🧾 Semantic parsing & meaning representation

A survey · about 20 minutes to read · every paper named here has its own page in this library

Semantic parsing is the task of turning a natural-language utterance into a formal representation of what it means: a logical formula, a graph of concepts and relations, a database query or a program that a machine can run. Where a syntactic parser tells us how the words of a sentence group together, a semantic parser tells us who did what to whom, under which conditions, and often what should be done about it. The field of semantic parsing and meaning representation covers both sides of that coin: the design of the target formalisms themselves, and the algorithms that map text into them and back out again.

The papers here span more than three decades. They run from early logic-based translation systems of the early 1990s, through statistical learners that induced grammars from pairs of sentences and logical forms, to the annotation of large "sembanks" such as Abstract Meaning Representation (AMR), the rise of cross-domain text-to-SQL benchmarks, and the current wave of large language models that write SQL from a prompt. Two broad communities meet on this page. One treats meaning representation as a linguistic object to be designed, annotated and compared across languages; the other treats semantic parsing as an engineering problem of building natural-language interfaces to databases, knowledge bases, assistants and robots.

The library's collection on this topic leans toward AMR and its descendants, frame semantics, and text-to-SQL, with a lively recent strand on Uniform Meaning Representation and on testing whether language models encode compositional meaning at all.

What semantic parsing is

At its core, semantic parsing maps a string of words to a structure in some meaning representation language (MRL). The MRL can be chosen for very different reasons, and the choice shapes the whole task:

• Executable representations. Logical forms over a fixed database (as in the classic GeoQuery and ATIS corpora), SQL queries, lambda-calculus expressions over a knowledge base, or action sequences for an agent. Here correctness can be checked by running the output and comparing the answer or behaviour, which opens the door to learning from answers rather than annotated forms.
• Broad-coverage semantic graphs. AMR, Discourse Representation Structures (DRS), Uniform Meaning Representation (UMR) and related formalisms aim to describe the meaning of any sentence, independent of a particular application. They are evaluated by comparing graphs with a gold annotation, most often with the Smatch metric.
• Shallow or intermediate representations. Semantic role labeling and frame-semantic parsing identify predicates and label their arguments without building a full logical form; task-oriented dialogue systems use intent-and-slot trees. These are sometimes called "shallow semantic parsing".

The field therefore borders several neighbours. It overlaps heavily with Semantic role labeling, which supplies the predicate-argument backbone of AMR and of frame-semantic parsing; with Syntax & grammar formalisms, because many parsers build meaning compositionally over a syntactic derivation; and with Code & programming languages, since SQL and other programs are increasingly the target. It differs from Question answering in that the intermediate formal structure is the object of interest, not only the final answer.

A recurring distinction is between parsing (text to meaning) and generation (meaning to text). Many of the papers here treat both directions, and a large share of the AMR literature is about generation from graphs, which makes Text generation another close neighbour.

History and milestones

Logic and transfer (1990s)

The oldest work here comes from the knowledge-based tradition of the 1980s and early 1990s. Translation by Quasi Logical Form Transfer used the Core Language Engine to translate between English and Swedish in a car-hire domain, analysing each sentence only as far as a "quasi logical form" and writing transfer rules over that level. The design goals it argued for — modularity, compositionality and reversibility — reappear throughout later work. A few years later, Generation that Exploits Corpus-Based Statistical Knowledge introduced the Nitrogen generator, which already accepted loosely specified meaning structures under the name "Abstract Meaning Representation" and let a corpus-trained language model pick among many candidate phrasings.

Statistical learning of semantic parsers (2000s)

In the 2000s the question became how to learn a parser from data. Two lines ran in parallel. The first was shallow semantic parsing on top of syntactic trees: Shallow Semantic Parsing using Support Vector Machines pushed PropBank role labeling forward with SVM classifiers and richer features, while Open Text Semantic Parsing Using FrameNet and WordNet built a rule-based parser from FrameNet examples. The second line learned to map sentences into full logical forms from paired data, chiefly on the GeoQuery geography corpus. Learning Synchronous Grammars for Semantic Parsing with Lambda Calculus brought machine-translation machinery — word alignment and synchronous grammars — to the problem, and the same authors showed in Generation by Inverting a Semantic Parser that Uses Statistical Machine Translation that the learned rules could be run backwards for generation. Unsupervised Semantic Parsing asked whether logical forms could be induced with no labels at all, clustering paraphrases with Markov logic.

