Understanding the effects of word-level linguistic annotations in under-resourced neural machine translation

Proceedings of the 28th International Conference on Computational Linguistics, pages 381–395 Barcelona, Spain (Online), December 8-13, 2020 381 Neural Approaches for Natural Language Interfaces to Databases: A Survey Radu Iacob1, Florin Brad2, Elena-Simona Apostol1, Ciprian-Octavian Truic˘a1, Ionel Hosu1, Traian Rebedea1 1University Politehnica of Bucharest, Romania 2Bitdefender, Bucharest, Romania {radu.iacob,elena.apostol,ciprian.truica}@upb.ro {ionel.hosu,traian.rebedea}@upb.ro fbrad@bitdefender.com Abstract A natural language interface to databases (NLIDB) enables users without technical expertise to easily access information from relational databases. Interest in NLIDBs has resurged in the past years due to the availability of large datasets and improvements to neural sequence-to-sequence models. In this survey we focus on the key design decisions behind current state of the art neural approaches, which we group into encoder and decoder improvements. We highlight the three most important directions, namely linking question tokens to database schema elements (schema linking), better architectures for encoding the textual query taking into account the schema (schema encoding), and improved generation of structured queries using autoregressive neural models (grammar-based decoders). To foster future research, we also present an overview of the most important NLIDB datasets, together with a comparison of the top performing neural models and a short insight into recent non deep learning solutions. 1 Introduction Semantic parsing is the task of mapping natural language (NL) utterances to a formal meaning represen- tation (MR), mainly with the purpose of querying an information source, i.e. a database or a knowledge graph. These machine-readable MRs are sometimes referred to as logical forms (LF). MRs can be ex- pressed using abstract, language-agnostic representations, such as λ-calculus expressions (Zettlemoyer and Collins, 2012) or abstract meaning representations (Banarescu et al., 2013). However, they can also be represented directly in a programming or query language, for example as Python code (Ling et al., 2016) or SQL (structured query language) commands (Zhong et al., 2017). The text-to-SQL semantic parsing task, also known as Natural Language Interfaces to Databases (NLIDB), has gathered strong research interest even in the first wave of AI (Androutsopoulos et al., 1995). Interest in NLIDBs resurfaced in the last five years due to the success of large end-to-end neural architectures for conditional text generation (Cho et al., 2014; Bahdanau et al., 2015) and the availability of larger datasets (Zhong et al., 2017; Brad et al., 2017; Yu et al., 2018c; Yu et al., 2019a). Early systems adopted pattern-matching, handcrafted grammars and rules, syntactic parsers, and in- termediate language representations, which were usually deployed in a multi-stage system (Androut- sopoulos et al., 1995). The main difficulties for the successful application of NLIDB systems include: fuzzy linguistic coverage (user does not know when the system truly understands the query), linguistic vs. conceptual failures (NLIDB failure caused either by linguistic coverage or the inability to resolve database entities), language ambiguity (e.g. incorrect quantifier scoping, incorrect attachment modifier, nominal compound problem, anaphora resolution, elliptical constructs), and out-of-domain queries. However, recent solutions treat the NLIDB problem as a sequence-to-sequence transduction task. In this case, a large end-to-end neural network processes the input sequence (NL utterance X, …

Content truncated to 500 words