I have a word embedding model with glove algorithm where I compare association of word X with Y over time, using fourr discreet periods. However, size of my data varies for each period. In Period one I have 82 texts, in period 2 I have 169 texts, in period 3 I have 329 texts and in period 4 I have 879 texts. Empirically, I find a clear increase in the association between X and Y over time. Can this be a function of different data sizes? If it could be, would averaging around bootstrapped samples of the corpus across models until the results stabilizes help? Even if I follow such an approach, can the data size still effect the association of X and Y over time even if I use a bootstrapped sampling approach?
Are there any other techniques I can use. For example, is t better to firstly randomly selecting 82 speeches from each period then running large number of models based on bootstrapped samples and avarage them?