Arguing for logical realism and discussing the logical structure and constitution of the world.
Non omnes formulae significant quantitatem, et infiniti modi calculandi excogitari possunt.
Please do not use material posted here without referencing and giving credit to this blog. Thank you.
Friday, September 22, 2023
On the most general concept of a formal logic
Text and meaning
Texts are strings $s \in T^*$ of tokens $T$. But there are different possible choices of tokens corresponding to different choices of 'scale'. Thus a set of tokens $T_2$ refines a set $T_1$ if we have a decomposition map \[ t: T_1 \rightarrow T_2^+\] which induces a map\[ t^* : T_1^* \rightarrow T_2^*\] For each set of tokens $T$ we denote by $S_T \subset T^*$ the corresponding set of texts. Now $S_T$ comes with natural partially defined operations: \[ \otimes_i : S_T \times S_T \rightarrow S_T \]
\[ \oplus_i : S_T \rightarrow S_T \times S_T\] for $i \in O$ which we call the anabolic and katabolic operations respectively. These include for instance the basic operation of combining sentences or breaking a sequence of two sentences into the corresponding sentences. If we consider a decomposition of the tokens then more phonetic or morphological type operations may become available. Now consider a text $s \in S_T$. Consider a multi-sorted first-order language with equality including Peano Arithmetic with a further predicate describing the enumeration of elements of $T^*$. Se use the standard notation $s(n) = t$ for a text $s$ and token $t$.
Then for a given text $s$ we can consider the set $ \mathbb{T}(s)$ of all sentences that hold in $s$ considered as a first-order model. This is done by considering as the underlying set $|s|$ of the model the tokens occurring in $s$ as well as the standard model for the natural numbers. Likewise we can consider $\mathbb{T}(S_T)$ of all sentences holding in all the texts $s \in S_T$. We can also consider formulas with a fixed number of free variables (let us say all of sort $T$) $\mathbb{L}^n$ and consider the subset $\mathbb{V}^n(s)$ elements of $|s|^ n$ that satisfy this formula. Since the part of the model corresponding to this sort is finite we can also assign probability distributions to each $\phi(t_1,...,t_n)$ in $\mathbb{L}^n$. Bigram models are a special case of this. Likewise we can consider formulas and consider satisfaction in a whole set of texts. Thus the 'information' contained in a set of texts $S_T$ can be described as an assignment of probability distributions on $|s| ^n$ to formulas of $\mathbb{L}^n$. We propose that these ideas be used for a new version of transformers.But what about meaning ? Meaning depends on the sum-total of interlocutor's past conscious experience (since birth). The temporal unit of the reception of meaning is conditioned by an atomic interval of conscious experience or attention which in general is capable of encompassing numerous tokens. There is also a theory that there is a space of meanings $M$ the elements of which can be constructed via several operations from a set of primitive meanings. Also that any meaning can be decomposed in a canonical way into such operations and primitives. The problem is describing the correspondence between texts and meanings and in particular the semantic counterpart of the operations $\otimes_i$ and $\oplus_i$. This correspondence is in general given by a pair $(\mathcal{M}, \mathcal{T})$\[ \mathcal{M} : S_T \longrightarrow \mathbf{P} M\]
\[ \mathcal{T} : M \longrightarrow \mathbf{P} S_T\]A given text can have more than one interpretation/meaning. A given meaning can be expressed in a text in more than one way. But the above remark suggests that the meaning(s) of a text must depend also on an additional parameter space $\Psi$ of states of consciousness, for instance\[ \mathcal{M}' : S_T \times \Psi \longrightarrow \mathbf{P} M\]The dogma of compositionality states that the meaning of a $s \otimes_i s'$ can somehow be canonically constructed from the meanings of $s$ and $s_i$.
The dogma of sequentialism states takes compositionality only at the level of elementary tokens considered as meaning-bearers. The meaning of each token depends both on itself and its entire previous context in the text (and indeed alternatively in a similar situation in all texts in $S_T$).\[ \mathcal{M} (s' \star s) = F (s', s)\]Self-attention can be adapted to characterize how $F$ operates on the context. In a previous note we proposed that we think of meaning-units as provided with transmitters and receptors (think of biochemistry or the $\pi$-calculus). So each token in the context can send out signals to subsequent tokens like 'I am a verb looking for an object' or 'I am a direct object, where is my verb ?'.But a better way to think of this is that the 'meaning' of each token will in general depend on whole set of tokens both previous and subsequent. We call this set the relevance context. This gives rise to great computational complexity as a text will give rise to a graph. If we consider the problem of polysemy and idiomatic expressions then it is clear that the enveloping context (both back and forward) is crucial to determine meaning. Perhaps the probability distributions associated to formulas mentioned above could be a candidate for a rough formal treatment of meanings ?
Quodlibet
1. René Thom called quantum mechanics 'the greatest intellectual scandal of the 20th century'. Maybe this was too harsh, but quantu...
-
We cannot leave out time to begin with. Nor possibility, multiple times. A first attempt would be as follows. A system involves an input $I...
-
Take standard first-order semantics. Consider a model $M$ of a certain theory $T$. Now suppose that we wish to include a subset a $\Psi \su...
-
By 'theory' we mean a body of knowledge capable of a formal or semi-formal presentation. But here we focus on the kind of theory w...