Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 133 OMTAT annotation tool: semantical enrichment for legal document search ∗ Sylvie Szulman François Lévy Eve Paul LIPN, Université Paris 13 LIPN, Université Paris 13 LuapLab Sorbonne Paris Cité & CNRS Sorbonne Paris Cité & CNRS France France France eve.paul@luaplab.com ss@lipn.univ-paris13.fr fl@lipn.univ-paris13.fr Abstract comment mainly based on jurisprudential cases). The work relies on the textual annotation standoff This paper describes a help system for le- format defined by the Brat tool ((Stenetorp et al., gal document searching. The proposed ap- 2012b; Stenetorp et al., 2012a)) which has been proach relies on creating specific annota- used in particular in the BioNLP domain. The ap- tions over a corpus of documents. A tool proach is supported by a tool that we have built, has been built which implements the visu- alization of annotations, texts and semantic which allows user to annotate documents with our resources, the creation of annotations and different kinds of annotations and to query the set their collation in resources. A search engine of documents and annotations. The tool thus en- has been implemented as well to query the ables user to rapidly find answers to a legal ques- set of annotated documents in order to an- tion in the domain. swer user questions. 1 Introduction General ideas are illustrated with an example where our approach may help a user to extract the This paper presents a system intended to facili- most relevant legal excerpts for his problem. Con- tate the access to French legal documents. Given sider the case of a professional newsman which the huge amount of existing legal documents, we regularly works for a journal as a freelance, so he developped a document search system for non is paid by the piece. Suddenly the journal ceases lawyer users (trade unionist, human resources giving him work. He wants to know if he has some manager, etc.) to find relevant information in the right to an indemnity. The base document to query overabundance of documents. Even if there are le- is the labour code, which is 1800 pages long. This gal databases and law data Banks for the French use case is cited as an example all along the paper, law, most of them target professional users, so and is specifically considered in section 6. they are hard to access for the non lawyer user. Our system should be a base to make the docu- mentation more accessible to non lawyers. The The rest of the paper is divided as follows. In domain of experiences is restricted to the Code du section 2, the approach is positioned with respect Travail (labour code) and collective labour agree- to other approaches for the access to legal docu- ments. mentation. Section 3) presents the types of an- The proposed approach relies on creating spe- notations that have been defined in order to give cific annotations over the reference documents. an account of legally significant properties of the “Annotation” is used here in an NLP sense (a spe- text. The OMTAT tool which has been built to vi- cific mark in the text) and not in its legal sense (a sualize, explore or add annotations is described in ∗ This work is part of the program "Investissements section 4. Section 5 is specifically devoted to the d’Avenir" overseen by the French National Research Agency, query engine. The 6th and last section contains a ANR-10-LABX-0083 short description of the use case. Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 134 2 Accessing Documentation in the legal role. Resources of this kind exist but, due to busi- Domain ness reasons, they are mostly not public. Among public thesauri, EuroVoc is a multilin- 2.1 Information Retrieval gual thesaurus produced by the European Union5 . Information Retrieval in the legal domain consists It describes the terminology used by the different in retrieving law articles and case law decisions re- domains of activity of this Union and is available lated to a given subject matter. The search can be in twenty three languages. It is used, among oth- by reference, for instance the “Loi Aubry” or “loi ers, by the European Parliament, the Publications sur les 35 heures” (35hours working week law). It Office of the European Union, some national and can also be a search of keywords in plain text, or regional parliaments in Europe, as well as by na- in some specific sites of a previously defined struc- tional administrations and private users in differ- ture. Advanced search options allow to search by ent states, some members of the Europeaan Union document descriptors (previously annotated with and some not members. the help of semantic ressources, like a thesaurus). Jurivoc6 is a public, trilingual thesaurus used Major law-editors (Dalloz1 , Editions Francis in the Swiss Confederation. It was produced by Lefebvre 2 , Lamy3 , Lexis-Nexis4 ) offer to clients the Swiss Federal Court and the Insurance Federal a large quantity of documents and an engine cus- Court. tomized for their documents and their descriptors. These resources cope with domains other than Their