Weak supervision, CCG and frames (2010–2014)

The early 2010s brought weaker forms of supervision and richer grammars. Combinatory Categorial Grammar (CCG) with lambda-calculus semantics became a standard backbone: Lexical Generalization in CCG Grammar Induction for Semantic Parsing factored the lexicon to generalise better, and Weakly Supervised Learning of Semantic Parsers for Mapping Instructions to Actions trained a parser for navigation instructions using only whether the executed actions reached the right place. Frame semantics matured into a practical tool with the SEMAFOR parser, described in Probabilistic Frame-Semantic Parsing and the journal article Frame-Semantic Parsing. Neural similarity models entered with Semantic Parsing for Single-Relation Question Answering.

The single most important event of the period was the release of AMR as an annotation scheme. Abstract Meaning Representation for Sembanking proposed one graph format for whole-sentence meaning and a manually annotated bank to go with it, explicitly modelled on how the Penn Treebank had driven syntactic parsing.

The AMR boom and neural parsing (2015–2019)

The second half of the 2010s is the busiest and most cited period here. AMR parsing became a shared task, with transition-based parsers (Boosting Transition-based AMR Parsing with Refined Actions and Auxiliary Analyzers, CAMR at SemEval-2016 Task 8, An Incremental Parser for Abstract Meaning Representation), syntax-based machine translation (Parsing English into Abstract Meaning Representation Using Syntax-Based Machine Translation) and, from 2017, sequence-to-sequence neural models (Neural AMR, Neural Semantic Parsing by Character-based Translation). AMR-to-text generation grew into its own subfield, and researchers began using AMR inside other applications: summarization, machine translation and paraphrase.

At the same time, the database side of semantic parsing reorganised itself around large cross-domain benchmarks. Improving Text-to-SQL Evaluation Methodology exposed how older splits rewarded memorised query templates, and Spider set a new standard by testing on databases never seen in training. Conversational follow-ups (SParC, CoSQL) and task-oriented dialogue parsing (Semantic Parsing for Task Oriented Dialog using Hierarchical Representations) followed.

Pretraining and language models (2020 onward)

From 2020 the dominant tools became pretrained Transformers. Table-aware pretraining (TaPas, Structure-Grounded Pretraining for Text-to-SQL), serialised-schema encoders (BRIDGE) and graph pretraining for AMR (Graph Pre-training for AMR Parsing and Generation) all appeared in quick succession. By 2023 large language models prompted with a handful of examples were competitive with or better than fine-tuned systems on text-to-SQL, as shown by DIN-SQL and Text-to-SQL Empowered by Large Language Models, while the harder BIRD benchmark showed how far they still were from real analytics work.

The most recent papers turn back to representation design. Uniform Meaning Representation, a cross-lingual successor to AMR, is the focus of a cluster of 2025–2026 papers, many from the workshop on Designing Meaning Representations, and others ask what the old formalisms can tell us about the internal representations of language models.

Main problems and tasks

The papers here address a family of related tasks:

• Sentence-to-graph parsing. Producing an AMR, DRS or UMR graph from a sentence. The hard parts are concepts with no explicit word alignment, reentrancy (one node with several parents), named entities and abstract concepts that no word expresses.
• Graph-to-text generation. Writing a fluent sentence that expresses a given graph. Because AMR omits tense, number, articles and word order, the generator must choose them.
• Text-to-SQL and natural-language interfaces to databases. Translating a question into a query over a relational schema, often one the system has never seen. Schema linking — deciding which words refer to which tables, columns and stored values — is consistently reported as the largest source of error.
• Question answering over knowledge bases and tables. Mapping a question to a knowledge-base lookup or table operation, sometimes without any explicit logical form.
• Instruction following and task-oriented parsing. Mapping commands to agent actions or to nested intent-and-slot trees for virtual assistants.
• Frame-semantic parsing and role labeling. Finding the words that evoke FrameNet frames or PropBank predicates and labelling their arguments.
• Alignment. Linking words to graph nodes, needed to train many parsers and generators.
• Evaluation and comparison of representations. Designing metrics that compare graphs, measuring how well a formalism transfers across languages, and testing its expressive power.

