When undertaking the development of a new NLP resource, particularly in the case of minority languages, it is important to build upon existing resources when possible and to leverage findings of previous studies.
In recent years, some progress has been made in the collection and development of linguistic resources for Irish. A 30 million word corpus of Modern Irish text (NCII)5
was developed in 2004 for Foras na Gaeilge.6 In addition, corpus annotation tools,
namely a morphological analyser (U´ı Dhonnchadha et al., 2003), a part-of-speech (POS) tagger (U´ı Dhonnchadha and van Genabith, 2006) and a shallow parser (U´ı Dhonnchadha and van Genabith, 2010) have been developed. A 3,000-sentence gold standard POS-annotated corpus was produced as a by-product of this work. These sentences were randomly selected from the NCII corpus and consist of text from books, newspapers, websites and other media, forming a solid representation of Modern Irish language data. We use this POS-tagged corpus as a basis for our treebank. The tags are based on the PAROLE Morphosyntactic Tagset (IT´E, 2002). U´ı Dhonnchadha (2009) also made available a small corpus of 225 chunked Irish sentences. These sentences represented a test suite for a shallow parser (U´ı Dhon- nchadha and van Genabith, 2010) which is based on Constraint Grammar Depen- dency Mapping Rules (Karlsson, 1995) and implemented using Xerox Finite State Tools.7 The dependency analysis for this parser was based on Constraint Gram-
mar (CG), developed by Karlsson (1995). We present an example Irish sentence in Example 25. Figure 3.8 shows how CG rules annotate tokens of this sentence
5New Corpus for Ireland – Irish. See http://corpas.focloir.ie
6A government body in Ireland responsible for the promotion of the Irish language –
http://www.forasnagaeilge.ie
with grammatical functions such as @SUBJ (meaning the token is the subject) and dependency relations such as @>V (meaning the token is dependent on the verb to the right). Figure 3.9 shows the same chunked output without morphological tags. Finite State regular expressions are then applied (using Xerox XFST) to the anno- tated text to mark linguistic chunks. Chunks are groups of words that represent phrases, and are labelled as NP (noun phrase), V (verb), PP (prepositional phrase), for example. Dependency relation tags identify links between tokens within chunks, not between chunks.
(25) D’ fhan siad ansin le fiche bliain PAST stay they there with twenty years ‘They stayed there for twenty years’
[S
[V D’ do+Part+Vb+@>V fhan fan+Verb+VI+PastInd+Len+@FMV] [NP siad siad+Pron+Pers+3P+Sg+Masc+Sbj+@SUBJ NP]
[AD ansin ansin+Adv+Loc+@ADVL]
[PP le le+Prep+Simp+@PP_ADVL [NP fiche fiche+Num+Card+@>N bliain bliain+Noun+Fem+Com+Sg+@P< NP] PP]
. .+Punct+Fin+<<< S]
Figure 3.8: Example of chunked output (with morphological tags) for D’fhan siad ansin le fiche bliain ‘They stayed there for twenty years’
[S [V D’ @>V fhan] [NP siad] [AD ansin] [PP le [NP fiche @>N bliain @P< ]] . < S]
Figure 3.9: Example of chunked output for D’fhan siad ansin le fiche bliain ‘They stayed there for twenty years’
The shallow nature of this chunking parser means that the dependency analy- sis does not extend to cover coordination, prepositional attachment, long-distance dependencies or clausal attachment. However, these 225 invented sentences cover
the major syntactic phenomena of Irish and provided a valuable starting point for this treebank development. Many of our attachment rules within phrases are closely aligned with this particular grammar.
Our first step involved reviewing the dependency analysis for the shallow parser and adapting it to fit our chosen dependency scheme, as discussed in Section 3.1.2. We modified and extended the parses in this small chunked corpus to produce deep, full syntactic parses. Figure 3.8 is example output from the shallow parser for D’fhan siad ansin le fiche bliain ‘They stayed there for twenty years’. The sentence is parsed into 4 chunks; V (Verb), NP (Noun Phrase), AD (Adverb) and PP (Prepositional Phrase). This output also indicates the kind of data available to us in the POS- tagged corpus that we use in our treebank. For example, the token D’ is tagged with surface form (D’ ), lemma (do), coarse-grained POS-tag (Part – particle) and fine-grained POS-tag (Vb – verbal ).
Using the chunked corpus as a starting point highlighted the type of linguis- tic analysis required for defining a new dependency annotation scheme. It also prompted discussions around the type of formalism our scheme should be based on, as discussed in more detail in Section 3.1.2. We then added these fully parsed sen- tences to the 3,000 gold standard POS-tagged corpus and subsequently randomised the data so that the relatively simple 225 sentences were dispersed throughout the corpus8.
vparticle subj advadjunct padjunct quant pobj
D’ fhan siad ansin le fiche bliain
PAST stayed they there with twenty years ‘They stayed there for twenty years’
Figure 3.10: The fully parsed sentence of Figure 3.8
Figure 3.10 presents our extended parse analysis for the same sentence as Fig-
8The hand-crafted nature of the 225 sentences rendered them more simple structures than
ure 3.8, showing in particular adverbial and prepositional attachment to the verb. A deeper syntactical analysis such as this provides a more informative linguistic description of Irish text. Consider, for example, the attachment of prepositional phrases. A shallow parse output does not identify the phrase that is being modified by the prepositional phrase. For example, an ambiguous sentence will have differing prepositional attachment – (i) to the verb as in Figure 3.11 or (ii) to the object of a prepositional phrase as in Figure 3.12.
pobj padjunct pobj adjadjunct padjunct pobj
ag obair le seirbh´ıs´ı poibl´ı sa Ghaeltacht at working with services public in-the Gaeltacht
‘working [in the Gaeltacht] [with public services]’
Figure 3.11: Prepositional phrase ambiguity, where the head of the preposition sa is obair.
pobj padjunct pobj adjadjunct padjunct pobj
ag obair le seirbh´ıs´ı poibl´ı sa Ghaeltacht at working with services public in-the Gaeltacht
‘working with [public services in the Gaeltacht]’
Figure 3.12: Prepositional phrase ambiguity, where the head of the preposition sa is seirbh´ıs´ı.
Figure 3.13 demonstrates how this type of ambiguity cannot be resolved with a shallow parser. We can see that there is no connection between the aspectual phrase ag obair and either of the prepositional phrases.