Griffitt, Ulf Hermjakob, KevinKnight, Philipp Koehn, Martha Palmer, and NathanSchneider. 2013. Abstract Meaning Representationfor sembanking. In Proceedings of the 7th Linguis-tic Annotation Workshop and Interoperability withDiscourse, pages 178–186, Sofia, Bulgaria. Associa-tion for Computational Linguistics.
Jan Buys and Phil Blunsom. 2017. Robust incremen-tal neural semantic graph parsing. In Proceedingsof the 55th Annual Meeting of the Association forComputational Linguistics (Volume 1: Long Papers),pages 1215–1226, Vancouver, Canada. Associationfor Computational Linguistics.
Shu Cai and Kevin Knight. 2013. Smatch: an evalua-tion metric for semantic feature structures. In Pro-ceedings of the 51st Annual Meeting of the Associa-tion for Computational Linguistics (Volume 2: ShortPapers), pages 748–752, Sofia, Bulgaria. Associa-tion for Computational Linguistics.
Ann Copestake. 2009. Invited Talk: slacker seman-tics: Why superficiality, dependency and avoidanceof commitment can be the right way to go. InProceedings of the 12th Conference of the Euro-pean Chapter of the ACL (EACL 2009), pages 1–9,Athens, Greece. Association for Computational Lin-guistics.
Jeffrey Flanigan, Sam Thomson, Jaime Carbonell,Chris Dyer, and Noah A. Smith. 2014. A discrim-inative graph-based parser for the Abstract Mean-ing Representation. In Proceedings of the 52nd An-nual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers), pages 1426–1436, Baltimore, Maryland. Association for Compu-tational Linguistics.
Carlos Gómez-Rodríguez and Joakim Nivre. 2010.A transition-based parser for 2-planar dependencystructures. In Proceedings of the 48th Annual Meet-ing of the Association for Computational Linguis-tics, pages 1492–1501, Uppsala, Sweden. Associa-tion for Computational Linguistics.
Jan Hajic, Eva Hajicová, Jarmila Panevová, Petr Sgall,Ondrej Bojar, Silvie Cinková, Eva Fucíková, MarieMikulová, Petr Pajas, Jan Popelka, Jirí Semecký,Jana Šindlerová, Jan Štepánek, Josef Toman, ZdenkaUrešová, and Zdenek Žabokrtský. 2012. Announc-ing Prague Czech-English Dependency Treebank2.0. In Proceedings of the Eighth InternationalConference on Language Resources and Evaluation(LREC’12), pages 3153–3160, Istanbul, Turkey. Eu-ropean Language Resources Association (ELRA).
Daniel Hershcovich, Omri Abend, and Ari Rappoport.2017. A transition-based directed acyclic graphparser for UCCA. In Proceedings of the 55th An-nual Meeting of the Association for ComputationalLinguistics (Volume 1: Long Papers), pages 1127–1138, Vancouver, Canada. Association for Computa-tional Linguistics.
Zhengbao Jiang, Wei Xu, Jun Araki, and Graham Neu-big. 2020. Generalizing natural language analysisthrough span-relation representations. In Proceed-ings of the 58th Annual Meeting of the Associationfor Computational Linguistics, pages 2120–2133,Online. Association for Computational Linguistics.
Alexander Koller, Stephan Oepen, and Weiwei Sun.2019. Graph-based meaning representations: De-sign and processing. In Proceedings of the 57th An-nual Meeting of the Association for ComputationalLinguistics: Tutorial Abstracts, pages 6–11, Flo-rence, Italy. Association for Computational Linguis-tics.
Marco Kuhlmann and Stephan Oepen. 2016. Towardsa catalogue of linguistic graph Banks. Computa-tional Linguistics, 42(4):819–827.
Christopher Manning, Mihai Surdeanu, John Bauer,Jenny Finkel, Steven Bethard, and David McClosky.2014. The Stanford CoreNLP natural language pro-cessing toolkit. In Proceedings of 52nd AnnualMeeting of the Association for Computational Lin-guistics: System Demonstrations, pages 55–60, Bal-timore, Maryland. Association for ComputationalLinguistics.
Stephan Oepen, Omri Abend, Lasha Abzianidze, Jo-han Bos, Jan Hajic, Daniel Hershcovich, Bin Li,Tim O’Gorman, Nianwen Xue, and Daniel Zeman.2020. MRP 2020: The second shared task oncross-framework and cross-lingual meaning repre-sentation parsing. In Proceedings of the CoNLL
2020 Shared Task: Cross-Framework Meaning Rep-resentation Parsing, pages 1–22, Online. Associa-tion for Computational Linguistics.
Stephan Oepen, Omri Abend, Jan Hajic, Daniel Her-shcovich, Marco Kuhlmann, Tim O’Gorman, Nian-wen Xue, Jayeol Chun, Milan Straka, and ZdenkaUresova. 2019. MRP 2019: Cross-framework mean-ing representation parsing. In Proceedings of theShared Task on Cross-Framework Meaning Repre-sentation Parsing at the 2019 Conference on NaturalLanguage Learning, pages 1–27, Hong Kong. Asso-ciation for Computational Linguistics.
Stephan Oepen and Jan Tore Lønning. 2006.Discriminant-based MRS banking. In Proceed-ings of the Fifth International Conference onLanguage Resources and Evaluation (LREC’06),Genoa, Italy. European Language ResourcesAssociation (ELRA).
Naomi Saphra and Adam Lopez. 2015. AMRICA: anAMR inspector for cross-language alignments. InProceedings of the 2015 Conference of the NorthAmerican Chapter of the Association for Computa-tional Linguistics: Demonstrations, pages 36–40,Denver, Colorado. Association for ComputationalLinguistics.
Linfeng Song, Yue Zhang, Zhiguo Wang, and DanielGildea. 2018. A graph-to-sequence model for AMR-to-text generation. In Proceedings of the 56th An-nual Meeting of the Association for ComputationalLinguistics (Volume 1: Long Papers), pages 1616–1626, Melbourne, Australia. Association for Compu-tational Linguistics.
Sheng Zhang, Xutai Ma, Kevin Duh, and BenjaminVan Durme. 2019. AMR parsing as sequence-to-graph transduction. In Proceedings of the 57th An-nual Meeting of the Association for ComputationalLinguistics, pages 80–94, Florence, Italy. Associa-tion for Computational Linguistics.
