Style Detection for Free Verse Poetry from Text and Speech

Timo Baumann, Hussein Hussein, Burkhard Meyer-SickendiekCOLING 2018

TL;DR

This paper introduces a novel method for automatically classifying free verse poetry styles based on a continuum of poetic fluency, integrating both textual and acoustic features. By developing a deep hierarchical attention network, they achieve a weighted F-measure of 0.73, demonstrating that the subtle nuances of poetic recitation are crucial for distinguishing these styles.

What Problem Does This Paper Address?

Modern and post-modern free verse poetry has moved away from traditional rhyme and meter, creating a complex spectrum of prosodic styles. This shift has led to a rich variety in how poems are structured and delivered, ranging from fluent, conversational styles to highly disfluent, experimental forms. Analyzing and classifying these styles computationally is challenging because it requires understanding not just the written text but also the subtle prosodic cues present in spoken delivery, including intonation, rhythm, and even pauses. Existing methods often focus solely on textual features or are limited to more structured poetic forms, leaving a gap in understanding the full spectrum of free verse.

Methods

The authors propose a deep hierarchical attention network that combines textual and acoustic features to classify poetic styles. The core innovation lies in treating poetic style as a continuum of fluency and disfluency.

  • Textual Encoding: Each line of the poem is encoded character-by-character using a bidirectional recurrent neural network (GRU) with attention. This approach is chosen over word-level encoding to better capture whitespace usage and special characters, which are crucial in free verse.
  • Acoustic Encoding: Speech features (MFCCs and FFVs) are extracted from the audio recordings of the poems. Similar to the text, these acoustic features are processed line-by-line.
  • Pause Encoding: Crucially, the pauses between lines are also encoded. These pauses are hypothesized to carry significant information about enjambments and the overall rhythm of the recitation.
  • Hierarchical Attention: Per-line representations (text, audio, and pause) are concatenated and then fed into a poem-level encoder. An attention mechanism is employed at both the line and poem levels to allow the model to focus on the most relevant parts of the input for classification.
  • Pre-training: To combat data sparsity with only 175 manually classified poems, the model's character embeddings and line encoders are pre-trained using recurrent autoencoders on a larger corpus of German poetry. This helps the model learn general linguistic and prosodic patterns before fine-tuning on the specific classification task.

Datasets

  • Primary Dataset: A corpus of approximately 2400 German post-modern poems from the website lyrikline.org. Of these, 175 poems were manually classified into six prosodic styles by a philologist. Corresponding audio recordings of authors reading these 175 poems were also collected, totaling 52 hours.
  • Pre-training Dataset: The primary dataset, along with additional text-only German poetry published between 1800 and 1930 from the deutschestextarchiv.de, was used for pre-training.

Key Insights and Ideas

  • Fluency as a Continuum: The paper frames free verse poetry classification along a spectrum of fluency, moving from "parlando" (fluent, conversational) to "lettristic decomposition" (highly disfluent, experimental). This provides a philologically grounded hypothesis for style differentiation.
  • Importance of Speech and Pauses: The results strongly suggest that acoustic features and, particularly, pause information are critical for differentiating between certain poetic styles, especially those characterized by enjambments. Text alone is insufficient for many distinctions.
  • Character-Level Encoding's Advantage: Using character-level encoding for text proves effective, as it can capture nuances like whitespace and indentation that word-level models might miss, which are vital for free verse prosody.
  • Attention Mechanism's Role: The attention mechanism not only improves classification performance but also offers a window into the model's decision-making process, allowing for analysis of which parts of the text and speech it deems most important.

Key Results

  • Overall Classification Performance: The model achieves a weighted F-measure of 0.73 when using all features (text, speech, and pauses) for the six-class classification.
  • Feature Importance:
    • Distinguishing between the most fluent styles (parlando vs. variable foot) requires speech but not necessarily pause information.
  • Enjambment-dominated styles (unemphasized enjambment vs. gestic rhythm) are heavily reliant on pause information and cannot be differentiated well using text alone.
  • Decomposition styles (syllabic vs. lettristic) are somewhat differentiable from text alone, but speech features further improve performance.
  • Misclassification Patterns: Misclassifications tend to occur between styles that are adjacent on the proposed fluency continuum, supporting the hypothesis. Fluent styles are rarely confused with decomposition styles.

Limitations and Future Work

  • Data Sparsity: Despite pre-training, the limited number of manually classified poems (175) is a significant constraint.
  • Model Complexity: The authors acknowledge that the model is not yet optimal and that further meta-parameter tuning could improve results.
  • Future Work:
    • Expanding the dataset by semi-automatically identifying more poems and refining existing classes through a "human in the loop" approach.
  • Integrating syntactic information (e.g., POS tags) to better understand the link between enjambments and textual structure.
  • Exploring connectionist temporal classification (CTC) for more detailed alignment of text and speech.
  • Investigating speaker-specific characteristics to normalize them out, focusing solely on stylistic variations.
  • Further analysis of the low-dimensional mappings to better understand the multi-dimensional nature of poetic fluency.

Summary updated

BibTeX
@inproceedings{baumann2018style,
  author = {Timo Baumann and Hussein Hussein and Burkhard Meyer-Sickendiek},
  title = {{Style Detection for Free Verse Poetry from Text and Speech}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure