-
Notifications
You must be signed in to change notification settings - Fork 24
Mappings Comparison
Knews extracts its frame types starting from the output of Boxer and UKB (or alternatively Babelfy). WordNet synsets from UKB are aligned to the event predicate symbol from Boxer, then the WordNet synsets are mapped to FrameNet frames using a pre-existent mapping file. However, this mapping is still small. Recently, STLab developed new mappings within Framester and made them available.
The main objective of this experiment is to evaluate which of these mappings offer better coverage in terms of number of frame instances, frame types, roles and elements.
In this experiment, we used Knews to extract frames over a collection of 1653 documents using the both mapping options (Knews mappings and Framester mappings). For this, we created a new branch where is possible to select the mapping to be used just changing the configuration in this file.
In Table 1 and Table 2 are showed some statistics in the identification of frame instances, frame types, roles and elements using both mappings. As we can see, Knews mappings offer a slightly better coverage in the identification of these items. The main reason why Knews has better coverage is that its mappings have more frames and synsets associated between them (see Table 3).
Table 1: Knews Mappings
| Item | All values | Unique values |
|---|---|---|
| Instances | 114517 | 114517 |
| Types | 97154 | 2274 |
| Elements | 154422 | 3830 |
| Roles | 154422 | 222 |
Table 2: Framester Mappings
| Item | All values | Unique values |
|---|---|---|
| Instances | 114212 | 114212 |
| Types | 95347 | 2013 |
| Elements | 154095 | 3824 |
| Roles | 154095 | 222 |
Table 3: Resources
| Item | Knews Mappings | Framester Mappings |
|---|---|---|
| Frames | 881 | 870 |
| Synsets | 9080 | 8169 |