How to interpret negative scores produced by SummaCzs

[chris-opendata]

I got negative scores from SummaCzs model following the example usage with my data. For example, scores look like the following.

        "scores": {
            "for_reference_summary": [
                -0.193,
                -0.746
            ],
            "for_generated_summary": [
                -0.259,
                -0.402
            ]
        },

Is it normal to have negative scores? How can I interpret the negative scores if so?

Thanks

[tingofurro]

Hello @chris-opendata,

Yes, SummaC-ZS scores in the default setting range between [-1, 1] as they are calculated as:

P(entail) - P(contradict)

(so the contradict component can be negative).
Optionally, you can set SummaCZS([...], use_con=False) which will remove the second component in the equation above and always return a positive number.

Experiments in the paper did confirm that the scores are slightly improved when considering the difference, rather than either on their own, which is why we chose this setting for the default.

I hope this helps,

Philippe

[chris-opendata]

Hi Tingofurro,

Thank you for your quick reply and clarification. I will do as you suggested.

Chris