English. This paper presents first results of an ongoing work to investigate the interplay between lexical complexity and syntactic complexity with respect to nominal lexicon and how it is affected by textual genre and level of linguistic complexity within genre. A cross-genre analysis is carried out for the Italian language using multi-leveled linguistic features automatically extracted from dependency parsed corpora.
Linguistic complexity is a multifaceted notion
which has been addressed from different
perspectives. One established dichotomy distinguishes a
“global” vs a “local” perspective, where the
former considers the complexity of the language as a
whole and the latter focuses on complexity within
each sub-domains, i.e. phonology, morphology,
syntax, discourse
An interesting question addressed by recent corpus-driven research is how language complexity is affected by textual genre. At syntactic level, the study by Liu (2017) on ten genres taken from the British National Corpus showed that genrespecific stylistic factors have an influence on the distribution of dependency distances and dependency direction. Similarly for Italian, Brunato and Dell’Orletta (2017) investigated the influence of genre, and level of complexity within genre, on a range of factors of syntactic complexity automatically computed from dependency-parsed corpora. Inspired by that work, we also intend to analyze the effect of genre on linguistic complexity. However, unlike the dominant local approach, where each subdomain is typically studied in isolation, our contribution intends to address the interrelation between different levels, i.e. lexicon and syntax. Specifically, we investigate the following questions: to what extent is lexical complexity influenced by genre? to what extent is lexical complexity influenced by the level of complexity within the same genre? is there a correlation between lexical complexity and syntactic complexity? Does it vary according to genre and level of complexity within the same genre?
To answer these questions, we conducted an indepth analysis for the Italian language based on automatically dependency parsed corpora aimed at assessing i) the distribution of simple and complex nominal lexicon in different genres and different language varieties for the same genre ii) the syntactic role bears by “simple” and “complex” nouns characterizing each corpus iii) the correlation between “simple” and “complex” nouns with features of complexity underlying the syntactic structure in which they occur.
In what follows we first describe the corpora considered in this study. We then illustrate how lexical and syntactic complexity have been formalized. In Section 4 we discuss some preliminary findings obtained from the comparative investigation across corpora. 2
Four genres were considered in this study: Journalism, Scientific prose, Educational writing and Narrative. For each genre, we chose two corpora, selected to be representative of a complex and of a simple language variety for that genre. The level of complexity was established according to the expected target audience.
The Journalistic corpora are Repubblica (Rep) for the complex variety, and Due Parole (2Par) for the simple one. Rep is a corpus of 232,908 tokens and it is made of all articles published between 2000 and 2005 on the newspaper of the same name; 2Par contains 322 articles taken from the easy-to-read magazine Due Parole1, for a total of about 73K tokens.
The corpora representative of Scientific writing are Scientific articles (ScientArt) for the complex language variety, and Wikipedia articles (WikiArt)
for the simple one. The former is made of 84 documents (471,969 tokens) covering various topics on scientific literature. The latter is made of 293 documents (about 205K tokens) extracted from the Italian web portal “Ecology and Environment” of Wikipedia.
For the Educational writing corpora we relied on two collections of school textbooks: the ‘complex’ one (EduAdu) contains 70 texts (48,103 tokens) targeting high school students, the ‘simple’ one (EduChi) a sample of 127 texts (48,036 tokens) targeting primary school students.
Finally, the Narrative corpora are composed by the original versions of Terence and Teacher (TTorig), for the complex pole, and the correspondent simplified versions for the simple pole. Terence, which is named after the EU Terence Project2, is made of 32 documents, covering short novels for children. Teacher contains 24 documents extracted from web sites dedicated to educational resources for teachers. All Terence and Teacher texts have a simpler version (TTsemp), which is the result of a manual simplification process as described by Brunato and Dell’Orletta (2017).
All corpora were automatically tagged by the
part-of-speech tagger described in
Assessment of Lexical Complexity
For each corpus we extracted all lemmas tagged as
nouns, without considering proper nouns, and we
classified them as ‘simple’ vs ‘complex’ nouns.
Such a distinction was established according to
their frequency, which is one of the most used
parameter to assess the complexity of vocabulary
(see Section 1). Frequency was here computed
with respect to a reference corpus, i.e. ItWac
To investigate our main research questions, that
is how lexical complexity affects syntactic
complexity and the possible influence of genre and
language variety on this relationship, we focused
on a set of features automatically extracted from
the sentence parse tree. These features were
chosen since they are acknowledged to be
predictors of phenomena of structural complexity, as
demonstrated by their use in different scenarios,
such as the assessment of learners’ language
development or the level of text readability (e.g.
For each corpus, all the considered features were computed for all occurring nouns, for the subset of complex nouns and for the subset of simple nouns. Specifically, we focused on the following ones:
The linear distance (in terms of tokens) separating the noun from its syntactic head (HeadDistance in all following Tables) The hierarchical distance (in terms of dependency arcs) separating the noun from the root of the tree (RootDistance) The average number of children per noun (AvgChildren) The average number of siblings per noun (AvgSibling) 4
To have a first insight into the effect of genre and language variety on the interplay between lexical and syntactic complexity, we compared the main syntactic roles that nouns play in the sentence by calculating the frequency of all dependency types linking a noun to its head. This is shown in Figure 1, which reports the percentage distribution of typed dependency relationships linking a noun to its syntactic head across all corpora. For each corpus there are three columns: the first one considers data for all nouns of each corpus without any complexity label, the second one only data for the simple noun subset and the last one only data for the complex noun subset.
