We used APGCv2.1 to train and evaluate the gender identification systems is in gender-id/multi_user. The train/dev/test.txt files are identical to train.arin.tokens/dev.arin.tokens/test.arin.tokens that are available with the release of APGC v2.0. The only difference is that the train/dev/test.txt contain the extended word-level gender annotations (i.e., base word gender + clitic gender). We also corrected two sentences in the train split of APGC v2.0 (sentences with ids: B-8397.1 and C-1225.2).
gender-id/multi_user/augmented_data contains the augmented training data which we used in our augmentation experiments.
The data we used to train and evaluate word-level gender identification for the first-person only version of the task is in gender-id/single_user.
The data we used to train our multi-step gender rewriting systems is in rewrite/apgc-v2.1.
The train.*.tokens/dev.*.tokens/test.*.tokens are identical to train.*.tokens/dev.*.tokens/test.*.tokens that are available with the current release of APGC v2.0. Again, the only difference is that the data we provide contains the extended word-level gender annotations (i.e., base word gender + clitic gender), in addition to the corrected sentences mentioned above. The extended word-level gender annotations were obtained by using the utils/clitic_and_form_tagger.ipynb script.
The data we used to train our out-of-context word-level neural rewriter component is in rewrite/apgc-v2.1/nn_token_data. To create this data we did the following:
- Duplicated rewrite/apgc-v2.1/train.arin.tokens four times and removed all the tokens that are marked as B+B. This results in rewrite/apgc-v2.1/nn_token_data/train.arin.tokens+train.arin.tokens+train.arin.tokens+train.arin.tokens.no_B+B.clean.words
- Concatenated rewrite/apgc-v2.1/train.ar.MM.tokens, rewrite/apgc-v2.1/train.ar.FM.tokens, rewrite/apgc-v2.1/train.ar.MF.tokens, and rewrite/apgc-v2.1/train.ar.FF.tokens and removed all tokens that are maked as B+B. This results in rewrite/apgc-v2.1/nn_token_data/train.ar.MM.tokens+train.ar.FM.tokens+train.ar.MF.tokens+train.ar.FF.tokens.no_B+B.clean.words and rewrite/apgc-v2.1/nn_token_data/train.ar.MM.tokens+train.ar.FM.tokens+train.ar.MF.tokens+train.ar.FF.tokens.no_B+B.clean.gender, where the .words file contains the words and .gender file contains the word-level target gender labels.
- We repeated steps 2) and 3) to create the same files for the dev and test splits.
The above steps are applied using the rewrite/apgc-v2.1/nn_token_data/get_nn_data.sh script.
The training data we used in our augmentation experiments can be found in rewrite/apgc-v2.1/augmentation. The augmented training data used to train the neural rewriter model can be found in rewrite/apgc-v2.1/augmentation/nn_token_data.
The word-level data of the Google Translate outputs which we used in our post-editing experiments are in rewrite/apgc-v2.1/google_MT.
The data we used to train our sentence-level joint baseline rewriting models are in rewrite/apgc-v2.1/joint. This data was created as follows:
- Duplicated the train.arin file that is part of APGC v2.1 four times to create rewrite/apgc-v2.1/joint/train.arin+train.arin+train.arin+train.arin.
- Concatenated train.ar.MM, train.ar.FM, train.ar.MF, and train.ar.FF that are part of APGC v2.1 to create rewrite/apgc-v2.1/joint/train.ar.MM+train.ar.FM+train.ar.MF+train.ar.FF.
- We repeated steps 2) and 3) to create the same files for the dev and test splits.
The .gender files contain the target genders we are modeling (i.e., MM, FM, MF, and FF). The .label files contain the sentence-level labels that are part of APGC v2.1.
The data we used to train our multi-step gender rewriting system on the first-person only version of the task can be found in rewrite/apgc-v1.0.
The gold M2 word-level annotations which we used to evaluate our systems are in m2_edits/. The files in m2_edits/[v1.0|v2.1]/edits/ were created by suing the m2_edits/create_m2_edits.sh script.
We used mlm/train.txt and mlm/dev.txt to fine-tune CAMeLBERT MSA on the MLM task. The train and dev data are identical to the ones train.arin (with the two corrected sentences mentioned above) and dev.arin that are available as part of APGCv2.1.