English. We report on experiments to validate and extend two language-specific connective databases (German and Italian) using a word-aligned corpus. This is a first step toward constructing a bilingual lexicon on connectives that are connected via their discourse senses.
An important part of discourse processing deals with uncovering coherence relations that hold between individual, “elementary” units of a text. The lexical items that can signal such a relation are referred to as discourse connectives, and examples of these relations, also called the connectives’ senses, are contrast (e.g., ‘but’), elaboration (e.g., ‘in particular’), or cause (e.g., ‘therefore’). Notice, however, that relations need not always be signalled in text, if the context or world knowledge is sufficient for the reader to infer it, as (1)(4) demonstrate: (1) We should hurry, because it’s late. (2) We should hurry. It’s late. (3) The red pen costs $2, while the blue one is $2.50. (4) The red pen costs $2; the blue one is $2.50. On the other hand, example (6) is a perfectly grammatical sentence but the meaning is different from (5), so for this case of a Concession relation, the connective is in fact indispensable. (5) Although it is late, we don’t need to hurry. (6) It is late; we don’t need to hurry.
Recognizing these relations, which can hold
within a sentence, between two sentences, or
between larger spans of text, is a central task for
uncovering the structure of a text, as it has been
studied in theories like Rhetorical Structure
Theory
From a syntactic viewpoint, ‘connective’ is not a
homogeneous class, as it contains conjunctions,
different kinds of adverbials, as well as certain
prepositions. Our underlying definition of
discourse connectives is based on
Following
Connectives can pose interesting challenges to translation and for language learners, as the differences in meaning between similar connectives can be quite subtle. For these reasons, we are interested here specifically in a bilingual Italian– German lexical resource, to be built on top of two existing single-language lexicons. As a case study, we focus on the subgroup of contrastive/concessive connectives, which we determined to comprise (in the existing lexicons) 31 German connectives and 12 Italian connectives; see Tables 3.2.2 and 3.2.2.
The main contributions of this paper are (1) suggestions for improving the existing languagespecific resources used in this study through the technique of cross-lingual projection in a parallel corpus, which reveals correspondences between connectives and can point to gaps in either of the resources; and (2) an overview of the distribution of connectives and their senses, to be used in a bilingual database. Section 2 explains the two monolingual lexicons we work with, and Section 3 describes the corpus. Section 4 reviews related work in this area. Section 5 elaborates the idea of bilingual connective databases, and Section 6 summarises our findings. 2
We extracted the German contrastive connectives
from DiMLex
The set of Italian contrastive connectives comes
from LICo
For the parallel German/Italian corpus we used
Europarl
We approach the problem from two sides: First we look up every German connective (31 in total) to get Italian alignments. 30 of them appeared in our Europarl corpus (with dementgegen missing). Then we look up every Italian connective to get German alignments (all 12 connectives present in the corpus). We end up with a list of target language words or phrases (or empty elements, since a source language connective can also be covert in the target language) that are candidate contrastive connectives. Note that the lookup procedure does not differ structurally between words and phrases. In both cases, single words (stand-alone or in a phrase) can correspond to zero, one or more target words. The target representation is collected in a key-value structure, where the key is the position in the sentence and the value the word. This list is then sorted by position to return the target word or phrase (which is potentially discontinuous). Because the word alignment is not guaranteed to be correct, to filter for unlikely translations we focus on only the 3 most frequent alignments for every connective. We expect to find at least a subset of the already known (contrastive) connectives (from DiMLex or LICo), potentially complemented by a set of words or phrases that can help filling gaps in either of the lexicons.
This procedure produces at least some incorrect results for the following two reasons: 1) discourse connectives often can appear in a text with a connective reading or with a non-connective reading; and 2) connectives can have multiple senses, so that a connective may not have the contrastive reading in the particular sentence. The candidates produced hence have to be evaluated manually. Resulting candidates that have a connective reading are added to the seed list, in order to repeat the step back from the target language to the source language3. 3.2
The results of the first step of the iteration using the 31 German seed connectives are displayed in Table 3.2.2, where an underscore indicates an empty string (meaning that the connective was not aligned to a particular word or phrase in the target language) and the number after the underscore represents the (normalised) frequency of the alignment.
For the evaluation, we asked a native speaker of Italian with expert knowledge in linguistics to validate the resulting top 3 bilingual mappings. Firstly, we identified several possible connec3Ideally going back and forth until a stable and exhaustive set of candidates is found. For this study, we only did the first step, and then projected the found Italian connectives back to German. tive candidates that were aligned to German contrastive connectives, but were not present in LICo, such as al contempo, solo che, doppo tutto. Secondly, we observed several possible orthographic variants of the already existing Italian connectives: contro or contrario (as possible variants of al contrario), and d’altro canto (as a variant of a discontinious connective da un canto...dall’altro). Finally, we found that several Italian connectives only had the concession sense, while the corresponding German connectives also had the Contrast sense, such as comunque, for which we found the German alignments aber, allerdings and doch, for example.
As an example of a visualisation (for a single connective) the above analysis is based on, consider Figure 2, showing the most frequent alignments of jedoch, which always has a connective reading, thus nullifying the first problem mentioned in 3.1.
