This repository provides a balanced dataset for training and evaluating English homograph disambiguation (HD) models, generated with Meta's Llama 2-Chat 70B model.

The dataset contains 3,260 sentences covering the same set of 162 homograph words as in Google's Wikipedia HD dataset. Most words have two pronunciations; two words (august and mobile) have three. For each pronunciation of each homograph word, this dataset contains 10 sentences in which the homograph word occurs exactly once. The 10 sentences are divided evenly into a training set and an evaluation set.

This dataset is intended to be balanced and diverse:

• Each pronunciation of each homograph word gets the same number of sentences. This is unlike Google's Wikipedia HD dataset, in which some pronunciations get very few or no sentences at all.
• For each pronunciation of each homograph, we tried to capture as many word senses as possible. We also ensured the sentences come from diverse domains (e.g. math, chemistry, linguistics), and the homograph words play different roles in a sentence (e.g. as the subject or the object; in the main clause or a relative clause).

The dataset comes in two files: llama_hd_train.tsv and llama_hd_eval.tsv. Each file contains five fields, separated by a tab:

• homograph: The homograph word, in lower case.
• wordid: A string identifying a pronunciation of the homograph word. For the full list of wordids, see wordids.tsv in the repo of the Wikipedia dataset.
• sentence: A sentence containing the homograph word exactly once.
• start, end: The start and end indices of the homograph word in the sentence.

Notes:

1. All the fields are enclosed in double quotes ("), and double quotes within the fields are escaped as two double quotes ("). You can read and unescape the fields in Python using csv.reader(..., delimiter="\t", quotechar='"').
2. The homograph word in the sentence may be capitalized, have surrounding punctuations, take the possessive suffix ('s), or be part of a hyphenated word (e.g. lead-free). You need to make sure your tokenizer can treat the homograph word itself as a token.
3. The start and end indices aren't really necessary, since the homograph word always occurs only once in the sentence. They are included to be consistent with the Wikipedia dataset. Note that the indices are counted in UTF-8 bytes, not Unicode characters, e.g. é (C3 A9 in UTF-8) counts as two bytes.

This dataset is released under the Creative Commons Attribution Non-Commercial (CC-BY-NC) 4.0 license.

This dataset is only intended for training models for homograph disambiguation (HD) and related elementary NLP tasks, such as text normalization (TN) and POS tagging.

If you use this data in a publication, please cite the following paper:

Wonjune Kang, Yun Wang, Shun Zhang, Arthur Hinsvark, Qing He. Multi-task learning for front-end text processing in TTS. In Proceedings of the 2024 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Seoul, Korea, April 2024.