Dependency Parser

The task of dependency parsing entails taking an input sentence (𝑤) and returning a parse tree of typed dependencies. A parse tree is simply a tree-like representation of the structure of a sentence. Several computational implementations of DGs exist such as Minipar, RASP, SUPPLE, and while the details of these implementations are beyond the scope of this research, Stanford Universal Dependencies implementation (SUD) (de Marneffe et al, 2014a; de Marneffe et al, 2014b) is used.

SUD is a typed dependency implementation. It is called typed because each link in the parse tree is labelled with a dependency type. For example, figure 12 shows the labelled link or dependency relationship between ‘Europe’ and ‘Union’ to be the expression ‘compound’ for the sentence “We will definitely be leaving the European Union”. For the linguistic units

‘European’ and ‘Union’, the governor or superior unit is ‘Union’ while the dependent is ‘European’. SUD is used specifically for its popularity and accessibility but most importantly

because of its broad universal taxonomy of relations which have been designed to represent the grammatical function of each linguistic unit relative to its dependent. For instance, in figure 12, the dependency relationship between ‘definitely’ and ‘leaving’ expressed typographically as 𝑎𝑑𝑣𝑚𝑜𝑑(‘𝑙𝑒𝑎𝑣𝑖𝑛𝑔’, ‘𝑑𝑒𝑓𝑖𝑛𝑖𝑡𝑒𝑙𝑦’) states that the dependent ‘definitely’ is the adverbial modifier of the governor ‘leaving’ while in figure 13, the dependent word

77

Figure 12: Illustration of a Typed Dependency for sentence ‘We will definitely be leaving the European Union’

Figure 13:Illustration of a Typed Dependency for sentence ‘The British players have arrived safely’

SUD consists of 42 relations centred around core arguments - These are the dependency relations that the sentence predicate partakes in - and non-core arguments - which generally covers the dependency relations of modifiers, nouns and function words. Although there are 42 relations in the SUD, not all relations are applicable to all grammar types. Some relations tend to occur more commonly in non-Germanic languages for instance the relation ‘clf37_{’ occurs predominantly in Asian languages. In addition, the}

evolving nature of language means that additional relations are being discovered with newer relations emerging from older ones. Nevertheless, SUD clearly delineates core dependency relations - which are relations between the verb e.g. the root verb and subjects, objects or clausal complements in the sentence – and other dependency relations such as modifiers. Using the word class priority illustration in figure 11 as a reference, the relations used in this research include all 8 core dependency relations (based on figure 11, the relations between the headword and verbs in the sentence and the actants). These relations

78

are listed in table 6. The second category of relations are oblique relations or modifiers which relate the actants and modifying expressions as seen in figure 11.

Table 6: Core Dependency Relations (Source: SUD v1.0)

Dependency Relation Full Meaning

nsubj Nominal subject

csubj Clausal subject

nsubjpass Passive nominal subject

cubjpass Passive clausal subject

dobj Direct object of a predicate

ccomp Clausal complement of a verb or adjective

xcomp Clausal complement of a verb

iobj Indirect object of a verb

Table 7: Modifier Relations of Nouns and Clausal Predicates (Source: SUD v1.0)

Dependency Relation Full Meaning

nmod Nominal modifier

advcl Adverbial clause modifier

advmod Adverbial modifier

neg Negation

acl Adjectival clause modifier of a nominal

amod Adjectival modifier

appos Appositional modifier of a noun

nummod Numeric modifier of a noun

SUD’s 8 non-core and modifying relations are used and outlined in table 7. The third category are relations which exist between content words and function words and relations which express coordination, disfluencies, multiword expressions and punctuation. Although this third category of relations are important they are less regular and their occurrence can be dependent on the type of document or language. For example, the relation ‘expl’ relates expletives, and is commonly seen in informal documents. Therefore, 14 relevant content-function word relations are used and outlined in table 8.

In addition to these relations, SUD includes a relation called ‘root’ which reflects the root of the sentence. The governor of this relation is a fake node ‘ROOT’ which marks the start of the dependency tree and its dependent is the root word. In figures 12, 13 and 14, the root relations are 𝑟𝑜𝑜𝑡(𝑅𝑂𝑂𝑇, 𝑙𝑒𝑎𝑣𝑖𝑛𝑔), 𝑟𝑜𝑜𝑡(𝑅𝑂𝑂𝑇, 𝑎𝑟𝑟𝑖𝑣𝑒𝑑), 𝑟𝑜𝑜𝑡(𝑅𝑂𝑂𝑇, 𝑑𝑒𝑓𝑒𝑎𝑡𝑒𝑑).

79

Figure 14: Illustration of a Typed Dependency for sentence “UKIP and Labour were defeated by the Conservatives”

Table 8: Content-Function Word Relations (Source: SUD v1.0)

Dependency Relation

Full Meaning

det Determiner relation between a nominal and a determiner

mwe Relates multiword expressions

goeswith Links two parts of a word that are separated in text that is not well edited

name Used to relate names e.g 𝑛𝑎𝑚𝑒(𝐶𝑎𝑚𝑒𝑟𝑜𝑛, 𝐷𝑎𝑣𝑖𝑑)

foreign Used to label sequence of foreign words

list Used to relate chains of comparative items

aux Relates an auxiliary of a clause

auxpass Relates passive auxiliary of a clause

mark Relates a marker, which is the word introducing a finite clause subordinate to another clause

cop Captures relation between the complement of a copular verb and the copular verb ‘to be’

cc Relates an element to a coordinating conjunction

conf Relates two elements connected by a coordinating conjunction

case Used for any case-marking element which is treated as a separate syntactic word (including prepositions, postpositions, and clitic case markers)

compound Relates compound expressions e.g.

𝑐𝑜𝑚𝑝𝑜𝑢𝑛𝑑(𝑤𝑎𝑟𝑟𝑎𝑛𝑡, 𝑎𝑟𝑟𝑒𝑠𝑡)