[Paper Review] Generating Thematic Chinese Poetry with Conditional Variational Autoencoder.

Xiaopeng Yang, Xiaowen Lin+2 more|arXiv (Cornell University)|Nov 21, 2017

Topic Modeling22 references14 citations

TL;DR

This paper proposes a conditional variational autoencoder with an augmented word2vec architecture to generate thematically coherent Chinese poetry by explicitly modeling topic information. By integrating keyword-guided latent representations and a novel RHYTHM score metric, the model improves thematic relevance and structural consistency, achieving a 45.24% human judgment rate of generated poems as indistinguishable from human-written works.

ABSTRACT

Computer poetry generation is our first step towards computer writing. Writing must have a theme. The current approaches of using sequence-to-sequence models with attention often produce non-thematic poems. We present a conditional variational autoencoder with augmented word2vec architecture that explicitly represents the topic or theme information. This approach significantly improves the relevance of the generated poems by representing each line of the poem not only in a context-sensitive manner but also in a holistic way that is highly related to the given keyword and the learned topic. The proposed augmented word2vec model further improves the rhythm and symmetry. We also present a straightforward evaluation metric RHYTHM score to automatically measure the rule-consistency of generated poems. Tests show that 45.24% generated poems by our model are judged by humans to be written by real people.

Motivation & Objective

• To address the lack of thematic consistency in existing sequence-to-sequence models for Chinese poetry generation.
• To improve poem relevance by explicitly modeling topic or theme information through conditional latent variables.
• To enhance rhythmic and structural consistency in generated poems using a novel RHYTHM score metric.
• To develop a method that balances context-sensitive line generation with holistic thematic alignment to a given keyword.
• To evaluate model performance with both automated metrics and human judgment, focusing on poetic authenticity.

Proposed method

• Employing a conditional variational autoencoder (CVAE) to model the joint distribution of poem lines given a topic keyword.
• Integrating an augmented word2vec architecture that encodes both semantic meaning and rhythmic/syntactic constraints of Chinese characters.
• Using a topic-aware latent space where each line is conditioned on a given keyword, ensuring thematic coherence across the poem.
• Applying a RHYTHM score metric to evaluate rule-consistency in tone patterns, tonal balance, and structural symmetry.
• Training the model end-to-end with reconstruction loss and KL divergence regularization to preserve both semantic and rhythmic fidelity.
• Generating poems by sampling from the learned latent space conditioned on the input keyword, followed by decoding with a language model.

Experimental results

Research questions

• RQ1 Can a conditional variational autoencoder effectively generate thematically consistent Chinese poems when guided by a keyword?
• RQ2 How does the augmented word2vec architecture improve rhythmic and structural consistency in generated poetry?
• RQ3 To what extent does the RHYTHM score metric correlate with human perception of poetic quality and rule adherence?
• RQ4 How does the model’s performance compare to baseline sequence-to-sequence models in terms of thematic relevance and human evaluation?
• RQ5 What proportion of generated poems can be indistinguishable from human-written poetry based on human judgment?

Key findings

• The proposed model achieves a 45.24% human judgment rate, indicating that over 45% of generated poems are perceived as authentic by human evaluators.
• The integration of topic-aware latent variables significantly improves thematic relevance compared to standard sequence-to-sequence models.
• The RHYTHM score metric effectively captures rule-consistency in tone, rhythm, and symmetry, enabling automated assessment of structural quality.
• The augmented word2vec component enhances both semantic coherence and phonetic balance in generated lines.
• The model demonstrates superior performance in generating poems that are contextually and thematically aligned with the input keyword.
• Human evaluation confirms that the model generates poems with higher perceived authenticity than baseline approaches.