Supplementary code for the work "Local context models for the clause-final verb prediction in Hindi"
-
The main corpus data is taken from here. This is used to calculate the HDMI values and do the further analysis with distance and word-order.
-
To get the animacy annotations of the nouns, we extracted the annotated nouns collected in the data used in this work.
-
The parser used in this work is taken from here.
Other than the parser mentioned in the above section, the code requires to install:
-
conllu (
pip install conllu), pandas (pip install pandas), numpy (pip install numpy)
bayes_model_hindi_train.py <max_sentences> <save_name> trains a n-gram model based on the first max_sentences sentences of the above-mentioned corpus and saves the saved json model as the specified save_name. By default, we use a 4-gram model on the first 5 million sentences. Sentences are simplified based on the scheme mentioned in the paper inside this.
We don't need to learn the probability values of these models in the same way as an n-gram model. Rather, we can infer these on the fly during testing itself. The probabilities of only the n-gram model and corresponding noise distributions (noise_distrs.py) need to be specified and the rest, one can infer from the Bayesian network itself.
bayes_model_hindi_test.py finds the top predictions made by the n-gram model for each condition listed in the file test_file.txt. Similarly, lossy_surp_model_test.py finds the top predictions for the lossy-surprisal models for the conditions in the test_file.txt. One can change the various hyperparameters (which noise distribution to use, minimum probability to consider, how many top predictions etc.) inside these files.
select_top.py selects the top 50 predictions of various models.
We have created a manually annotated database verb_class_database.json and listed the valid grammatical completions valid_completions_per_condition.json. Using this, one can annotate the predictions into various verb classes.
analysis.py <model_name> finds the percentage of grammatical completions per condition, distribution of error types per condition, and the KL-divergence and recall values per condition (and over all) of the predictions of the "model_name" model. It stores them into various csv files.
Then, plots.R reads these csv files and plots various figures.
Note that statistics on human predictions (from the completion study) are not provided in this code.