English. Lexicalizing compounds, in addition to treating them dynamically, is a key element in giving us idiomatic translations and detecting compound errors. We present and evaluate an e-dictionary (NDS) and a grammar checker (GramDivvun) for North Sámi. We achieve a coverage of 98% for NDSqueries and of 96% for compound error detection in GramDivvun.
In this paper1, we discuss the use and necessity of the lexicalization of compounds – in addition to the dynamic approach to compounding – in two rule-based Natural Language Processing (NLP) applications, a grammar checker GramDivvun and an electronic dictionary NDS (short for Neahttadigisánit). We argue for a dual approach and support this view with an evaluation of these tools. For comparison, we also look at a third application, a corpus tool (Korp) for the North Sámi corpus SIKOR. SIKOR, the Sámi International KORpus, is the collection of texts in different Sámi languages compiled by UiT The Arctic University of Norway and the Norwegian Sámi Parliament.
In the past, we have mostly focussed on the dynamic approach to morphological analysis. This means that we have a lexicon with lemmata and stems, which in a finite-state manner are combined 1Copyright ©2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). with inflectional and derivational affixes and other stems and modified when morpho-phonological processes apply. In this way the linguistic processes inflection, derivation and compounding are modelled in a dynamic way, i.e. by means of concatenation and composition as opposed to listing of all forms. Lexicalization, i.e. listing compounds or inflected word forms as such, is the alternative approach to the dynamic one. In addition to these two approaches we also use guessers for certain tasks, i.e. proper name guessing in morphosyntactic parsing. Our approach is entirely rulebased and open source. Within our 20 year experience with language tools for the Sámi languages and other languages with complex morphology, we have achieved good results and produced reliable tools.
There are a number of approaches to error
detection of a few errortypes for morphologically
complex - although less complex than North Sámi
languages like Latvian
The Russian system focusses on more advanced error types - case, number agreement, gender agreement, preposition and aspect. However, the results show that the system is still in its initial phase with low precision and recall for most error types (precision is between 22% and 56%, only gender agreement reaches 68%, and recall is significantly lower, between 9% and 36%). None of these approaches deals with compound error detection.
For neural network approaches, large corpora with error mark-up are necessary, which are not available for North Sámi. The error marked-up corpus contains 120 459 words, and when looking at specific error types – as in this case compound errors – the corpus is even smaller. The Russian system is based on an error-marked corpus of 200k words (deemed too small by its authors), the Latvian system works with artificial errors, an approach that can be problematic as it does not reflect real text errors.
In compounding, two or several words are combined to form a new word. In Sámi, Finnic and Germanic languages, compounding is a productive process and new compounds like in (1) can be made on the fly.2 In Romance languages, these compounds typically correspond to prepositional constructions (ital. ‘la federa del cuscino del divano’).3 (1) soffájguoddájolggoža (North Sámi) sofajputejtrekk (Norwegian) ‘sofa pillow cover (English)’
The initial motivation for extensive lexicalization of compounds of North Sámi goes back to adapting the spellchecker to users’ needs, i.e. avoiding false alarms in Ávvir newspaper’s texts.
North Sámi is a Uralic language spoken in
Norway, Sweden and Finland by approximately 25 700
speakers
North Sámi has seven morpho-syntactic cases, i.e. nominative (Nom.), genitive (Gen.), accusative (Acc.), illative (Ill.), locative (Loc.), comitative (Com.), and essive (Ess.). Case plays a more central role in Sámi than in preposition-based case languages, since here syntactic functions are identified based on case only. In addition, nouns can bear possessive suffixes. Verbs are inflected 2To avoid confusion with hyphenated compounds, “j” is used to mark word boundaries in compounds
3Although there are a number of real compounds in Italian, such as fruttivendolo, as well. for person, number (singular, dual, plural), tense (present and past tense) and mood (indicative, conditional, and potential). Derivational processes (passive, causative, inchoative, diminutive, reflexive, to name only some of them) enhance the combinatory possibilities of each verb.
Table 1 illustrates that compounding in North Sámi is by no means restricted to noun noun combinations, but includes a number of other parts-ofspeech (PoS) as well, also as heads.4
Type N N A.Attr N Adv N Pron A Pron N Adv Pcle Adv V PrfPrc N Num Num Num N Num A Num A
Example láhkajrievdadusat boahttejáigi dáppejolmmoš iešguđetjlágan eanetjlohku duššejfal
Gloss and translation lawjchange.pl ‘law changes’ comingjtime ‘future’ herejperson ‘person from here’ eachjalike ‘different kinds of’ morejnumber ‘majority’ onlyjreally ‘just’ vuostáijváldojuvvo againstjtake.pass.3sg
‘received’ mearridanjfápmu decide.prfprcjpower
‘authority’ oktajnuppejlohkái onejsecondjten.ill
‘eleven’ 1978j-láhka 1978j-law ‘1978 law’ 3j-ivnnat 3j-colored ‘3-colored’ golmmajivnnat threejcolored ‘three colored’
In North Sámi, compounds are formed without
a hyphen, except for those involving a proper noun,
a digit, or an acronym like Davvi-Norgii
‘Northern Norway (Ill.)’, 3-juvllatsykkel ‘tricycle’, and
ILO-álgoálbmotsoahpamuš ‘ILO-indigenous
people agreement’ (
The North Sámi tools described in this
article – NDS, Korp for SIKOR and
GramDivvun
The computational modeling of the language is
done using finite-state morphology
Lexicalized four-element compounds are quite common in the noun lexicon, e.g. davvisámegielterminologiija ‘North Sámi language terminology’. Even six-element compounds (sáivačáhceguollevuostáiváldindilli ‘fresh water fish receive situation’) can be found.
