We gratefully acknowledge the work of Felix Dombek, who implemented the web app, built many scripts for converting lexicon formats, and also performed various manual mappings and error corrections in some of the source lexicons. Later, Peter Bourgonje took over the job of technical administration of the app, and contributed to the integration of new lexicons. Further, we would like to thank the two anonymous reviewers for their helpful comments. This work was partially funded by the Deutsche Forschungsgemeinschaft (DFG) [German Research Foundation] – Project number 317633480 – SFB 1287, Project A03.

1. Introduction

2The relationship between connective and relation/sense is by no means simple, however. While there are cases of widespread agreement (although signals a Concession), many connectives can be seen as ambiguous (since can signal a temporal connection or a causal one) or as vague (but can signal Contrast or Concession, but what is the difference between these two, exactly, and can it be distinguished in every context of use?). Further, it is well-known that some connectives can operate on different linguistic “levels” (cf. Sweetser, 1990), as exemplified by a causal connective that relates in fact to a speech act rather than to a proposition: Are you busy now? Because you should read this great article in The Times. Besides these issues of semantics and pragmatics, there are also many open questions concerning the exact borderline of the class of “connective”: which terms fit in, and which do not, and why? We believe that progress on these matters is more easily made when inventories, i.e., lists of connectives with linguistic descriptions, are freely available for as many languages as possible, so that researchers can inspect, consider, possibly disagree with boundaries of inventories or with linguistic descriptions; provide arguments for changes; and contribute to further improving our understanding of what these interesting lexical items actually do, and how that should be documented.

4In short, for the near future, Connective-Lex should be expanded not only in terms of breadth (by adding new languages) but also in terms of depth (by providing more, or more exact information). However, the two goals are not trivial to pursue in parallel: any changes on “depth” need to be done carefully in order to preserve – or even enhance – the comparability between languages.

5While the various lexicons assembled in Connective-Lex have already been individually introduced in earlier publications, the main contribution of the present paper is to present, for the first time, the joint online database project. We explain what the system does, and how it can be used, so that potential users can check whether it is helpful for their research, and prospective developers of new lexicons know what needs to be done to add them to the system. In addition, as mentioned earlier, we take a step back and reflect on what has been done so far, and what should be done next.

6In Section 2, we first provide background information by reviewing the origin of the first connective lexicons, and by summarizing some earlier research on the crosslinguistic comparison of connectives. Section 3 then introduces Connective-Lex, both from the perspective of a user and that of a potential lexicon developer. In Section 4, we provide a more detailed investigation of the underlying linguistic questions, viz. the definition and demarcation of the class of connectives, and the information provided in lexical entries. Here, we go beyond the status quo and sketch some possible directions for extensions. Section 5 turns to the multilingual perspective and examines possibilities for improving the support for contrastive studies. Finally, in Section 6 we draw some conclusions and provide an outlook.

2. Related research

2.1. Lexicons of connectives

7The first machine-readable lexicon of discourse connectives was the German DiMLex, whose version 1 was introduced by Stede and Umbach (1998). The main motivation was to use a declarative resource as a component for software that needs knowledge about discourse relations in order to:

• generate text: when a system produces natural language from underlying structured data or a knowledge representation system, and it needs to express a temporal sequence relation between events, it would select from a set of similar connectives. The first application was Techdoc (Rösner & Stede, 1994), which produced maintenance instructions of car manuals. Example: Park the car on level ground and switch the engine off. Then, check the engine oil;

• understand text: when a system is supposed to construct a discourse representation, it can exploit connectives as hints on what relations may hold between adjacent units of text. The first applications were the parsers by Hanneforth et al. (2003) and by Reitter (2003), which worked on German text and built trees in the spirit of Rhetorical Structure Theory (RST – Mann & Thompson, 1988).

8Shortly after presenting the first version of DiMLex, interactions started with the group of linguistics researchers at the Institut für Deutsche Sprache (Mannheim), who were working towards publishing the voluminous Handbuch der deutschen Konnektoren (Pasch et al., 2003). As a result of this collaboration, some changes were made to DiMLex, and more connectives were added, resulting in a set of 175; this is documented in Stede (2002). From the outset, DiMLex had been confined to what is nowadays called primary connectives, i.e., the inclusion/exclusion decision was made largely in line with the definition by Pasch et al. (2003), which we discuss in detail in Section 4. In contrast, multiword units that are compositional, modifiable and inflectable (e.g., for this reason), which today are often called secondary connectives, are deliberately not covered (see, e.g., Danlos et al. [2018], or the discussion of alternative lexicalizations by Prasad et al. [2008]).

9The contents of a DiMLex entry in that early version were:

• orthography: list of alternative spellings (if any);

• syntax: category (subordinating conjunction, coordinating conjunction, adverbial, preposition); phrasal status (one or more words; continuous or not); possible positions within the sentence (this can be quite elaborate, especially for German adverbials; we followed the feature-based approach of Pasch et al. [2003]); possible linear orders of the two arguments;

• semantics: coherence relation or sense; modifiability by focus particles; certain idiosyncratic features applying only to certain connectives (presuppositions, style);

• ambiguity information: whether the word also has a non-connective reading;

• examples for the various readings.

10To support automatic processing, DiMLex was represented in an XML (Extensible Markup Language) format, which we will illustrate in Section 3 (in the slightly different form as used today in Connective-Lex).

11Also in the early 2000s, a tool for the semi-automatic annotation of connectives was presented (Stede & Heintze, 2004): ConnAnno reads a reduced version of DiMLex, spots candidate words in the input text and highlights them; if the user confirms that one is a connective, the system uses simple surface rules to suggest possible arguments (on the basis of the syntactic type of the connective), which the user can in turn either confirm or correct.

