Licenses of digital resources describe rights and duties for users. If the licenses are expressed in natural language, as it is frequently the case, it is hard to reason and verify the license compatibility to specific uses. We propose an automatic end-to-end workflow for extracting Open Digital Rights Language (ODRL) representations from textual license documents. This process uses AMR semantic representations as an intermediate; it adapts a tool that performs a semantic transduction analysis, using formal rules. This work focuses on deontic modalities expressing the permissions and obligations of the user. We provide a proof of concept and discuss experiments.
most are sentences that describe the permissions and obligations of the user. Such sentences typically contain One may wish to attach a license to a work (e.g. a pro- a modal verb such as "may" or "must", and an action that gram, a picture or a dataset) before publishing it on the is an object of this modal verb, such as "distribute" or Web. A license expresses in natural language (NL) the "give credit". Additionally, the action of the sentence rights and duties that the users of the work must com- refers to a subject (e.g. the user) and a target (e.g. the ply with. However, although some licenses such as the work ofered under the terms of the license). These parts licenses provided by Creative Commons1 are also avail- of information all correspond to specific ODRL classes. able in a machine readable format, most licenses are only For instance, in the sentence "You may reproduce the provided as human readable texts, which are written in Work", the modal verb "may" expresses a permission, NL. Machine readable licenses enable a much easier au- and controls the action "reproduce". The whole sentence tomatic extraction of the rights expressed by the licenses, means that the action of reproduction appears among which is useful for tasks such as license synthesis [1], the user’s permissions. or compatibility and compliance inference [2], [1]. Per- Thus, our approach aims to extract ODRL representaforming such tasks using only NL license texts would be tions for specific content, with a focus on deontic modalimuch more arduous and ambiguous. ties expressing authorized actions and unauthorized ones. Minimizing errors in a fully automated workflow is also an important issue.
aims at automatically computing machine-readable versions of NL license texts. More specifically, our target Contribution. Our main contribution is the proposal representation format is the Open Digital Rights Lan- of an automatic end-to-end workflow for extracting guage (ODRL [3]), an RDF vocabulary meant to express ODRL representations from textual license documents. rights and duties. In order to realize this translation, This process integrates a pre-existing analysis tool [5], Abstract Meaning Representation (AMR), a graph-based implementing the principles of semantic transduction semantic representation proposed by [4], is used as an analysis [6]. This tool automatically produces ontologies intermediary representation. from semantic graphs, using formal rules. We adapted it
The information we aim at extracting from the license by developing specific rules to deal with deontic modalitexts is rather specific. Indeed, what interests us the ties and to capture specific semantic content corresponding to the target representation. In addition, an ODRL Proceedings of the Sixth Workshop on Automated Semantic Analysis of graph generation step has been added to the workflow. Information in Legal Text (ASAIL 2023), June 23, 2023, Braga, Portugal. At this stage, we propose a proof of concept evaluated a$urMeliaelno..Rlaemveelr@ceirriies@a.ftret(rMas.-Rliebvreel.)f;r (A. Lamercerie); on a dataset composed of a hundred typical sentences Annie.Foret@irisa.fr (A. Foret); Zoltan.Miklos@irisa.fr (Z. Miklos) with a moderate complexity. Our experimentation, which © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License is intended to be preliminary, was notably oriented by 1 hCPWrEooUrctkReshtdoinpgpssIhStpN:/c1e:6u1r3-w/-0s.o7r3g/crACettaEritbUiuvtRioencW4o.0moInrtmekrnsoahtnioosnpa.ol(PCrgCro/BYce4.0e).dings (CEUR-WS.org) following ODRL implementation best practices presented
in the report [7]. Meaning Representation (AMR) introduced by [12], in the form proposed by [4].
