Notes

The translation approach we propose is implemeted by no modifications on the DA classification solutions being tested. Instead, the mechanism proposed attempts to train and test those models in the target language, by modifying the data fed to it.

Before training the DA classification models, using the translation model, the capability of mapping the monolingual word embedding spaces of all the target languages to the word embedding space of the source (i.e. English) language is obtained by training the translation matrix for each pair of languages. Following that, all the words in the SwDA and MRDA datasets are obtained, in their source language. When training a DA classification model for a target language with our translation approach, using the translation matrices trained beforehand, translation of each word in the dataset to that target language is found. When feeding the word embeddings of those words to the DA classification model, instead of feeding the model with the monolingual embedding of each word in English vector space, the embedding of the translated counterpart is used. Consequently, despite having lost some of the sentence-level information, the model is actually being trained with utterances which have words, and embeddings, in the target language.

Clearly, using the same methodology for both training and testing processes can not yield objective results. Therefore, for testing, a different route is taken. Prior to the training process, for each dataset, each utterance used for testing the performance of the models is translated to the target languages using Google Translate API. [TODO ADD LINK/CITATION] This time, to obtain a more complete translation of the utterance that preserves the sentence-level information, each utterance is considered as a standalone text and is translated as a single text, as opposed to translating each word separately. During the testing phase, the translated version of the testing data is used. The word embeddings of each word occurring in the translated text is collected from the monolingual vector space of the target language.

As a result, a mechanism that transforms both the training and the testing data is achieved, and a valid evaluation strategy of the methods with a target language is devised.