Methods and schools

Grammar-based and logical approaches

The oldest school builds meaning compositionally over a grammar. Rule-based systems such as the Core Language Engine relied on hand-written lexicons and transfer rules. Later work kept the compositional machinery but learned it from data: synchronous context-free grammars whose meaning side is a lambda function (λ-WASP), probabilistic CCG lexicons induced by higher-order unification (UBL and its factored successor FUBL), and weighted CCG parsers trained against execution. The strength of this school is interpretability and principled handling of variables and scope; its weakness has been coverage on open text. The Groningen pipeline behind The Parallel Meaning Bank is a large-scale descendant: CCG parsing, semantic tagging and Boxer combine into DRS annotations that are then projected into translations.

Translation-style and statistical approaches

A strikingly productive idea, visible in many papers here, is to treat semantic parsing as machine translation from natural language into a meaning language. Wong and Mooney aligned words to grammar productions with standard MT aligners; Pust and colleagues retrained a full syntax-based MT pipeline to output AMR trees; Aligning English Strings with Abstract Meaning Representation Graphs adapted IBM alignment models to graph–string pairs. Generation was framed the same way: Generating English from Abstract Meaning Representations learned how to order a linearised graph before phrase-based translation, and Linguistic realisation as machine translation compared phrase-based and neural MT for the job.

Transition-based and graph-based parsing

Borrowing from dependency parsing, transition-based AMR parsers build the graph through a sequence of actions over a stack and buffer, sometimes starting from a dependency tree. They are fast — often linear in sentence length — and easy to enrich with features from other analysers such as role labelers and coreference resolvers. Joint and incremental decoders (AMR Parsing with an Incremental Joint Model) tried to avoid the error propagation of pipelines in which concepts are fixed before relations.

Neural sequence and graph models

From 2017 neural models took over. One camp linearises the target graph and trains an encoder-decoder, relying on tricks such as anonymising named entities, scope brackets, silver data from existing parsers and, later, pretrained models such as BART. Another camp argues that linearisation throws away structure and designs encoders that read the graph directly: graph-state LSTMs (A Graph-to-Sequence Model for AMR-to-Text Generation), graph recurrent networks, and attention restricted to graph neighbours (AMR-To-Text Generation with Graph Transformer). For parsing, AMR Parsing via Graph-Sequence Iterative Inference builds the graph node by node while concept and relation decisions refine each other. The debate between "linearise and use a strong pretrained model" and "respect the graph" runs through the whole recent literature; Investigating Pretrained Language Models for Graph-to-Text Generation found that plain linearisation plus BART or T5 could beat specialised graph encoders, while warning that such models may lean on memorised facts instead of the input structure.

Weak and indirect supervision

Because gold logical forms are expensive, much work learns from cheaper signals: the answer a query returns, the end state of an agent, paraphrase clusters, or annotations produced for other tasks. TaPas skips the logical form entirely and trains a table reader from answers. emrQA generated a large clinical question-answering corpus by attaching logical-form templates to existing medical annotations. The trade-off, noted repeatedly, is spurious programs that reach the right answer by accident.

Prompting large language models

The newest school treats a large language model as the parser and focuses on prompt construction: how the schema is written into the prompt, which examples are shown, and whether the task is decomposed into steps such as schema linking, difficulty classification, generation and self-correction. The papers here report that such pipelines reach top leaderboard positions on Spider, but that messy real databases and the need for outside knowledge still leave a large gap to human experts.

Datasets, resources and benchmarks

Smatch, which scores two graphs by the best overlap of their triples, is the standard measure for AMR; several papers here argue it should be complemented by fine-grained sub-scores, as in the evaluation suite released with AMR-EAGER, or by metrics with graded matches. For text-to-SQL the two common measures are exact set match against the gold query and execution accuracy, the latter crediting any query that returns the right result.

Landmark papers in this library

Abstract Meaning Representation for Sembanking (2013)

Abstract Meaning Representation for Sembanking defined AMR: a rooted, directed graph with labelled edges whose nodes are PropBank frames, English words and special entity concepts. It deliberately drops articles and tense, gives the same graph to sentences that differ only in form, and disclaims being an interlingua. The accompanying sembank and guidelines turned whole-sentence semantics into a supervised learning problem and spawned the parsing, generation and evaluation work that fills much of this page.

Spider (2018)

Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task paired hand-written questions with complex SQL across many databases and, crucially, held out whole databases for testing. Existing parsers performed very poorly under this split. Spider became the reference benchmark for cross-domain text-to-SQL, followed by SParC, CoSQL and many derived datasets.

TaPas (2020)

TaPas: Weakly Supervised Table Parsing via Pre-training extended BERT with row, column and rank embeddings, pretrained it on Wikipedia tables, and trained it to select cells and an aggregation operator directly from answers. It showed that question answering over tables could skip the explicit logical form while matching or beating parsers that produced one.

