Robust Semantic Parsing with Adversarial Learning for Domain Generalization

Abstract

This paper addresses the issue of generalization for Semantic Parsing in an adversarial framework. Building models that are more robust to inter-document variability is crucial for the integration of Semantic Parsing technologies in real applications. The underlying question throughout this study is whether adversarial learning can be used to train models on a higher level of abstraction in order to increase their robustness to lexical and stylistic variations. We propose to perform Semantic Parsing with a domain classification adversarial task without explicit knowledge of the domain. The strategy is first evaluated on a French corpus of encyclopedic documents, annotated with FrameNet, in an information retrieval perspective, then on PropBank Semantic Role Labeling task on the CoNLL-2005 benchmark. We show that adversarial learning increases all models generalization capabilities both on in and out-of-domain data.

MaskParse@Deskin at SemEval-2019 Task 1: Cross-lingual UCCA Semantic Parsing using Recursive Masked Sequence Tagging

“…In prior experiments we introduced an adversarial objective similar to (Kim et al, 2017;Marzinotto et al, 2019) to build a language independent representation. However, the language imbalance on the training data did not allow us to take advantage from this technique.…”

Adapting a FrameNet Semantic Parser for Spoken Language Understanding Using Adversarial Learning

“…In Natural Language Processing tasks, this approach has been used to build crosslingual models, doing transfer learning from English to low resource languages for POS tagging [4] and sentiment analysis [14], by using language classifiers with an adversarial objective to train task-specific but language agnostic representations. This technique is not only useful in cross-lingual transfer prob-lems, as it has been used to improve generalization in a document classification [15], Q&A systems [16], duplicate question detection [17] and semantic parsing [5] in a monolingual setup.…”

“…We use a BIO label encoding in all our experiments. On inference, we apply the coherence filter [5] that selects the most probable Frame for the LU and filters all incompatible FE. To ensure that output sequences respect the BIO constrains we implement an A * decoding strategy as the one proposed by [21].…”

“…Adversarial domain training was initially proposed in [3] and adapted for semantic parsing in [5]. We start from our biGRU semantic parser and on the last hidden layer, we stack a CNN with a decision layer to implement a domain classifier (called adversarial task).…”

“…A recent approach attempting to tackle domain biases and build robust systems consists in using neural networks and adversarial learning to build domain independent representations [3]. In the NLP community, this method has been mostly used for cross-lingual transfer learning [4] and more recently in monolingual setups in order to alleviate domain bias in semantic parsers [5].…”