This study uses data-driven methods to detect and interpret differences between the High German used as standard language of written communication in Switzerland, and German High German. The comparison is based on a comparable web corpus of two million sentences, one million from Switzerland and one million from Germany. We describe differences at the levels of lexis, morphosyntax, and syntax, and compare to previously described differences. We show that data-driven methods manage to detect a wide range of differences.
While the various spoken dialects of Swiss German
differ considerably from High German, and also from
each other, the differences between Swiss High
German and German High German are relatively small,
and mostly concern the level of lexis and
morphosyntax. Many of the differences have been described in
lexica, see e.g. Meyer (1989). Also some
syntactic differences have been reported
We apply a data-driven method to the detection of
differences. In data-based approaches, existing
hypothesis are tested, whereas in data-driven approaches,
hypotheses arise from the data. Data-driven methods
have the advantage that previously unnoticed
differences may be detected, thus improving the recall of
the phenomena under observation, potentially
showing new patterns that one is not yet aware of, and it
also allows one to put the differences into a
quantitative perspective. Data-driven methods also have
disadvantages, in particular that they depend directly on
the corpus and its sampling
The Wortschatz Leipzig corpus1 consists of a collection of news and web-derived corpora, each comprising one million sentences. For the comparison of Swiss High German to German High German, we use their Swiss and matching German web corpus, which contain random texts from the year 2002. The Swiss corpus contains 15.817.004 words, the German one 16.850.144 words. 2.2
As pre-processing steps, we reduced full-forms to
their lemmas and applied part-of-speech tagging. For
these two steps we used Treetagger
which employs the STTS tagset
We use a syntactic dependency parser
Document classification is a supervised method
generally used to assign each document, whether a
newspaper article, a web page, a book, a paragraph, a
tweet, or a similar discourse unit, to a class. Classes
can, for example, be broad topics divided into the
binary classes of relevant or irrelevant documents for an
Information Retrieval task (see Jurafsky and Martin
(2009, chapter 23.1)) or Manning and Schu¨tze (2001,
chapter 25)) for an introduction). In the majority of
the implementations, the words in the documents are
used as discriminators between the classes, typically
without respecting their sequence or syntactic
context, which is why the method is called a
“bag-ofwords” approach. Since every word type (as soon as
it reaches a token frequency above a certain
threshold) is a feature, there are often thousands of features.
Each feature in isolation is usually neither a good
descriptive feature nor a good discriminator between the
classes. The simplest approach, Na¨ıve Bayes, simply
gives equal weight to each feature. More advanced
algorithms, for example logistic regression, which we
use in the present study, give optimal weight to each
feature. While most features in isolation are bad
discriminators, some are better and logistic regression
automatically finds the optimal weight (also called
influence) for each feature. Those features that obtain
a high weight are relatively good discriminators and
therefore they can be considered typical of their class.
These words can be interpreted as keywords, because
document classification is also a possible keyword
extraction algorithm
For the detection of English National Dialects, Lui and Cook (2013) have tested a range of methods. They conclude that document classification performs best on the task of detecting the originating nation of a text, they state that this is probably because of the very large feature set of this method. The main interest of Lui and Cook (2013) was to obtain a high classfication accuracy, they were not mainly interested in lingustic interprations of the features. 2.4
As overuse metric we use O/E and derived measures.
O/E stands for Observed divided by Expected, where
Expected is the homogenous distribution over the
entire corpus. The value gives a direct, and easily
interpretable effect size. O/E is often affected by sparse
data problems which can lead to inflated values for
items with low counts. One thus sometimes gets a
more accurate impresssion by adding a frequency
factor, for example O2/E or O ∗ log(O)/E. O/E is well
known from research on collocations
In a first classification task, we applied document
classification to the raw texts. Our tool of choice for
applying the method is LightSide2 because it is easy to
use, includes tokenisation, and offers a wide range
of machine learning algorithms, including logistic
regression from the LIBLINEAR library
While the performance of the system is nearperfect (only one document was misclassified), the strongest features are dominated by place names and proper names. Therefore, as a second classfication task, we thus removed all proper names (tag NE), and we also replaced full forms by lemmas. The classification is still very accurate (99.97% accuracy, 8 documents are misclassified), and a large subset of the top
Comment Relative Pronoun zu¨richer baseler (dialect word not recognized as proper name) zufolge veranstaltung etwaig selbst binnen, innerhalb jahreshauptversammlung vorsitzende/r krankenhaus ansonsten aufgrund neben; ausserdem unterschiedlich anwenden lehrgang anders erfordernis kostenlos gesamtschule urlaub schnell hinsichtlich versta¨rkt, geha¨uft zum vorteil von begehren abschliessend, bestimmt fahrrad satzung vero¨ffentlichen
O ∗ logO/E 4.009 4.696 4.872 5.503 4.325 5.570 4.450 3.969 5.942 5.624 6.149 5.694 5.644 6.004 4.860 4.863 4.833 6.378 5.856 2.949 4.944 4.476 5.029 2.934 1.248 3.218 2.655 2.521 1.150 1.651
Comment Relative pronoun welche/r/s Present perfect Contraction wo, wobei Present perfect Hedging phrases, Swiss indirectness? Present perfect det+proper name paratactic style Present perfect synthetic, genitive drop, fewer postpositions Hedging phrases, Swiss indrectness? shorter sentences more postpositions in German High German Acronyms such as KFZ fewer wo, wobei, fewer relatives with welcher German directness? / short forms absent in CH, rare in DE German directness? / short forms features are linguistically meaningful, exhibiting helvetisms. Table 1 lists the top 72 features (of a total of 21789). We have manually filtered non-linguistic features (such as the adjective Schweizer) in this list3, but give the position in the original list in the first column, to give an impression of the level of noise. The last column gives our interpretation, explanation, or near-synonyms which are overused in the German High German corpus, and at least partly explain the high weight of the feature in the Swiss German corpus.
The top entry is the lemma welch, stemming from the full forms welche, welcher, welches which are strongly overrepresented in Swiss High German, while the relative pronoun forms der, die, das are used less often than in German High German. The preposition gema¨ss is overrepresented in Switzerland, while the semantically largely corresponding pre- or postposition zufolge is a strong feature of German High German. The majority of the features can be explained, but data-driven approaches also lead to some results which are difficult to explain. Words related to education and schooling (for example Schu¨lerin) seem generally overrepresented in the Swiss data. This may be due to a bias in the corpus collection or due to the importance that Switzerland gives to education.
Further down in the list than shown in Table 1, we can still find many lexical differences, for example the Swiss Velo for Fahrrad, Offerte for Angebot, benu¨tzen for anwenden, Gesuch for Antrag, selber for selbst, Reservation for Reservierung, Mitgliederbeitrag for Mitgliedsbeitrag, Ferien for Urlaub, and hundreds more.
Most of the reported lexemes are not exclusive to one variety, but exploring the feature weights offers an exciting resource to the lexicographer. We can equally browse the strongest German High German features and learn for example that German High German (over-)uses Personalausweis for Swiss Identita¨tskarte, gezahlt for bezahlt, zeitnah for bald, Stadtmitte for Zentrum, PKW for Auto, Festplatte for Harddisk, Rundfunk for Radio, Renovierung for Renovation, etc. Many of these differences are known, but a data-driven approach allows us to verify lists and complete dictionaries.
3zu¨rcher and basler are kept, because they are linguistically meaningful, German High German would use zu¨richer and baseler 3.2
Some morphosyntactic differences are also well known. For example, due to the fact that the Swiss dialects do not use the simple past tense, overuse of the present perfect can be expected. To obtain a more complete picture, we have sorted all part-of-speech tags by overuse metrics. The results sorted by O/E are given in Table 2. The expected overuse of present perfect is mirrored by more auxiliary verbs and participles (VAINF, VAFIN, VVPP). The table reflects relative pronouns with welch again (PWAT), and with wo (PWAV). The increased frequency of PWAV is partly also due to wobei, which is described as a Swiss feature in Du¨rscheid et al. (2015, 228). Determiners (ART) are more frequent in Swiss texts because proper names are often preceded by determiners. The fact that infinitive verbs with particlezu (PTKZU, VVIZU) are more frequent in the Swiss corpus is due to the frequent use of fixed semi-modal, hedging phrases: (1) Es PPER ist VAFIN anzunehmen VVIZU , $, dass KOUS das ART Quartier NN etappenweise ADJD u¨berbaut VVPP wird VAFIN . $. (2) Zuna¨chst ADV ist VAFIN festzuhalten VVIZU , $, dass KOUS der PDS vermeintlich ADJD ” $( normale ADJA ” $( Zustand NN nicht PTKNEG existiert VVPP . $. (3) Der ART Name NN ist VAFIN zuru¨ckzufu¨hren VVIZU auf APPR ein ART Treffen NN von APPR Veteranen NN der ART American NN Legion NN 1976 CARD in APPR den ART USA NE . $.
Semi-modal patterns of the type es ist VVIZU are 50% more frequent in the Swiss than in the German corpus.
It is tempting to interpret these phrases in contrast to the fact that imperatives (VAIMP, VMIMP, VVIMP) are among the strongest German High German features. Inspecting the data reveals, however, that the majority of verbs tagged as imperative are in fact short forms, not imperatives, such as (4) hab VAIMP ich PPER einen ART von APPR der ART
Sicherheit NN getroffen VVPP . $.
Data-driven approaches are relatively susceptible to skews and systematic errors in the data and the annotation process.
The Swiss texts have shorter sentences, therefore the full stop tag $. is also overrepresented in the Swiss data. As Swiss sentences are shorter, the Swiss corpus is also a bit smaller, which explains why most O/E values are slightly below 1. 13 12 12 17 21 8 11 4 3 3 8 3 1 1 0 0
Prepositions seem to show important dfferences. Contraction of prepositions plus article (am, im, beim etc.) are more frequent in the Swiss variety, while postpositions (APPO) are a German High German feature. We have already seen in the comparison of lexis that the preposition gema¨ss has the semantically largely corresponding pre- or postposition zufolge in German High German, there may be an interdependence. As prepositions (APPR) appear among the overused words, we list results of the uses of individual prepositions in Table 3. Some contractions are more typical for Swiss (im, am, vom, zum, zur, beim), while others are typically German (hinterm, unterm, vorm, u¨bern, u¨berm).
