SymbolicAI: A framework for logic-based approaches combining generative models and solvers

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

• Bridging Logic and Learning: A Neural-Symbolic Approach for Enhanced Reasoning in Neural Models (ASPER) (2023)
• Probabilistic Abduction for Visual Abstract Reasoning via Learning Rules in Vector-symbolic Architectures (2024)
• On the verification of Embeddings using Hybrid Markov Logic (2023)
• Transformer-Based Models Are Not Yet Perfect At Learning to Emulate Structural Recursion (2024)
• LangProp: A code optimization framework using Language Models applied to driving (2024)

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

This work confuses the notion of linguistic competence, claims that LLMs demonstrate linguistic competence and - after grounding itself in GPT embeddings - proceeds to ``anchor'' itself in formal language theory by discussing formal grammars and compositionality in the context of domain-specific grammars and functional composition of data streams from non-language modalities into computational graphs, which are also used to evaluate the system.

Their definition of a symbol: it's a vector!

A Symbol object marks a non-reducible atomic unit.
Furthermore, each Symbol object contains valued and vector-valued representations, obtained through value' and embedding' attributes.

They claim to adhere to a throw-it-in-the-bucket analogy of language that sounds like something Steve Jobs would have said, perhaps inspired by semiotics, and the following working definition of language:

We consider language as a central processing module, distinct from other cognitive processes such as reasoning or memory

Languages have these things called coordinating conjunctions that function as logical operators. You may have heard of them. Some examples in English are and' and or'.

Acquisition is a different process than learning. LLMs don't acquire a semantics. They don't even learn one. They cannot reliably handle core aspects of any rational theory of meaning. They crudely and inconsistently capture semantic properties of language that are frequently useful for information retrieval systems and are evaluated with information retrieval benchmarks. Sometimes new ones that the system was designed for, as is the case here.

The focus on LLMs as ``semantic parsers'' (that do their domain parsing after a crude semantic parse has already been obtained) and misplaced discussion of formal grammars and compositionality completely misses the point of any meaningful integration of symbols and numeric representations. To this end it would have been better to stick with one solver with symbols that mean something. And aren't numeric! Cool IR system, tho.

When you plant the flag then realize you're on the wrong hill.

We thank you for your critical assessment and address your points in the following way:

“This work confuses the notion of linguistic competence, claims that LLMs demonstrate linguistic competence and - after grounding itself in GPT embeddings - proceeds to”
…
“Their definition of a symbol: it's a vector!”
…
“To this end it would have been better to stick with one solver with symbols that mean something. And aren't numeric! Cool IR system, tho.”

1. The reader conflates methodology and implementation details.
2. We introduce the notion of symbols and expressions in our Design Principles section. Their respective realization is detailed in our Framework section, which offers auxiliary properties, such as helper methods, vector-valued data attributes, etc.
3. Please further see our section Design Principles:
“As posited by Newell & Simon, symbols are elemental carriers of meaning within a computational context”
…
“However, this language-centric model does not inherently encompass all forms of representation, such as sensory inputs and non-discrete elements, requiring the establishment of additional mappings to fully capture the breadth of the world. This limitation is manageable, since we care to engage in operations within this abstract conceptual space, and then define corresponding mappings back to the original problem space.“

For the realization of the above design principles, we then refer to our Framework section.
4. We do not ground our work in GPT embeddings. We actually assume that we do not have access to embeddings of the LLMs (see Section 6, Performance Measure, second paragraph after Equation 1), therefore we use a separate embedding model (all-mpnet-base-v2) to obtain vector-valued representations.
5. The embeddings are used to measure the similarities in our benchmark for the evaluation of our VERTEX score, and not required for defining symbols.

“They claim to adhere to a throw-it-in-the-bucket analogy of language that sounds like something Steve Jobs ...“
…
“We consider language as a central processing module, distinct from other cognitive processes such as reasoning or memory”

We do not understand the reader's critic and blending of our work with “Steve Jobs.” But we consider language as a central processing module based on the phrase prior to the quote introduced by the reader, which clearly expresses its meaning:
“In designing the architecture of SymbolicAI, we drew inspiration from the body of evidence that suggests the human brain possesses a selective language processing module (Macsweeney:02, Fedorenko:10, Menenti:11, Regev:13, Terri:16, Deniz:19, Hu:21), prior research on cognitive architectures (Newell:56, Newell:57, Newell:72, Newell:90, Laird:22), and the significance of language on the structure of semantic maps in the human brain (Huth:16).“

“Acquisition is a different process than learning. LLMs don't acquire a semantics. ...”