13DiMLex and its XML format inspired the design of the French LexConn (Roze et al., 2012), for storing the orthography, syntactic category, and sense of French connectives. The first version of the lexicon was obtained by compiling a list of elements belonging to the syntactic categories described above, and this list was manually filtered. This resulted in a lexicon with 325 entries. The sense inventory was modeled on the basis of Segmented Discourse Representation Theory (SDRT – Asher & Lascarides, 2003). Later on, a corpus annotation project (Afantenos et al., 2012) extended the lexicon with an additional 30 entries. The underlying corpus contained about 18,000 sentences, in which roughly 10,000 connective tokens were annotated; on this basis, the authors now consider the lexicon to be complete. In contrast, DiMLex so far has only been applied to a relatively small corpus of 175 German newspaper editorials (Stede & Neumann, 2014) with 1,100 connective tokens. Since that corpus is also genre-specific, the empirical basis of LexConn is on the whole considerably broader.

14Enabled by a TextLink short-term scientific mission, Colinet (2015) undertook a detailed comparison of DiMLex and LexConn. She pointed out a difference in the criteria used to demarcate the class of connectives: while the criteria used for DiMLex also treat certain nominalized verbs as potential arguments for connectives, LexConn restricts arguments to full clauses. A consequence is that DiMLex accepts more prepositions than LexConn, which includes only prepositions taking infinitival clauses (of which French has many more than German). A second difference is that LexConn has a number of relatively-lexicalized prepositional phrases (for example, à l’instant où [at the instant when]), whose German correspondents are compositional phrases which therefore do not pass the criteria of non-inflectability and non-modifiability (see Section 4.1).

2.2. Multilingual/contrastive studies on connectives

15In linguistics, it seems that the interest in connectives increased considerably in the 1990s, which may be related to the publications on approaches to defining and taxonomizing coherence relations starting in the 1980s. In this section, we point to some studies that took a decidedly multilingual (or contrastive-linguistic) perspective – with much recent work having originated in the TextLink network – and distinguish them according to the primary goals of the work. These goals are of course not disjoint, as many studies may follow multiple lines; here we try to group approaches by their predominant motivation.

2.2.1. Insight into coherence relations

16For research on coherence relations, it is methodologically not easy to demonstrate the cognitive or linguistic “reality” of such relations. While it is generally accepted that connectives serve as (more or less clear) signals of such relations, some researchers argued that parallel findings on different languages can be taken as stronger evidence for their (language-independent) role in cognition. In this vein, Knott and Sanders (1998) employed the connective substitution test of Knott and Dale (1994) for producing taxonomies of English and Dutch connectives in four semantic areas (positive and negative causality, positive and negative addition). They examined in what way the non-/substitutability in certain contexts can be explained by cognitive-linguistic features and argue that those features that had been used in the Cognitive Approach to Coherence Relations (CCR – Sanders et al., 1992) are appropriate for the task. Recently, working in the same framework, Hoek et al. (2017) studied the translations of English connectives in four languages of the Europarl corpus (Koehn, 2005) in order to determine factors for the decision whether to also use a connective in the target language or to leave the relation implicit. This question was also investigated in a study by Zufferey (2016). There are also various studies that center on one relation specifically and examine variants of its realization in different languages, for instance the corpus-based work by Grote et al. (1997) on concessions in English and German, which led the authors to propose three different subclasses of the Concession relation.

2.2.2. Translation of connectives

17Quite a lot of work focused explicitly on the crosslinguistic mapping of connectives, without resorting to coherence relations as a possible inter-lingual backbone. Two recent examples are the investigation of translating however into Lithuanian by Mazeikiene and Vaiciuniene (2016) and the study by Nedoluzhko and Lapshinova-Koltunski (2018), who used a parallel corpus to determine the distribution of translation equivalents of a range of German pronominal adverbs in English, Czech and Russian. In machine translation (MT), Meyer and Poláková (2013) presented an MT experiment from English to Czech using a gold-standard of manually annotated discourse connectives, and achieved better translations. They illustrate difficult cases of translation, such as meanwhile, which is ambiguous between a temporal and a contrastive sense, and was wrongly translated, in a contrastive context, with a Czech connective that has only a temporal sense.

2.2.3. Semantics and pragmatics of connectives

18A similar line of research also compared semantically similar connectives across languages, but focused more on gaining insights into the linguistic description of connective meaning, with translation equivalence being only a secondary aspect. One example is Behrens and Fabricius-Hansen (2002), who examined Norwegian, German and English connectives signaling an Elaboration, analyzed their syntactic context, and found, for example, that English by-clauses and German indem-clauses have different scope properties: indem-clauses fall outside the scope of the negator in the matrix clause, while by-clauses have narrow scope, i.e., they attach to the propositional nucleus of the matrix clause. Using the same corpus-based research paradigm, Mortier and Degand (2009) focused on French en fait [in fact] and Dutch eigenlijk [actually], gathered statistics on their collocations, signaled relations, etc., and obtained a detailed “relational semantic field” that contributes to explaining under what circumstances the two connectives have the same or a different meaning. In a similar way, Zufferey and Cartoni (2012) examined English and French causal connectives and found a number of fine-grained features implicated in their conditions of (monolingual) use and their translation.

