Grounded Unsupervised Semantic Parsing
Poon, Hoifung

Article Structure

Abstract

We present the first unsupervised approach for semantic parsing that rivals the accuracy of supervised approaches in translating natural-language questions to database queries.

Introduction

Semantic parsing maps text to a formal meaning representation such as logical forms or structured queries.

Background

2.1 Semantic Parsing

Grounded Unsupervised Semantic Parsing

In this section, we present the GUSP system for grounded unsupervised semantic parsing.

Experiments

4.1 Task

Conclusion

This paper introduces grounded unsupervised semantic parsing, which leverages available database for indirect supervision and uses a grounded meaning representation to account for syntax-semantics mismatch in dependency-based semantic parsing.

Topics

semantic parse

Appears in 43 sentences as: semantic parse (17) semantic parser (1) semantic parses (5) Semantic Parsing (3) Semantic parsing (2) semantic parsing (17)
In Grounded Unsupervised Semantic Parsing
  1. We present the first unsupervised approach for semantic parsing that rivals the accuracy of supervised approaches in translating natural-language questions to database queries.
    Page 1, “Abstract”
  2. Our GUSP system produces a semantic parse by annotating the dependency-tree nodes and edges with latent states, and learns a probabilistic grammar using EM.
    Page 1, “Abstract”
  3. Semantic parsing maps text to a formal meaning representation such as logical forms or structured queries.
    Page 1, “Introduction”
  4. Recently, there has been a burgeoning interest in developing machine-leaming approaches for semantic parsing (Zettlemoyer and Collins, 2005; Zettlemoyer and Collins, 2007; Mooney, 2007; Kwiatkowski et al., 2011), but the predominant paradigm uses supervised learning, which requires example annotations that are costly to obtain.
    Page 1, “Introduction”
  5. Poon & Domingos (2009, 2010) proposed the USP system for unsupervised semantic parsing , which learns a parser by recursively clustering and composing synonymous expressions.
    Page 1, “Introduction”
  6. In this paper, we present the GUSP system, which combines unsupervised semantic parsing with grounded learning from a database.
    Page 1, “Introduction”
  7. GUSP starts with the dependency tree of a sentence and produces a semantic parse by annotating the nodes and edges with latent semantic states derived from the database.
    Page 1, “Introduction”
  8. Unlike USP, GUSP predetermines the target logical forms based on the database schema, which alleviates the difficulty in learning and ensures that the output semantic parses can be directly used in querying the database.
    Page 1, “Introduction”
  9. We evaluated GUSP on end-to-end question answering using the ATIS dataset for semantic parsing (Zettlemoyer and Collins, 2007).
    Page 2, “Introduction”
  10. 2.1 Semantic Parsing
    Page 2, “Background”
  11. The goal of semantic parsing is to map text to a complete and detailed meaning representation (Mooney, 2007).
    Page 2, “Background”

See all papers in Proc. ACL 2013 that mention semantic parse.

See all papers in Proc. ACL that mention semantic parse.

Back to top.

