grounded language understanding: speakers: from the world

Post on 27-Jul-2022

1 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Grounded language understanding:Speakers: From the world to language

Christopher Potts

Stanford Linguistics

CS224u: Natural language understanding

1 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Task formulation and data

Color Utterance

xxxx greenxxxx purplexxxx grapexxxx turquoisexxxx moss greenxxxx pinkish purplexxxx light blue greyxxxx robin’s egg bluexxxx british racing greenxxxx baby puke green

2 / 8

McMahan and Stone 2015

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

x1

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

x1

h1

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

x1

h1

dark

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

x1

h1

dark

light

error signal

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

x1

h1

dark

light

error signal

embedding

derived from x1 and color rep, the initial hidden state

predicted probability distribution over the vocab

one-hot encoding for next word

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s>

x1

h1

dark

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s>

x1 x37

h1

dark

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s>

x1 x37

h1 h2

dark

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s>

x1 x37

h1 h2

dark blue

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s>

x1 x37

h1 h2

dark blue

blue

error signal

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s> blue

x1 x37

h1 h2

dark blue

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s> blue

x1 x37 x11

h1 h2

dark blue

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s> blue

x1 x37 x11

h1 h2 h3

dark blue

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s> blue

x1 x37 x11

h1 h2 h3

dark blue green

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Training with teacher forcing

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

light<s> blue

x1 x37 x11

h1 h2 h3

dark blue green

</s>

error signal

3 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

x1

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

x1

h1

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

<s>

x1

h1

dark

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

dark<s>

x1 x20

h1 h2

dark

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

dark<s>

x1 x20

h1 h2

dark blue

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

dark<s> blue

x1 x20 x11

h1 h2 h3

dark blue

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer: Prediction

DecoderEncoder

208.3, 60, 88.2

color embedding

color rep

dark<s> blue

x1 x20 x11

h1 h2 h3

dark blue </s>

4 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Miscellaneous design choices

• The Encoder and Decoder could have more hiddenlayers. We would expect the layer counts to match tofacilitate the hand-off between Encoder and Decoder,though pooling or copying might work too.

• It seems very common at present for researchers to tiethe embedding and classifier parameters (Press and Wolf2017)

• During training, one might drop teacher forcing a smallpercentage of the time to encourage the model toexplore.

5 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Color describer of Monroe et al. (2016)

DecoderEncoder

HSV

Fourier transform

color rep

light<s> blue

x1 x37 x11

h1 h2 h3

w2 w3 w4

6 / 8

Task Training Prediction Design choices Monroe et al. (2016) Related tasks

Related tasks

Non-linguistic representation ⇒ Language

• Image captioning

• Scene description

• Visual Question Answering(Image + Question-text ⇒ Answer-text)

• Instruction giving (State ⇒ Language)

• . . .

7 / 8

References

References I

Brian McMahan and Matthew Stone. 2015. A Bayesian model of grounded color semantics. Transactions of the Associationfor Computational Linguistics, 3:103–115.

Will Monroe, Noah D. Goodman, and Christopher Potts. 2016. Learning to generate compositional color descriptions. InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2243–2248,Stroudsburg, PA. Association for Computational Linguistics.

Ofir Press and Lior Wolf. 2017. Using the output embedding to improve language models. In Proceedings of the 15thConference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages157–163, Valencia, Spain. Association for Computational Linguistics.

8 / 8

top related