It can be noted that the distribution of nouns used as prepositional complements (prep) is the higher one across all corpora although with differences ranging from the lowest percentage (35.5%) in the ‘easy’ version of the narrative corpus (i.e. TTsemp) to the highest one (49.9%) in ScientArt (i.e. the complex language variety for the scientific writing genre). The syntactic role of prepositional complement is especially played by simple nouns compared to complex nouns. This is particularly evident in ScientArt and Repubblica, where the difference between simple and complex nouns occurring as prepositional complements is equal respectively to 20 and 15 percentage points. Conversely, complex nouns are more widely used as modifiers than simple nouns, especially in Repubblica. The percentage of nouns occurring in the subject and object position is less than 20% in all corpora. Interestingly, the higher occurrence of nominal subjects is attested in DueParole and ChildEdu (14.1 and 16, respectively). This might suggest that simpler language varieties, independently from genre, make more use of explicit subjects than implicit or pronominal ones. Besides, the likelihood of a noun to be simple or complex does not particularly affect the overall presence of nominal subjects, unless for ScientArt and Rep which both show a higher percentage of simple nouns in the subject position.
A deeper understanding of the relationship
between lexical and syntactic complexity was
provided by the investigation of the syntactic
features described in Section 3.2. Table 1 shows the
average value of the monitored features with
respect to all nouns (All), to the subset of complex
nouns (Comp) and to the subset of simple nouns
(Simp) extracted from all corpora. We assessed
whether the variation between these feature
values was statistically significant in a three different
comparative scenarios: i) between the two corpora
of the same genre, ii) between the complex
corpora of each different genre and ii) between the
simple corpora of each different genre. Table 2
shows linguistic features varying significantly for
all the considered comparisons according to the
Wilcoxon rank-sum test, a non parametric
statistical test for two independent samples
If we compare the two language varieties within each genre, it can be seen, for instance, that nouns are hierarchically more distant from the root in the complex than in the simple version. Such a variation, which is highly significant for all genres, affects more the Journalistic genre (DueParole: 2.969; Rep: 4.197) and, to a lesser extent, the Educational one (EduChi: 3.408; EduAdu: 4.269). However, for the other monitored syntactic features, the Wiki corpus appears as slightly more difficult than its complex counterpart: it has nouns that are less close to their head (Wiki: 2.531; ArtScient: 2.162) and have a richer structure in terms of number of children (Wiki: 1.363; ArtScient: 1.229). With the exception of root distance, variations concerning other features within the Narrative genre are not statistically significant. This can be possibly due to the particular composition of the two selected corpora: indeed, both Terence and Teacher texts in their original version were already conceived for an audience of children and young students, and they were not greatly modified in their simplified version.
We finally assessed whether the variation of these features was statistically significant comparing the simple and the complex noun subset of the same corpus (Table 3). According to this dimension, we can observe that complex nouns have, on average, less dependents (AvgChildren feature) than simple ones, independently from the internal distinction within genre; on the contrary, they tend to occur more distant from the root, especially in the complex variety of Scientific prose (ArtScient Comp: 5.132; ArtScient Simp: 4.598). 5
While language complexity is a central topic in linguistic and computational linguistics research, it is typically addressed from a local perspective, where each subdomain is investigated in isola2Par vs Rep Wiki vs ArtScient EduChild vs EduAdu TTsempl vs TTorig ArtScient vs EduAdu Rep vs ArtScient Rep vs EduAdu Rep vs TTorig TTorig vs ArtScient TTorig vs EduAdu 2Par vs EduChild 2Par vs TTsemp 2Par vs Wiki TTsemp vs EduChild TTsemp vs Wiki Wiki vs EduChild tion. In this preliminary work, we have defined a method to study the interplay between lexical and syntactic complexity restricted to the nominal domain. We modeled the two notions in terms of frequency, with respect to lexical complexity, and of a set of parse tree features formalizing phenomena of syntactic complexity. Our approach was tested on corpora selected to be representative of different genres and different levels of complexity within each genre, in order to investigate whether noun complexity differently affects syntactic complexity according to the two dimensions. We observed e.g. that nouns tend to appear closer to the root in simple language varieties, independently from genre, while the effect of genre and linguistic complexity is less sharp with respect to the other considered features.
To have a deeper understanding of the observed tendencies we are currently carrying out a more in depth analysis focusing on fine-grained features of syntactic complexity, such as the depth of the nominal subtree. Further, we would like to enlarge this approach to test other constituents of the sentence, such as the verb.
The work presented in this paper was partially supported by the 2–year project (2017-2019) PERFORMA – Personalizzazione di pERcorsi FORMativi Avanzati, funded by Regione Toscana (Progetti Congiunti di Alta Formazione – POR FSE 2014-2020 Asse A – Occupazione) in collaboration with Meta srl company.