dependency tree

Appears in 18 sentences as: dependency tree (14) dependency trees (5)
In Grounded Unsupervised Semantic Parsing
  1. GUSP starts with the dependency tree of a sentence and produces a semantic parse by annotating the nodes and edges with latent semantic states derived from the database.
    Page 1, “Introduction”
  2. Poon & Domingos (2009, 2010) induce a meaning representation by clustering synonymous lambda-calculus forms stemming from partitions of dependency trees .
    Page 2, “Background”
  3. USP defines a probabilistic model over the dependency tree and semantic parse using Markov logic (Domingos and Lowd, 2009), and recursively clusters and composes synonymous dependency treelets using a hard EM-like procedure.
    Page 2, “Background”
  4. Top: the dependency tree of the sentence is annotated with latent semantic states by GUSP.
    Page 3, “Background”
  5. GUSP produces a semantic parse of the question by annotating its dependency tree with latent semantic states.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  6. Second, in contrast to most existing approaches for semantic parsing, GUSP starts directly from dependency trees and focuses on translating them into semantic parses.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  7. To combat this problem, GUSP introduces a novel dependency-based meaning representation with an augmented state space to account for semantic relations that are nonlocal in the dependency tree .
    Page 3, “Grounded Unsupervised Semantic Parsing”
  8. GUSP’s approach of starting directly from dependency tree is inspired by USP.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  9. a dependency tree and annotating it.
    Page 4, “Grounded Unsupervised Semantic Parsing”
  10. Let d be a dependency tree , N (d) and E(d) be its nodes and edges.
    Page 4, “Grounded Unsupervised Semantic Parsing”
  11. At the core of GUSP is a joint probability distribution P9 (d, 2) over the dependency tree and the semantic parse.
    Page 4, “Grounded Unsupervised Semantic Parsing”

See all papers in Proc. ACL 2013 that mention dependency tree.

See all papers in Proc. ACL that mention dependency tree.

Back to top.

logical forms

Appears in 14 sentences as: logic forms (1) logical form (1) logical forms (12)
In Grounded Unsupervised Semantic Parsing
  1. Semantic parsing maps text to a formal meaning representation such as logical forms or structured queries.
    Page 1, “Introduction”
  2. However, although these methods exonerate annotators from mastering specialized logical forms , finding the answers for complex ques-
    Page 1, “Introduction”
  3. While their approach completely obviates the need for direct supervision, their target logic forms are self-induced clusters, which do not align with existing database or ontology.
    Page 1, “Introduction”
  4. Unlike USP, GUSP predetermines the target logical forms based on the database schema, which alleviates the difficulty in learning and ensures that the output semantic parses can be directly used in querying the database.
    Page 1, “Introduction”
  5. (2011) used the annotated logical forms to compute answers for their experiments.
    Page 1, “Introduction”
  6. GUSP is unsupervised and does not require example logical forms or question-answer pairs.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  7. The ZC07 dataset contains annotated logical forms for each sentence, which we do not use.
    Page 7, “Experiments”
  8. Since our goal is not to produce a specific logical form , we directly evaluate on the end-to-end task of translating questions into database queries and measure question-answering accuracy.
    Page 7, “Experiments”
  9. Both ZC07 and FUBL used annotated logical forms in training, whereas GUSP-FULL and GUSP++ did not.
    Page 8, “Experiments”
  10. The numbers for GUSP-FULL and GUSP++ are end-to-end question answering accuracy, whereas the numbers for ZC07 and FUBL are recall on exact match in logical forms .
    Page 8, “Experiments”
  11. (Note that this ambiguity is not present in the ZC07 logical forms , which use a single predicate from (f, c) for the entire relation paths.
    Page 8, “Experiments”

See all papers in Proc. ACL 2013 that mention logical forms.

See all papers in Proc. ACL that mention logical forms.

Back to top.

