Conflict between motion extraction and question naturalness determination

2021-02-09 pKeicho-done pKeicho The context before and after the key phrase is retrieved in “Determining the naturalness of a question”. When a keyphrase containing a verb is extracted by “Motion Extraction”, it may have been converted to its original form, in which case a string search for the keyphrase from the previous input will not find a match.

solution

• A: Change string-based processing to word-based processing.
  • Teacher data created in the past will no longer be available, so training data must be created again.
• B: Putting the information of “not found” on the feature
  • New situations can be taken care of while preserving past teacher data.
  • However, none of the study data on this new situation is available to begin with
• C: Upstream with a naturalness of 0.5 when an exception occurs in feature extraction.
  • If there is no data, I’ll just leave it at 0.5.

I’ll go with C this time.

Devise features at a time in the future when more data will be available.

• Active learning with existing data while accumulating data for new feature calculation methods
• Discard old format data when new data is accumulated.

---

This page is auto-translated from /nishio/動きの抽出と質問自然度判定の衝突 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.