2.2.4. Generating lists of connectives

19Finally, during the last 10 years another research direction has become quite prominent: inducing lists of connectives from parallel corpora, where connectives are (manually or automatically) annotated in one language, and then automatically projected to the other one. Versley (2010) trained a connective classifier on the English PDTB, then automatically tagged part of the English Europarl corpus (Koehn, 2005), and projected the connectives to the (automatically-aligned) German section, so that candidates for German connectives were induced. In similar ways, also starting from existing labeled data for English, Zhou and Xue (2012) built a list of Chinese connectives, Hajlaoui and Popescu-Belis (2012) an Arabic one, and Laali and Kosseim (2014) a French one. Recently, Bourgonje et al. (2017) used Europarl to extend the information in the existing German and Italian lexicons (DiMLex, LICO): they found additional connectives that ought to be added to the respective lexicons, and they also studied the mapping between the annotated senses, in order to find areas of overlap in readings, and to compare the degrees of connective ambiguity in the two languages.

3. Design and implementation of Connective-Lex

20The web database Connective-Lex was designed with two central goals in mind: it should provide a search functionality that allows users to browse related connectives both within and across languages; and it should allow for easily adding new lexicons (i.e., new languages) to the system. The latter goal on the one hand translated into a purely technical requirement on the web app, which should notice the presence of a new lexicon in the resource directory and smoothly integrate it into the search front end. On the other hand, since at this stage the emphasis is on encouraging the development of new lexicons, we kept the contents of entries to a minimum (in order to reduce the effort of building a lexicon for a new language from scratch). At present, lexicons for the following languages are available in Connective-Lex: Arabic (Keskes et al., 2014), Bangla, Czech (CzeDLex; Mírovský et al., 2017), Dutch (DisCoDict; Bourgonje et al., 2018), English (Eng-DiMLex; Das et al., 2018), French (LexConn; Roze et al., 2012), German (DiMLex; Stede, 2002), Italian (LICO; Feltracco et al., 2016), and Portuguese (LDM-PT; Mendes et al., 2018). Pointers to the original source lexicons are provided on the Connective-Lex website (see the “About” links).

3.1. Minimal connective descriptions for the search facility

21The goals of relative simplicity and of language-neutrality led us to define only a small set of obligatory attributes that need to be known about a connective in order to ensure compatibility with the database. This is, first, the syntactic category, where we use the following inventory: coordinating conjunction, subordinating conjunction, adverbial, preposition (see Section 4 for details). Second, we require that connectives come with one or more senses from the PDTB3 tagset (Webber et al., 2016; Webber et al., 2019). This choice was made because the PDTB corpus is the largest resource available with annotated connectives, and the tagset has therefore been extensively tested on empirical data and in agreement studies. The database uses the full three-level hierarchy of tags, and connectives may be described on any level. The additional attributes (“+Belief” and “+SpeechAct”) in PDTB corpus annotation are associated with one of the arguments of the relation; nevertheless, these features are also being used to describe connectives in some lexicons, and therefore we also account for them.

22Syntax and PDTB3 sense are two of the filter settings of the web app that allow for constraining a search; the other two are the language(s) to be considered and a free text term that is matched against substrings of connectives. For instance, when the language is set to Italian and the string altri [other] is entered in the search field, the system returns the entries for the conjunctions altrimenti [otherwise] and in altre parole [in other words]. The latter results from the fact that LICO provides orthographic variants of that canonical form, one of which is in altri termini [in other terms]. Such variants can be displayed if any are available; likewise, there may be synonyms, either in the same or in a different language. For example, since the Italian LICO was modeled closely after the German DiMLex, it also provides pointers to closely related German connectives. (Hence, the term “synonym” is used in a rather wide sense throughout this paper, also covering near-synonyms or “plesionyms”, as well as translation correspondents.)

23A different search scenario checks for connectives in multiple languages that express the same relation. For illustration, Figure 1 shows a screenshot of Connective-Lex with part of the result found for “Equivalence” in Italian, Dutch and Arabic.

3.2. Integrating lexicons

25The binary <focuspart> tag specifies whether the connective can be in the scope of a focus particle (especially because) or not (? especially while). Then, we embed the semantic description within the syntax entries, of which there can be many when a connective has multiple syntactic functions; this decision will be explained in Section 4. Besides the categories mentioned in the previous subsections, the syntax part can specify ordering constraints for the argument: i.e., whether Arg2 can precede Arg1 (<ante>), follow it (<post>), or be embedded in it (<insert>).

26Obviously, the integration of a new lexicon is particularly easy when it is modeled from scratch in a very similar way (as done, for example, with the Italian LICO). Otherwise, we have mapped existing resources: for the case of Portuguese LDM-PT, the original source formats were an Excel sheet and a structurally-equivalent XML (produced via a Perl script); the latter was then mapped to the DiMLex XML format by means of an XSL (Extensible Stylesheet Language) sheet. Details can be found in Dombek (2017). For illustration, Figure 3 shows an entry from LDM-PT in the original XML format. Note that the creators made a different design decision for the relation between syntax and semantics: both are on the same level of embedding, i.e., one entry always gives one combination of syntax and semantics. The XSL script thus needs to merge different entries into a single one, iterating over entries for identical words and over their syntactic categories. A very similar situation held for the French LexConn, where it was also important to have the example sentences and (within-language) synonyms appear in the right place in our target structure.

27Besides this merging, the other central task for the mapping scripts is to convert syntactic categories (or part-of-speech tags) and sense information. The original DiMLex, for example, uses a slightly more elaborate syntax tagset (konnadv, padv, konj, subj, v2emb, postp, appr, appo, apci, einzel), which is being reduced to that of Connective-Lex. Relating different semantic/pragmatic relation sets is more difficult, and we discuss this in Section 4. Technically, we used declarative tables for mapping tagsets, so they can be easily re-used and adapted for the integration of a new lexicon.