The results of the first step of the iteration using the 12 Italian seed connectives are displayed in Table 3.2.2. For 11 of the 12 contrastive connectives from LICo, the top 3 alignments yielded an existing DiMLex entry. The only connective without a DiMLex entry in the top 3 was al contrario, for which a possible new German connective candidate im Gegenteil was found through alignment.
Upon further investigation of the lower-ranked alignments (not included in Table 3.2.2), we were able to identify several other gaps in the German lexicon. Firstly, we observed that the Italian connective invece is frequently aligned to the German word anstelle, which is not in DiMLex (but anstelle dessen is). After examining the corresponding examples, we conclude that anstelle should be added to DimLex as a separate entry (similarly to the already existing aufgrund vs. aufgrund dessen). Also, we found that DiMLex lacks statt dessen as an orthographic variant of the more canonical stattdessen.
Finally, we identified two interesting cases that are DiMLex candidates: umgekehrt and (ganz) im Gegenteil, which we found aligned to the Italian viceversa and al contrario, respectively, but more corpus evidence is required to decide whether they can indeed serve as connective in the German language.
As an example visualisation, consider Figure 3, showing the most frequent alignments of invece, which always has a connective reading.
For Italian–German, we repeated the steps above with the candidates found using the German seed list (projecting the resulting Italian list back to German) to see if any additional connectives or orthographic variants would be found. We again found im Gegenteil through alignment of al contrario and a few alternative lexicalisations for DiMLex connectives4, but no new candidates. 4Not listed here for reasons of space. Parallel corpora have been successfully exploited before in order to automatically generate or induce connective lexicons in different languages. In particular, Versley (2010) projected discourse connectives across an English–German parallel corpus to train a discourse parser capable of disambiguating connective and non-connective readings. Similarly, Zhou et al. (2012) used an English–Chinese parallel corpus in order to build a Chinese connective lexicon via cross-lingual projection, and Hajlaoui and Popescu-Belis (2013) relied on parallel data to automatically retrieve Arabic counterparts for a subset of English connectives.
Since our goal was not to build a connective lexicon from scratch, but to extend the connective lists and refine the inventory of senses for the already existing lexicons, the closest approach to ours is the one adopted by Laali and Kosseim (2014), who aimed at automatically inducing a French connective lexicon via English–French parallel corpora using additional filtering rules. Similar to ours, their results have shown that using parallel translations can improve the coverage of the connective lists in both languages; however, since their lexicons used different sets of discourse relations, they were not able to extend their connective database in respect to senses, as opposed to our work. 5
Our study is meant as a step toward moving from single-language connective lexicons to a bilingual one that provides information about the relationships between the language-specific entries. Both monolingual lexicons are already publicly available on GitHub and in addition an interface allowing bilingual search has been made public in a related project5. Below we sketch additional plans for providing this information on the levels of connective tokens, and senses (coherence relations). 5.1
One central purpose of a bilingual database is to assist translators (human or machine) or (human) language learners. For most connectives, there is a complicated m:n mapping between languages, which standard dictionaries do not cover, and the relevant features for making choices are not systematically known yet. A corpus-based inventory of mappings – ideally supplemented by pointers to the corpus instances and their context – can be a very useful resource for undertaking contrastive lexical investigations. 5.2
The PDTB
A bilingual connective database can shed light on the distribution of senses over different languages and the degree of ambiguity that individual connectives exhibit. While we consider such conclusions premature for the current stage of the language-specific resources, we include Figure 4, which shows groups of connectives that share the same sense (or group of senses for ambiguous connectives) and their alignment to similar groups on the target side. The 12 Italian connectives (on the left), when grouped together based on their sense(s), form 4 sets, whereas for German (right side), fewer connectives (11 that were found in DiMLex among the top 3 alignments of the 12 source connectives) group into more sets (10). This suggests more ambiguity in Italian connectives, with less different senses represented by a larger set of connectives.
In addition, we observed that Italian connectives with a sense Contrast or Concession are frequently aligned to their German counterparts with a sense Substitution, such as anstelle-invece. Having examined the parallel examples more closely, we conclude that assigning both senses would be valid for both German and Italian, although they are placed distantly in the PDTB hierarchy of senses. These findings are confirmed by Feltracco et al. (2016), who acknowledge that the distinction between the two senses was one of the main cases of the inter-annotator disagreement. We conclude that both lexicons could benefit from adding additional senses gained via comparing parallel translations. 6
We present, to the best of our knowledge, the first Italian–German investigation of discourse connective lexicons. For the subclass of Contrast (in a wide sense), we were able to identify several missing entries in both lexicons, and provided a start on identifying AltLex items for the two languages (future work). Once the information is organized in a complete bilingual database, it can assist translation and conclusions can be drawn regarding connective distribution, sense distribution and ambiguity in the different languages.
As prominent steps for future work, we note the disambiguation of connective- and non-connective readings, the implementation of more sophisticated filtering strategies to retrieve more reliable connective candidates and repeating this study for different languages pairs.
We are grateful to the Deutsche Forschungsgemeinschaft (DFG) for funding this work in the project ‘Anaphoricity in Connectives’. We would like to thank the anonymous reviewers for comments on an earlier version of this manuscript. We also thank Flavia Adani for her help with translation and interpretation of the Italian results, and Tatjana Scheffler for the recent work on DiMLex.