services are accessed mostly by profession- the labour code and cannot be used as such for an- als, due to their commercial offers. Some smaller notating the French Code du Travail. A terminol- law editors make available, in the Code du Travail ogy of French employment law is in construction domain, mementos and/or fact-sheets that can be by one of the authors who is a legal expert. Cur- bought by non-lawyers economic players (named rently the built terminology is restricted to the use NLEP in the following): trade-unionists, human case. It allows to create a set of terminological resources managers, managers of small business, annotations. etc. NLEP can then refer back to these sum- maries to answer practical questions. Our present 2.3 Standard based Initiatives approach proposes to make sources of law avail- XML (eXtensible Markup Language) is a meta- able for NLEP, through a semantic and structural language allowing to define for a particular do- search in documents which have been previously main a set of semantic and structural marks. Sev- semantically and structurally annotated. eral XML standards have been defined for the encoding of legal documents, like MetaLex and 2.2 Semantic Approach AkomaNtoso. Information Retrieval in the legal domain has re- cently interested the academic community. (Berry • MetaLex has been devised by the CEN as a et al., 2012) consider how different language mod- Workshop Agreement which standardizes the els of the collection succeeded or failed to be used way in which sources of law and references by domain specialists, (Mimouni, 2015) studies to sources of law are to be represented in two approaches for querying a collection, viewed XML. It involves an Open XML Interchange as a network of documents. The first is based on Format for Legal and Legislative Resources, Formal Concept Analysis, the second on semantic so as to avoid locking a client to a provider web technologies, namely an ontology to annotate (Boer et al., 2008). the collection and Sparql to query it. In a semantic approach as the one adopted here, • AkomaNtoso (Palmirani et al., 2005) defines semantic resources, i.e. ontologies, thesauri or ter- a set of simple technology-neutral electronic minologies in the legal domain, play an important representations in XML format of parliamen- tary, legislative and judiciary documents. 1 http://www.editions-dalloz.fr/ 2 5 http://www.efl.fr/ http://eurovoc.europa.eu/drupal/?q=fr 3 6 http://www.wkf.fr/accueil.html http://www.bger.ch/fr/index/juridiction/jurisdiction- 4 http://www.lexisnexis.fr/ inherit-template/jurisdiction-jurivoc-home.htm Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 135 Note that XML standards are not exclusive of that different words may receive the same our approach, for semantic and structural markups meaning and that, most often, one of them can be converted into annotations. An advantage is chosen to represent this meaning. In the of our approach is the possibility to create rela- Code du travail , salaire, sanction or contrat tional annotations, that is annotations bearing on de travail are terms (the meaning of the first other annotations which are difficult to represent can also be supported by e.g. traitement, ap- in XML. pointements, paie). As in many domains, le- gal terms are often gathered in specialized re- 3 Annotations sources like terminologies or ontologies. From a NLP point of view, an annotation links Relations (named AnnotationR) allow to link two data to a text fragment or to other annotations. terms by a specific relation. For instance, The data can be of many kinds, syntactic (POS, an authority_for relation links entreprise de grammatical roles, etc.), semantic (lexical en- presse to journaliste in Est journaliste pro- try, terminological element, ontological entity, fessionnel toute personne qui a pour activité anaphora. . . ), discursive (focus, concession, re- principale, régulière et rétribuée, l’exercice striction, emphasis, etc.), and free comment. The de sa profession dans une ou plusieurs entre- text fragment can be all in one segment, or can in- prises de presse et en tire l’essentiel de ses volve several segments. In our approach as in most ressources (Is a professional journalist any works focussed on technical contents, the annota- person whose main, regular and paid activity tions considered have constrained data (enumer- consists in exercising his profession in one or ated a priori in a closed list). The text fragments more newspaper companies and who obtains to which they are attached can be made of separate this way the main part of its resources). pieces, and they can be structured by other (lower level) annotations. References (named AnnotationL) annotations Different kinds of knowledge help legal texts mark fragments which refer to other frag- users to search in the text, and so OMTAT uses ments of a legal text, most often with the different types of annotations to reflect the partic- help of a standard identifier, e.g. La pré- ular role of each. somption de salariat prévue à l’article L. Keywords (named AnnotationK) are