Construction Of EPEC: A Reference Corpus For Basque Natural Language Processing.pdf

Motivation:

The authors, drawing on Sarasola's (2000) strategic priorities for minority language technology, emphasize the interconnectedness of language foundations, tools, and applications in NLP. They aim for EPEC to serve as a foundational "reference" corpus for Basque NLP.

Corpus Composition:

Source Data: EPEC comprises 50,000 words split evenly between:
A subcorpus from the Statistical Corpus of 20th Century Basque, a rich resource covering various dialects and genres.
Articles from Euskaldunon Egunkaria, a daily newspaper in standard Basque.
Text Selection: Texts were chosen to represent both literary and non-literary prose, reflecting the diversity of written Basque.

Tagset Design:

EPEC employs a four-level morphosyntactic tag system:

1. General POS tags: 20 broad categories (noun, adjective, verb, etc.).
2. Refined POS tags: Each general category is further subdivided into specific subcategories (e.g., "pronoun" includes common, emphatic, interrogative, etc.).
3. Morphological Features: This level captures basic grammatical information like case, number.
4. Full Morphosyntactic Analysis: Combines morphological and syntactic annotations, providing detailed insights into word functions within sentences.

Construction Process:

The creation of EPEC involves several stages:

1. Text Compilation: Gathering and selecting texts from reliable sources.
2. Morphological Disambiguation:

Initial disambiguation using MORFEUS, a Basque morphological analyzer.
Manual review and refinement of MORFEUS outputs based on linguistic expertise.

3. Shallow Syntax Tagging: Assigning basic syntactic tags (subject, object, etc.) to words in the text.
4. Phrase Tagging: Identifying and tagging phrases like noun phrases and verb phrases.
5. Deep Syntax Tagging (Future): Utilizing a dependency parsing approach for more nuanced syntactic analysis.

Figure 1 illustrates the interconnected stages of EPEC construction, highlighting manual and automated components.

Future Work:

The authors plan to continue refining EPEC through:

Manual review and disambiguation of remaining words.
Integration of dependency parsing for deeper syntactic annotation.
* Development of tools leveraging EPEC's rich annotations for tasks like machine translation, information extraction, and text summarization.