From 31d2e3e70251e9fb56517186d23f804659f3fed2 Mon Sep 17 00:00:00 2001 From: Carel van Niekerk <40663106+carelvniekerk@users.noreply.github.com> Date: Thu, 23 Feb 2023 10:19:32 +0100 Subject: [PATCH] Update BERTNLU README (#137) --- convlab/nlu/jointBERT/unified_datasets/README.md | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/convlab/nlu/jointBERT/unified_datasets/README.md b/convlab/nlu/jointBERT/unified_datasets/README.md index baf7bc12..0ba28589 100755 --- a/convlab/nlu/jointBERT/unified_datasets/README.md +++ b/convlab/nlu/jointBERT/unified_datasets/README.md @@ -7,6 +7,10 @@ We support training BERTNLU on datasets that are in our unified format. ## Usage +#### Important note! + +The BERTNLU codebase utilizes the speaker parameter to identify the entity that is speaking. If a model is trained with `speaker=user`, it should be utilized by the system agent to anticipate user actions. When setting up a dialogue system, it is crucial to select a checkpoint that has been trained for `speaker=user` for the NLU of the system agent and visa-versa. It is important to differentiate between the `speaker` and the agent's name to avoid confusion. The zipped downloadable model names have the following format `bertnlu_unified_<dataset_name>_<speaker>_context<context_len>.zip`. + #### Preprocess data ```sh -- GitLab