Weakly supervised parsing of instructions (2013)

Weakly Supervised Learning of Semantic Parsers for Mapping Instructions to Actions used a CCG parser with lambda-calculus meanings and let execution in a simulated environment guide both interpretation and learning. Validating only the path or the end position replaced annotated logical forms, a key demonstration of grounded learning.

Nitrogen (1998)

Generation that Exploits Corpus-Based Statistical Knowledge built a word lattice of many possible realisations from an underspecified meaning structure and ranked them with bigram statistics. It is an early ancestor both of statistical generation and, in name and spirit, of AMR.

Shallow semantic parsing with SVMs (2004)

Shallow Semantic Parsing using Support Vector Machines improved PropBank role labeling with SVMs and new features such as verb clusters and partial tree paths, and documented a sharp drop on text from outside the training newswire. It anchored the role-labeling stage that later AMR parsers drew on.

λ-WASP (2007)

Learning Synchronous Grammars for Semantic Parsing with Lambda Calculus allowed a learned synchronous grammar to output logical forms with bound variables, reaching the best recall on GeoQuery among the systems it compared and working comparably on several languages.

Single-relation question answering (2014)

Semantic Parsing for Single-Relation Question Answering split each question into an entity mention and a relation pattern and scored both with convolutional similarity models, outperforming a rule-based system. It was an early example of neural matching replacing hand-induced lexicons.

SEMAFOR and frame-semantic parsing (2010–2014)

Frame-Semantic Parsing consolidated the SEMAFOR parser: rule-based target detection, a latent-variable frame classifier extended to unseen predicates by graph-based semi-supervised learning, and joint argument decoding with dual decomposition. It made FrameNet structures a practical input for other systems; Frame-Semantic Parsing with Softmax-Margin Segmental RNNs and a Syntactic Scaffold later showed that a neural parser could do without a syntactic parser at test time.

Neural AMR (2017)

Neural AMR: Sequence-to-Sequence Models for Parsing and Generation linearised AMR graphs and trained encoder-decoder models in both directions, overcoming data scarcity by self-training on millions of unlabelled sentences. It set the template for much subsequent neural AMR work.

Large language models for text-to-SQL (2023)

DIN-SQL: Decomposed In-Context Learning of Text-to-SQL with Self-Correction and Text-to-SQL Empowered by Large Language Models: A Benchmark Evaluation showed that carefully designed prompts — decomposition and self-correction in one case, systematic choice of schema format and examples in the other — let general-purpose language models top the Spider leaderboard. Can LLM Already Serve as A Database Interface? introduced BIRD and a query-efficiency metric, and found the best model still far behind human annotators on realistic databases.

Improving text-to-SQL evaluation (2018)

Improving Text-to-SQL Evaluation Methodology cleaned and standardised older corpora and showed that splitting by question lets the same query appear in training and test, so that a simple template-filling baseline looks competitive. Its call for query-based splits shaped the design of later benchmarks.

Unsupervised semantic parsing (2009)

Unsupervised Semantic Parsing converted dependency trees into quasi-logical forms and clustered their parts into paraphrase classes with Markov logic, evaluating the result by answering questions over biomedical abstracts. It remains a reference point for inducing meaning without annotated targets.

AMR in applications (2015–2019)

Toward Abstractive Summarization Using Semantic Representations merged sentence-level AMR graphs and selected a summary subgraph, an early proof of concept for summarising through meaning. Semantic Neural Machine Translation Using AMR added a graph encoder of the source AMR to a translation model and reported gains, larger when training data was scarce. Together with A Graph-to-Sequence Model for AMR-to-Text Generation, these papers made the case that semantic graphs are useful beyond parsing benchmarks.

People

Nathan Schneider (Georgetown University) is the most prolific author on this topic. His papers here run from the SEMAFOR frame-semantic parser (Probabilistic Frame-Semantic Parsing, Frame-Semantic Parsing) through the original AMR paper to extensions of AMR, such as Annotation of Tense and Aspect Semantics for Sentential AMR.

Kevin Knight (now at Texas Christian University) connects the two ends of the AMR story: the Nitrogen generator of 1998 and the 2013 sembank, together with AMR alignment and generation work with Ulf Hermjakob (University of Southern California). Hermjakob also co-authored the syntax-based MT approach to AMR parsing.

