A Rule-based Approach for Karmina Generation

FrankyNAACL 2013

TL;DR

This paper presents a rule-based system for generating Karmina, a two-line Malay poetic form. The authors tackle the challenge of creating poems with a distinct, non-discursive relationship between the first line (hook) and the second line (message), achieving this by generating each component separately using predefined schemas and a manually built knowledge base.

What Problem Does This Paper Address?

The paper addresses the challenge of computationally generating poetry, specifically focusing on Karmina, an old Malay poetic form popular in Indonesia. Karmina is characterized by its brevity (two lines) and a unique poetic structure where the first line (hook or sampiran) serves as an engaging opening without a direct semantic link to the second line (message or isi). This lack of discourse relation between the hook and message presents a distinct challenge compared to other poetic forms. The authors aim to explore computational creativity in language by generating Karmina, which requires not only grammatical correctness and meaningfulness but also adherence to poetic characteristics like rhyme and structure. They also see this work as a way to conserve and introduce this poetic form.

Methods

The core of the approach is a rule-based system that generates Karmina in two distinct phases: hook generation and message generation.

  • Hook Generation: The authors analyzed existing Karmina examples and identified recurring syntactic and semantic patterns, leading to the creation of approximately 19 schemas for hook generation. These schemas define the structure of the hook, often involving relationships between nouns, adjectives, and verbs. A manually built knowledge base, categorized by concepts like "fish," "flower," and "tree," was used to populate these schemas.
  • Message Generation: Similar to hook generation, messages were also generated using predefined schemas, primarily drawing from two identified structures. The key innovation here is the use of sentiment analysis by employing a manually curated lexicon of Indonesian positive and negative sentiment words, derived from English equivalents and filtered for relevance.
  • Syllabification: A crucial component for rhyme matching, a rule-based syllabification algorithm was developed. It iteratively cuts syllables from words based on vowel-consonant patterns and letter combinations.
  • Rhyme Matching: The system supports both perfect and imperfect rhymes. For perfect rhymes, it matches the last syllables of cue words. For imperfect rhymes, it considers either the last vowel of the last syllable or the entire last syllable if it ends in a consonant.
  • Karmina Construction: The final Karmina is constructed by pairing a generated message with a hook that rhymes with it, prioritizing perfect rhymes. The system ensures that the hook and message share a similar structural schema.

Datasets

  • Subjectivity Lexicon: The authors created an Indonesian subjectivity lexicon by translating an English lexicon using Google Translate. This lexicon was manually filtered and contains 740 positive and 1500 negative words.
  • Corpus for Message Generation (Future Work): The paper mentions considering a corpus-based approach for message generation, specifically mentioning a blog corpus for its informal and conversational style.

Key Insights and Ideas

  • Decoupled Generation: The most significant insight is the successful decoupling of hook and message generation. By treating them as independent entities governed by separate schemas and knowledge bases, the authors effectively address the non-discursive relationship characteristic of Karmina.
  • Schema-driven Poetry: The identification and application of specific schemas are central to generating grammatically sound and structurally coherent poetic lines, demonstrating the power of structured rules in creative generation.
  • Sentiment-driven Messages: The use of a sentiment lexicon to generate messages adds a layer of semantic control, allowing for the creation of messages with either positive or negative connotations, or shifts between them.
  • Syllabification for Rhyme: The development of a rule-based syllabification algorithm is a practical solution for enabling rhyme matching, a fundamental aspect of poetry.

Key Results

The paper reports on the generation of Karmina using their rule-based approach.

  • Evaluation: 100 randomly selected Karminas were evaluated by native speakers.
  • Acceptable Karminas: 10 Karminas were deemed to have acceptable hooks and messages.
  • Hook Issues: Improper hooks were often due to the use of uncommon names or words that, while grammatically correct, lacked common cultural resonance (e.g., "tomato and sugar-apple" in a fruit context).
  • Message Issues: Improper messages were frequently caused by incomplete constraints on word categories (e.g., combining an adjective with a noun in a way that sounds unusual) or words that don't naturally fit together.
  • Schema Dominance: The evaluation showed a tendency for generated Karminas to fall into Schema 2, with difficulties in generating Schema 1 Karminas due to a lack of rhyming hook-message pairs.

Limitations and Future Work

  • Incomplete Constraints: The authors acknowledge that the constraints within their schemas are not exhaustive, leading to some grammatically or semantically awkward outputs. This is identified as a primary area for improvement.
  • Manual Knowledge Base Creation: The process of building the knowledge base and sentiment lexicon is manual and time-consuming. Automating this knowledge acquisition is a significant future direction.
  • Automated Schema Creation: Developing methods to automatically extract or validate schemas from text is also suggested.
  • Corpus-based Message Generation: The authors plan to explore corpus-based approaches for message generation, leveraging existing text to create more natural and varied messages.
  • Filtering Knowledge: The paper notes that having too much knowledge can sometimes lead to less poetic results, suggesting a need for better filtering mechanisms.

Summary updated

BibTeX
@inproceedings{franky2013rule,
  author = {Franky},
  title = {{A Rule-based Approach for Karmina Generation}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2013},
}
Figure