The different types of North Sámi compounds in Table 1 are not treated equally in the morphological analyzer. Only the compounds in the first two lines can be derived dynamically. All others need to be lexicalized, i.e. listed in the lexicon, to receive a compound analysis. Numeral compounding is not treated dynamically in the FST. The dynamic compounds are generated from the dictionary by concatenating word forms (such as a genitive or nominative noun followed by other noun) and adding a compound tag +Cmp. The main dynamic compounds are (derived and non-derived) noun + noun pairs. One feature of the underlying technology is that the compounding mechanism is capable of modeling infinitely long compounds: for example nouns of any magnitude are compounds and modeled by the finite-state automaton. Since the compounding mechanism of an FST is very powerful, it also leads to ambiguity. When we allow arbitrary lexemes to combine to form compounds, 5The table is based on the dictionary size at the time of the writing (September 2020); it is actively developed daily. Further abbreviations are Adp=adposition, Conj=conjunction. some will overlap other existing lexemes, cf. ex. (2). (2)
North region;direction.oven ‘The northern region’ Here, regiuvdna ‘region’ has a typical spelling error, o>u. The FST analyzes it as a misspelling of regiovdna ‘region’, but also as a compound with the elements regi, a common wrong form of regiija ‘direction’, and uvdna ‘oven’. While this example has only two possible analyses, twenty or more different analyses are not uncommon.
N Num
A V Adv Adp Conj 2
We present three applications, an e-dictionary, a
corpus tool, and a grammar checker tool.
3.1
The North Sámi – Norwegian dictionary contains
25 000 lemmata and uses an FST. The e-dictionary
was first implemented in 2013 with no use of
relational databases (all linguistic resources are
contained within static files and external
commandline tools)
In the FST 90% of the 100 000 nouns, and in the dictionary 75% of the 25 000 nouns are compounds. 3.2
The web application and corpus search tool
Korp
Table 3 shows the statistics for compounds in SIKOR.6 The results are obtained using the scripts that can be found in GiellaLT.7 According to our analyses 8.6% of the tokens in corpus are compounds, and 86% are lexicalized. The rest is mainly composed of 2-elements compounds (13.4%) and a very small part of 4-7 elements (0.5%).
Many of the longer compounds in SIKOR are quite creative and are hyphenated as the one in ex. (3). (3) suoidne-varra-bleahkka-mála-bihkka-senet-dielku hay-blood-ink-paint-tar-mustard-stain mu báiddis lei dušše lihkohisvuohta. my shirt.loc was only mishap ‘The hay-blood-ink-paint-tar-mustard-stain on my shirt was only a mishap.’
N Prop
2 96.2 3.8
3 98.9 1.1
4 89.2 10.8
5
The current public version of the Sámi corpus
SIKOR
7https://github.com/giellalt/ conf-clicit2021 that does not distinguish between lexicalized and dynamic compounds. The (non-public) version of SIKOR used in this article makes this distinction, though, as will future versions in Korp.
A search for compound tags only returns split compounds, i.e. the first coordinated hyphenated nominal element, cf. in ex. (4), i.e. riddo- ‘coast-’. (4) riddo- ja vuotnaguovlluin coast- and fjordregion.loc.pl ‘in coastal and fjord regions’
GiellaLT has already produced a solution, i.e. a tag for cohorts with a dynamic compound (<with-dynamic-compound>) added by a Constraint Grammar module. However, this tag does not provide any information about the number of elements and the beginning and ending of each element. 3.3
GramDivvun, the North Sámi grammar
checker
Lexicalization of compounds is relevant for grammar checking within compound error detection. One common error that cannot be resolved by a spellchecker is the spelling of compounds as two or more words. GramDivvun performs this type of error detection as part of the tokenization. The tokenization is done in two steps. In the first step potential compounds are tokenized ambiguously (either as one or as two words, the first of which is accompanied by an errortag). In the second step, a Constraint Grammar module8 selects or removes the error reading. Two conditions need to be met to find the compound error: 1. the compound needs to be lexicalized, and 2. the syntactic context needs to support the compound reading.
The syntactic context is specified in handwritten Constraint Grammar rules. The 8https://github.com/giellalt/lang-sme/blob/ 3a43911929458fd39da309ed23178bf5dbd04bcd/ tools/tokenisers/mwe-dis.cg3 REMOVE-rule below removes the compound error reading (identified by the tag Err/SpaceCmp) if the head is a 3rd person singular verb (cf. l.2) and the first element of the potential compound is a noun in nominative case (cf. l.3). The context condition further specifies that there should be a finite verb (VFIN) somewhere in the sentence (cf. l.4) for the rule to apply.
