Event/entity filtering from AMR output

[elizlee]

• [ ] Events: ignore events with uninformative PB frames, e.g. "to be" (be-02), "to have" (have-03) Example:

  "event": {
                  "text": "have",
                  "mention_id": null,
                  "entity": null,
                  "doc_id": "L0C04CA23",
                  "span": [
                      304,
                      338
                  ],
                  "confidence": 0.6699219935060171,
                  "qnode_id": "Q76664561",
                  "description": "to experience something, to be a target of something",
                  "from_query": "have-org-role-91"
              },

• [x] Events: prevent duplicate events in the same claim sentence Example:

  {
              "event": {
                  "text": "saying",
                  "mention_id": null,
                  "entity": null,
                  "doc_id": "L0C04CA23",
                  "span": [
                      1153,
                      1161
                  ],
                  "confidence": 0.8630922945557871,
                  "qnode_id": "Q3026787",
                  "description": "short, usually meaningful phrase, such as a maxim, proverb, motto, or adage",
                  "from_query": "say-01"
              },
              "args": {}
          },
          {
              "event": {
                  "text": "saying",
                  "mention_id": null,
                  "entity": null,
                  "doc_id": "L0C04CA23",
                  "span": [
                      1153,
                      1161
                  ],
                  "confidence": 0.8630922945557871,
                  "qnode_id": "Q3026787",
                  "description": "short, usually meaningful phrase, such as a maxim, proverb, motto, or adage",
                  "from_query": "say-01"
              },
              "args": {}
          },

• [ ] Entities: ignore arguments with PB labels, which imply that they are events and not entities Example:

  "A0": {
                      "type": {
                          "text": "research",
                          "mention_id": null,
                          "entity": null,
                          "doc_id": "L0C04CA23",
                          "span": [
                              283,
                              293
                          ],
                          "confidence": 1.0,
                          "qnode_id": "Q42240",
                          "description": "systematic study undertaken to increase knowledge",
                          "from_query": "research-01"
                      }
                  }

  Examples are from /nas/gaia/users/elee/phase3_test/cdse_dryrun_id_match/WORKING/en/isi_ttl_output/output.json. Reproduce by running the cdse-covid pipeline with input /nas/gaia/users/elee/phase3_test/cdse_dryrun_id_match/WORKING/en/ttl_output/claim_all.json through all steps until convert_claims_to_json.py is finished. Observe the json output.

[elizlee]

Thoughts so far:

Strategy for event task 1: create a dict with all PB frames that I don't believe we'd be interested in extracting events for.

Event task 2: I still need to investigate why there are duplicate events.

Entity task: Several "event" arguments yield nominal qnodes, like in the provided example ("research"). How much would we be losing if we filtered these out? Is there some other check we can use instead?