Verb Reduction

For each prediction topk replacements $V^i_{S^j}$, we normalise the prediction scores between [0-1]. We exclude the original verb from the normalisation. We do normalise using the following equation:

$$ s_k = \frac{s_k - s_{min}}{s_{max} - s_{min}}$$

For each source verb $v_i$ across all prediction sentences, we add the normalised prediction scores and divide the source verb occurrence to create a list of potential replacements with the scores functioning as how confident the model is that the replacement is

In the case of induce it will look like as follows:

• Now we introduce a [threshold parameter] $\lambda$. We ignore the verb replacements that is lower than $\lambda$:

induce: {produce, express, mediate, cause, induce, stimulate}

We repeat this process for all the verbs in the corpus:

induce: {produce, express, mediate, cause, induce, stimulate}
reveal: {'indicate, determine, display, show, exhibit, suggest, reveal, demonstrate, depict, confirm}, 
block: {block, inhibit, prevent}
...

Find the smallest subset so that at least one verb is selected from each list. This fits the HittingSet (NP-Complete) problem. We a SAT algorithm to extract the subset.