Related works are explored in Section 2, while Sec- AMR is a readable expressive form of graph-based setion 3 introduces the notions and knowledge on which mantic representation at the sentence level, for which we rely in this paper. The workflow and the implemented datasets and (semantic) parsing tools have been develmethodology are detailed in Section 4. The experimenta- opped [13]. This scheme abstracts away from syntax, as tion carried out is finally presented in Section 5. a single AMR may correspond to several sentences (paraphrases). AMR relies heavily on the PropBank 2 semantic roles inventory that is verb-oriented.
ifgure 1 for sentence (1) "You may distribute the work". Every AMR has a unique root. It has variables (p, etc.), events and concepts (such as permit-01, determine-01) and roles (such as ARG0, ARG1) that label the edges. In a node, a slash indicates that the variable on the left denotes an instance of the concept on the right. The edge relations follow verb frame descriptions in PropBank, in the case of "determine-01" : ARG0 denotes "distributor", ARG1 denotes "thing distributed".
is quite a specific task that has not been extensively addressed yet. A state of the art for this task is provided by Cabrio et al. [8], whose goal is similar to ours, but whose method difers from ours.
In order to extract RDF specifications from NL texts, Cabrio et al. [8] first pre-process the input license by tokenizing, lemmatizing and processing the Parts of Speech of the text, then use classifiers based on Support Vector Machines (SVM) to generate a CC REL [9] or an ODRL [3] graph. CC REL and ODRL are two Resource Description Framework (RDF) vocabularies dedicated to digital rights specification, and particularly license specification.
Cabrio et al. [8] evaluated their system on a dataset of 37 licenses that they annotated in RDF. The classifiers achieved an overall precision of approximately 0.77 and a recall of 0.43, varying depending on the classified action.
These results are not totally satisfying and show that the framework of Cabrio et al. [8] must be considered as Figure 1: AMR Graph of sentence (1) "You may distribute the a first step towards automatic NL license analysis rather work" than a definitive solution. This article [ 8] shows that tackling this challenge with the use of SVM is not trivial, and we hope our approach will obtain better results. Negation is expressed in AMR with a polarity relation.
Havur et al. [1] present DALICC, which is both a li- The relation is between the concept marked as negated cense library that indexes several license texts along with and the constant "-". For this example (2) "you are not their ODRL representations, and a system to reason on allowed to distribute the work", we get the graph (in these RDF graphs and for instance compose customized PENMAN notation 3): licenses. Moreau et al. [2] propose CaLi, a model that compares and partially orders licenses based on compati- ( v:2p o/l aarliltoyw −−01 bility and compliance relations between the licenses. To : ARG1 ( v3 / d i s t r i b u t e −01 do so CaLi arranges formal representations of licenses : ARG1 ( v4 / work − 0 1 ) ) based on ODRL in lattices over which CaLi reasons. Vu : ARG1 ( v1 / you ) ) et al. [10] introduce JCivilCode, an AMR dataset that is specific to legal texts, and they evaluate several AMR parsers with this dataset.
3.1. Semantic Representations, AMR
several annotation languages or formats to represent 2see the PropBank sites: https://verbs.colorado.edu/~mpalmer/ semantic information [11]. In this study we use Abstract projects/ace.html and https://propbank.github.io/ 3described for exemple here https://penman.readthedocs.io/en/ latest/notation.html
here is that for texts in the domain of digital licenses, AMR forms are good intermediate representations, from which we can derive the deontic information to be expressed in languages such as ODRL (The Open Digital Rights Language). Obvioulsy at a practical level, a given AMR parser may output errors (a wrong AMR). But im- representations. This workflow is composed of several portantly, at the level of format and guidelines, the way steps (figure 3): (1) document splitting and selection of AMR highlights verbs (with their arguments) and the way key sentences, (2) conversion of NL statements into seAMR expresses and attach modals and negation seems mantic representations (AMR graph), (3) RDF serializaappropriate to extract ODRL-like information. While tion of the resulting representations, (4) transduction some general limitations have been noted: AMR does parsing to extract the semantic content, and (5) generanot explicitly deal with universal quantification, and has tion of the ODRL representations. no tense (future, etc.), we think this should not create problems for the task and texts considered in this study. 3.2. Targeted ontology
Open Digital Rights Language (ODRL). ODRL is an RDF vocabulary dedicated to the representation of statements describing rights and duties, and license texts are made of statements of this kind. An ODRL Policy is composed of one or many Rules, which are either a Permission (what one may do), a Duty (what one must do) or a Prohibition (what one is prohibited to do). These rules are linked to one or many Action which describe the terms of use of an Asset - e.g. software or pictures.
