The RORIC-LING Bulletin

months 13 - 18

General Questions Concerning the Two Discussed Approaches to Morphology

The grammatical dictionary is a representative summary of all basic word forms in a certain language accompanied by their grammatical characteristics. These features determine the generation of all word forms which are derived from the basic one (its word-formation) and provide the basic information for the results of the text analysis. Grammatical dictionaries are among the first NLP applications and represent a basic tool for collecting and organizing the linguistic data.

The morphological dictionary is a database which provides a wide range of data about the morphological characteristics and the forms of a certain word. It also allows for quick retrieval of grammatical information coming simultaneously from different templates (paradigm tables). The main purpose of a grammatical dictionary is to identify the relations between a concrete word form and its invariant (lemma). The purpose of the morphological dictionary is therefore to identify the word form and its characteristics and to classify it with regard to its lemma.

For more information on grammatical (or morphological) dictionaries, see the Bulgarian page of the BALRIC-LING project, at
                                                                          http://www.larflast.bas.bg/balric/index/index_eng.htm

The terms stem, theme or thema designate the root + affixes/infixes. Therefore the stem represents the inflexional base of a word to which other elements, such as thematic vowels and consonants, inflections etc. are added. Obviously, in many cases, the stem of a word can be identical to its root. For example, in Greek, the root lip underlies the present theme leip to which the inflection ein is added to form the present infinitive of the verb léipein (to leave).

The main benefits are:

1. One avoids once and forever the painful, endless inflexional discussion at morphemic level and focuses directly on the bricks-words (but note, all word forms are considered in isolation). Inflexion concerns the word-formation language particularities. However, since today HLT has moved to text analysis for most European languages, one should be preoccupied in highlighting the word -> text focuses, a task which can be fulfilled by using the full-form lexicon.

2. Mapping any text onto such a lexicon allows one to start discussing the POS-disambiguation problems, which today are considered the real problems of text analysis (at "word-level").

The main criteria for inclusion of a feature can be expressed by the following question: "Is the feature to be included important for the production and distinction of paradigm members?"

Specific Questions Concerning the Full-Form Lexicon and the Corresponding Implementation

We don't think that it can be used directly. However, taking into consideration that, within such a dictionary, all inflexional forms of a word are present, we think that it can serve (when being complete, not as in our case a mere sample) for the creation of a list of words that could, in turn, be used in order to design a spell checker. Such a solution would only make use of a small part of the information which is included in a morphological dictionary.

The main difference is that in a full-form lexicon one doesn't have the representation of the word's structure. You have only the relevant information about the word plus the form as such.

Just because the language in newspapers is representative for contemporary Romanian. This is a routine practice in work on corpora.

The reason is that, from a morphological point of view, transitivity is not relevant in Romanian since transitive verbs do not show special inflected forms.

We would like to, but this depends on the opportunities regarding a new project.

To the best of our knowledge, no. But I also know that different attempts (and ongoing projects) exist in Bucharest (Romanian Academy Institute for Artifical Intelligence) and Cluj.

Prescriptive grammars view the passive voice as a morphological category. We have serious doubts concerning this point of view. We preferred not to consider the passive voice a morphological category. We don't have passive constructions in our lexicon.

I am repeating the explanation given in my web presentation. Suppose we have a word form such as fly. The lemma features help us to disambiguate the part of speech. So we have two lemmas, one for the noun fly and one for the verb fly. As for the verb, it is registered twice, with the following information: fly pr12sg; fly123pl.

This difference is needed because there are morphological differences between proper names and common nouns and our criterion has been the following: treat as morphologically relevant any feature (be it semantic or syntactic) which has morphological (i.e. inflectional) consequences.

Adjectives in Romanian have (definite) article indeed. Roughly speaking, this happens when they are placed in prenominal position. For instance, the adjective frumos (nice, beautiful) is non-articled in postnominal position (copilul frumos, literally child-the beautiful, that is "the beautiful child"), but articled in prenominal position (frumosul copil, literally beautiful-the child).

Of course it would, the only problem is building up such a program. This remains one of our tasks for the future.

Such a dictionary could serve, for instance, as a tool in second language teaching.

No, it isn't. All it can do is give the word form along with its relevant information (for instance gender, number, case).

Yes. Adjectives with comparison degrees in Romanian are treated as words composed of other words.

The difference between lemma features and word form features has been adoped for reasons of uniform description. To a certain extent, it is a theoretical distinction, too, but we believe that a description made without it works equally well.

Yes, we do, but this is supposed to be a part of another project.

This category is fairly heterogeneous, indeed, but there is no other way to deal with words that are neither adverbs nor another part of speech. So, the solution we adopted was one of "high emergency".

The two so-called 'articles' do not occur in the lexicon, probably because the corpus does not contain them. But there is no difficulty to extend the lexicon with these categories.

It's very easy to check out this relationship (provided, of course, that you know Romanian a little bit!). But to be more specific, I will tell you that the lexicon is richer than the corpus. The corpus contains about 1500 word forms, while the lexicon contains the full paradigm of a word form occurring in the corpus.

This is because, in one case it only characterizes the word form, while in the other case it is specific to the lemma itself. For instance, the gender of the noun is a lemma feature, because it does not determine the inflection. Nevertheless, the gender of the adjective does determine the inflection, and so it is a word form feature.

Am dormit is taken to be a compound word - a collocation. This is the analysis provided for all compound verbal forms.

I'm not sure I understand what you mean. If you are referring to the words you can find in the dictionary, they are there along with the relevant (morphological) information. But if you are referring to the way this information has been assigned to the word form, I have to say that this has been done in an automatic way.

The extension of the lexicon has been performed manually.

Leaving aside the vocative, there are four cases in Romanian: nominative, accusative, genitive, and dative.

As far as I know, the address is the same. Try again!

The analyzer provides the information required in connection with a given word form.

The tokenizer is language-independent in the sense that, if you give it a training corpus from a language different from Romanian, it will be able to perform the same task as the one for Romanian.

The tokenizer helps you to extract lexical items from a text faster and easier than in the manual way.

A compound word is considered a single lexical item, however composed of other words. We mark compound words with underscore: nici_un (no one).

This is because, in one case a feature only characterizes the word form, while in another case it is specific to the lemma itself. For instance, the gender of the noun is a lemma feature, because it does not determine the inflection. Nevertheless, the gender of the adjective does determine the inflection, so it is a word form feature.

Good question! We have to incorporate this pair of features, too, because it determines the inflection of the nouns.