3.3. Implementation of the web app

29As explained briefly in Scheffler et al. (2018) and in detail by Dombek (2017), the web app consists of frontend and backend, and is built on the principles of the so-called single-page web application (SPA) technology, which has several advantages: the frontend resides completely in a HTML page, so there are no page reloads, and all the rendering is done on the side of the client (browser). Accordingly, client-server communication is kept to a minimum, and it is done by sending JSON (JavaScript Object Notation) data. When the client loads some data from the PHP (Hypertext Preprocessor) backend, it updates the document by directly editing the HTML structure (DOM – Document Object Model). All these measures serve to speed up the browsing experience of the user.

30One area of potential improvement is responsive design. So far, the system is built for large screens, and enabling a smooth experience with handheld devices has not been our focus. But this step is on the agenda for the near future.

31Importantly, the system is designed such that adding a new lexicon is very simple: the lexicon itself (in DiMLex XML), a metadata file and, if necessary, mapping tables for syntactic or semantic tags are added to the data directory in the backend, and the app then automatically integrates the new resource and makes it available to the users.

3.4. Creating compatible connective lexicons

32Connective-Lex can easily integrate lexicons for additional languages if they adhere to the described basic XML format. Such lexicons can be created in a variety of ways: (i) by manual compilation, (ii) by translating and editing an existing lexicon, or (iii) by extracting connectives from an annotated corpus of discourse structure (e.g., in PDTB format).

33The most straightforward way for lexicon creation for a new language is the manual compilation of an inventory of discourse connectives, which then has to be completed by adding relevant syntactic and semantic properties. This process can be aided semi-automatically by starting from a list of lexical items that fall into the major syntactic classes for connectives (conjunctions, sentence adverbs, adpositions). For example, the creators of the French LexConn (Roze et al., 2012) started with a list of connective candidates based partly on syntactic criteria.

34Manual lexicon creation is a very time-consuming task. The process can be speeded up by starting from an existing lexicon for another (possibly related) language, and translating the items to obtain a draft lexicon of connectives that includes not only the items but their potential syntactic and semantic properties, as well. The Italian lexicon LICO (Feltracco et al., 2016) enhanced existing lists of Italian discourse connectives by translating the example sentences in the German DiMLex lexicon and keeping those candidate translations that preserve the sense of the German connective.

4. Taking stock: toward a multilingual connective database

36As presented in the previous sections, Connective-Lex provides a multilingual interface to connective lexicons in several languages, but was built on the model of the German DiMLex lexicon. In this section, we will assess the effects of this design and evaluate the fundamental limits of such a multilingual resource. In addition, we compare our “machine lexicon” with human-created resources and identify potential information types that can be added systematically to the existing platform.

4.1. Connective: definition and delimitation

37The first important decision in the creation of a specific lexicon is which items to include in it. For a lexicon of connectives or discourse markers, it is essential to define exactly what is meant by these terms, since individual researchers have proposed different ways of limiting the space of the phenomena under discussion. Definitions also typically follow language-specific considerations (or are focussed on a small set of languages, as discussed in Section 2). As the starting point for the multilingual interconnected lexicons described here is the German DiMLex, we largely follow their definition of “discourse connectives”, based on the work by Pasch et al. (2003). Our definition has four main criteria:

[1]

Definition of connective A lexical item or phrase x is a connective if and only if: 1. x cannot be inflected or modified; 2. the meaning of x is a two-place relation; 3. the arguments of this relation are abstract objects (propositions, events, states, or processes); 4. these arguments are typically expressed as clauses or sentences (but they can also be expressed as groups of sentences, or as noun phrases that denote abstract entities).

38In addition to single words which have these properties, fixed phrases can be included as connectives if they cannot be modified (see condition 1) and if their internal semantics is thus not compositional. The intention of this condition is to exclude the potentially very large number of freely modifiable phrases that also denote functions over abstract entities from a finite and focussed resource such as a connective lexicon. Freely modifiable phrases include for example many verbs that have denotations which also occur as discourse relations, such as cause, lead to, contrast with, etc. These verbs contribute their relations to the regular sentence semantics. They contribute to the discourse structure of a text (Danlos, 2006), but should not be included in a connective lexicon.

41Specific syntactic restrictions, on the other hand, lead to problems and grey areas when applied simultaneously to many different languages. Fine-grained part-of-speech categories often do not have close matches in another language. Thus, the decisions on what to include in a given corpus annotation or connective lexicon have often been made based on language-specific criteria, such as the restriction to adverbials and conjunctions, and the ability to take clausal syntactic arguments, in the PDTB. In CzeDLex, prepositions are excluded as primary connectives, because connective’s arguments are required to have a verbal core. In contrast, many connective lexicons in other languages allow pre- or postpositions, since these items otherwise fulfil the semantic requirements of a connective. An adposition may have a nominalized internal argument, which can often express an abstract object. For example, the Arabic discourse treebank includes prepositions that take nominalized arguments (in the Al-Mazdar form), as well as clitics; the French LexConn explicitly includes prepositions as candidates; the Chinese lexicon includes a type of postposition called localizers; and the German DiMLex allows for pre- or postpositions with noun phrase arguments that denote abstract objects. In many cases, these types of connectives would correspond to items in English that take gerund arguments and thus may be connectives according to the PDTB definition as well.

42Completely removing any syntactic constraints from the definition of “connective” will also lead to conceptual problems, however. For example, there are certain types of verbs that can take clauses as arguments. Consider the English verb depend on, as in the following sentence:

[2]	Whether we will go skiing depends on what the weather is like.

