According to Wing 2011, #abstraction in natural language is the act of referring to a set of elements via their shared properties, discarding irrelevant distinctions.
It can be a shape they make ("object"), order of traversal ("control structure") or a condition they all meet>>
It can be a shape they make ("object"), order of traversal ("control structure") or a condition they all meet>>
From a human communication perspective, speakers make ample use of abstractions expressed in NL, in order to convey complex non-trivial instructions efficiently.
Crucially, the use of such abstract structures in NL is done informally, intuitively, and w/o any formal training >>
Crucially, the use of such abstract structures in NL is done informally, intuitively, and w/o any formal training >>
We set out to systematically address the ability of #NLProc models to detect, ground and execute non-trivial NL instructions that contain multiple levels of #abstraction.
Our data are spontaneous narrative by novices, non-CS-trained humans, elicited via an online game 🌸 >>
Our data are spontaneous narrative by novices, non-CS-trained humans, elicited via an online game 🌸 >>
The Hexagons is a (fun!) online game where one can draw images on a hexagons board:
nlp.biu.ac.il/~royi/hexagon-…
Do try it at home!
Based on this game, we devised our data collection protocol as follows >>
nlp.biu.ac.il/~royi/hexagon-…
Do try it at home!
Based on this game, we devised our data collection protocol as follows >>
The key idea is that the app contains a single predicate: *paint(position,color)* but one can create increasingly complex images.
The task is simply to describe the image to a remote friend with an empty board:
nlp.biu.ac.il/~royi/hexagon-…
Do try that too!
>>
The task is simply to describe the image to a remote friend with an empty board:
nlp.biu.ac.il/~royi/hexagon-…
Do try that too!
>>
To describe such images, humans refrain from simply repeating the predicate (cf. #Grice maxims), and they naturally resort to abstractions.
We collect their utterances, verify their executability with a (disjoint) set of crowdworkers, and let models predict the execution.
>>
We collect their utterances, verify their executability with a (disjoint) set of crowdworkers, and let models predict the execution.
>>
Our experiments show that while contextualize models are great at uncovering *concrete* pred-arg structures in utterances, they fall short of processing & grounding such #abstraction leaps in NL -- making it an understudies yet fascinating regime for grounded semantic parsing >>
We offer our
🌸 game
nlp.biu.ac.il/~royi/hexagon-…
🌸 protocol
nlp.biu.ac.il/~royi/hexagon-…
🌸 methodology & results
arxiv.org/abs/2106.14321
For the #NLProc community as a foundation to probe further into the expression & interpretation of #abstraction in #NL
[At @Nlp4Prog-#ACL]
🌸 game
nlp.biu.ac.il/~royi/hexagon-…
🌸 protocol
nlp.biu.ac.il/~royi/hexagon-…
🌸 methodology & results
arxiv.org/abs/2106.14321
For the #NLProc community as a foundation to probe further into the expression & interpretation of #abstraction in #NL
[At @Nlp4Prog-#ACL]
All the kudos in the world to my co-authors: Royi Lachmy and @valentina__py, for making all that happen!
<EOT>
<EOT>
• • •
Missing some Tweet in this thread? You can try to
force a refresh
