preview

Morfette

Decent Essays

Morfette, the morphological analyzer and lemmatizer we included in our annotation pipeline, reaches an accuracy of 98.38% for morphological features and of 97.21% for lemmas on the test set (cfr. Table 13). However, in Morfette a concatenation of features is treated as a single string and not in an atomic way. We computed accuracy for single feature on the test set and we report the results in Table 23. It is possible that features with a low accuracy have a negative impact during pars- ing. We compared the results of Table 23 to the results of our experiment on morphological features for parsing (Table 17). PronType, NumType, Poss, Reflex are the features that were deemed useless for parsing; however, they present a high prediction accuracy, …show more content…

Table 17). This feature refers to the form of the verb, which in Italian can be finite, infinitive, participle or gerund. The second best feature is Mood. In linguistics, mood is one of three major grammatical categories of the verbal conjugation system and is used to indicate the attitude with which the speaker presents the action expressed by the verb. In Italian this can be indicative, imperative, conditional or subjunctive. The two most helpful features are both re- lated to verbs. In the dependency grammar framework, the root of a sentence is usually its main verb and all other words are directly or indirectly connected to it. Therefore, the verb is a key component to the whole sentence. We believe that this is the reason why Mood and VerbForm represent the best morphological features to use during …show more content…

Along with Gender, they form the paradigm of agreement features. However, Person and Number are beneficial during parsing where Gender is not. In fact, when Gender was added, the accuracy dropped of 0.3% points. If it is true that the most helpful features are those concerned with verbs, this could explain why Gender is not beneficial. In fact, only a subsets of Italian verbs differ- entiate for gender; most of them differentiate only for mood, verb form, number and person. We conclude that between agreement features, only Person and Number are useful for parsing Italian, while Gender is actually harmful. Previous works within the MaltParser framework on other MRLs have showed similar results (see for example Ambati et al. [2010] for

Get Access