• Corpus ID: 41138518

Improving G2p from wiktionary and other (web) resources

  title={Improving G2p from wiktionary and other (web) resources},
  author={Steffen Eger},
We consider the problem of integrating supplemental information strings in the grapheme-to-phoneme (G2P) conversion task. In particular, we investigate whether we can improve the performance of a G2P system by making it aware of corresponding transductions of an external knowledge source, such as transcriptions in other dialects or languages, transcriptions provided by other datasets, or transcriptions obtained from crowd-sourced knowledge bases such as Wiktionary. Our main methodological… 
Do we need bigram alignment models? On the effect of alignment quality on transduction accuracy in G2P
The need for bigram alignment models and the benefit of supervised alignment techniques in graphemeto-phoneme (G2P) conversion are investigated and a highly significant impact of alignment quality on overall G2P transcription performance is estimated.
A Comparison of Four Character-Level String-to-String Translation Models for (OCR) Spelling Error Correction
This work investigates four general string-to-string transformation models that have been suggested in recent years and applies them within the spelling error correction paradigm and finds that a simple ‘k-best decoding plus dictionary lookup’ strategy performs in this context and can significantly outdo baselines.
Multi-Task Learning for Argumentation Mining
It is found that dataset properties such as entropy or the size of the label inventory are good indicators for a potential multi-task learning success and that multi- task learning is particularly useful if the task at hand suffers from data sparsity, i.e. a lack of training data.


Leveraging supplemental representations for sequential transduction
A unified reranking approach is applied to both grapheme-to-phoneme conversion and machine transliteration demonstrating substantial accuracy improvements by utilizing heterogeneous transliterations and transcriptions of the input word.
WFST-Based Grapheme-to-Phoneme Conversion: Open Source tools for Alignment, Model-Building and Decoding
This paper introduces a new open source, WFST-based toolkit for Grapheme-toPhoneme conversion. The toolkit is efficient, accurate and currently supports a range of features including EM sequence
Multiple Many-to-Many Sequence Alignment for Combining String-Valued Variables: A G2P Experiment
This work investigates multiple many-to-many alignments as a primary step in integrating supplemental information strings in string transduction and applies the approach to boosting baseline G2P systems using homogeneous as well as heterogeneous sources of supplemental information.
Joint-sequence models for grapheme-to-phoneme conversion
A novel estimation algorithm is presented that demonstrates high accuracy on a variety of databases and studies the impact of the maximum approximation in training and transcription, the interaction of model size parameters, n-best list generation, confidence measures, and phoneme-to-grapheme conversion.
Latent-Variable Modeling of String Transductions with Finite-State Methods
A conditional loglinear model is presented for string-to-string transduction that employs overlapping features over latent alignment sequences, and which learns latent classes and latent string pair regions from incomplete training data, and it is demonstrated that latent variables can dramatically improve results, even when trained on small data sets.
Stochastic Contextual Edit Distance and Probabilistic FSTs
This work shows how to construct and train a probabilistic finite-state transducer that computes stochastic contextual edit distance and model typos found in social media text to illustrate the improvement from conditioning on context.
Learning String-Edit Distance
The stochastic model allows us to learn a string-edit distance function from a corpus of examples and is applicable to any string classification problem that may be solved using a similarity function against a database of labeled prototypes.
Substring-Based Transliteration
It is shown that the substring-based transducer not only outperforms a state-of-the-art letterbased approach by a significant margin, but is also orders of magnitude faster.
Joint Generation of Transliterations from Multiple Representations
This work shows that transliterations incorporate information from both spelling and pronunciation, and proposes an effective model for joint transliteration generation from both representations, and generalizes this model to include transliterated from other languages, and enhances it with reranking and lexicon features.
Robust LTS rules with the Combilex speech technology lexicon
A loose comparison with other studies indicates Combilex is a superior quality lexicon in terms of consistency and size.