Sounds Wilde. Phonetically Extended Embeddings for Author-Stylized Poetry Generation

А. М. Тихонов, Ivan P. Yamshchikov2018 · DOI

Abstract

This paper addresses author-stylized text generation. Using a version of a language model with extended phonetic and semantic embeddings for poetry generation we show that phonetics has comparable contribution to the overall model performance as the information on the target author. Phonetic information is shown to be important for English and Russian language. Humans tend to attribute machine generated texts to the target author.

Introduction

Most modern Natural Language Processing (NLP) models are essentially "deaf." They represent words as high-dimensional vectors based on their semantic context—what they mean—but they ignore how they sound. While this works for translating a technical manual, it often fails in poetry, where the "acoustic structure" (rhyme, rhythm, and euphony) is as critical as the literal message.

Researchers Tikhonov and Yamshchikov tackle the challenge of author-stylized text generation by arguing that to truly capture a poet’s "voice," a model must account for the phonetic patterns unique to that writer. Their work bridges the gap between models that generate coherent text and those that capture the specific stylistic fingerprint of a human author.

Methodology: The Concatenated Embedding

The researchers built upon a standard Long Short-Term Memory (LSTM) language model but introduced a novel concatenated embedding strategy. Instead of a single vector representing a word, the model uses a composite representation consisting of four distinct parts:

  1. Semantic Embedding: A standard vectorized representation of the word’s meaning.
  2. Character-level Bi-LSTM: A sub-network that processes the word letter-by-letter to capture morphological patterns.
  3. Phoneme-level Bi-LSTM: The "secret sauce." The researchers used heuristics to transcribe words into the International Phonetic Alphabet (IPA) and processed these sounds through a bidirectional LSTM.
  4. Author/Document Projections: Information about the target author and the specific document is projected into the model’s state space to steer generation toward a specific style.
The concatenated word representation architecture, combining semantic, character-level, phoneme-level, and author/document embeddings.

By combining these, the model "knows" that "night" and "bright" are semantically different but phonetically similar, allowing it to maintain the rhythmic and rhyming integrity of a specific author's style.

Datasets

The study utilized two large datasets of poetry:

  • English Dataset: Approximately 110,000 documents from 19,000 authors (including Shakespeare, Lewis Carroll, and lyrics from Bob Marley).
  • Russian Dataset: Approximately 330,000 documents from 1,700 authors (including Pushkin and Esenin).

Key Results and Insights

The researchers found that adding phonetic information improved the model’s performance more than adding explicit author identity. This suggests that "style" is deeply encoded in the sounds a writer chooses.

Quantitative Performance

The "Full Model" (incorporating phonetics) achieved a BLEU score of 29.5%, a significant improvement over the vanilla LSTM baseline of 29.0%. While a 0.5% absolute increase might seem modest, the relative contribution of phonetics (+1.7% over baseline) was nearly double the contribution of author information alone (+1.0%).

BLEU score comparison across different model configurations, highlighting the performance gain from phonetic and author information.

Qualitative Human Evaluation

In a "Turing Test" style survey with 140 English speakers, the model demonstrated impressive mimicry:

  • Shakespeare was recognized 46% of the time (2.5x better than random chance).
  • Neuro-Shakespeare (the AI) successfully "tricked" readers into thinking it was the real Bard in 37% of cases.
  • Lewis Carroll proved the hardest to mimic, with identification rates roughly equal to random guess (20%).

Conclusion and Future Work

By proving that an AI needs an "ear" to write like a human, this paper opens the door for more nuanced, stylistically accurate creative AI. However, challenges remain regarding data sparsity for less prolific authors and long-term narrative coherence. Future work may involve adding "pseudo-semantic rules" to ensure generated poems maintain logical sense alongside their musicality.

Summary updated

Abstract

This paper addresses author-stylized text generation. Using a version of a language model with extended phonetic and semantic embeddings for poetry generation we show that phonetics has comparable contribution to the overall model performance as the information on the target author. Phonetic information is shown to be important for English and Russian language. Humans tend to attribute machine generated texts to the target author.
BibTeX
@misc{2018sounds,
  author = {А. М. Тихонов and Ivan P. Yamshchikov},
  title = {{Sounds Wilde. Phonetically Extended Embeddings for Author-Stylized Poetry Generation}},
  year = {2018},
  doi = {10.18653/v1/w18-5813},
}
Figure