Big Data Mining and Analytics  2019, Vol. 2 Issue (4): 217-239    DOI: 10.26599/BDMA.2019.9020011
Statistical Learning for Semantic Parsing: A Survey
Qile Zhu, Xiyao Ma, Xiaolin Li*
Qile Zhu, Xiyao Ma, and Xiaolin Li are with National Science Foundation Center for Big Learning, University of Florida, Gainesville, FL 32608, USA. E-mail: valder@ufl.edu; maxiy@ufl.edu.

Abstract

A long-term goal of Artificial Intelligence (AI) is to provide machines with the capability of understanding natural language. Understanding natural language may be referred as the system must produce a correct response to the received input order. This response can be a robot move, an answer to a question, etc. One way to achieve this goal is semantic parsing. It parses utterances into semantic representations called logical form, a representation of many important linguistic phenomena that can be understood by machines. Semantic parsing is a fundamental problem in natural language understanding area. In recent years, researchers have made tremendous progress in this field. In this paper, we review recent algorithms for semantic parsing including both conventional machine learning approaches and deep learning approaches. We first give an overview of a semantic parsing system, then we summary a general way to do semantic parsing in statistical learning. With the rise of deep learning, we will pay more attention on the deep learning based semantic parsing, especially for the application of Knowledge Base Question Answering (KBQA). At last, we survey several benchmarks for KBQA.

Received: 17 September 2018      Published: 06 January 2020
Corresponding Authors: Xiaolin Li
 Table 1 Examples for utterance-action pairs. Fig. 1 Semantic parsing taxonomies by supervisory signal and techniques (names are corresponding to Section 4.1). Table 2 Examples of entity linking. 49]. To narrow down the space of logical predicates, they use (1) course alignment based on Freebase and a text corpus and (2) a bridging operation that generates predicates compatible with neighboring predicates."> Fig. 2 An example to answer a question through a knowledge base[49]. To narrow down the space of logical predicates, they use (1) course alignment based on Freebase and a text corpus and (2) a bridging operation that generates predicates compatible with neighboring predicates. 50]. For each candidate logical form (red), it generated canonical utterances (purple). The model is trained to paraphrase the input utterance (green) into the canonical utterances associated with the correct denotation (blue)."> Fig. 3 An example of paraphrasing for semantic parsing[50]. For each candidate logical form (red), it generated canonical utterances (purple). The model is trained to paraphrase the input utterance (green) into the canonical utterances associated with the correct denotation (blue). 60]."> Fig. 4 Query graph of question "who first voiced Meg on Family Guy" [60]. 60]."> Fig. 5 Legitimate actions to grow a query graph[60]. 70]."> Fig. 6 A SEQ2TREE decoding example for the logical form "AB(C)" [70]. ⩽ and $⩾$ are defined on numbers and dates[71]. 𝕂 is the knowledge base and ε denotes a set of entities."> Fig. 7 Interpreter functions of the NSM. r represents a variable, p is a predicate in Freebase. and are defined on numbers and dates[71]. 𝕂 is the knowledge base and ε denotes a set of entities. R𝟏" ). The memory bridges these two steps to achieve compositionality[71]."> Fig. 8 Semantic parsing with NSM. A special token "GO" indicates the start of decoding, and "Return" indicates the end of decoding. Due to the fact that the decoding model never sees the values in the encoder ( "US" ) here, so it only references them with the name of the variable ( "R𝟏" ). The memory bridges these two steps to achieve compositionality[71]. Table 3 List of domain-general predicates[73]. 73]."> Fig. 9 Actions taken by the transition system for generating the ungrounded meaning representation of the example. Symbols in red indicate domain-general predicates[73]. 79]."> Fig. 10 Framework of semantic parsing via paraphrasing. Firstly, the logical forms are converted deterministically into canonical utterance in natural language. Combing canonical utterance with input utterance as the input, paraphrase model is trained to learn and transfer from the source-domain to the target-domain. External language resources are applied to facilitate domain adaptation[79]. Table 4 Comparison between different word embedding initializations. ES: per-example standardization. FS: per-feature standardization. EN: per-example normalization. Cosine similarity is computed on a randomly selected. Table 5 An overview of all datasets ( "-" means there is not official split for this dataset).