Training VAE on PubTator text to use for transfer and fine tune on the corpora at hand. VAE as input to ranking function.

To Do Now:
1) start writing paper as you do each module!
2) Add corpora overview (i.e. unique documents, cancer types, ratio 0/1 for about_cancer and clinical tasks)
3) Describe preparation (link to BioNLP paper)
4) Desribe training process for a) DL models (used FastText, 75 percentile for sentence length, used masking) with 5-fold CV b) sklearn models (custom tokenizer, tf-idf, no chi2, RandomizedSearchCV with 1- folds, char and word n-gram, hyperparam optimization with 10 combinations)

To Do Future:
1) reindex VIST with new models (after selecting best one)
2) extend ranking function evaluation metrics with MAP, MAR
3) try Logistical Regression and (V)AE as ranking function

Developed VAE for document retrieval.

To Do Now:
1) start writing paper as you do each module!
2) Add corpora overview (i.e. unique documents, cancer types, ratio 0/1 for about_cancer and clinical tasks)
3) Describe preparation (link to BioNLP paper)
4) Desribe training process for a) DL models (used FastText, 75 percentile for sentence length, used masking) with 5-fold CV b) sklearn models (custom tokenizer, tf-idf, no chi2, RandomizedSearchCV with 1- folds, char and word n-gram, hyperparam optimization with 10 combinations)

To Do Future:
1) reindex VIST with new models (after selecting best one)
2) extend ranking function evaluation metrics with MAP, MAR
3) try Logistical Regression and (V)AE as ranking function

Developed VAE for document retrieval.

To Do Now:
1) start writing paper as you do each module!
2) Add corpora overview (i.e. unique documents, cancer types, ratio 0/1 for about_cancer and clinical tasks)
3) Describe preparation (link to BioNLP paper)
4) Desribe training process for a) DL models (used FastText, 75 percentile for sentence length, used masking) with 5-fold CV b) sklearn models (custom tokenizer, tf-idf, no chi2, RandomizedSearchCV with 1- folds, char and word n-gram, hyperparam optimization with 10 combinations)

To Do Future:
1) reindex VIST with new models (after selecting best one)
2) extend ranking function evaluation metrics with MAP, MAR
3) try Logistical Regression and (V)AE as ranking function

To Do Now:
1) start writing paper as you do each module!
2) Add corpora overview (i.e. unique documents, cancer types, ratio 0/1 for about_cancer and clinical tasks)
3) Describe preparation (link to BioNLP paper)
4) Desribe training process for a) DL models (used FastText, 75 percentile for sentence length, used masking) with 5-fold CV b) sklearn models (custom tokenizer, tf-idf, no chi2, RandomizedSearchCV with 1- folds, char and word n-gram, hyperparam optimization with 10 combinations)

To Do Future:
1) reindex VIST with new models (after selecting best one)
2) extend ranking function evaluation metrics with MAP, MAR
3) try Logistical Regression and (V)AE as ranking function

To Do:
1) each corpus (excluding Onco+/- and PubMed (used as clinically not relevant)
2) concat all datasets
3) leave on out (ablation)

Models:
1. HATT
2. MTL
3. Linear SVM
4. Random Forrest

To Do:
1) each corpus (excluding Onco+/- and PubMed (used as clinically not relevant)
2) concat all datasets
3) leave on out (ablation)

Models:
1. HATT
2. MTL
3. Linear SVM
4. Random Forrest

To Do: Evaluate on GS data, seperate for each model.

Not stratified but no unseen train/test classes.
Next: tokenize, convert to sequences, labelize, save all models and store data as such.
Implement MTL models.
Evaluate.

Rewrote data preparation/loading, to support both HATT as well as MTL schema. Added recent_pubmed.p for NotCancer exmples.
Next: create labeling schemata for MTL.
Input: training doc [embedded]
Output { [0,1], [0,1], [0,..,n ]

Hierarchical text classification works. minor issues with tokenization (performed twice, unnecessary).

ToDo:
implement P/R/F1 score on validation data, to compare with stat Ml performance.