Machine Translation Evaluation Dataset for Amharic

Machine Translation Evaluation Dataset for Amharic 

The dataset contains source sentences in Amharic and English and their corresponding reference translations that were collected using crowd sourcing. These ground-truth sentences are from across different domains such as news headlines, social media, Wikipedia and everyday conversation.

Metadata of files in the dataset

amen.tsv
  - Domain: news | wiki | twitter | convo
  - Source Sentence: Amharic sentence
  - Reference Translation: English translation
  - Google Translate: output of Google Translate
  - Yandex Translate: output of Yandex Translate

enam.tsv
  - Domain: news | wiki | twitter | convo
  - Source Sentence: English sentence
  - Reference Translation: Amharic translation
  - Google Translate: output of Google Translate
  - Yandex Translate: output of Yandex Translate

Amharic source and reference translations across domains:

News: These are news headlines from Ethiopian news websites.
Wikipedia: A random sample of sentences from the Amharic Wikipedia.
Twitter: Amharic Twitter posts on consumer products.
Conversational: Everyday conversational expressions from Amharic native speakers.

English source and reference translations across domains:

News: These are news headlines from Wikipedia current events portal.
Wikipedia: A random sample of sentences from the English Wikipedia.
Twitter: English Twitter posts on global events from Wikipedia current events portal.
Conversational: Everyday conversational expressions from English native speakers.

Evaluation of two systems that provide Amharic translation

The dataset also contains evaluation of two commercial systems: [Google
Translate](https://translate.google.com/) and [Yandex
Translate](https://translate.yandex.com/). Both systems provide free APIs that
users can sign up and get access keys to. The translations for Amharic to English were generated on 14th
February 2020. The translations for English to Amharic were generated on 30th March 2020.