Martha Palmer, a creator of PropBank, co-authored the AMR proposal and the cross-lingual comparison Not an Interlingua, But Close, and appears in recent work mapping AMR to UMR. Nianwen Xue (Brandeis University) led the CAMR line of transition-based AMR parsers and the comparison of English AMRs with Chinese and Czech.

Noah A. Smith (Allen Institute) appears on SEMAFOR, on the segmental-RNN frame parser, on AMR-based summarization and on AMR generation with tree transducers. Dipanjan Das was lead author of the SEMAFOR papers, including the semi-supervised extension to unknown predicates. Jeffrey Flanigan (University of California, Santa Cruz) worked on AMR generation, summarization and the multi-sentence extension DocAMR.

Luke Zettlemoyer (University of Washington and Meta) is a central figure of the learning-based school, appearing here on CCG grammar induction, weakly supervised instruction following, Neural AMR and crowdsourced question–answer meaning representations; Yoav Artzi and Tom Kwiatkowski were his co-authors on the grounded and CCG work. Raymond Mooney and Yuk Wah Wong brought statistical MT techniques to parsing and generation with WASP.

Tao Yu, Rui Zhang, Michihiro Yasunaga and Dragomir Radev (Yale University) are behind Spider, SParC and CoSQL, which define much of modern text-to-SQL; Xi Victoria Lin co-authored the conversational datasets and the BRIDGE parser.

Yue Zhang (Westlake University), with Linfeng Song, Daniel Gildea and Zhiguo Wang, developed graph encoders for AMR-to-text generation and AMR-aided translation, and later graph pretraining for AMR.

Johan Bos and Rik van Noord (University of Groningen) represent the formal-semantics tradition: the Parallel Meaning Bank, character-level neural AMR parsing and Bos's analysis Squib: Expressive Power of Abstract Meaning Representations, which showed what AMR cannot express about quantifier scope.

Shira Wein (Amherst College) is among the most active recent authors, studying AMR across languages and using it to measure meaning differences in translation, as in Assessing the Cross-linguistic Utility of Abstract Meaning Representation. Maxime Amblard and Rémi de Vergnette (CNRS and Université de Lorraine) contributed several of the newest papers, on UMR parsing, extracting first-order logic from semantic graphs and using semantic parsing to evaluate language models.

Institutions and places

The topic is strongly American. The University of Colorado Boulder is the current home of authors behind the largest group of papers here, reflecting the PropBank, AMR and UMR work of its computational semantics group. Georgetown University follows, through Schneider and his students' work on AMR generation and evaluation. The University of Southern California, home of the Information Sciences Institute where AMR was born, the University of Washington, Stanford University, Carnegie Mellon University, the Allen Institute, Brandeis University and the University of South Florida are also prominent. Industry contributions come from Amazon and Google, the latter the origin of TaPas.

Outside the United States, China is the second centre, with Westlake University and the Huazhong University of Science and Technology among the listed institutions. In Europe, the University of Groningen anchors the DRS tradition, and France is represented by the Centre National de la Recherche Scientifique, Inria and the LORIA laboratory in Nancy. Czech researchers appear through the comparison of English AMRs with Czech ones.

The papers are spread across the main ACL venues — the annual ACL meeting, EMNLP, NAACL, EACL — together with Transactions of the Association for Computational Linguistics and Computational Linguistics. The single venue with the most papers is the 2026 workshop on Designing Meaning Representations, a sign of how much current effort goes into the representations themselves.

Open questions

• How universal can a meaning representation be? AMR was explicitly English-centred, and cross-lingual studies found that English, Chinese and Czech graphs often diverge. Uniform Meaning Representation, the Parallel Meaning Bank and studies such as Mapping AMR to UMR are attempts to answer this, but agreement on a truly cross-lingual scheme is still open.
• What should a representation include? AMR leaves out tense, aspect, quantifier scope and sentence-crossing links; proposals add tense and aspect, document-level graphs, and quantification. Each extension makes annotation harder. Recent work also asks how idioms and other non-compositional meaning fit, as in Modelling Idiomatic Expressions in Abstract Meaning Representation.
• How should graphs be compared? Smatch is the standard but treats all triples alike and ignores near-synonyms. Graded, decomposed and kernel-based metrics have been proposed, and benchmarks show none of them is robust to every kind of meaning change.
• Does structure still matter? Pretrained language models given a flat string often match or beat graph-aware architectures, yet they can ignore the structure they are given. Whether explicit meaning representations remain useful inside large models — for translation, faithful generation or evaluation — is debated, and papers such as Compositional Meaning Representations in LLMs review the probing evidence.
• Robust database interfaces. Text-to-SQL systems still struggle with schema linking, messy values, outside knowledge, multi-turn context and efficiency, and benchmarks have so far been mostly English and SQLite-based.
• Generalisation. From the coverage gaps of FrameNet documented in Evaluating FrameNet-style semantic parsing to unseen databases in Spider, parsers repeatedly falter on what their training data did not cover.