meaning representation

Appears in 11 sentences as: meaning representation (11)
In Grounded Unsupervised Semantic Parsing
  1. Semantic parsing maps text to a formal meaning representation such as logical forms or structured queries.
    Page 1, “Introduction”
  2. To handle syntax-semantics mismatch, GUSP introduces a novel dependency-based meaning representation
    Page 1, “Introduction”
  3. The goal of semantic parsing is to map text to a complete and detailed meaning representation (Mooney, 2007).
    Page 2, “Background”
  4. The standard language for meaning representation is first-order logic or a sublanguage, such as FunQL (Kate et al., 2005; Clarke et al., 2010) and lambda calculus (Zettlemoyer and Collins, 2005; Zettlemoyer and Collins, 2007).
    Page 2, “Background”
  5. Poon & Domingos (2009, 2010) induce a meaning representation by clustering synonymous lambda-calculus forms stemming from partitions of dependency trees.
    Page 2, “Background”
  6. In this problem setting, a natural-language question is first translated into a meaning representation by semantic parsing, and then converted into a structured query such as SQL to obtain answer from the database.
    Page 2, “Background”
  7. To combat this problem, GUSP introduces a novel dependency-based meaning representation with an augmented state space to account for semantic relations that are nonlocal in the dependency tree.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  8. However, GUSP uses a different meaning representation defined over individual nodes and edges, rather than partitions, which enables linear-time exact inference.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  9. Their approach alleviates some complexity in the meaning representation for handling syntax-semantics mismatch, but it has to search over a much larger search space involving exponentially many candidate trees.
    Page 4, “Grounded Unsupervised Semantic Parsing”
  10. In the remainder of this section, we first formalize the problem setting and introduce the GUSP meaning representation .
    Page 4, “Grounded Unsupervised Semantic Parsing”
  11. This paper introduces grounded unsupervised semantic parsing, which leverages available database for indirect supervision and uses a grounded meaning representation to account for syntax-semantics mismatch in dependency-based semantic parsing.
    Page 9, “Conclusion”

See all papers in Proc. ACL 2013 that mention meaning representation.

See all papers in Proc. ACL that mention meaning representation.

Back to top.

question answering

Appears in 7 sentences as: question answering (7)
In Grounded Unsupervised Semantic Parsing
  1. We evaluated GUSP on end-to-end question answering using the ATIS dataset for semantic parsing (Zettlemoyer and Collins, 2007).
    Page 2, “Introduction”
  2. Despite these challenges, GUSP attains an accuracy of 84% in end-to-end question answering , effectively tying with the state-of-the-art supervised approaches (85% by Zettlemoyer & Collins (2007), 83% by Kwiatkowski et al.
    Page 2, “Introduction”
  3. (2003, 2004) proposed the PRECISE system, which does not require labeled examples and can be directly applied to question answering with a database.
    Page 2, “Background”
  4. Figure 1: End-to-end question answering by GUSP for sentence get flight from toronto to san diego stopping in dtw.
    Page 3, “Background”
  5. Figure 1 shows an example of end-to-end question answering using GUSP.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  6. The numbers for GUSP-FULL and GUSP++ are end-to-end question answering accuracy, whereas the numbers for ZC07 and FUBL are recall on exact match in logical forms.
    Page 8, “Experiments”
  7. Table 2: Comparison of question answering accuracy in ablation experiments.
    Page 9, “Experiments”

See all papers in Proc. ACL 2013 that mention question answering.

See all papers in Proc. ACL that mention question answering.

Back to top.

end-to-end

Appears in 6 sentences as: End-to-end (1) end-to-end (5)
In Grounded Unsupervised Semantic Parsing
  1. We evaluated GUSP on end-to-end question answering using the ATIS dataset for semantic parsing (Zettlemoyer and Collins, 2007).
    Page 2, “Introduction”
  2. Despite these challenges, GUSP attains an accuracy of 84% in end-to-end question answering, effectively tying with the state-of-the-art supervised approaches (85% by Zettlemoyer & Collins (2007), 83% by Kwiatkowski et al.
    Page 2, “Introduction”
  3. Figure 1: End-to-end question answering by GUSP for sentence get flight from toronto to san diego stopping in dtw.
    Page 3, “Background”
  4. Figure 1 shows an example of end-to-end question answering using GUSP.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  5. Since our goal is not to produce a specific logical form, we directly evaluate on the end-to-end task of translating questions into database queries and measure question-answering accuracy.
    Page 7, “Experiments”
  6. The numbers for GUSP-FULL and GUSP++ are end-to-end question answering accuracy, whereas the numbers for ZC07 and FUBL are recall on exact match in logical forms.
    Page 8, “Experiments”

See all papers in Proc. ACL 2013 that mention end-to-end.