For instance, figure 2 shows an ODRL graph (serialized in Turtle syntax) that expresses the information contained in the sentence (1) "You may distribute the Work" (where "You" and "the Work" are terms that are defined within the license text) : : l i c e n s e a o d r l : P o l i c y ; o d r l : p e r m i s s i o n [ a o d r l : P e r m i s s i o n ; o d r l : t a r g e t " Work " ; o d r l : a s s i g n e e " You " ; o d r l : a c t i o n c c : D i s t r i b u t e
cessing. It aims at highlighting diferent sets of sentences from the whole license document, and to select in particular the sentences that correspond to definitions of terminology or expressions of legal rules.
Indeed the only sentences that we want to consider are the ones that describe rights and duties. The first approach would be to consider all the sentences and only keep at the end the ones that describe rights and duties, but this method may result in performance issues, since the semantic parsing takes quite a long time to process. Another approach would be to do a first preprocessing step before the semantic parsing, which filters out some "useless" sentences. This way less sentences are to be parsed, and the remaining "useless" sentences can be ignored in a later step of the analysis.
Moreover, most license texts contain a section (usually the first section of the document) that defines some terms used in the text. Here is an example of a definition of "You" (from [14]) : "You means the individual or entity exercising the Licensed Rights under this Public License. Your has a corresponding meaning.". Being able to automatically recognize this section and analyze its definitions would make the extraction of the deontic sentences easier and less ambiguous, because the deontic sentences make use of the defined words.
This preprocessing step is currently purely hypothetic: so far the only computation done on a text before the semantic parsing step is the sentence splitting. As a consequence, in this article we only consider isolated sentences that express rights or duties. 4.2. Semantic Graph Construction
struct the corresponding AMR graphs. This step is done by amrBatch 4, which makes use of two modules: • AMRLib 5 is a module that among others parses
NL sentences to create AMR graphs. This parser may rely on diferent pretrained models, and Figure 4: AMR Graph with semantic nets we chose the parse_xfm_bart_large model, which reached a SMATCH score of 83.7 in 2022. • AMR-LD 6 is a utilitary module that is used by
AMRBatch to convert AMR graphs from the PEN- satisfy certain typing or relation criteria with a formula. MAN format provided by AMRLib into an AMR By linking such a formula to a constructive method, a graph. This RDF version of the AMR is a direct mechanism makes possible to generate new nets. These translation, and is still far from the ODRL graph rules are called transduction rules, and can be structured that we aim to reach. using a schema, introducing recursion if necessary.
4.3. Strategies on Semantic Graph
Patterns Semantic Transduction Analysis aims at extracting the semantic content to generate the desired representations. This approach, similar to graph transformation techniques, was initiated in the thesis of A. Lamercerie [6]. We take up here the general idea, which we have adapted to an application on licenses in order to extract elements of digital rights.
The notion of Semantic Net is introduced to support its implementation. A semantic net is an abstract object that covers graph nodes. It is a way to "capture" a meaning interpretation of nodes in relation. Each net can be typed and associated with diferent useful semantic data. For example figure 4 shows the nets A and B, which allow to capture respectively a modality and a property.