All possible compounds written apart are considered to be errors by default, unless the lexicon specifies a two or several word compound or a syntactic rule removes the error reading. There are numerous syntactic contexts where the potential parts of compounds make perfectly sense. In the case of noun-noun compounds, the second element can for example be a simple adverbial, as in ex. (5). The second element can be homonymous with another PoS, it can be a finite verb or an infinitive. (5) son lea boarráseamus mánná joavkkus. s/he is oldest child group.loc ‘s/he is the oldest child in the group.’ 4
We evaluate the e-dictionary (coverage) and the grammar checker (precision, recall) for compounding (errors). The corpus search tool does not exhibit compounding information and is therefore not evaluated. 4.1 We analyzed the logs for NDS (Neahttadigisánit) for 2019, and found that 12.6% of the types in the user queries are compounds. The results are obtained using the scripts that can be found in GiellaLT 7. The amount of lexicalized compounds in the logs (72.1%) is approximately the same as in the dictionary, where it is 75% (cf. Section 3.1 above). As much as 98% of the compound queries get a translation, either a lexicalized one or of its parts. Thus dynamic compounding contributes with a substantial improvement to dictionary coverage. If the alternatives are “getting no help from the dictionary” and “getting help to translate the parts” then the latter is to be preferred, even though the correct translation would be different from just joining the parts. For example, the compound word ruhtahearrá ‘rich man’ is not lexicalized in NDS but it does get a translation of its parts ruhta ‘money’ and hearrá ‘man’, which can help the user to understand the meaning of the compound word itself.
Most of the non lexicalized compounds are composed of 2 elements (96% in the logs and 93% in the entries). When analyzing the entries in the dictionary, we found that 24.8% are compounds and of those 97.6% are lexicalized. Table 4 shows PoS for compounds in NDS logs and entries.
Parts PoS N A Prop V Adv
Logs
Entries
L
We evaluate error detection for syntactic
compound errors (i.e. words that are written apart
and should be a compound) in GramDivvun in two
ways. Firstly, we compare last year’s results in
Wiechetek (2019a) with a newer version of
GramDivvun, from now on referred to as the
Nodalida-corpus. Last year’s results are based on
version r183544
However, as the focus in the last analysis was a different one, i.e. we evaluated other error types as well, we ran a second evaluation on a 2 363 wordcorpus11 specifically made to test compound error detection, i.e. every sentence contains a potential compound. These sentences are hand-selected from SIKOR.
The results of the evaluation are presented in Table 5. We can see that precision has gone significantly up, i.e. the average precision is 95.5%.
9https://github.com/giellalt/lang-sme/ releases/tag/nodalida-2018 on 2019-09-26 10https://github.com/giellalt/lang-sme/ releases/tag/clicit on 2020-09-07
11http://gtsvn.uit.no/freecorpus/orig/sme/ odda_mahppa/compounds.correct.txt However, the recall has gone down to average 46%. We are investigating the reasons for that. But in general, a high precision is desirable in grammar checking, even at the cost of a lower recall.
The results of the evaluation of GramDivvun compound grammar checking are shown in Table 5.
False negatives are typically due to the lack of lexicalization. Many of those are proper noun combinations which are very productive, e.g. Murmánska-aviisa ‘Murmansk newspaper’, Várggát-festiválas ‘at the Várggát festival’, kmgalba ‘km sign’ and Divttasvuotna-regiovnna ‘Divttasvuotna region’.
Other reasons are certain (unlikely) analyses of especially the first element, e.g. that generally suggest a syntactic construction rather than a compound as in ex. (6). Here the first element duorastat ‘Thursday’ has a finite verb reading as well. dán duorastat veaiggi. this.gen Thursday twilight.gen ‘this Thursday evening’ (6) (7)
The false positive is due to an error in the recognition of the span of the target. In ex. (7), lulli sámi guvlui is concatenated, but it should only be lulli sámi.
dohko lulli sámi guvlui. thither South Sámi area.ill ‘thither towards the South Sámi area.’ 5
We have shown that the lexicalization of compounds – in addition to their dynamic treatment – is useful and necessary for two language applications for North Sámi, an e-dictionary (NDS) and a grammar checker (GramDivvun). The evaluation of NDS shows that we get a good coverage: 98% of the compounds logged do get a translation and 72% are lexicalized in the FST. The evaluation of GramDivvun has shown that we manage to identify compound errors with a precision of 98% and a recall of 49% utilising a combination of information from the lexicon and syntax.
We conclude that there are perfectly good reasons for lexicalizing compounds, i.e. providing idiomatic translations for when it cannot be derived from the parts, and to support compound grammar checking. At the same time, lexicalization can dissimulate word formation information in corpus tools. This can be resolved and we have already implemented a solution in Constraint Grammar to make the information available in a future version of the corpus tool. As dynamic compounding is limited to few PoS at the moment, in the future we want to investigate and model compounding of other PoS (in the FST). Also experiments with neural network approaches and a comparison of the results to our rule-based grammar checker could be an interesting future project.