+ All Categories
Home > Documents > Coupling distributed and symbolic execution for natural...

Coupling distributed and symbolic execution for natural...

Date post: 05-Oct-2020
Category:
Upload: others
View: 4 times
Download: 0 times
Share this document with a friend
36
Coupling distributed and symbolic execution for natural language queries Lili Mou Zhengdong Lu Hang Li Zhi Jin
Transcript
Page 1: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Coupling distributed and symbolic execution for natural language queries

Lili Mou Zhengdong Lu Hang Li Zhi Jin

Page 2: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Outline

• Learning the semantics of a question from its execution

• Neural vs. Symbolic

• Our Proposal: coupling the two views

• Conclusion

Page 3: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Outline

• Learning the semantics of a question from its execution

• Neural vs. Symbolic

• Our Proposal: coupling the two views

• Conclusion

Page 4: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Learning the semantics of a question from its execution

NL questionSemanticParser

somethingexecutable

Page 5: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

+ KB

Learning the semantics of a question from its execution

NL questionSemanticParser

somethingexecutable answer

Page 6: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

+ KB

Learning the semantics of a question from its execution

NL questionSemanticParser

somethingexecutable answer

AnswerChecker

groundtruth

correctional signal

Page 7: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

select Duration wherearea = max(area)

Two parsing choices: (1)

Semantic parsing

① Question into a composite SQL-like command (Liang et al.ACL-17)

It is essentially a sequence-to-sequence model, while the output sequence is executable

Page 8: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

STEP-1: Row selection:argmax(area)

Two parsing choices: (2)

① Question into a composite SQL-like command (Liang et al.ACL-17)② Question into a sequence of “primitive” operations (Neelakantan et al. ICLR-16, Yin et al. IJCAI-16)

STEP-2: Value selection:select_value(Duration)

Semantic parsing

Page 9: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Question as a sequence of operations

NL question

executor

controller ① Neural net controller sequentially emitsoperations (with argument)

② Each operation has its own semantics, sothe operation of each step can bepotentially supervised

③ Each operation is applied on KB andmemory from previous operation, with thenew result saved in the current memory

S1 S2 S3 S4

Memoryat time 1

KB

Memoryat time 2

Memoryat time 3

Memoryat time 4

We will use this as our base models

O1 O2 O3 O4

Page 10: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Outline

• Learning the semantics of a question from its execution

• Neural vs. Symbolic

• Our Proposal: coupling the two worlds

• Conclusion

Page 11: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

We choose to parse questions into a sequence of operations

NL question

executor

controllerS1 S2 S3 S4

Memoryat time 1

KB

Memoryat time 2

Memoryat time 3

Memoryat time 4

O1 O2 O3 O4

Page 12: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Symbolic executor vs. Neural executor

Again we have two modeling choices

① Symbolic Executor: the execution is purely symbolic, while the controller is neural net-based, whose optimization objective is non-differentiable

② Neural Executor: everything is “neuralized”, including the executor and the intermediatememory, so the objective is naturally differentiable. Although it is easy to learn, it suffersfrom low execution efficiency and low generalization ability

The choice of Neural Programmer (Neelakantan et al. ICLR-16) is an interesting middlecourse, but we don’t consider it due its limited potential for complex operations

Page 13: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Choice-I: Symbolic executor

• Learning is hard (with reinforcement learning):• relatively big action space: primitive operators x argument• only final reward (when the executions return the correct result)

Bingo!

NL question S1 S2 S3 S4

Memoryat time 1

KB

Memoryat time 2

Memoryat time 3

Memoryat time 4

O1 O2 O3 O4

Page 14: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Examples of symbolic operators

• We limit ourselves with knowledge-base with a single table• Each execution on a table is specified by a primitive operator with an argument

Example: argmax(year) selects the row with the field year having the greatest value

Page 15: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Choice II: Neural executor

NL question S1 S2 S3 S4

Distributedmemory at

time 1

world

a distribution ofpossible predictionsDistributed

memory attime 2

Distributedmemory at

time 3

Distributedmemory at

time 4

• Neural Enquirer (Yin et al. IJCAI-16) as the example: Learning is typically easy through normalback-propagation. It can learn to deal with quite complicated questions

• Its execution efficiency is low due to its fully neural architecture, and the accuracy on parsingcomplex questions is not satisfying

Page 16: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Neural Enquirer: Overall diagram

l Embed the table: keep the table structure, but embed the value and fieldl Fully “neuralized” execution (matrix/vector processing with gating and pooling )l Stacked layers of (Excutor, Memory) pairs to mimic the sequence of operations, while the

memory saves the intermediate result of each layer of execution

Each execution step in Neural Enquirer includes- Soft column attention (this part is naturally interpretable)- Distributed row annotation

Page 17: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Neural vs. Symbolic

Symbolic Neural Wanted

Learning Efficiency Very low High High

Execution efficiency High Low High

Interpretability High Low High

Accuracy Low Low High

Page 18: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Neural vs. Symbolic

Can we (sort-of) have the best of both worlds?

