declare-lab/dialogue-understanding

This repository contains PyTorch implementation for the baseline models from the paper Utterance-level Dialogue Understanding: An Empirical Study

dialogue-understanding's Issues

CNN extractor Prob.

In your paper, the convoluted features are then max-pooled with a window size of 2(consists with the annotations). But in your code the max-pooled window size is the second dimension length of the Conv1d output. Is that actually what you want?

dialogue-understanding/glove-end-to-end/model.py

Line 294 in 4fef597

pooled = [F.max_pool1d(c, c.size(2)).squeeze() for c in convoluted]

I have a question during the training process.

First of all, thank you for doing interesting research.
I have a question in the process of training the roberta-end-to-end model (DialogueRNN).

1. How can i use multi-gpu?

2. Is it possible to use the latest transformers version (4.3.2)?

About DialogueRNN performance Prob.

Yours DialogueRNN performance:（in your endtoend implement, Bidirection & attention on emotion strategy are used.）

Original DialogueRNN performance:

From Above, your implement of DialogueRNN is worse than the original paper?

Performance for DailyDialog

Hi. I have a question for reproducing performance for dailydialog.

1. In the photo, @best Valid F1 values ​​are Test F1 values ​​when validation F1 is the highest?

2. I train the model with batch size =1 due to the computing power problem,.
   Can this be the cause of the difference between the performance of the paper (59.50) and my performance (57.5)?

the Label Shift part is quite confusing to me and the code does not contain this part

Hi, I read about the An Empirical Study paper and don't understand what you did when during the Intra- or Inter- Speaker Shift experiment. Did you select the specific samples only with shifts happen or use some other techniques?
I want to see the detail in the code but apparently they are not published. Would you please kindly explain a little?

A question about DialogRNN

Hi, First, I would like to express my appreciation for your outstanding works in multimodal sentiment analysis, which have greatly inspired me.
But I have one problem about "Party GRU" of your DialogRNN code:

As shown in the figure above, my understanding is Speaker A's party state at time t-1 as input to the Party GRU. right？

Typo in model.py "featutes" instead of "features"

There's a typo in model.py in lines 655 and 836, "featutes" should be "features".

A problem.

Hi, I encounter an error when executing the codes using the command "python train.py --dataset iemocap --classify emotion --cls-model lstm --residual".

Is there something wrong with the codes?

How to run NN and model on novel input data?

Hi,

This looks to be a great system!

First, I am relatively new to NNs, and I've noticed that often times models are not distributed with code, nor is there an easy entry point for using the model on novel data. I am assuming I am missing something involved in the process, but I don't know what that would be. Could you please explain to me why oftentimes ML codebases don't come with pre-trained models or evaluation capabilities?

Second, I have used the following command to begin training a model:

cd roberta-end-to-end && python train.py --dataset dailydialog --classify act --cls-model lstm --residual --batch-size 2

I don't know Python very well. I am wondering is there an easy way to use this trained model, when completed, in order to process texts? I can definitely preprocess the text into any format using Perl. I just don't know how to run/evaluate the NN on specific input data for use in classifying texts (in my case I'm trying to extract commissives from multi-party dialogues (however it looked like these dialogues were two-person dialogues, so perhaps I should have trained the utterance model)).

Thank you for making this software available, it may be a great help to us,

Andrew