Thread by @hadyelsahar

Talk3: @raquelfdzrovira talking about representations shaped by dialogue interaction data.

#ACL2019nlp #rep4nlp

"Task-oriented dialogue" is the setup we are discussing now because it gives us success notion to the dialogue analyse

The plan:

Instead of pre-defined symbolic representations for dialogue systems lets model visually grounded agents that learn to "see, ask and guess"

Using a joint architecture for Asking questions using (QG) and Guesser (Img clf) that uses a couple of learning approaches:
* supervised multitask learning
* cooperative learning (using gen. dialogues only to let the agent learn to guess).

arxiv.org/abs/1809.03408

Adding dialogues has a success on test scores of guessing the correct class of the picture.

However Qualitatively we can find that this success is
not attributed to really well-formed dialogues

However, the representations learned by the dialogue generators don't not seem to be useful for other tasks.

"Evaluating the Representational Hub
of Language and Vision Models" Shekhar et al.
aclweb.org/anthology/pape…

Continual learning via Visual question answering.

Inspired by learning order in children (learning WH questions before other questions) for curriculum learning for VQA systems. This enhances results.

Claudio Greco et al. paper from #ACL2019nlp
aclweb.org/anthology/P19-…

still, this isn't continual learning as skills are learned independently. When they try to learn skills sequentially the model suffers less forgetting.

see Table2

The final part of the talk:

"The photobook dataset": new dialogue dataset that exhibits new features of conversations

Participants talk together to see which pictures they have in common. Looks fun!!

The data has some nice trends from psycholinguistics utterances

"The photobook dataset"

See the paper from #ACL2019nlp
arxiv.org/abs/1906.01530

Conclusions & collaborators see pics :)

👆👆👆 👆👆👆👆👆👆👆👆👆

Summary of @raquelfdzrovira talk #3 in #rep4nlp

Thanks, for the great talk that tackles dialogue from a new fresh perspective.

#acl2019nlp

@threadreaderapp unroll