Visual Analytics for Distributional Semantic Model Comparisons

Distributional semantic models have shown to be a successful technique for Word Sense Disambiguation and Word Sense Induction tasks. However, these models, and more specifically the token-level variants, are extremely parameter-rich. We are still in the dark on how the different parameters can be efficiently set and even more on how to evaluate the outcome when no gold standard is readily available. To gain a better insight, we are developing a visual analytics approach which shows these models in two ways: a scatterplot matrix for inter-model parameter comparison and zoomable individual scatter plots allowing for more details on-demand. More specifically, we first use a scatterplot matrix to compare models with different parameter settings in a single view. This enables us to track selections of tokens over different models. On top of this, we create a scatter plot for each individual model, enriched with both model dependent and model independent features. This way, we can have a more in-depth visual analysis of what is going on and visualise the distinct properties or parameters of the individual model.