Transduction Mechanics defines the methodology used to direct the analysis. It aims to bring out semantic nets until one or more nets containing all the data needed to generate the expected ODRL statements are obtained. For its implementation, we define a set of net types, possibly structured as in figure 5. Using this typing, it becomes possible to check if a net has a given type with a simple formula as classNet(x). Similarly, two nets can be related, and we can check this using a formula as arg1(x, y). We can thus check if one or more nets
The Analysis Procedure is initialized by analyzing each node of a graph, and by associating an atomic net (a net that covers a single node) to it. The rest of the procedure consists in the analysis of the generated nets, and in the application of diferent rules permitting to build new nets by composition of diferent nets. The analysis of the nets A and B of the example of the previous figure (figure 4) reveals the existence of an action (corresponding to the property associated to B). A new net is created, E on the figure 6, covering the node associated with B, and extended by analyzing the relations starting from B (arg0 which points to C, and arg1 which points to D). Finally, A can be combined with this new net to produce another net, covering all the nodes and highlighting the existence of a digital rights rule.
The implementation uses composition rules as described in [6] or [5]. For this work on license rights, rules have been revised or specifically developed. These
recognize in a sentence, because there are a lot of different actions, and actions may be composed of a lot of AMR concept nodes that are not necessarily close together in the graph. Moreover, some actions have very similar meanings and are hard to diferentiate in the AMR and even in the NL sentence. For instance [8], cc:shareAlike is textually described as "Redistributions must reproduce the above copyright notice" and Duty:attachPolicy as "Redistributions must retain the copyright notice".
sentence can express a lot of information (for instance several actions) through coordinating conjunctions.
the hypotheses above and we show its AMR and ODRL graph. The sentence (3) is the following: "You may not reproduce the Work". In this sentence, the modality is a prohibition, the assignee is "You", the target is "the Work" and the action is "reproduce".
Here is an AMR translation for this sentence (in PENMAN notation): have in particular targeted the extraction of modalities and actions, by trying to take into account the variability of the structures to be highlighted, and the relationships between these structures.
Finally, the last step exploits the result of the semantic transduction analysis to generate the expected ODRL representations. 4.4. Hypotheses on sentence structures
a certain structure, which corresponds to the semantic graph patterns that are recognized by TENET 7. This structure is also determined by the ODRL ontology classes and properties. Here are the features that we expect to encounter in sentences that express digital rights.
pear: permission, obligation and prohibition. The simplest way to express these operators in NL is through the usage of "may" (which is usually translated to the AMR concept permit-01) and "must" (translated to And here is an ODRL translation for this sentence (in obligation-01), although deontic modality can be ex- TURTLE syntax): pressed through much more complex structures in license texts (e.g. "the Licensor hereby grants You a worldwide, @@ pp rr ee ff ii xx cocd:r l <: h<t thpt t:p/ /: /c /r weawtwi v. wec3o. mormg o/ nnss /. oordgr /l n/2s /# >>.. royalty-free, non-sublicensable, non-exclusive, irrevocable license to exercize the Licensed Rights in the Licensed " L i c e n s e " a o d r l : P o l i c y ; Material to" 8 expresses a permission). In addition, nega- o d r l : p r o h i b i t i o n [ tion can be applied to the operator, which changes its o d r l : t a r g e t " Work " ; meaning. For instance "You are not prohibited to <ac- oo dd rr ll :: aa sc st ii og nn e ec c ": YRoeup"r o ;d u c t i o n tion>" expresses a permission, while "You are prohibited ] . not to <action>" expresses an obligation. # : : s n t You may n o t r e p r o d u c e t h e Work . ( p / p e r m i t −01 : p o l a r i t y − : ARG1 ( r / r e p r o d u c e −01 : ARG0 ( y / you ) : ARG1 (w / work − 1 2 ) ) )
Entities Some entities can appear in the sentence, 4.5. Sources of error and can be actors (odrl:assignee) or targets (odrl:target) of the action. The main entities are often defined in the definitions section of the text, and they often start with a capital letter.
the lack of precision of AMRLib parsing tool for some sentences that we consider. For instance, let us consider the previous sentence (3) again: "You may not reproduce the Work".
