build your own website

BroadSem: Induction of Broad Coverage Semantic Parsers

BroadSem is a research project hosted at the University of Edinburgh and University of Amsterdam and led by Ivan Titov. The project is supported by an ERC starting grant (678254).   The key bottleneck for natural language processing is the lack of accurate methods for producing meaning representations of texts and reasoning with these representations. BroadSem aims at addressing these shortcoming by providing effective methods for semantic parsing, relying deep generative modeling, and developing methods for incorporating meaning representations in downstream applications (e.g., machine translation and question answering).

Our team

Mobirise

Ivan Titov


Team leader

Mobirise

Phong Le

BroadSem post-doc
(2017 -)

Mobirise

Chunchuan Lyu

BroadSem PhD student
(2017 - )

Mobirise

David Hodges

BroadSem PhD Student
(2017 - )

Mobirise

Bailin Wang

BroadSem 
PhD Student
(2018 - )

Mobirise

Yanpeng Zhao

BroadSem
PhD student
(2018 - )

Mobirise

Diego Marcheggiani

Post-doc
Associated Researcher
(2016 - )

Mobirise

Caio Filippo Corro

Post-doc
Associated Researcher
(2017 - )

Mobirise

Michael Schlichtkrull

PhD student
Associate Researcher
(2017 - )

Mobirise

Serhii Havrylov

PhD student
Associated Researcher (2017 - )

Mobirise

Lena Voita

PhD student 
Associated Researcher
(2017 - )

Mobirise

Xinchi Chen


BroadSem postdoc
(2018 - )

Mobirise

Nicola De Cao

PhD student
Associated Researcher
(2018 - )

Mobirise

Joost Bastings

PhD student
Associated Researcher 
(2015 - )

Mobirise

Arthur Brazinskas

PhD student
Associated 
Researcher
(2018 - )

Mobirise









Mobirise













Mobirise










Our papers




Software

  1. SEMANTIC ROLE LABELER - our depependency-based semantic role labeler achieveing state-of-the-art results on CoNLL 2008, 2009 and 2012 benchmarks for Chinese and English.  Both syntax-agnostic and syntax-aware versions (relying on graph convolutional networks) are available. Described in Marcheggiani and Titov (EMNLP 2017) and Macheggiani et al (CoNLL 2017). [Code]
  2. AMR SEMANTIC PARSER  - abstract meaning representations (AMR) are broad-coverage sentence-level semantic representations. AMR encodes, among others, information about semantic relations, named entities, co-reference, negation and modality. Our semantic parser. Our AMR parser parser achieves the best reported result on the standard benchmark (74.4% on LDC2016E25).  Described in Lyu and Titov (ACL 2018). [Code]
  3. DIFFERENTIABLE PERTURB-AND-PARSE (DPP) - our framework for modeling latent dependency trees.  We introduced novel latent-variable generative model for semi-supervised dependency parsing. As exact inference is intractable, we used differentiable relaxation to obtain approximate samples and compute gradients with respect to the parser parameters. Our DPP relies on differentiable dynamic programming over stochastically perturbed edge scores. The code support semi-supervised learning, the implementation supporting task-specific structure induction can be provided by request. Described in Corro and Titov (ICLR 2019).  [Code]
  4. Relational GCN -  link prediction with relational graph convolutional neural networks (R-GCNs).  Described in Schlichtkrull et al. (ESWC 2018, best student paper).  [Code]
  5. GRAPH-2-TEXT - our framework for generating text from structured data relying on graph neural networks. Described in Macheggiani and Perez-Beltrachini (NLG-2018, best short paper). [Code]
  6. Syntax-aware NMT - neural machine translation system which relies on graph convolutional networks to inject information about predicted syntactic parses. Described in Bastings et al. (EMNLP 2017). [Code]
  7. Bayesian Context-Dependency Word Embedings (BSG) - We introduce a method for embedding words as probability densities in a low-dimensional space. Rather than assuming that a word embedding is fixed across the entire text collection, as in standard word embedding methods, in our Bayesian model we generate it from a word-specific prior density for each occurrence of a given word. Described in Bražinskas et al (COLING 2018).  [Code]