43The verb denotes a two-place semantic relation between two abstract objects, which are expressed as clauses. Still, one would not like to include such a verb as a discourse connective, because it does not relate independent clauses and its meaning is part of sentence semantics and not discourse. For this reason, we add condition 1 to our definition of connectives. It captures the crosslinguistic generalization that connectives typically come from closed word classes that encompass function words. Open class items and phrases such as nouns or noun phrases, and verbs, are to be excluded. These open class items are inflected in many languages and generally allow for free modification. Several individual language connective lexicons try to capture facets of this constraint by excluding items that are partly substitutable, variable, or modifiable (e.g., French LexConn). CzeDLex and LDM-PT distinguish between non-modifiable, fully grammaticalized “primary” connectives, and freer secondary connectives. Only the primary ones match our definition.

44The chosen definition of the notion “discourse connective” is thus language-neutral as far as possible. It references mainly semantic/pragmatic notions that are independent of language-specific morphosyntactic instantiations (conditions 2 and 3). Condition 4 is not strictly a constraint since it merely describes typical shapes of arguments (without properly restricting them). The only morphosyntactic constraint (condition 1) is deliberately chosen to be broadly applicable. It will lead to specific ranges of categories of items included as discourse connectives based on the properties of each language, but it guarantees a common core and as much overlap as possible.

4.2. Information on lexicon entries

45As detailed in Section 3, Connective-Lex currently indexes only limited information for the connectives in its lexicons: mainly orthography, basic syntactic information such as part of speech, and the semantic relation expressed. Some manually compiled lexicons are much more comprehensive. Here, we briefly discuss the additional types of information included in linguistic connective lexicons and their possible inclusion in an extension of our machine lexicons.

4.2.1. Syntax

47In addition, many connective lexicons include certain fixed phrases which can be used as connectives. We can distinguish three main categories of such phrasal connectives:

1. phrasal connectives that are syntactically equivalent to single-word connectives, such as adverbial connective phrases: on the other hand (English), abgesehen davon ([despite that/other than that], German);

2. paired connectives;

3. other grammaticalized phrases.

48Many phrasal connectives can easily be mapped onto one of the four major syntactic categories included in Connective-Lex, because they play the same syntactic role as single-word connectives. Thus, English on the other hand often matches also, and German abgesehen davon [despite that] behaves like nevertheless. They are both listed as adverbials in the multilingual lexicon. Sometimes, complex phrasal connectives also behave like other syntactic types, for example the German abgesehen davon, dass (also “despite the fact that”, but including the complementizer), which has the syntactic role of a subordinating conjunction.

49A special phenomenon present in many languages is the case of paired connectives. These connectives always or typically occur paired, one connective introducing each of the two arguments of the relation expressed by the connective pair. Many lexicons in Connective-Lex feature these kinds of connectives, for example English (either… or, if… then) and German (sowohl… als auch [both… and]). In Chinese, paired connectives dominate the range of possible explicit connectives. Since the paired connectives typically exhibit the same syntactic category for both parts, we categorize them the same way we would their individual parts. That is, sowohl… als auch is a coordinating conjunction, because each part behaves like one.

50Finally, each language exhibits idiosyncratic grammaticalized expressions that function as discourse connectives. These expressions frequently consist of more than one word. Sometimes, the two words do not even form a phrase/constituent in the syntactic sense, as with the German connective sei es [be it = if], where es [it] is arguably the syntactic subject. Since such cases occur on an individual case by case basis in each language, without allowing for crosslinguistic generalizations, we provide the syntactic category “other” for these kinds of (single-word or phrasal) connectives.

4.2.2. Semantics/pragmatics

51The other central type of information included in the lexicons on Connective-Lex is the possible discourse relations that can be signaled by each connective. Since one discourse connective can have different syntactic instantiations, we subordinate the semantic readings under each syntactic instantiation separately. This is justified by the fact that many discourse connectives denote different semantic relations when they are different parts of speech. Each syntactic option for each connective therefore lists the set of readings available for this connective-part-of-speech combination (see Figure 1).

52Discourse relations are semantic/pragmatic entities and applicable crosslinguistically. Still, there are many options for inventories of discourse relations ranging from Knott (1996), through the relation taxonomies developed in the RST and SDRT frameworks, to the different variants of PDTB relations (see Section 2.1). Since Connective-Lex assumes a lexical approach to the representation of discourse relations, we chose the PDTB3 relation hierarchy (Webber et al., 2016), which is an improved variant of the well-attested PDTB2 hierarchy. The PDTB3 relations have been used in an annotation project for a diverse range of languages, the TED-Multilingual Discourse Bank corpus (TED-MDB – Zeyrek et al., 2018).

54In mapping the relation schemas, some relations lack corresponding PDTB3 senses. In addition, some connectives may have lacked a clear sense assignment already in the original monolingual connective lexicon. These cases are assigned the relation sense “other” in order to make them searchable on Connective-Lex. For example, the German connective so oder so [in any case/anyway] denotes an unconditional relation that is not represented in the PDTB3 sense inventory. Another example is au moins [at least] in French.

4.2.3. Lexicographic information

5. Linking the languages

57Individual lexicons of connectives are useful for discourse studies and NLP (Natural Language Processing) tasks related to the automatic identification of connectives (Stede & Heintze, 2004; Mendes & Río, 2018) and discourse processing (Lin et al., 2014; Stede, 2014), but the possibility of linking these resources provides additional advantages for research areas such as machine translation (Meyer et al., 2011; Meyer & Poláková, 2013), to develop statistical translation models that are able to operate above the sentence and/or phrase level and that correctly model discourse connectives and coherence relations. Linked resources also provide a very helpful support for manual translation as well as language learning and teaching in CALL (Computer-Assisted Language Learning) systems (Meurers & Dickinson, 2017).