See all papers in Proc. ACL that mention end-to-end.

Back to top.

ime

Appears in 4 sentences as: ime (4)
In Grounded Unsupervised Semantic Parsing
  1. arrivaLt ime ).
    Page 4, “Grounded Unsupervised Semantic Parsing”
  2. departure_t ime or ticket price fare .
    Page 4, “Grounded Unsupervised Semantic Parsing”
  3. departure_t ime , and so the node state P : flight .
    Page 5, “Grounded Unsupervised Semantic Parsing”
  4. departure_t ime : argmin is created and can be assigned to superlatives such as earliest.
    Page 5, “Grounded Unsupervised Semantic Parsing”

See all papers in Proc. ACL 2013 that mention ime.

See all papers in Proc. ACL that mention ime.

Back to top.

semantic relations

Appears in 4 sentences as: semantic relations (4)
In Grounded Unsupervised Semantic Parsing
  1. by augmenting the state space to represent semantic relations beyond immediate dependency neighborhood.
    Page 2, “Introduction”
  2. In particular, dependency edges are often indicative of semantic relations .
    Page 3, “Grounded Unsupervised Semantic Parsing”
  3. To combat this problem, GUSP introduces a novel dependency-based meaning representation with an augmented state space to account for semantic relations that are nonlocal in the dependency tree.
    Page 3, “Grounded Unsupervised Semantic Parsing”
  4. GUSP only creates edge states for relational join paths up to length four, as longer paths rarely correspond to meaningful semantic relations .
    Page 5, “Grounded Unsupervised Semantic Parsing”

See all papers in Proc. ACL 2013 that mention semantic relations.

See all papers in Proc. ACL that mention semantic relations.

Back to top.

development set

Appears in 3 sentences as: development set (3)
In Grounded Unsupervised Semantic Parsing
  1. In preliminary experiments on the development set , we found that the naive model (with multinomials as conditional probabilities) did not perform well in EM.
    Page 6, “Grounded Unsupervised Semantic Parsing”
  2. We used the development set for initial development and tuning hyperparameters.
    Page 8, “Experiments”
  3. For the GUSP system, we set the hyperparame-ters from initial experiments on the development set , and used them in all subsequent experiments.
    Page 8, “Experiments”

See all papers in Proc. ACL 2013 that mention development set.

See all papers in Proc. ACL that mention development set.

Back to top.

latent semantic

Appears in 3 sentences as: latent semantic (3)
In Grounded Unsupervised Semantic Parsing
  1. GUSP starts with the dependency tree of a sentence and produces a semantic parse by annotating the nodes and edges with latent semantic states derived from the database.
    Page 1, “Introduction”
  2. Top: the dependency tree of the sentence is annotated with latent semantic states by GUSP.
    Page 3, “Background”
  3. GUSP produces a semantic parse of the question by annotating its dependency tree with latent semantic states.
    Page 3, “Grounded Unsupervised Semantic Parsing”

See all papers in Proc. ACL 2013 that mention latent semantic.

See all papers in Proc. ACL that mention latent semantic.

Back to top.

syntactic parsing

Appears in 3 sentences as: syntactic parsers (1) syntactic parsing (2)
In Grounded Unsupervised Semantic Parsing
  1. However, in complex sentences, syntax and semantic often diverge, either due to their differing goals or simply stemming from syntactic parsing errors.
    Page 5, “Grounded Unsupervised Semantic Parsing”
  2. Upon manual inspection, many of the remaining errors are due to syntactic parsing errors that are too severe to fix.
    Page 9, “Experiments”
  3. This is partly due to the fact that ATIS sentences are out of domain compared to the newswired text on which the syntactic parsers were trained.
    Page 9, “Experiments”

See all papers in Proc. ACL 2013 that mention syntactic parsing.

See all papers in Proc. ACL that mention syntactic parsing.

Back to top.

PDF available from www.aclweb.org.
Send us an email