Learning compositional structures for semantic graph parsing

Groschwitz, Jonas; Fowlie, Meaghan; Koller, Alexander

doi:10.18653/v1/2021.spnlp-1.3

Cited by 3 publications

(6 citation statements)

References 13 publications

(17 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We follow and rely on the dependency parsing model of Kiperwasser and Goldberg (2016), which scores each dependency edge by feeding neural represen- tations for the two tokens to an MLP. We train the parser using the setup of Groschwitz et al (2021), which does not require explicit annotations with AM dependency trees.…”

Section: The Am Parsermentioning

confidence: 99%

“…Hyperparameters. For the AM parser, we primarily copy hyperparameter values from the AMR experiments of Groschwitz et al (2021). This helps prevent overfitting on COGS, but we also note that hyperparameter tuning for compositional generalization datasets can be difficult anyways since one can typically easily achieve perfect scores on an indoman dev set.…”

Section: A Training Details Of the Am Parsermentioning

confidence: 99%

“…Copied values include for instance the number of epochs (60 due to supervised loss for edge existence and lexical labels), the batch size, the number and dimensionality of neural network layers and not using early stopping (but selecting best model based on per epoch evaluation metric on the dev set). Choosing 3 sources has worked well on other datasets (Groschwitz et al, 2021) and we adopt this hyperparameter choice. We note that with ditransitive verbs (i.e.…”

Section: A Training Details Of the Am Parsermentioning

confidence: 99%

“…"hedgehog") occur 100 times. For word embeddings, we either use BERT-Largeuncased (Devlin et al, 2019) like Groschwitz et al (2021) or learn embeddings from the dataset only (embedding dimension 1024, same as for the BERT model). We decrease the learning rate from 0.001 to 0.0001: we observed that the learning curves are still converging very quickly and hypothesize that COGS training set might also be easier than the AMR one used in Groschwitz et al (2021).…”

Section: A Training Details Of the Am Parsermentioning

confidence: 99%

“…"touch" would have sources at the nodes b and a (incoming 'agent' or 'theme' edge respectively). These nodes can receive any source out of the three available (S0,S1,S2) 5 , so the tree automaton build as part of Groschwitz et al (2021)'s method would allow any combination of source names for the unfilled 'arguments'. Because there is only one input token, alignment is trivial.…”

Section: Additional Informationmentioning

confidence: 99%

See 4 more Smart Citations

Compositional generalization with a broad-coverage semantic parser

Weißenhorn¹,

Donatelli²,

Koller³

2022

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics

Self Cite

View full text Add to dashboard Cite

We show how the AM parser, a compositional semantic parser (Groschwitz et al., 2018), can solve compositional generalization on the COGS dataset. It is the first semantic parser that achieves high accuracy on both naturally occurring language and the synthetic COGS dataset. We discuss implications for corpus and model design for learning human-like generalization. Our results suggest that compositional generalization can be best achieved by building compositionality into semantic parsers.

show abstract

Section: The Am Parsermentioning

confidence: 99%

Section: A Training Details Of the Am Parsermentioning

confidence: 99%

Section: A Training Details Of the Am Parsermentioning

confidence: 99%

Section: A Training Details Of the Am Parsermentioning

confidence: 99%

Section: Additional Informationmentioning

confidence: 99%

See 3 more Smart Citations

Compositional generalization with a broad-coverage semantic parser

Weißenhorn¹,

Donatelli²,

Koller³

2022

Proceedings of the 11th Joint Conference on Lexical and Computational Semantics

Self Cite

View full text Add to dashboard Cite

show abstract

Compositionality in Computational Linguistics

Donatelli

Koller

2023

Annu. Rev. Linguist.

View full text Add to dashboard Cite

Neural models greatly outperform grammar-based models across many tasks in modern computational linguistics. This raises the question of whether linguistic principles, such as the Principle of Compositionality, still have value as modeling tools. We review the recent literature and find that while an overly strict interpretation of compositionality makes it hard to achieve broad coverage in semantic parsing tasks, compositionality is still necessary for a model to learn the correct linguistic generalizations from limited data. Reconciling both of these qualities requires the careful exploration of a novel design space; we also review some recent results that may help in this exploration. Expected final online publication date for the Annual Review of Linguistics, Volume 9 is January 2023. Please see http://www.annualreviews.org/page/journal/pubdates for revised estimates.

show abstract

SLOG: A Structural Generalization Benchmark for Semantic Parsing

Li,

Donatelli,

Koller

et al. 2023

Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

View full text Add to dashboard Cite

The goal of compositional generalization benchmarks is to evaluate how well models generalize to new complex linguistic expressions. Existing benchmarks often focus on lexical generalization, the interpretation of novel lexical items in syntactic structures familiar from training. Structural generalization tasks, where a model needs to interpret syntactic structures that are themselves unfamiliar from training, are often underrepresented, resulting in overly optimistic perceptions of how well models can generalize. We introduce SLOG, a semantic parsing dataset that extends COGS (Kim and Linzen, 2020) with 17 structural generalization cases. In our experiments, the generalization accuracy of Transformer models, including pretrained ones, only reaches 40.6%, while a structure-aware parser only achieves 70.8%. These results are far from the near-perfect accuracy existing models achieve on COGS, demonstrating the role of SLOG in foregrounding the large discrepancy between models' lexical and structural generalization capacities. * * This work was conducted during Bingzhi Li's visit to NYU. The middle authors are listed in alphabetical order. Training Generalization COGS Emma saw the dog. ; * dog(x3); see.agent(x1,Emma) ∧ see.theme(x1, x3) The cat ran. ; * cat(x1); run.agent(x2, x1) The dog ran. ; * dog(x1); run.agent(x2, x1) SLOG Emma saw the dog that Max held. ; * dog(x3); see.agent(x1,Emma) ∧ see.theme(x1, x3) ∧ dog.nmod(x3, x6) ∧ hold.agent(x6,Max) ∧ hold.theme(x6, x3) The cat ran. ; * cat(x1); run.agent(x2, x1) The dog that Max saw ran. ; * dog(x1); dog.nmod(x1, x4) ∧ see.agent(x4,Max) ∧ see.theme(x4, x1) ∧ run.agent(x5, x1)

show abstract

Learning compositional structures for semantic graph parsing

Cited by 3 publications

References 13 publications

Compositional generalization with a broad-coverage semantic parser

Compositional generalization with a broad-coverage semantic parser

Compositionality in Computational Linguistics

SLOG: A Structural Generalization Benchmark for Semantic Parsing

Contact Info

Product

Resources

About