exact de- 7121-3, or: Lorsque le travail du journaliste nominations (except morphological varia- donne lieu à publication dans les conditions tions) which have a specific meaning in the définies à l’article L. 132-37 du code de la domain and can be by themselves a clue. propriété intellectuelle. When the reference For instance, Partie (Part), Titre (Title7 ), is relevant for the question at hand, the text Chapitre (Chapter) Article (Article) are key- under consideration must be extended with words : using a synonym of Part as Fragment the fragment referreed to by the annotation. (Fragment) is impossible, as the meaning of Three main specific relations can link an these words is only defined by the hierarchy annotation to a reference : defined_in, in the table of contents. Note that the heading decided_in, mentioned_in of a division is under the scope of the relevant keyword and is attached as an attribute of the Events (named AnnotationE) link sets of argu- annotation. ments around a central predicate, the trig- ger. For instance, in Le salaire perçu par un Terms (named AnnotationT) represent significant mannequin pour une prestation donnée (The entities of the domain. Terms can be sup- salary received by a model for a given perfor- ported by words or multiwords, but their sig- mance), the trigger of the event is perçu, its nificance relies on the attached meaning, rep- theme is salaire, its agent is un mannequin resented in the term-annotation label. Note and its cause is une prestation donnée. 7 In French law, Title is a level in the hierarchy of divi- sions, not to be mistaken for the text of a heading of any level, Context (named AnnotationC) annotations cover named its title a possibly large fragment having a functional Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 136 Figure 1: Main annotation view role by which interpretation has some speci- tions use the same kind of encoding and have been ficity. In the Code du Travail, context an- devised for the specific needs of legal annotation. notations correspond to entities in the ta- ble of contents, because the functional view 4 OMTAT Tool is reflected in this table. For instance, ar- In this section, we present the OMTAT (One More ticle L7111-3 states Si l’employeur est à Text Annotation Tool) system, built as an E4 l’initiative de la rupture, le salarié a droit à Eclipse application to implement annotations as une indemnité qui ne peut être inférieure à described above. Many functionalities have been la somme représentant un mois, par année defined to create and/or visualize different annota- ou fraction d’année de collaboration, des der- tions. The main window involves four views (see niers appointements8 . It cannot be accurately fig 1): understood without accounting for its posi- tion: in the legislative part, part VII (Rules • Text view: it shows the text of the document specific to some professions), Livre 1 Title (and its name). All the annotations are em- 1 (journalists), Chapter 2 (employment con- phasized with different text colors according tract) Section 2 (breach of contract) - hence to their type. “the salaryman” need to be a professional • Semantic resource view: it shows the seman- journalist. Note that, in other texts as case law tic resource (a thesaurus or an ontology). Any decisions for instance, judgments have func- semantic resource in SKOS format or OWL tional parts which remain implicit, while they format may be loaded. A SKOS semantic re- must nevertheless be recognized. source may be enriched when an annotationT Terms, relations and events have been widely is created and its semantic value does not ex- used in the BioNLP challenge, thanks to the outer ist in the resource. A OWL semantic resource encoding provided by the Brat tool9 ((Stenetorp et cannot be modified. In the figure, the seman- al., 2012b; Stenetorp et al., 2012a)). Other annota- tic resource is a thesaurus. It is built on the 8 labour code and restricted to the use case. Translation: If the breach is initiated by the employer, the salaryman has a right to an indemnity which cannot be less • Text annotation view: it displays all the an- than the amount of money representing one month per year of fragment of a year of collaboration, of the last salary notations defined on the document showed in 9 http://brat.nlplab.org/about.html In the Brat tradition, the text view. A click on an annotation selects Terms are named Types the corresponding text in the text view. Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 137 • Annotation graph view: it shows the selected possible arguments. Last, Brat is designed to be annotations in the form of a graph which used as a centralized collaborative tool through a nodes correspond to values and edges corre- web server. Its user interface combines text and spond to labels. drawings to visualize or add all categories of an- notations, while the annotation schema is centrally Other global functionalities include: controlled. A search tool locates annotations of a given category according to the conjunction of • Text and sets of annotationK(s) may be conditions on their text and on their attributes. extracted from XML files provided that In comparison, OMTAT is a single user tool markups are defined in a configuration file. under Java as Gate (but Gate has a collaborative • A semantic resource provided with preferred extension). It has the three categories of Bratt, labels may be cast on a text to create termi- it can read its standoff format and defines some nological annotations (AnnotationT). more categories to account for the structure of le- gal documents. It accepts to dynamically manage From the annotation point of view, OMTAT has semantic resources (Gate also does) and includes a common points with well known tools, GATE 10 search engine which is described in the following (Cunningham, 2002; Cunningham et al., 2011) section. and BRAT11 (Stenetorp et al., 2012b), but has also significant differences. GATE considers one 5 The Search Engine single category of annotations of the form (type, An experimental search engine has been devel- span, features) ; the type is a single word, fea- oped for the need of exploring the annotated cor- tures are key-value pairs. Gate can model struc- pus currently in memory. Its main role is to build tured annotations with the help of a “constituent” so called w-tuples, tuples of elements (i.e. annota- feature which records lists. GATE documentation tions, sentences and documents) constrained by a states that “No special operations are provided set of conditions. At the engine level, a query has a in the current architecture for manipulating con- simple Select . . . From . . . Where . . . form and stituents” 12 . After version 7, the provided library returns s-tuples, tuples of attribute values. For ex- allows modeling relations through a members[] ample, figure 2 shows a plain graphic interface to array13 . The 8.1 basic version does not offer a user the engine and a query returning the list of pairs interface to manually add relations as it does for (S.text, Aart.text) where S.text is the text of a plain annotations. A search tool locates the next sentence containing a Term annotation Presump- occurrence of a string or regex and can restrict the tion of Salary, and Aart.text is the number of the search to the scope of existing annotations. article containing the sentence14 . The function of Brat has three categories of annotations: types, each clause is shortly explained here: relations and events. Types are the basic blocks associating a (possibly discontinuous) portion of • The From clause provides a set of open (in text with a type. A Relation involves two typed memory) documents in which the search will annotations and a labelled link. An Event is made happen. of a main typed annotation (its trigger) linked • The Where clause describes the form of w- to a variable number of arguments. The last two tuples and the conditions that they must sat- can be represented in Gate by relations, but Brat isfy. For that purpose, every element in the emphasizes their difference from a user point of tuple is named and typed (the type may be an- view: the relation need no lexical support (e.g. notation, sentence or document; for the sake there is no lexical link between a bacterium and of conciseness, the first letter of the name the mouth which is its localization) while the trig- involves the type). Conditions are then re- ger of an event is its lexical support and restricts lations between attributes of the elements 10 General Architecture and Text Engineering, and possibly constant values. For example, http://gate.ac.uk S.numsent < 5 is a condition requiring that 11 Brat rapid annotation tool, http://brat.nlplab.org/ 12 14 Gate inline documentation, section 5.4.2 Hovering the mouse over truncated sentences shows the 13 Documentation, section 7.7 full text. The OK button saves the result. Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 138 Figure 2: The Query view the element S (a sentence) be at the begin- from inheritance. Attributes are provided by a spe- ning of the document (in the first five sen- cific inner mechanism of annotation classes allow- tences), and A_obj.label = A_agent.label is ing to define computed attributes which are not ex- a condition requiring that the two elements plicit in the Brat format (the size of the text, the A_obj and A_agent (both annotations) share number of arguments of an event, etc.). The kind the same label. of annotations is identified by the type attribute. Some more attributes are common to all anno- • The Select clause describes what s-tuple tations, and others are specific to one or several is returned for each built w-tuple. Only ele- types. In the present version, common attributes ment.attribute names can be used here, and are the identifier of the annotation, its type, its all the element names must be defined in the label, its size, its start and end position, its sen- w-tuples, i.e. used in the Where clause. Any tence number and the identifier of its document. attribute of a defined name can be used in the Other attributes may be for instance the subject Select. Last, a