3.1 Anaphoric Connectives and PoS

Surprisingly, among the 20 most frequent Czech connectives with a non-adjacent external argument, 10 are coordinating conjunctions,⁶ which are structural connectives and should not accept non-adjacent external arguments.

There are several possible explanations for this behaviour. First, the issue may lie in the definition of a coordinating conjunction itself in different languages. There is a well-known tendency in the diachronic development of some adverbs, possibly in connection with demonstrative pronouns, towards sentence adverbs and gradually to conjunctions (see e.g. ^[18], p. 153-155).⁷ In contrast to English grammar, where the strict coordinating conjunction category only contains and, but and or (e.g. ^[14], p. 920), the tradition of Czech PoS categorization also includes historically adverbial/pronominal expressions, the syntactical behaviour of which is nevertheless in contemporary Czech equal to those of conjunctions.

Second, for the task of Arg1 detection in ^[11], the sentence-initial But-adverbial was introduced, as also the annotations confirm long-distance relations for even the basic coordinating conjunctions. In the PDT 3.0, a very frequent coordinating conjunction však [but, however] is to our surprise more frequently used as an inter-sentential (1,120) than intra-sentential (236) connective. Moreover, in absolute numbers it is the most frequent connective with non-adjacent external argument (113 tokens) in the corpus.

Third, according to ^[17], structural discourse connectives allow "stretching", similarly as syntactic dependencies within a sentence allow long-distance by embedding constituents. The interpretation may also be that structural connectives allow non-adjacent external arguments via (syntactic) stretching, not via anaphora resolution. Also another study of (German) ACs reports that the absence of an explicitly-anaphoric morpheme in the connective does not exclude its anaphoric behaviour ^[16].

As a practical application here, we suggest (and the more for experiments with non-English data) to also work with coordinating conjunctions as possible anaphoric connectives and to be critical to the outcomes of a PoS tagger. Also, detection of such inter-sententially used conjunctions might be not trivial, as, at least in Czech, they may not stand at the sentence-initial position, see Example 3.

(3) "Já to nevyhrál za svých šestnáct let závodění, já totiž žádné peníze nikdy nedostával. Za různé prémie a etapová vítězství jsem ovšem měl tolik aktovek a necesérů, že bych je mohl prodávat. Také nějaké ty tepláky jsem vyhrál," vzpomněl Veselý. "Na jednu stovku si ale přece jen dobře pamatuji.

["I did not win it in my sixteen years of racing, I never got any money at all. For various bonuses and stage victories, I have won so many briefcases and washbags that I could sell them. I've also won some sweatpants," Veselý remembers. "But those hundred crowns, I still remember them well.]

Lit.: On one hundred reflex.pron but still well I-remember.

3.2 Types of "Gaps"

For a more detailed insight, we analyzed 245 tokens of the most frequent connectives with non-adjacent discourse arguments manually (70 tokens of však and all tokens of ani, dále, také, ale, přesto, proto and přitom), according to their relative frequencies and across semantic classes. We concentrated on their positions with respect to paragraph boundaries, reported speech zones and we classified the nature of the "gaps", i.e. the text segments left out of the relation. Our observations are displayed in Table 2.⁸ The detailed corpus analysis reveals that long-distance relations in the PDT 3.0 can be divided into two general groups of thematic patterns (or progressions): First, it is mostly a general statement/claim in the external argument, a certain type of elaboration in the gap, and a return or strong link to the first topic in the internal argument. Often, the elaboration in the gap zooms in to a specific detail or background information or gives an example.

The second group are digressions in the gaps. It is marked parentheses (in brackets, dashes), but much more often unmarked, and so difficult to detect, comments on the topic by the writer or other person, switching between the plan of the writer and the plan of reported content (reported speech appears in the journalistic data of the PDT often without quotation marks), and also technical digressions like author names, photo captions, subheadings.

The practical difference between these two types of gapping is their referential linkage to their closest text environment. For digressions, less coreference and associative anaphora is expected, sometimes even none (see Section 3.4 below).

3.3 Local Coherence and Higher Discourse Structure

It can be supposed that arguments of connectives in paragraph-initial sentences are more likely to be distant, but also to be represented by larger blocks. For the long-distant relations in the PDT 3.0, a connective in paragraph-initial (ParInit) sentence takes another ParInit sentence as its argument in 15.1% (96/636), and 18.4% (53/288) in the subset described in Table 2. In these specific cases it can be very difficult to decide, whether they are indeed long-distance discourse relations or whether to interpret them as relations between higher discourse segments (paragraphs) that are in fact adjacent. The issue in the local coherence annotation in the PDT may be the annotation rule called the minimality principle: annotators were instructed to include in an argument as many clauses and/or sentences as are minimally required and sufficient for the interpretation of the relation. In the PDT, no supplementary information was annotated (compare ^[13], p. 14), which could potentially lead to misinterpretation of cases of paragraph coherence. It is nevertheless a problem of analytical perspective, a point where local and global discourse analyses clash and each such case should be judged individually. In the studied dataset, at least 9 relations had both relevant interpretations and there may be more.

3.4 Non-Adjacency across Semantic Classes

The distribution of the four main semantic classes (Temporal, Expansion, Comparison, Contingency) in long-distance relations in the PDT 3.0 is very uneven. There are only 42 (6.6%) Temporal and 71 (11.2%) Contingency relations, whereas the relations of Expansion and Comparison with 261 occurrences (41%) and 262 (41.2%) are much more frequent. Additive and contrastive connectives are thus much more likely to take part in these relations, but also, from the viewpoint of a global analysis (e.g. RST, ^[7]), these types of connectives can be expected more often in ParInit positions or even relating individual paragraphs. These findings correspond to the nature of the relations: causal, conditional or temporal relations require proximity of their arguments. This is often secured by syntax and by the use of subordinating conjunctions, and inter-sententially by adjacency. Although long-distance is also possible, these relations appear, at least in the studied data, less flexible to embedded contents. Furthermore, arguments of the additive connectives in our survey daile [next, further] and také [also, too] show specific patterns which relate to semantics of the relation: they have parallel syntactic patterns with referential identity of subjects (that might be interrupted in the gap) or they contain identical or synonymous verbs forms. Dále takes part in 14 cases of the type He said - He further commented and in 13 enumerative-like structures with sequences like First - then - next.