Compound splitting and lexical unit recombination for improved performance of a speech recognition system for German parliamentary speeches

Martha A. Larson, Daniel Willett, Joachim Köhler, Gerhard RigollInterspeech 2000 · DOI

TL;DR

This paper tackles the challenge of building effective speech recognition systems for "compounding" languages like German, where words can be very long and complex. They propose a novel data-driven method to intelligently split these long words into meaningful sub-units and then strategically recombine them to optimize the system's vocabulary, achieving a significant reduction in out-of-vocabulary (OOV) words without severely impacting recognition accuracy.

What Problem Does This Paper Address?

Speech recognition systems typically rely on a predefined lexicon of words. For languages like German, which frequently form compound words (e.g., "Friedenspolitik" - peace policy), a standard lexicon quickly becomes enormous to cover all possible compounds, leading to a high rate of "out-of-vocabulary" (OOV) words – words the system doesn't recognize. This is a major hurdle for accurate speech recognition. While splitting compounds into their constituent parts can reduce OOV, it can also strip away valuable contextual information, potentially harming recognition performance. Conversely, recombining word parts into phrases can improve context but might reintroduce OOV issues. The paper addresses the need for a principled approach to balance these competing factors: how to optimally decompose and recombine lexical units for a specific domain (German parliamentary speeches) to maximize coverage while maintaining recognition accuracy.

Methods

The authors introduce a two-stage, data-driven approach:

  1. Data-Driven Compound Splitting: Instead of relying solely on linguistic rules, they analyze a large corpus of German parliamentary speeches. They use a statistical method to identify potential splitting points within compound words. This involves calculating how many words in the training data begin or end with specific letter sequences. Splitting points are chosen where these counts show local maxima in both forward and backward directions, suggesting statistically relevant sub-units. They also filter out very short sub-units (less than 3 letters) to avoid excessive acoustic confusability. This process creates a "maximally split" baseline system.
  2. Iterative Phrase Recombination: Starting from the maximally split system, they iteratively recombine sub-units into larger lexical units (phrases). This recombination is guided by the frequency of co-occurrence of these sub-units in the training data. They experiment with different frequency thresholds (e.g., combining bi- and trigrams occurring more than 100, 200, 500, or 1000 times). The goal is to find a balance that minimizes language model perplexity (a measure of how well the model predicts unseen text) while also considering the OOV rate.

Datasets

  • Training and Evaluation Data: A corpus of 15 million transcribed words from the German parliament. This data was pre-processed to be lower-case ASCII, with abbreviations and numbers replaced by their written forms.
  • Test Set: A subset of 200 recorded sentences (approximately 4.5K words) from the same domain for recognition experiments.

Key Insights and Ideas

  • Domain-Specific Lexicon Design is Crucial: The paper strongly advocates that the optimal lexicon for a speech recognition system is not universal but must be tailored to the specific domain of the speech data.
  • Data-Driven Splitting Outperforms Purely Morphological: Their statistical approach to splitting compounds is presented as more robust and relevant to the recognition task than relying solely on linguistic morphology, which might not capture domain-specific semantic relationships.
  • Recombination Can Mitigate Splitting's Downsides: While splitting reduces OOV, it can hurt accuracy. Recombining frequently co-occurring sub-units back into phrases is shown to be a viable strategy to recover some of the lost contextual information.
  • Perplexity as a Key Optimization Metric: Language model perplexity is used as a primary criterion to evaluate the quality of different recombination strategies, aiming for models that are better at predicting unseen text.

Key Results

  • Significant OOV Reduction: The best experimental system, achieved by recombining sub-units that occurred 500 times or more, demonstrated a 35% reduction in OOV rate compared to the baseline unsplit system.
  • Trade-off in Accuracy: While OOV was reduced, the recombined systems generally showed a slight decrease in recognition accuracy compared to the baseline unsplit system. The best recombined system achieved an accuracy of 61.58%, compared to 65.39% for the baseline unsplit system.
  • Optimal Recombination Threshold: The study found that recombining bi- and trigrams occurring 500 times or more yielded the best balance between perplexity and OOV reduction, outperforming thresholds of 100x and 1000x.

Limitations and Future Work

  • Accuracy vs. OOV Trade-off: The paper acknowledges that the recombined models, despite improved OOV rates, did not surpass the baseline system in recognition accuracy.
  • Phonemization Challenges: The authors suggest that difficulties in automatically phonemizing split words and recombined word sequences might be a contributing factor to the observed accuracy drop. This area is identified as a promising direction for future investigation.
  • Further Optimization: The paper establishes the effectiveness of their data-driven splitting and optimization criteria, suggesting that further refinement of these methods could lead to even better results.

Summary updated

BibTeX
@inproceedings{larson2000compound,
  author = {Martha A. Larson and Daniel Willett and Joachim Köhler and Gerhard Rigoll},
  title = {{Compound splitting and lexical unit recombination for improved performance of a speech recognition system for German parliamentary speeches}},
  booktitle = {Interspeech},
  year = {2000},
  doi = {10.21437/icslp.2000-690},
}
Figure