58As shown in Section 3.1, Connective-Lex enables the users to retrieve, in multiple languages, connectives that express the same sense, or connectives of a specific category. For instance, using the search options “sense:Purpose (Arg2-as-goal)” and “category:csu” (subordinating conjunction) over the German and the Portuguese lexicons retrieves four Portuguese connectives (a fim de que, de forma que, de modo a que, para que) and two German connectives (bis dass and damit). Of these, the first is ambiguous between a temporal and purpose interpretation, and the second has only a purpose interpretation. The search options are therefore quite useful for a contrastive approach. They are nevertheless limited. One such limitation is the result of different inventories of discourse relations, as mentioned in 4.2.2. For instance, the sense “CONTINGENCY:Purpose” is new in the PDTB3 hierarchy, while only “CONTINGENCY:Result” is found in PDTB2. When mapping from PDTB2 to PDTB3, the new relations, such as “Purpose”, could not be automatically mapped. As a result, a search in Connective-Lex using “CONTINGENCY:Purpose:Arg2-as-goal” as a search option will retrieve six connectives in LDM-PT and nine in DiMLex (both using PDTB3), but none in Eng-DiMLex (see Appendix for the mapping table and Section 4.2.2 for discussion; in future versions of Connective-Lex the goal is to overcome such differences in the sense hierarchy). Also, when connectives have been automatically extracted from annotated corpora, the listing of senses might need revision, and this will have negative effects when searching multilingual data.

59Some lexicons in Connective-Lex are further linked through “synonym” pointers, provided during the manual creation process of those lexicons. For instance, the connective a fim de que is labeled in Portuguese as a subordinating conjunction (“csu”) with the PDTB3 sense “CONTINGENCY:Purpose:Arg2-as-goal”. It is linked to the entry so that in Eng-DiMLex, labeled as “CONTINGENCY:Result:Arg2-as-result”. Although the senses of each connective differ, the linking establishes that they are near-synonyms.

60One important future step is to systematize this linking between the lexicons in Connective-Lex and to enable a “top-down” view on the similarities and differences between the languages. This task raises the issue of the way in which the linking of connectives should be put into place. Two possible approaches are: (i) to link all pairs of languages or (ii) to use English as the pivot language. Linking the lexicons through English as a pivot language is possibly the best solution in Connective-Lex, considering that the set of lexicons is quickly enlarging and that it would prove difficult to link all pairs of languages. In Connective-Lex, the Portuguese lexicon includes a synonym linking to the English lexicon, while the Italian links to a synonym in the German lexicon (the Italian lexicon was translated from the German DiMLex, cf. Section 3.3). In the following subsections, we will explore the linking of the Portuguese and English connectives, as a source of information for the general goal of using English as the pivot language in Connective-Lex.

5.1. Linking connectives in the DiMLex structure

61Considering the set of features and the internal structure of the connective’s entry, what information should be linked? The word form of the connective is clearly not enough because most connectives are ambiguous between one or more rhetorical senses. For instance, in 2.2.2, we refer to the ambiguity of meanwhile and the difficulty in translating to Czech (Meyer & Poláková, 2013). It is thus not enough to map the word forms in two languages, but rather it is necessary to map both the word form and its sense (or one of its senses).

62As mentioned in Section 3.2, in LDM-PT, each entry corresponds to an association of word form/category/sense. So, the same word form will occur in two different entries if it has two different categories or two different meanings, similarly to the structure of LexConn. The Portuguese lexicon in Connective-Lex includes one or more English synonyms for each word form/sense pair. We choose, when applicable, one of the entries of the Eng-DiMLex, compiled from data from the PDTB, and provide the unique identifier of the connective in the English lexicon (Das et al., 2018). The English synonym has an equivalent sense but should also occur in the same context, so it should be the same category when applicable. As the near-synonyms are attached to a specific sense, they are inserted into the DiMLex structure as children of the attribute “sem”. The synonym relation then connects two specific word forms/categories/senses.

63When mapping LDM-PT to the DiMLex format, the individual entries for ambiguous connectives were merged by grouping them, first by word form, then by word class (see Section 3.2). As a result, the synonyms are now presented under the main entry of the Portuguese connective, but they are still linked to a specific category and sense of the connective in LDM-PT. For instance, the adverb consequentemente with the sense “CONTINGENCY:Cause:Result” is linked to the English adverb consequently in the Portuguese lexicon (see Figure 4). The entry in the English lexicon (see Figure 5) shows that the connective has the same category and the same sense as the Portuguese one. Additionally, both examples in the two languages illustrate the use of the connective in a similar position, introducing the proposition interpreted as result.

5.2. Linking ambiguous connectives

64The linking is a bit more complex when the connectives have different senses. In these cases, each sense of the Portuguese connective is linked to a specific English synonym. This is illustrated by desde que, which is linked to since and as long as (see Figure 6). The connective is linked to the English entry and not to a specific sense of the English connective. In most cases, the same sense is listed in the Portuguese and English entries, so the linking with the proper sense is straightforward. However, in the future we should consider the possibility of linking to a specific sense of the English connective, rather than to the full entry. This would guarantee that the linking is fully bidirectional, as the information on word form and sense could be retrieved in both directions.

65An important aspect in the linking process is the underlying perspective towards ambiguity that is adopted by each lexicon. The interpretation of the connectives in context may vary, as a result of the linguistic co-text. On the one hand, a lexicon fully derived from corpus annotation will tend to present a larger set of senses per connective, as a reflection of the decisions made by the annotator to capture the interpretation of the sentence (see Section 3.3 for different methods of creating lexicons of connectives). On the other hand, a lexicon based on corpus data and reference grammars and dictionaries will involve the difficult task of deciding when to include an additional sense. The Portuguese lexicon was derived from the TED-MDB corpus but was then revised to avoid listing a high number of senses in the entries, and the perspective was conservative. Therefore, many entries that are labeled with a single sense in LDM-PT will link to a connective that is highly ambiguous in the English lexicon. For instance, the Portuguese subordinating conjunction ainda que is labeled with the single sense “Arg2-as-denier” and is linked to the subordinating conjunction although, which has four different values in the lexicon (see Figure 7). Notice, however, that the high number of senses reflects the difficulty in deciding between “Contrast” and “Concession”, which leads, in some contexts, to the choice of the higher-level sense “COMPARISON”.

66Another example is the highly ambiguous coordinate conjunction e [and]. It is labeled with three senses in Portuguese (“Conjunction”, “Result”, “Precedence”), while the English equivalent connective and is listed with the same three senses, plus five additional ones (“Condition”, “Contrast”, “Instantiation”, “Concession”, “Arg1-as-detail”) and two higher level senses (“EXPANSION” and “level-of-detail”). Consequently, the Portuguese entry only links to a few of the senses of the connective and.

5.3. Linking using the category attribute

67The examples above have shown that the word form is not enough to ensure the linking between connectives and that it had to be paired with the sense. One question is whether there are any other constraints to be taken into consideration when linking connectives.

68Linking does not strictly involve the association of connectives of the same category. However, these categories reflect certain syntactic properties that may be important to keep in the two languages. For instance, a subordinating conjunction will be followed by a finite clause, while prepositions introduce nominal phrases (and non-finite clauses in Portuguese) and adverbs are usually syntactically detached and may occur in different positions. To ensure that the connective is used in the same context, keeping the same category is the preferred option. The importance of the attribute category in the lexicons was already pointed out when discussing the retrieval of equivalents in Portuguese and German, two lexicons that are not explicitly linked in Connective-Lex. It makes it possible to retrieve sets of equivalent connectives in terms of sense and category. When explicitly linking two near-equivalent connectives, the goal is to select lexical items from the same category, when applicable. For instance, the subordinating conjunction ainda que, labeled as “COMPARISON:Concession:Arg2-as-denier”, is linked to the subordinating conjunction although, labeled as “Concession”. This ensures that they both occur in the same set of contexts and in the same positions: they introduce Arg2, they cannot occur inside Arg2, and Arg2 can occur after (Example [3]) or before Arg1 (Example [4]).

[3]	a.	Este fármaco é usado há já alguns anos de uma forma rotineira nos hospitais portugueses na indução do trabalho de parto e nalguns abortos ao abrigo da lei, ainda que esta indicação não esteja prevista na respectiva bula. (LDM-PT)
	b.	Ms. Levine had never been fired, although she had stopped working at the restaurant. (Eng-DiMLex)

[4]	a.	Ainda que eu não seja ninguém, não valho menos do que esses tais grandes apóstolos.
	b.	Although started in 1965, Wedtech didn’t really get rolling until 1975. (Eng-DiMLex)

69The goal was then to select a connective with an equivalent category in Portuguese and in English, but this was not always possible. For instance, de tal maneira que [literally “in a such a way that”] is linked to as a result. While the Portuguese connective is a subordinating conjunction, highly integrated in the syntactic structure, as a result is an adverb and is used in a parenthetical position. The lexicon provides the possible alternative so, listed as “csu”, among other categories. However, the best match would be so that, labeled with the category “other”. In this case, keeping the category would not provide the best equivalent connective.

70Another mismatch is related to the category “preposition”. Prepositions introduce infinitive clauses with many different senses in Portuguese, while the set of senses of infinitive clauses seems more restricted in English. Purpose infinitive clauses (introduced by a preposition such as para [to], a fim de [in order to]) are easily linked to near-equivalent prepositions in Eng-DiMLex. However, infinitive clauses introduced by a preposition may also convey the sense “Reason” in Portuguese, as illustrated in [5] – the connective is the multiword preposition devido a. There is no available preposition in English in this context, and the translation of [5] uses instead the conjunction because followed by a finite clause. This is another case where the two languages use a different category to express the same sense (and, as a result, a different type of clause in Arg2).

[5]	E, no entanto, o ciclo das fases da Lua demonstrava que esta, tal como a Terra, não produzia qualquer luz e apenas brilhava devido a reflectir a luz solar. (LDM-PT)
	‘And, nevertheless, the cycle of the Moon phases showed that, just like the Earth, it didn’t produce any light and only shone because it reflected the sun’s light.’ (literally: ‘because of reflect_inf the sun’s light’)

5.4. Syntactic, semantic and pragmatic properties

71Other properties may also be important for the linking process. For example, two synonyms with the same sense and category can differ in terms of the positions in which they may occur in the sentence. For instance, the adverb mais is linked to the adverb moreover and both are labeled with the sense “EXPANSION:Conjunction”. They are also both parenthetical elements, separated by a pause (orthographically marked by a comma). However, while mais only occurs as position-initial, moreover may occur in initial position and also internal to Arg2.

72Other restrictions may apply to synonyms. For instance, the two words ainda que and although share the same category, sense and position in the sentence, but they differ in terms of mood restrictions. While the Portuguese conjunction requires the use of the subjunctive mood in Arg2, there is no such restriction holding on the Arg2 of although. This restriction on mood is not visible in Connective-Lex but the attribute is listed in the LDM-PT lexicon. Restrictions on mood are extremely frequent with Portuguese subordinating conjunctions and would have to be dealt with in automatic processing systems, for instance for machine translation.

