A Statistical Model of Word Rank Evolution

The availability of large linguistic data sets enables data-driven approaches to study linguistic change. This work explores the word rank dynamics of eight languages by investigating the Google Books corpus unigram frequency data set. We observed the rank changes of the unigrams from 1900 to 2008 and compared it to a Wright-Fisher inspired model that we developed for our analysis. The model simulates a neutral evolutionary process with the restriction of having no disappearing words. This work explains the mathematical framework of the model - written as a Markov Chain with multinomial transition probabilities - to show how frequencies of words change in time. From our observations in the data and our model, word rank stability shows two types of characteristics: (1) the increase/decrease in ranks are monotonic, or (2) the average rank stays the same. Based on our model, high-ranked words tend to be more stable while low-ranked words tend to be more volatile. Some words change in ranks in two ways: (a) by an accumulation of small increasing/decreasing rank changes in time and (b) by shocks of increase/decrease in ranks. Most of the stopwords and Swadesh words are observed to be stable in ranks across eight languages. These signatures suggest unigram frequencies in all languages have changed in a manner inconsistent with a purely neutral evolutionary process.

Grid Search Hyperparameter Benchmarking of BERT, ALBERT, and LongFormer on DuoRC

The purpose of this project is to evaluate three language models named BERT, ALBERT, and LongFormer on the Question Answering dataset called DuoRC. The language model task has two inputs, a question, and a context. The context is a paragraph or an entire document while the output is the answer based on the context. The goal is to perform grid search hyperparameter fine-tuning using DuoRC. Pretrained weights of the models are taken from the Huggingface library. Different sets of hyperparameters are used to fine-tune the models using two versions of DuoRC which are the SelfRC and the ParaphraseRC. The results show that the ALBERT (pretrained using the SQuAD1 dataset) has an F1 score of 76.4 and an accuracy score of 68.52 after fine-tuning on the SelfRC dataset. The Longformer model (pretrained using the SQuAD and SelfRC datasets) has an F1 score of 52.58 and an accuracy score of 46.60 after fine-tuning on the ParaphraseRC dataset. The current results outperformed the results from the previous model by DuoRC.