The higher frequency of am can also be explained
by the frequent use of the progressive form am, e.g.
Ich bin am Laufen
As Swiss German has no genitive form, genitive drop is probably also frequent in Swiss High German, explaining the overuse of von and vom. But in order to investigate this question, syntactic information is needed, which we provide in the following subsection. 3.3
The overuse of the tags VVIZU and PTKZU in Swiss High German, and the frequent conjunctions (KON) are stylistic features. Also many of the syntactic differences can be seen as differences in style. On the surface, it is already noticeable that German High German sentences are considerably longer (mean of 15.85 words per sentence) than Swiss High German ones (mean of 14.82 words). Sentence length and complexity are usually strongly correlated. One feature of sentence complexity is the use of subordinating and coordinating clauses. On the level of tags, we can see in Table 2 that conjunctions (KON) are even overused in Swiss High German, which is surprising. Looking at the parsed data can give more deVariety DE CH tailed answers. A comparison of important syntactic relations is shown in Figure 1. Subordinate clauses are expressed by the dependency relations (konj)neb for adjunct clauses and (konj)objc for complement clauses. Particularly adjunct clauses are indeed underused in the Swiss texts. There seems to be a slight trend towards a more paratactic style at the expense of hypotactic style.
Due to the fact that Swiss High German uses more present perfect forms instead of the simple past, the higher frequency of aux dependencies is expected.
There is an overuse of det, which is due to proper names with determiners, as seen in overused POS tags.
Subjects (subj) are distributed homogeneously. Accusative objects (obja) and dative objects (objd) are almost as frequent in Swiss High German as in German High German, but genitive objects (objg) are considerably rarer. Prepositions and verbs governing a genitive object, as in the following German High German examples, are rarer in the Swiss texts. (5) Trotz APPR des ART Zugewinns NN ging VVFIN
Karl NE Braun NE leer ADJD aus PTKVZ . $. (6) Da ADV machte VVFIN sich PRF auch ADV ans APPRART Werk NN das ART Ratsgymnasium NN , $, denn KON es PPER gedachte VVFIN des ART alten ADJA Schlagers NN ” $( Fu¨r APPR Gaby NE tu VVFIN ich PPER alles PIS ” $( . $.
Verbs governing genitive objects are also receding in Standard German (see e.g. Ueberwasser (2014); Scha¨tzle (2013)), but prepositions governing genitive case remain stable across time. Nouns can be modified by genitive NPs (dependency labelgmod), a phenomenon which is stable in Standard German, but less frequent in the Swiss data, where genitive modifcation is partly replaced by prepositional phrases with von, as in the following example: (7) Der ART Bruder NN von APPR Frau NN Dreifuss NN ist VAFIN u¨brigens ADV Tierexperimentator NN ! $.
The fact that the genitive case after prepositions is often replaced by a dative (e.g. wegen, trotz, see Table 4) in Swiss High German could lead one to expect higher counts of dative objects (objd), which is not the case. An important reason for the lack of increase is the fact that the dative object is itself under threat in German dialects, it is often replaced by the preposition an. German (like English) has a dative shift alternation, see e.g. Adler (2011). The difference is often seen as formal vs. informal, due to type of event or dialectal influence, also in German High German, but we find significantly higher counts for structures like the following in the Swiss corpus: (8) Hier ADV koennen VVFIN Sie PPER Ihre PPOSAT Feriengruesse NN an APPR die ART Welt NN senden VVINF . $.
Although more research is needed on this question, such factors can explain the higher frequency of prepositions (tag APPR, dependency label pn). Coupled with the increased use of auxiliary verbs, Swiss High German also seems to be slightly more synthetic while German High German is slightly more inflectional.
Another observation related to language and dialect typology is illustrated in example (3), which shows a slightly unexpected verbal brace, where zuru¨ckzufu¨hren does not appear at the end of the clause. While this choice can be caused by end-weight constraints, we wondered if there is a difference between Swiss High German and German High German. We counted how often objects and PPs of the main verb occur to the left or the right of the main verb. Frequencies to the right are considerably higher in the Swiss corpus, but closer inspection revealed that this is mainly due to the fact that there are fewer subordinate clauses in the Swiss data. In subordinate clauses, the default position of the main verb is clause-final, while in main clauses the position is after the subject or fronted elements (verb-second). 4
We have shown that automated data-driven methods manage to detect a wide range of the differences between Swiss High German and German High German described in the literature, by applying document classification and overuse metrics to a large web corpus, and shown that automatic part-of-speech tagging and syntactic dependency annotation detects patterns beyond lexis. Candidate lists from the relatively easy levels of lexis, but also from the more intricate levels of morphosyntax, syntax and style have been illustrated, also unveiling subtle stylistic differences. We have pointed out strengths and possible pitfalls of the method. In future work, we plan to analyse the reported candidate lists in further detail.