The AMR described above is manually annotated, and the AMR generated by AMRLib tool with this sentence as input is actually quite diferent: # : : s n t You may n o t r e p r o d u c e t h e Work . ( p / p o s s i b l e −01 : ARG1 ( r / r e p r o d u c e −01 : p o l a r i t y − : ARG0 ( y / you ) : ARG1 (w / work − 0 1 ) ) )
issue here is the fact that the negation is misplaced in the AMR. This AMR means "You are authorized not to reproduce the work": the negation is placed on the action, which changes the meaning of the modality.
This remark does not point to a weakness of the AMR format but a weakness of some current parsers, that we hope will improve in the future.
5.1. Dataset Our dataset is made of a hundred simple sentences expressing deontic policies, inspired by ODRL best practices examples [7]. About 20% of these sentences have been manually crafted in order to convey diverse and interesting linguistic phenomena. The other are automatically crafted using a simple grammar. These sentences must contain a modality (which may be negated), one or several actions that may be expressed in several ways. Actions can be simple or composite, and associated with one or two targets and an assignment. Moreover, 10 additional negative entries are added to the dataset. These 10 sentences are grammatically correct but do not express any deontic policy, and should not be recognized as deontic policies by our system. Indeed, many license text sentences are not relevant for the ODRL extraction and may be ignored during the analysis. 5.2. Evaluating experiments
dataset described previously. The results are given by the ifgure 7. Precision (P) and recall (R) were evaluated for the modality and action extraction task. This dataset is used to show that our approach can work on simple cases. In this context, classification of modalities is decent but not excellent, partly since some negated modalities are incorrectly parsed by AMRLib. However, we have a very good precision and recall on the action classification.
Modalities
R 0.640
Actions
R 0.810
A classic F-score evaluation was chosen, because we were primarily interested in evaluating the system’s ability to extract the expected semantic content. The results obtained confirm the direction followed. That said, our workflow also covers a formalization task, which it would be more relevant to evaluate with a semantic structure evaluation metric such as Smatch [15]. This measure will be particularly interesting to refine our evaluations in future work. The goal remains the processing of real documents, with complex sentences. Additional experimentation is therefore targeted on broader data, such as the dataset of Cabrio et al [8].
In this paper, we present a first step towards a workflow that automatically extracts ODRL representations from natural text licenses, with AMR as an intermediate representation. A formal representation of the rights could enable to automatically verify the licence compability in specific situations. Our first experiment provided a proof of concept, with suficient results on a simple dataset. In particular, we obtained good performance figures for the extraction of modalities, which express permissions, requirements and prohibitions, and actions, which are what the user can, must or is forbidden to do. Some linguistic phenomena are taken into account, such as negations and coordinating conjunctions. Another outcome of the process is that it provides an indirect evaluation of AMR parsers.
Thus, although our full system is far from being complete, our method is achieving encouraging results so far and we plan to continue working on it. First, the efective implementation of a pre-processing phase (Figure 3) would improve runtime performance and help recognize entities and actions. Also, our rule system is still incomplete and more AMR patterns need to be covered. Some evolutions and alternatives can also be considered, such as the use of another pivot representation as an alternative of AMR or as a complementary information.
dataset, Artif. Intell. Law 30 (2022) 221–243. URL: https://doi.org/10.1007/s10506-021-09292-6. doi:10. [1] G. Havur, S. Steyskal, O. Panasiuk, A. Fensel, 1007/s10506-021-09292-6.