Distinct modifier allows to and the object of a relation (type "R") or the di- merge identical returned tuples. vision of a context annotation (type "C"), i.e. the kind of textual unit (chapter, paragraph) which For instance, the query defines its borders. Elementary conditions in the Select distinct A1.numsent Where clause are only combined by conjunction; From current a limited form of disjunction is available through Where A1.numsent = A2.numsent inequalities and string relations starts with, ends and A1.idannot != A2.idannot with. and A1.label = A2.label returns the sentence number of those sentences in Parsing provides a control of the query and of its the current document in which the same label oc- conditions. Elementary conditions are sorted ac- curs twice. cording to the names in their left and right hand The different kinds of annotations described in part. w-tuples are then built recursively on the section 3 are implemented as classes, benefiting set of names involved. In the application of con- Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 139 ditions, the validity of many attributes can only on by requiring a K annotation with a Partie Légis- be decided when the type is known; furthermore, lative label. Requiring a supplementary annotation some of them are optional in the type (e.g. the on Contrat de Travail only leaves 5 articles. arguments of an event). So the validity of at- Among these, article L 7112-1 states Toute tributes used by conditions is dynamically deter- convention par laquelle une entreprise de presse mined while the w-tuples are built. s’assure, moyennant rémunération, le concours 6 Use Case d’un journaliste professionnel est présumée être un contrat de travail. Cette présomption subsiste 6.1 Presentation quels que soient le mode et le montant de la ré- We have considered annotation of the French munération ainsi que la qualification donnée à la Code du Travail, which has 4644 articles. In a le- convention par les parties15 . Note also that the sec- gal sense, a French Code groups and organizes the tion is entitled Présomption de salariat (Presump- legal and regulatory texts produced along time for tion of wage relation), implying that wages are a a given domain. Jurisprudential knowledge comes form of payement specifically attached to employ- from other sources. A legal user of this text must ment contracts. discover as quickly and easily as possible which This means that, provided he is a professional, fragments are relevant to his problem, and may the freelance is presumed to be governed by an have to examine many excerpts. employment contract and hence to benefit of a Following a well established layout of French right to an indemnity. The user must enjoy a min- codes, we stored each article in a separate doc- imal understanding of legal reasoning to supple- ument which includes a reminder of its position ment this basic information with the help of some in the table of content: the document starts with more queries. Namely, he must know that, beside headings of all levels in the scope of which it employment, other kinds of contracts may be rele- is. At the moment, a small set of annotations vant in order to “remunerate the support” of some have been marked in the text: Contrat de tra- individual, and that employment contracts them- vail (employment contract), Journaliste (journal- selves may have different categories. Searching ist), Journaliste professionnel (professional jour- for Journaliste and Remuneration annotations in nalist), Présomption (presumption), Rémunéra- the same article gives six results, tion (remuneration). One of them is L. 7113-3: Lorsque le travail du Recall the example legal problem of section 1, journaliste professionnel donne lieu à publication for which our approach may help a NLEP to find dans les conditions définies à l’article L. 132-37 useful excerpts of the labour code answering to his du code de la propriété intellectuelle, la rémuné- problem. A freelance journalist used to work for a ration qu’il perçoit est un salaire16 . It means that newspaper, and at a moment the newspaper does it is possible to argue that the payment is of the not ask him anymore article. Being freelance, the kind specific to intellectual property, i.e. an au- journalist is paid for each provided product (arti- thorship rights remuneration, and not an employ- cle, photograph, drawing, etc..). In the case, he has ment contract. The conditions of this possibility never signed any written contract whatsoever. The are to be found in the Intellectual Property Code. stakes are if he has a right to an indemnity. To maintain the legal complexity inside reason- 6.2 A first Analysis able bounds, this path is not followed here. Articles can be selected according to the pres- ence of annotations Journaliste in the document, 15 Any agreement by which a press company obtains, including titles. This yield 336 distinct answers, through remuneration, the support of a professional journal- due in particular to one of the titles of level Livre, ist, is presumed to be an employment contract. This presump- which enumerates various professions, so for in- tion remains whatever can be the mode and the amount of the stance articles about (theater, movies, . . . )-players remuneration or how participants qualify the agreement. 