Hasil Pencarian  ::  Simpan CSV :: Kembali

Abstrak :
This study aims at describing syntactic behavior of congruence and rection in Arabic by examining the syntactic element attached to them. This study also aims at determining types of congruence and rection in Arabic that have not been revealed so far. Syntactic element that is involved in the process of congruence and rection is the secondary category such as person, number, gender, definiteness, case, mood, and aspect. Both of congruence and rection focused on relation between constituents in one construction. The two term can be distinguished by looking at the category involved. Congruence focuses on the relation between constituents in one construction that is marked by the appearance of the same category, for example, the category of person, number, and gender play an important role in the relation between a subject and its verbal predicate, while rection focuses on relation that is marked by the appearance of a certain category resulted from the relation between one constituent and another constituent, for example, accusative case in the objective function is a result of relation between an object and its predicate. This study concludes that in Arabic, congruence and rection can be found at the clause and phrase level.

Abstrak :
Pendekatan transfer learning telah digunakan di beragam permasalahan, khususnya low-resource language untuk meningkatkan performa model di masing-masing permasalahan tersebut. Fokus pada penelitian ini ingin menyelidiki apakah pendekatan cross-lingual transfer learning mampu meningkatkan performa pada model constituency parsing bahasa Indonesia. Constituency parsing adalah proses penguraian kalimat berdasarkan konstituen penyusunnya. Terdapat dua jenis label yang disematkan pada konstituen penyusun tersebut, yakni POS tag dan syntactic tag. Parser model yang digunakan di penelitian ini berbasis encoder-decoder bernama Berkeley Neural Parser. Terdapat sebelas macam bahasa yang digunakan sebagai source language pada penelitian ini, di antaranya bahasa Inggris, Jerman, Prancis, Arab, Ibrani, Polandia, Swedia, Basque, Mandarin, Korea, dan Hungaria. Terdapat dua macam dataset bahasa Indonesia berformat Penn Treebank yang digunakan, yakni Kethu dan ICON. Penelitian ini merancang tiga jenis skenario uji coba, di antaranya learning from scratch (LS), zero-shot transfer learning (ZS), dan transfer learning dengan fine-tune (FT). Pada dataset Kethu terdapat peningkatan F1 score dari 82.75 (LS) menjadi 84.53 (FT) atau sebesar 2.15%. Sementara itu, pada dataset ICON terjadi penurunan F1 score dari 88.57 (LS) menjadi 84.93 (FT) atau sebesar 4.11%. Terdapat kesamaan hasil akhir di antara kedua dataset tersebut, di mana masing-masing dataset menyajikan bahwa bahasa dari famili Semitic memiliki skor yang lebih tinggi dari famili bahasa lainnya. ......The transfer learning approach has been used in various problems, especially the low-resource languages, to improve the model performance in each of these problems. This research investigates whether the cross-lingual transfer learning approach manages to enhance the performance of the Indonesian constituency parsing model. Constituency parsing analyzes a sentence by breaking it down by its constituents. Two labels are attached to these constituents: POS tags and syntactic tags. The parser model used in this study is based on the encoder-decoder named the Berkeley Neural Parser. Eleven languages are used as the source languages in this research, including English, German, French, Arabic, Hebrew, Polish, Swedish, Basque, Chinese, Korean, and Hungarian. Two Indonesian PTB treebank datasets are used, i.e., the Kethu and the ICON. This study designed three types of experiment scenarios, including learning from scratch (LS), zero-shot transfer learning (ZS), and transfer learning with fine-tune (FT). There is an increase in the F1 score on the Kethu from 82.75 (LS) to 84.53 (FT) or 2.15%. Meanwhile, the ICON suffers a decrease in F1 score from 88.57 (LS) to 84.93 (FT) or 4.11%. There are similarities in the final results between the two datasets, where each dataset presents that the languages from the Semitic family have a higher score than the other language families.