Scylla: domain adaptation using frames and qualia

API usage

1. Set up a Google API Key and enable it for the translation service.

• This key can be generated by creating the credentials on Google Cloud console, see the “Google Cloud Console page” at (https://cloud.google.com/cloud-console) for more information.
• Access the URL:

http://server3.framenetbr.ufjf.br:9600/inject?set_gtrans_key=

e.g: http://server3.framenetbr.ufjf.br:9600/inject?set_gtrans_key=AIzaSyCeBay2LRY7DVB_4FUZpaxeq6rsGMJhVEZ

2. Testing Scylla . Use API service "inject", e.g.

http://server3.framenetbr.ufjf.br:9600//inject?sentence=o%20atacante%20marcou%20um%20gol&from_lang=pt&to_lang=en&method=post

Parameters for experiments:

• sentence: the sentence in the source language;
• from_lang: source language - "pt" (for Brazilian Portuguese);
• to_lang - target language - "en" (for English);
• method - define the method for injection:
  • Scylla-T: method = post
  • Scylla-S: method = pre

The result of the request is a JSON Array containing the top results from the injection. Each JSON Object has the following properties:

• original_sentence - sentence in the source language given as input for the injection;
• translation_sentence - translation in the target language for the sentence given as input;
• injected_sentence - resulting translation after passing through the injection process;
• rank - rank of the resulting translation from injection;
• injections - number of injections made by the algorithm in the resulting translation.

Dataset

Dataset contains the files used for experiments:

• File A – Source sentences: 50 Sentences of the Sports domain in Portuguese
• File B - Gold standard translations: Reference Translations of the source sentences translated into English by a professional native speaker translator
• File C - Baseline System output: MT translations of the source sentences, translated into English by the Baseline System (NMT API)
• File C1 - Baseline system output sentences edited by Editor 1 for calculating HTER
• File C2 - Baseline system output sentences edited by Editor 2 for calculating HTER
• File C3 - Baseline system output sentences edited by Editor 2 for calculating HTER
• File D - Scylla-S output: MT translations of the source sentences, translated into English by Scylla-S
• File D1 - Scylla-S output sentences edited by Editor 1 for calculating HTER
• File D2 - Scylla-S output sentences edited by Editor 2 for calculating HTER
• File D3 - Scylla-S output sentences edited by Editor 2 for calculating HTER
• File E - Scylla-T output: MT translations of the source sentences, translated into English by Scylla-T
• File D1 - Scylla-T output sentences edited by Editor 1 for calculating HTER
• File D2 - Scylla-T output sentences edited by Editor 2 for calculating HTER
• File D3 - Scylla-T output sentences edited by Editor 2 for calculating HTER

Database

Daisy processing access directly a FrameNet database to get the wordforms, lexemes, lemmas, LUs and frames. For sake of completion, a database dump is available in this repository.

License

GNU GPLv3 - See the COPYING file for license rights and limitations.