Talk:Semantic parsing

Wiki Education assignment: INFO 505 - Foundations of Information Science

This article was the subject of a Wiki Education Foundation-supported course assignment, between 22 August 2023 and 11 December 2023. Further details are available on the course page. Student editor(s): AyeshaBD (article contribs). Peer reviewers: SummerNightmare2023, Waveformleaf.

— Assignment last updated by SummerNightmare2023 (talk) 03:54, 26 October 2023 (UTC)

Thank you for the review. AyeshaBD (talk) 03:04, 7 November 2023 (UTC)

Asking for a complete rewrite of section "Types of semantic parsing"

I'd like to suggest to more strongly disentangle: (a) the type of meaning representation, (b) type of models used and (c) properties of the two. Here are two examples to show you what I mean with that:

(1) "compositional": As someone who has published about semantic parsing and compositional generalization, the current paragraph under 'deep semantic parsing' irked me:

Deep semantic parsing, also known as compositional semantic parsing, is concerned with producing precise meaning representations of utterances that can contain significant compositionality. Shallow semantic parsers can parse utterances like "show me flights from Boston to Dallas" by classifying the intent as "list flights", and filling slots "source" and "destination" with "Boston" and "Dallas", respectively. However, shallow semantic parsing cannot parse arbitrary compositional utterances, like "show me flights from Boston to anywhere that has flights to Juneau". Deep semantic parsing attempts to parse such utterances, typically by converting them to a formal meaning representation language. Nowadays, compositional semantic parsing are using Large Language Models to solve artificial compositional generalization tasks such as SCAN.

While this places compositional semantic parsing as a synonym for deep semantic parsing, I've never understood the two terms as synonyms so far and was surprised to read this. I'd appreciate a source that equates both terms. The question is: what makes semantic parsing compositional? Is it the parser that has to follow a compositional process (as used by the AM parser presented Groschwitz et al. 2018 for instance)? Is it that the parser has proved to generalize compositionally (more on this below)? The cited texts suggest to me it is rather the meaning representation that has to be compositional, although the example isn't really clear to me: is it the fault of the meaning representation that a sentence 'show me flights from Boston to anywhere that has flights to Juneau' cannot be parsed? That the meaning representation is just not powerful enough to do it? That it lacks lexical or structural elements to express everything that we find in natural language? Here I'd guess it is about quantifiers like 'anywhere' which pose problems for first-order logic? But that wouldn't render anything 'not (fully) compositional', it is just not powerful enough to derive a compositional meaning. The starting sentence "is concerned with producing precise meaning representations of utterances that can contain significant compositionality" also doesn't make sense to me: what does it mean for a MR of an utterance to "contain significant compositionality"? The meaning of an utterance can be compositionally derived or maybe not (and represented as a MR), but how can MRs or MR-utterance pairs contain compositionality?

Conflating compositional semantic parsing with deep semantic parsing is even more weird, when it comes to the last sentence, where compositional generalization is mentioned. In a compositional generalization task, it is tested whether a model can generalize in accordance with the principle of compositionality (i.e. parse new unseen utterances correctly because both the words in it and the structure how they were combined were both seen during training, just not in this combination). If a model fails to generalize compositionally, it is -by design of the task- not the fault of the meaning representation or the sentence in question.

(2) It continues with the Neural Semantic Parsing section. The juxtaposition with sections on shallow and deep suggests that it is a third type distinct from the other two. But this is already confusing since neural approaches are mentioned in the two preceeding sections (RNNs in shallow, LLMs in deep semantic parsing), and I don't get how parsing e.g. into AMR (mentioned in this third section) cannot be considered deep semantic parsing. (btw when neural models weren't the standard yet 'deep' in front of something sometimes also just meant "I'm using deep learning here").

Here is my suggestion on a new structure: I'd suggest to only keep the distinction between shallow and deep parsing to distinguish simple slot filling/semantic role labeling from parsing into more complex structures. These structures can be logical forms of some sort, graphs or query languages like SQL or boxes like in Discourse Representation Theory (DRT) or ...

A separate topic is how this 'semantic output', be it semantic roles (shallow) or logical forms (deep parsing) can be achieved. No one stops you from training and using a neural network to get this output, or using a non-neural approach or a neuro-symbolic one. Depending on how you represent the output - as a sequence, as a graph or something else, you can use different kinds of neural network architectures. For instance, AMR can be represented both as a sequence and therefore tackled with a seq2seq approach (e.g. by Konstas et al 2017 or Bevilacqua et al. 2021) or as a graph and then addressed by a seq2graph approach (e.g. Ivan & Titov 2018). The examples given in the current text seem to still focus a lot on RNNs, especially looking at sources cited, while in recent years Transformers have superseeded RNNs, so I guess this also needs updating. By the way, subtle ways in how to represent the meaning representation can have an effect on how easy or hard it is for a system to learn (e.g. how you linearize the AMR).

Last but not least, some sentences in this section seem off, but as I am not a native speaker I might be wrong when it comes to grammar:

• "This parsing follow an unsupervised learning techniques." at the end of the shallow parsing section
• "Nowadays, compositional semantic parsing are using Large Language Models to solve artificial compositional generalization tasks [...]" at the end of the deep parsing section
• "A well-established field of study, semantic parsing finds use in voice assistants, question answering, [...]" in the neural semantic parsing section
• "Since Neural approaches have been available for two years," - why hard-code the "two years" and why capitalize "Neural" there but not four sentences later?
• "We'll give a summary of contemporary neural approaches to semantic parsing and discuss how they've affected the field's understanding of semantic parsing." - that's not very Wikipedia-style, is it?

If I find time I might rewrite the section myself, but so far I haven't done major edits, so I'm a bit hesitant here. PlusMinuscule (talk) 05:10, 20 June 2025 (UTC)