73Finally, pragmatic properties are also to be accounted for when linking connectives. Section 4.2.1 provided examples of connectives specialized in terms of mode (spoken and written) or medium. Indeed, some connectives are preferentially used in formal or informal registers. For instance, the two connectives pois and porque have the sense “CONTINGENCY:Cause:Reason” but differ because the first one is typically used in formal, written registers. Pois also differs because it expresses a justification rather than a cause, but in informal contexts speakers would still rather use porque [because] for justification. The same semantic distinction holds between since and because, and this led us to link pois to since, and porque to because (although the difference of register in English would have to be confirmed). Similarly, the prepositions devido a, devido ao facto de, em virtude de [because/due to] all convey a “Reason” relation, but em virtude de is clearly used in more formal registers.

5.5. Summary: considerations on linking

74To conclude, although the search options in Connective-Lex can already provide important insights for a multilingual and contrastive approach, connectives have specific properties that make linking the lexicons more efficient. We discussed the method for such linking and, considering the high number of lexicons in Connective-Lex, propose to use English as the pivot language.

75As most connectives are ambiguous between different senses, linking merely on the basis of the word form is not enough. The linking of the lexicons has thus to apply to a word form/sense pair. As the lexicons do not all share the same sense inventory, the best solution would be to link directly the sense of the connective in one language to a specific sense of the English connective. In terms of the XML structure of DiMLex this would mean that the linking is established at the level of the attribute “sem”. This was not applied in the linking of LDM-PT and Eng-DiMLex: the word form/sense pair in the Portuguese lexicon links to the top-level entry of the English connective, with no specific indication of the sense that it links to. This is not an issue when both lexicons use the same hierarchy of senses, nor if there is in the future a mapping solution between different sense inventories.

76Another type of information that proves to be important for the linking process is the category of the connective. Since the category restricts many properties of the connective, the linking should also try to keep the same category to guarantee that the connectives do occur in the same syntactic contexts. The examples above show nevertheless that languages differ in terms of the syntactic structures that are used to express some senses and that the category of the synonym in English might differ. While word form and sense are required elements for the linking, the category attribute is a preferred but not required element. This proves, however, that the category is no doubt an essential attribute for the lexicons. In many cases the category will determine the position of the connective, although not always. This is especially true for adverbs, which may be more or less restricted in terms of their position (initial, internal or final in the second argument). But, as mentioned in 4.2.1, sentence positions of the connectives are language-specific and it would be difficult to establish a common descriptive system for all languages in Connective-Lex. Nevertheless, if possible, this would prove very useful for the linking process. Finally, linking might also consider additional information such as mood and register. Although not required, adding this level of granularity would certainly improve the linking. The recommendation for the linking process would then be to consider the levels of word form and sense as the minimum required information, plus the category when applicable.

6. Conclusion

77Even though connectives do not form a syntactically homogeneous class, they can be meaningfully analyzed by means of the standard repertoire of relations in lexical semantics: we find synonymy (e.g., German obwohl, obzwar), plesionymy (e.g., although, though), antonymy (e.g., if-unless or before-after), hyponymy (e.g., but, although), and polysemy (e.g., while). This is one indication why taking a broad lexical perspective on these items can be productive. Connective-Lex.info is intended as a starting point for collecting information about these lexical items, and for additionally adopting a multilingual perspective and building bridges between languages. While our initial set of common attributes to describe connectives across languages is relatively small, in this paper we have hinted at various possible directions for adding “depth” to the descriptions. For example, we mentioned register preferences (Section 4.2) and mood restrictions (Section 5.4).

78On balance, we found that the definition of “connective”, stemming from the work on German by Pasch et al. (2003), appears to cover what we found in other languages too, and thus functions as our language-neutral working definition. Likewise, the minimal set of attributes for syntax and semantics, derived from a reduced version of the original DiMLex, also appears to be useful for the other languages.

79From a methodological viewpoint, lexical description and corpus-based research can mutually benefit from each other: as pointed out in Section 2.2, several approaches have been proposed to derive connective lists (which can be the starting point for a lexicon) from mono- or bilingual corpora (e.g., Versley, 2010); likewise, an existing list or lexicon can be verified and extended by mining distributionally similar words from corpora (e.g., Bourgonje et al., 2017). Moving beyond this initial phase of creating inventories, features of lexical description can be tested with corpora and lexical entries in turn be improved. Finally, we plan to include frequency information into the lexical entries, as one way of reflecting degrees of ubiquity of connectives, their different senses, and possible genre/register dependencies. While we have compiled some information of this kind for the German DiMLex already, it is not straightforward to generalize this across languages: as soon as similar kinds of numbers appear in the different languages of Connective-Lex, one has to be clear whether and in what way they can be meaningfully compared.

80One central question for extending Connective-Lex is the means of linking the entries across languages. The principal alternatives are establishing pairwise correspondences, or declaring one language as the pivot to which the others are related. Both approaches have their advantages, depending on the use case: precise comparisons of related connectives in two languages (cf. Section 2.2) can yield detailed insights which could considerably improve the usefulness of Connective-Lex for purposes of translation studies, if it is possible to systematically represent such information in a complex linkage scheme. The pivot approach – where presumably English is the most useful candidate – on the other hand would make the system more useful for language learners, since corresponding connectives of any two languages can in principle be found via a commonly understood “interlingua”. Of course, one has to be aware of limitations: the mapping from language x to English will be only an approximation; the mapping from English to language y will also be approximative; and then the comparison between languages x and y is bound to be more approximative.

81For both the direct connections or the pivot connections, as we pointed out in Section 5, the meaningful units for a mapping need to be defined, which is not trivial due to the ambiguity of connectives, and its specific representation in the lexicons. We argued that word form/sense pairs are the most promising units, and we plan to use these in our future work on enhancing the language links in Connective-Lex. The overall goal is to provide more fine-grained correspondences than the current implementation of the search interface does (i.e., the ability to retrieve connectives across languages that share the same sense and/or the same syntactic category).