Neighbouring topics

Semantic role labeling is the closest neighbour: PropBank and FrameNet roles are the core of AMR and of frame-semantic parsing, and some of the earliest work here was called shallow semantic parsing. Code & programming languages shares the text-to-SQL literature, while Evaluation & benchmarks shares the dataset and metric papers. Language models & pretraining supplies the encoders and the prompting methods of the newest work. Machine translation is linked twice over: MT techniques are used to build parsers and generators, and meaning representations are proposed as a pivot for translation. Text generation and Summarization cover AMR-to-text and AMR-based summarization.

Further links run to Syntax & grammar formalisms (CCG, synchronous grammars), Universal Dependencies (combining dependency trees with AMR), Multilingual & cross-lingual NLP (MTOP, cross-lingual AMR), Dialogue systems (task-oriented parsing, CoSQL), Coreference resolution (reentrancy, cross-document event coreference with AMR), Question answering, Information extraction, Idioms & multiword expressions, Interpretability, Knowledge graphs & ontologies, Lexical semantics & word embeddings and Semitic languages, where Arabic text-to-SQL appears.

On the shelves, this topic stands mostly on Machine translation, Natural language processing and Linguistics & languages, with a few papers on AI & machine learning.

Key works in this library

• Abstract Meaning Representation for Sembanking — the definition of AMR and its sembank.
• Spider — cross-domain text-to-SQL with unseen test databases.
• TaPas: Weakly Supervised Table Parsing via Pre-training — table QA from answers, no logical forms.
• Weakly Supervised Learning of Semantic Parsers for Mapping Instructions to Actions — grounded CCG parsing for navigation.
• Generation that Exploits Corpus-Based Statistical Knowledge — the Nitrogen lattice generator.
• Shallow Semantic Parsing using Support Vector Machines — PropBank role labeling with SVMs.
• Learning Synchronous Grammars for Semantic Parsing with Lambda Calculus — λ-WASP on GeoQuery.
• Unsupervised Semantic Parsing — logical forms induced without labels.
• Lexical Generalization in CCG Grammar Induction for Semantic Parsing — factored CCG lexicons.
• Frame-Semantic Parsing — the SEMAFOR parser for FrameNet.
• Semantic Parsing for Single-Relation Question Answering — neural matching for knowledge-base questions.
• Improving Text-to-SQL Evaluation Methodology — query-based splits and cleaned corpora.
• Neural AMR: Sequence-to-Sequence Models for Parsing and Generation — seq2seq AMR with self-training.
• A Graph-to-Sequence Model for AMR-to-Text Generation — graph-state LSTM encoder.
• AMR Parsing via Graph-Sequence Iterative Inference — concept and relation decisions refined jointly.
• Graph Pre-training for AMR Parsing and Generation — self-supervised pretraining on graphs.
• Semantic Parsing for Task Oriented Dialog using Hierarchical Representations — the TOP tree format for assistants.
• MTOP: A Comprehensive Multilingual Task-Oriented Semantic Parsing Benchmark — six-language compositional assistant data.
• SParC: Cross-Domain Semantic Parsing in Context — multi-turn text-to-SQL.
• Bridging Textual and Tabular Data for Cross-Domain Text-to-SQL Semantic Parsing — BRIDGE, a BERT encoder over serialised schemas.
• DIN-SQL: Decomposed In-Context Learning of Text-to-SQL with Self-Correction — prompted decomposition for SQL.
• Can LLM Already Serve as A Database Interface? A BIg Bench for Large-Scale Database Grounded Text-to-SQLs — BIRD and the efficiency score.
• The Parallel Meaning Bank: Towards a Multilingual Corpus of Translations Annotated with Compositional Meaning Representations — DRS annotations across four languages.
• Squib: Expressive Power of Abstract Meaning Representations — AMR in first-order logic and its limits.
• Weisfeiler-Leman in the Bamboo: Novel AMR Graph Metrics and a Benchmark for AMR Graph Similarity — comparing AMR similarity metrics against human judgments.