Thread by @hadyelsahar

#rep4nlp Invited talk 2: @mohitban47 "Adversaially robust Representation Learning"

#acl2019nlp

@mohitban47 Adv. examples can break Reading comprehension systems.

"Adversarial Examples for Evaluating Reading Comprehension Systems" Jia and liang 2017
arxiv.org/abs/1707.07328

To fix that: "AddSentDiverse" a mod of AddSent (Jia et al. 2017), aimed at producing adversarial examples for robust training purposes based on rule-based semantic rules

Robust Machine Comprehension Models via Adversarial Training (NAACL2018 short)
arxiv.org/pdf/1804.06473…

Reasoning shortcuts: are non-preferred model behaviour that tries to avoid compositional reasoning

i.e. reducing Multi-hop into Single hop QA

Again solution: adv examples using what it seems to be rule-based semantic distractors.

(Jiang et al. 2019)
arxiv.org/pdf/1906.07132…

Adversarial dialogue: User-Error Robustness
(Niu and Bansal CoNLL 2018)

The responses of dialogue models change with small perturbations in the input words.

Solutions: Rule based generation of adv. examples and paraphrasing.

arxiv.org/abs/1809.02079

NLI models suffer from lower compositionality awareness.

This paper is looking into if the high performance of models is attributed to focusing only on lexical information or also compositionality.

"Analysing compositionality sensitivity"
arxiv.org/abs/1811.07033

Some rule-based semantically motivated adversaries to test a set of SOTA NLI models "SOSwap".

Lexically misleading score (LMS) in brief: the more lexically misleading an example is, the more compositional information is required to solve it.

LMS to select examples from existing evaluation to create stronger evaluation dataset.

#rep4nlp #ACL2019nlp

There are some other additional tasks that I missed in the end.

Summary of key Mohit's talk in #rep4nlp #ACL2019nlp points here 👆👆👆👆

@threadreaderapp unroll