Planning, Inference and Pragmatics in Sequential Language Games

Khani, Fereshte; Goodman, Noah D.; Liang, Percy

doi:10.1162/tacl_a_00037

Cited by 20 publications

(23 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In other words, a model that could generate a sequence of questions and closed “yes/no” answers, which progressively reduce uncertainty about the subject of conversation (i.e., contextual knowledge). These sorts of sequential communication games have extensively been tackled in the literature: including one round of question-answer ‘whisky pricing’ interaction ( Hawkins et al, 2015 ), playing restricted ‘cards corpus’ with one-off communication ( Potts, 2012 ), sequential ‘info-jigsaw’ game ( Khani et al, 2018 ), ‘hat game’ where agents learn to communicate via observing actions ( Foerster et al, 2016 ) and conversations about visual stimulus ( Das et al, 2017 ). Having specified the generative model for our “Twenty Questions” paradigm, we made no further assumptions—we used off-the-shelf (marginal) message passing to simulate neuronal processing ( Dauwels, 2007 ; Friston et al, 2017c ; Parr and Friston, 2018 ; Winn and Bishop, 2005 ).…”

Section: Generative Models Of Languagementioning

confidence: 99%

Generative models, linguistic communication and active inference

Friston

Parr

Yufik³

et al. 2020

Neuroscience & Biobehavioral Reviews

View full text Add to dashboard Cite

show abstract

Section: Generative Models Of Languagementioning

confidence: 99%

Generative models, linguistic communication and active inference

Friston

Parr

Yufik³

et al. 2020

Neuroscience & Biobehavioral Reviews

View full text Add to dashboard Cite

show abstract

“…Our proposed extension maintaining a (belief) state could be used to define a Markov Decision Process where different question or answer selections lead to anticipated transitions in beliefs about goals or the world, respectively. A related approach was recently taken by Khani, Goodman, and Liang (2018), who used a forward-looking recursive utility to define an action policy in a sequential language game. In this task, pragmatic agents took turns saying one-or two-word utterances, informing their partner about the world state, but were not permitted to ask questions.…”

Section: Discussionmentioning

confidence: 99%

Why do you ask? The informational dynamics of questions and answers

Hawkins¹,

Goodman²

2017

Preprint

Self Cite

View full text Add to dashboard Cite

Asking questions is one of our most efficient and reliable means of learning about the world. Yet we do not often pose these questions to an impartial oracle; we ask cooperative social partners, in dialogue. In this paper, we aim to reconcile formal models of optimal question asking and answering with classic effects of social context. We begin from the observation that question-answer dialogue is motivated by a two-sided asymmetry in beliefs: questioners have a private goal but lack goal-relevant information about the world, and answerers have private information but lack knowledge about the questioner's goal. We formalize this problem in a computational framework and derive pragmatic questioner and answerer behavior from recursive social reasoning. Critically, we predict that pragmatic answerers go beyond the literal meaning of the question to be informative with respect to inferred goals, and that pragmatic questioners may therefore select questions to more unambiguously signal their goals. We evaluate our pragmatic models against asocial models in two ways. First, we present computational simulations accounting for three classic answerer effects in psycholinguistics. We then introduce the Hidden Goal paradigm for experimentally eliciting questioner and answerer behavior in scenarios where there is uncertainty about the questioner's goal. We report data from three experiments in this paradigm and show how our core computational framework can be composed with more sophisticated question semantics, hierarchical goal spaces, and a persistent state over which extended dialogue can unfold. We find that social inference is needed to account for critical aspects of the data.

show abstract

“…The rational speech act (RSA) model raised by Golland et al [24] and developed in [22,54,25,2] accommodates the idea of using not only the utterance but also the selection of the utterance to understand the speaker. Previous works in these areas are mainly from human action interpretation [21,60,36], language emergence [69,35], linguistics [33,2,13] and cognitive science [25,8] perspectives. To our knowledge, our work is the first to relate pedagogy and recursive reasoning to generic parameter learning and shows a provable improvement.…”

Section: Related Workmentioning

confidence: 99%

Iterative Teacher-Aware Learning

Yuan¹,

Zhou²,

Shen³

et al. 2021

Preprint

View full text Add to dashboard Cite

In human pedagogy, teachers and students can interact adaptively to maximize communication efficiency. The teacher adjusts her teaching method for different students, and the student, after getting familiar with the teacher's instruction mechanism, can infer the teacher's intention to learn faster. Recently, the benefits of integrating this cooperative pedagogy into machine concept learning in discrete spaces have been proved by multiple works. However, how cooperative pedagogy can facilitate machine parameter learning hasn't been thoroughly studied. In this paper, we propose a gradient optimization based teacher-aware learner who can incorporate teacher's cooperative intention into the likelihood function and learn provably faster compared with the naive learning algorithms used in previous machine teaching works. We give theoretical proof that the iterative teacher-aware learning (ITAL) process leads to local and global improvements. We then validate our algorithms with extensive experiments on various tasks including regression, classification, and inverse reinforcement learning using synthetic and real data. We also show the advantage of modeling teacher-awareness when agents are learning from human teachers.35th Conference on Neural Information Processing Systems (NeurIPS 2021).

show abstract

Planning, Inference and Pragmatics in Sequential Language Games

Cited by 20 publications

References 21 publications

Generative models, linguistic communication and active inference

Generative models, linguistic communication and active inference

Why do you ask? The informational dynamics of questions and answers

Iterative Teacher-Aware Learning

Contact Info

Product

Resources

About