Joint parsing of syntactic and semantic structures

TODO: good performance, transition-based system: Swayamdipta et al. (2016), multi-task model: Shi et al. (2016)

TODO: Li et al. (2010): "it shows that incorporating semantic role-related information into the syntactic parsing model significantly improves the performance of both syntactic parsing and semantic parsing."

TODO: Morante et al. (2009)

TODO: selectional constraint for dependency parsing (Mirroshandel and Nasr, 2016)

Synchronization

Ignores semantic arcs not corresponding to single syntactic arcs (Thompson, Levy, and Manning 2003; Titov and Klementiev 2011)

Resorts to pre/post-processing strategies which modify semantic or syntactic structures (Lluís and Màrquez 2008; Lang and Lapata 2011; Titov and Klementiev 2012)

Li, Zhou, and Ng (2010) explore different levels of coupling of syntax and semantics, and find that only explicit interleaving or explicit feature selection yield improvements in performance. Instead of synchronising individual steps,

Henderson et al. (2013): (1) decompose both the syntactic derivation and the semantic derivation into subsequences, where each subsequence corresponds to a single word in the sentence, and then (2) synchronise syntactic and semantic subsequences corresponding to the same word with each other.

Evaluation

CoNLL 2008

Henderson et al. (2008), who implemented a generative history-based model (Incremental Sigmoid Belief Networks with vectors of latent variables) where syntactic and semantic structures are separately generated but using a synchronized derivation (sequence of actions);

Samuelsson et al. (2008), who, within an ensemble-based architecture, implemented a joint syntactic-semantic model using MaltParser with labels enriched with semantic information;

Lluıs and Marquez, who used a modified version of the Eisner algorithm to jointly predict syntactic and semantic dependencies; and finally,

Sun et al. (2008), who integrated dependency label classification and argument identification using a maximum-entropy Markov model.

Additionally, Johansson and Nugues (2008), who had the highest ranked system in the closed challenge, integrate syntactic and semantic analysis in a final reranking step, which maximizes the joint syntactic-semantic score in the top k solutions.

In the same spirit, Chen et al. (2008) search in the top k solutions for the one that maximizes a global measure, in this case the joint probability of the complete problem.

CoNLL 2009

Most of the observations from the 2008 shared task still hold; namely, the best systems overall do not use joint learning or optimization (the best such system was placed third in the Joint task, and there were only four systems where the learning methodology can be considered “joint”). Therefore, most of the observations and conclusions from 2008 shared task hold as well for the current results

This article is in an early stage. Action is needed to make it more useful.