We specifically address in section Design Principles, Formal Languages, paragraph 2, how we consider the intersection between natural languages and formal languages.

The reader conflates methodology and implementation details.

How so? From the intro:

recent research also highlights the limitations of LLMs in functional linguistic competence despite their proficiency in formal linguistic competence

We introduce the notion of symbols and expressions in our Design Principles section.

This is where you describe language as ``the convex hull of the knowledge of our society,'' which to my ears sounds like something Steve Jobs would say.

Their respective realization is detailed in our Framework section, which offers auxiliary properties, such as helper methods, vector-valued data attributes, etc.

Whether the vectors constitute the primary value of this object or an auxiliary value isn't relevant to my point: your symbol contains numeric information, which in my view doesn't make it much of a symbol. Additionally, in my view, an object that contains multiple attributes doesn't make it much of a non-reducible atomic unit.

Please further see our section Design Principles:
“As posited by Newell & Simon, symbols are elemental carriers of meaning within a computational context”
…
“However, this language-centric model does not inherently encompass all forms of representation, such as sensory inputs and non-discrete elements, requiring the establishment of additional mappings to fully capture the breadth of the world. This limitation is manageable, since we care to engage in operations within this abstract conceptual space, and then define corresponding mappings back to the original problem space.“

For the realization of the above design principles, we then refer to our Framework section.

Yet you describe conveying sensory input as ``functional linguistic competence.''

We do not understand the reader's critic and blending of our work with “Steve Jobs.” But we consider language as a central processing module based on the phrase prior to the quote introduced by the reader, which clearly expresses its meaning:
“In designing the architecture of SymbolicAI, we drew inspiration from the body of evidence that suggests the human brain possesses a selective language processing module (Macsweeney:02, Fedorenko:10, Menenti:11, Regev:13, Terri:16, Deniz:19, Hu:21), prior research on cognitive architectures (Newell:56, Newell:57, Newell:72, Newell:90, Laird:22), and the significance of language on the structure of semantic maps in the human brain (Huth:16).“

Re: the cognitive linguistics line of thinking in Huth et al, voxel modeling a semantic system'' into semantic maps'' or a ``semantic atlas'' in the brain is not going to help you obtain a semantics of natural language. Especially when you start with representations from computer vision. If anyone would agree that language is a more fundamental cognitive process than language it's the cognitive linguists. How could anyone suggest that reasoning is a cognitive process distinct from language and keep a straight face?

We do not ground our work in GPT embeddings. We actually assume that we do not have access to embeddings of the LLMs (see Section 6, Performance Measure, second paragraph after Equation 1), therefore we use a separate embedding model (all-mpnet-base-v2) to obtain vector-valued representations.
The embeddings are used to measure the similarities in our benchmark for the evaluation of our VERTEX score, and not required for defining symbols.

Thanks for the correction. When I saw

the initial development of SymbolicAI started with the GPT family of models

I incorrectly assumed the GPT-2 token-based BPE embeddings as was used in Mamba. I see now that you use sentence-level embeddings, not token-level embeddings. In any case, a vector is a vector. It has directionality and can only have some interpretation placed on it when, in some space of arbitrary dimensionality that also carries no meaning, it is near some other thing that also has directionality. It's an odd, but convenient, representation of language. In my view working at the sentence level takes you even farther away from where an integration of symbolic and numeric information should begin: a non-reducible atomic term, e.g. the English plural morpheme (/ɪz/, /s/ or /z/). Working with vectors means multi-modal integrations are what's important, not a representation or a semantics that's suitable for language, as evidenced by the closing section of the evaluation and the VERTEX score.

We specifically address in section Design Principles, Formal Languages, paragraph 2, how we consider the intersection between natural languages and formal languages.

My comment about acquisition being a different process than learning was targeting at the author of the LinkedIn post where I first encountered this work. They claimed that "neural nets acquired knowledge, language and reasoning". I was triggered and forced to respond to your work.

how we consider the intersection between natural languages and formal languages

My biggest issue is the designation of a multi-attribute object as a symbol. My second biggest issue is that by glossing over the foundational representation and focusing the discussion of formal grammar and compositionality in the context of domain-adaptation and multi-modal integrations, respectively, you fail to address this intersection in any meaningful way. I'm also bothered by the claim that LLMs demonstrate (formal) linguistic competence, the semantic parsing as domain parsing angle, the evaluation metric that focuses on multi-modal integrations, and the general land-grabbiness throughout.

Take your coverage of type theory and formal grammars to the part of the system that matters.