Symbolic Neural Wanted

Learning Efficiency Very low High High

Execution efficiency High Low High

Interpretability High Low High

Accuracy Low Low High

Page 19: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Outline

• Learning the semantics of a question from its execution

• Neural vs. Symbolic

• Our Proposal: coupling the two views

• Conclusion

Page 20: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

General intuition

• Neural models and symbolic models are like two different views of the same complex semanticparsing process

• We can maintain both views in the same system, and let them talk to each other, to encouragesome consistency between the two views

• It is a bit like Multi-view Learning, while in this work the contrast of views come fromintrinsic representation choices, instead from different given aspects or features of thesame object

Page 21: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

The diagram

Neural view (Neural Enquirer)

Symbolic view

① We have both neural andsymbolic view in the samesystem

② There is informationexchange between the twoviews during the training

③ We will use only symbolicview for testing after thetraining is done, for highexecution efficiency

Page 22: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Coupling the two views

General idea: (distributed symbolic)• STEP-1: Train the neural model as in (Yin et al. IJCAI-16) in an end-to-end fashion• STEP-2: Pre-train the field selection part of the symbolic model with the prediction of the

neural model trained in STEP-1 in a step-by-step way• STEP-3: Train the symbolic model with REINFORCE with the execution accuracy as reward

Page 23: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Coupling the two views

General idea: (distributed symbolic)• STEP-1: Train the neural model as in (Yin et al. IJCAI-16) in an end-to-end fashion• STEP-2: Pre-train the field selection part of the symbolic model with the prediction of the

neural model trained in STEP-1 in a step-by-step way• STEP-3: Train the symbolic model with REINFORCE with the execution accuracy as reward

(distributed symbolic distributed)• STEP-4: (Feedback step) Use the symbolic model to train the attention of the neural

model in a step-by-step way

Page 24: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Coupling the two views

General idea: (distributed symbolic)• STEP-1: Train the neural model as in (Yin et al. IJCAI-16) in an end-to-end fashion• STEP-2: Pre-train the field selection part of the symbolic model with the prediction of the

neural model trained in STEP-1 in a step-by-step way• STEP-3: Train the symbolic model with REINFORCE with the execution accuracy as reward

(distributed symbolic distributed)• STEP-4: (Feedback step) Use the symbolic model to train the attention of the neural

model in a step-by-step way

STEP-2 and STEP-4 approximately maintain the consistency of the two viewson field selection

Page 25: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

l Let m be the number of actions to pre-train, J is the function to be maximizel Only the parameters associated with field selection is trained in this phrase, the other parts are

left dangling

l We used supervised learning for pre-training, but many other ways (eg, some smart sampling)may also work

Step-by-step supervision

Imperfect supervision signal fromNeural Enquirer MAP prediction

Pre-training with supervision form neural view

Page 26: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Policy improvement with REINFORCE

• Gradient:

• Reward R : 1 for correct result, 0 otherwise

• Tricks- Exploring with a small probability (0.1)- Subtracting the mean (reinforcement comparison)- Truncate negative reward (reward-inaction

Page 27: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Experimental setting

l Dataset: from (Yin et al. IJCAI-16 )- Synthesized data: table has 10 fields (columns) and 10 rows, about Olympic games

- 25k samples (different queries and tables): tables are randomly generated, the questions andanswers are generated accordingly.

- Many questions are extremely complicated, eg ”How long is the last game which has smaller country size than the game whose host country GDP is 250?”

Page 28: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Pasupat & Liang, ACL-16Compositional semantic parsing on semi-structured tables.

Experimental results: Accuracy

Page 29: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

accuracy on givingthe right answer

Experimental results: Accuracy

Page 30: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

accuracy on being righton every execution

Experimental results: Accuracy

Page 31: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Experimental results: Learning efficiency

Page 32: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Experimental results: Execution efficiency

Page 33: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Experimental results: with feeding-back

Page 34: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Outline

• Learning the semantics of a question from its execution

• Neural vs. Symbolic

• Our Proposal: coupling the two views

• Conclusion

Page 35: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Conclusion and future work

• Coupling the symbolic view and distributed view in one model might be better than either oneworking alone, especially on hard problems

• We are looking for broader more profound ways to combine symbolic model and neuralmodels in real-world semantic parsing tasks

Page 36: Coupling distributed and symbolic execution for natural ...sei.pku.edu.cn/~moull12/resource/coupling-ICML.pdf · Neuralview (NeuralEnquirer) Symbolicview ①Wehavebothneuraland symbolicviewinthesame

Thank you

Poster #36(today)for more questions and job positions at DeeplyCurious


Recommended