16 When the work of the professional journalist gives rise to are also under this title. Restricting the annotation a publication in such conditions as defined in article L. 132- to occur in the body of the article reduces to 45 ar- 37 of the Intellectual Property Code, the remuneration that he ticles, 17 of which are in laws and can be focused receives is a salary Proceedings of the conference Terminology and Artificial Intelligence 2015 (Granada, Spain) 140 6.3 More on the Contract Thanks Every French worker knows that employment con- We thank Jorge Garcia Flores and Nadi Tomeh for their tracts belong to one of two categories : contracts friendly review of our writings. with an indeterminate duration (CDI) or with a de- terminate one (CDD), and that they do not involve the same rights. More, article L. 7112-2, one of References the five obtained from the first query, considers a Michael W. Berry, R. Esau, and Bruce Keifer. 2012. breach of the CDI, while no mention is made of The use of text mining techniques in electronic dis- a CDD. Definitions can only be found considering covery for legal matters. In C. Jouis, I. Biskri, Jean- articles which apply to the generic case. Searching Gabriel Ganascia, and M. Roux, editors, Next Gen- eration Search Engines: Advanced Models for Infor- for titles of the highest possible level annotated mation Retrieval, pages 174–190. IGI Global. with employment contract yields Part 1 (individ- A. Boer, A. Winkels, and F. Vitali, 2008. Metalex xml ual employment relations) Livre II (Employment and the legal knowledge interchange format., vol- contract). ume Computable Models of the Law - Lecture Notes Searching in this Livre for articles annotated in Computer Science -4884, pages 21–41. Springer. both with CDI and CDD provides two basic texts. Hamish Cunningham, Diana Maynard, Kalina In L. 1221-2 it is stated that the CDI is the default Bontcheva, Valentin Tablan, Niraj Aswani, Ian case: Le contrat de travail à durée indéterminée est Roberts, Genevieve Gorrell, Adam Funk, Angus Roberts, Danica Damljanovic, Thomas Heitz, la forme normale et générale de la relation de tra- Mark A. Greenwood, Horacio Saggion, Johann vail. Toutefois, le contrat de travail peut comporter Petrak, Yaoyong Li, and Wim Peters. 2011. Text un terme fixé avec précision dès sa conclusion ou Processing with GATE (Version 6). résultant de la réalisation de l’objet pour lequel il H. Cunningham. 2002. Gate - a general architecture est conclu dans les cas et dans les conditions men- for text engineering. In Computers and the Human- tionnés au titre IV relatif au contrat de travail à ities, Volume 36, pages 223–254. durée déterminée17 . And in L. 1242-12, this is en- Nada Mimouni. 2015. Interrogation d’un réseau forced by requirements on the form of the CDD: sémantique de documents: l’intertextualité dans l’accès à l’information juridique. Ph.D. thesis, Paris Le contrat de travail à durée déterminée est établi 13 - Sorbonne Paris Cité, janvier. par écrit et comporte la définition précise de son M. Palmirani, R. Brighi, and M. Massini. 2005. Au- motif. A défaut, il est réputé conclu pour une du- tomated extraction of normative references in legal rée indéterminée18 . texts. In NY USA ACM, editor, Proccedings of the 9th international conference on Artificial intel- 7 Conclusion ligence and Law, ICAIL, pages 105–106. Pontus Stenetorp, Sampo Pyysalo, Goran Topić, This article has described the use of different tex- Sophia Ananiadou, and Akiko Aizawa. 2012a. Nor- tual annotations to help legal documents search malisation with the BRAT rapid annotation tool. In take advantage of the document structure. An ex- Proceedings of the 5th International Symposium on Semantic Mining in Biomedicine, Zürich, Switzer- perimental use case demonstrates the utility of the land, September. approach. An implementation has been carried out Pontus Stenetorp, Sampo Pyysalo, Goran Topić, which allows to explore, query and add annota- Tomoko Ohta, Sophia Ananiadou, and Jun’ichi Tsu- tions. Future work will allow us to deepen the jii. 2012b. brat: a web-based tool for nlp-assisted study of links between semantic annotations and text annotation. In Proceedings of the Demonstra- semantic document search. tions at the 13th Conference of the European Chap- ter of the Association for Computational Linguistics, pages 102–107, Avignon, France, April. Association 17 The CDI is the normal and standard form of employment for Computational Linguistics. relation. However, the employment contract may involve an end date precisely fixed at start or determined by the achieve- ment of the object in view of which it is decided, in cases and conditions mentioned in Titre IV related to CDD 18 The CDD is set up in writing and contains a precise defi- nition of its motive. Failing that, it is deemed to be agreed for an indeterminate duration