V. Mireles, T. Pellegrini, T. Thurner, A. Polleres, [11] O. Abend, A. Rappoport, The state of the art in seS. Kirrane, Automatic license compatibility check- mantic representation, in: Proceedings of the 55th ing, in: M. Alam, R. Usbeck, T. Pellegrini, H. Sack, Annual Meeting of the Association for ComputaY. Sure-Vetter (Eds.), Proceedings of the Posters tional Linguistics (Volume 1: Long Papers), Assoand Demo Track of the 15th International Confer- ciation for Computational Linguistics, Vancouver, ence on Semantic Systems co-located with 15th Canada, 2017, pp. 77–89. URL: https://aclanthology. International Conference on Semantic Systems (SE- org/P17-1008. doi:10.18653/v1/P17-1008. MANTiCS 2019), Karlsruhe, Germany, September [12] I. Langkilde, K. Knight, Generation that exploits 9th - to - 12th, 2019, volume 2451 of CEUR Work- corpus-based statistical knowledge, in: 36th Anshop Proceedings, CEUR-WS.org, 2019. URL: https: nual Meeting of the Association for Computa//ceur-ws.org/Vol-2451/paper-13.pdf . tional Linguistics and 17th International Confer[2] B. Moreau, P. Serrano-Alvarado, M. Perrin, ence on Computational Linguistics, Volume 1, AsE. Desmontils, Modelling the compatibility of li- sociation for Computational Linguistics, Montreal, censes, in: P. Hitzler, M. Fernández, K. Janowicz, Quebec, Canada, 1998, pp. 704–710. URL: https:// A. Zaveri, A. J. Gray, V. Lopez, A. Haller, K. Hammar aclanthology.org/P98-1116. doi:10.3115/980845. (Eds.), The Semantic Web, Springer International 980963.
Publishing, Cham, 2019, pp. 255–269. [13] J. May, J. Priyadarshi, SemEval-2017 task 9: Ab[3] W3C Recommendation, ODRL Information stract Meaning Representation parsing and generModel 2.2, 2018. URL: https://www.w3.org/TR/ ation, in: Proceedings of the 11th International odrl-model/. Workshop on Semantic Evaluation (SemEval-2017), [4] L. Banarescu, C. Bonial, S. Cai, M. Georgescu, Association for Computational Linguistics, VanK. Grifitt, U. Hermjakob, K. Knight, P. Koehn, couver, Canada, 2017, pp. 536–545. URL: https: M. Palmer, N. Schneider, Abstract Meaning Repre- //aclanthology.org/S17-2090. doi:10.18653/v1/ sentation for sembanking, in: Proceedings of the S17-2090. 7th Linguistic Annotation Workshop and Interop- [14] Creative Commons, Attribution-NonCommercial erability with Discourse, Association for Computa- 4.0 International Public License, ???? URL: tional Linguistics, Sofia, Bulgaria, 2013, pp. 178–186. https://creativecommons.org/licenses/by-nc/4.0/ URL: https://aclanthology.org/W13-2322. legalcode. [5] A. Lamercerie, D. Rouquet, Construction [15] S. Cai, K. Knight, Smatch: an evaluation metric for d’ontologies à partir de textes : démonstration semantic feature structures, in: Proceedings of the d’une approche basée sur l’analyse de graphes amr, 51st Annual Meeting of the Association for CompuRevue des Nouvelles Technologies de l’Information tational Linguistics, Association for Computational Extraction et Gestion des Connaissances, RNTI-E- Linguistics, Sofia, Bulgaria, 2013, pp. 748–752. 39 (2023) 589–596. [6] A. Lamercerie, Principe de transduction sémantique pour l’application de théories d’interfaces sur des documents de spécification, Thèse, Université Rennes 1 ; Rennes 1, 2021. URL: https: //tel.archives-ouvertes.fr/tel-03366457. [7] W3C ODRL Community Group, ODRL Implementation Best Practices, 2023. URL: https://w3c.github.
io/odrl/bp/. [8] E. Cabrio, A. P. Aprosio, S. Villata, These are your rights - a natural language processing approach to automated rdf licenses generation, in: Extended
Semantic Web Conference, 2014. [9] W3C Member Submission, ccREL: The Creative
Commons Rights Expression Language, 2008. URL: https://www.w3.org/Submission/ccREL/. [10] S. T. Vu, M. Le Nguyen, K. Satoh, Abstract meaning representation for legal documents: An empirical research on a human-annotated