unicode-fol-kit 0.2.0

Parser and toolkit for first-order logic formulas using Unicode operators

unicode-fol-kit

A Python toolkit for parsing and working with first-order logic (FOL) formulas written with Unicode operators. The single parser class MSFLParser supports three modes — classical FOL, many-sorted FOL (MSFOL), and many-sorted fuzzy logic (MSFL, Łukasiewicz) — selected by constructor flags.

Features

• Three parser modes — FOL, many-sorted FOL (MSFOL), and many-sorted fuzzy/Łukasiewicz logic (MSFL), all from one class
• Unicode surface syntax — natural symbols (∀ ∃ ∧ ∨ ¬ → ↔ ⊕ ⊗ = ≠ ≤ ≥) with no ASCII fallbacks needed
• Sorted quantifiers and constants — ∀x:Human P(x), P(alice:Human) in MSFOL and MSFL modes
• Łukasiewicz operators — weak ∧ / ∨ (min/max), strong ⊗ / ⊕ (t-norm/t-conorm), and Łukasiewicz ¬ → ↔ in MSFL mode
• Full AST — all standard FOL constructs plus MSFL-specific nodes, all as Python dataclasses
• Reductions — to_msfol() lowers Łukasiewicz operators to classical nodes; to_fol() further eliminates sorts via relativisation
• Serialisation — convert formulas to/from JSON dictionaries; round-trip safe
• Tree view — render any formula as a readable ASCII tree
• Z3 export — translate formulas to Z3 expressions for SMT solving
• Prover9 export — translate formulas to Prover9 syntax for automated theorem proving
• TPTP export — translate formulas to TPTP syntax
• Equivalence checking — check if two formulas are logically equivalent via Z3
• Entailment checking — check if a conclusion follows from premises via Prover9
• Lambda abstraction — λx. φ syntax in all three parser modes; parameters can be variables (λx.), named constants (λfoo.), or predicate symbols (λP.); body extends rightward through all connectives
• Higher-order predicate application — (func)(arg) explicit application; λP. P(x) writes the body naturally and is automatically scope-resolved to Application(LambdaVar("P"), Variable("x"))
• Lambda-calculus operations — free-variable computation, capture-avoiding substitution, beta-reduction (normal-order, step-limited), eta-reduction, combined beta-eta normalisation to fixpoint, and lexical scope resolution

Installation

Via pip

pip install unicode-fol-kit

Via git clone

git clone https://github.com/felixvossel/unicode-fol-kit.git
cd unicode-fol-kit
pip install .

Parser modes

MSFLParser is instantiated with two boolean flags:

MSFLParser(many_sorted=False, fuzzy=False)   # FOL   (default)
MSFLParser(many_sorted=True,  fuzzy=False)   # MSFOL
MSFLParser(many_sorted=True,  fuzzy=True)    # MSFL
MSFLParser(many_sorted=False, fuzzy=True)    # → raises ValueError

many_sorted	fuzzy	Mode	Quantifiers	Constants	Connectives
False	False	FOL	unsorted ∀x	unsorted	classical ∧ ∨ ⊕ ¬ → ↔
True	False	MSFOL	sorted ∀x:Sort	sorted alice:Sort	classical ∧ ∨ ¬ → ↔ (no ⊕)
True	True	MSFL	sorted ∀x:Sort	sorted alice:Sort	weak ∧ ∨, strong ⊗ ⊕, Łuk ¬ → ↔

Usage

FOL mode (default)

from unicode_fol_kit import MSFLParser

parser = MSFLParser()
formula = parser.parse("∀x (Human(x) → Mortal(x))")

MSFOL mode

Quantifiers and ground constants must carry a sort annotation. The colon can be written with or without a space before it:

parser = MSFLParser(many_sorted=True)

# Sorted quantifier
q = parser.parse("∀x:Human (Mortal(x) ∧ ¬Immortal(x))")
# SortedQuantifier(type='∀', variable=Variable('x'), sort='Human', formula=…)

# Sorted constant — both spacing forms are accepted
parser.parse("P(alice:Human)")
parser.parse("P(alice :Human)")

MSFL mode

Uses Łukasiewicz logic: ∧/∨ become weak (min/max), ⊗/⊕ become strong (t-norm/t-conorm), and ¬/→/↔ map to their Łukasiewicz counterparts:

parser = MSFLParser(many_sorted=True, fuzzy=True)

parser.parse("P(x) ∧ Q(x)")    # WeakConjunction (min)
parser.parse("P(x) ⊗ Q(x)")    # StrongConjunction (t-norm: max{0, x+y−1})
parser.parse("P(x) ⊕ Q(x)")    # StrongDisjunction (t-conorm: min{1, x+y})
parser.parse("¬P(x)")           # LukNegation (1−x)
parser.parse("P(x) → Q(x)")    # LukImplication (min{1, 1−x+y})
parser.parse("∀x:Human P(x)")  # SortedQuantifier

ASCII tree view

formula = MSFLParser().parse("∀x (Human(x) → Mortal(x))")
print(formula.tree_str())
# ∀ x
# └── →
#     ├── Atom: Human
#     │   └── Variable: x
#     └── Atom: Mortal
#         └── Variable: x

Exporting to other formats

formula.to_prover9()   # '(all x (Human(x) -> Mortal(x)))'
formula.to_tptp()      # '(![X]: (human(X) => mortal(X)))'
formula.to_dict()      # JSON-serialisable dict

Serialisation

from unicode_fol_kit import Node

d = formula.to_dict()
formula2 = Node.from_dict(d)  # round-trip

Lambda-calculus

All three parser modes support lambda abstraction and application. parse() automatically applies scope resolution, so the returned AST is always fully resolved.

from unicode_fol_kit import (
    MSFLParser,
    LambdaVar, Lambda, Application,
    free_variables, substitute,
    beta_reduce, eta_reduce, beta_eta_normalize,
    resolve_lambda_scope,
    ReductionLimitError,
)

parser = MSFLParser()

# Parse — scope resolution is applied automatically
term = parser.parse("λP. λx. P(x)")
# Lambda(LambdaVar("P"), Lambda(LambdaVar("x"), Application(LambdaVar("P"), LambdaVar("x"))))

# Application
app = parser.parse("(λP. P(x))(Q)")
# Application(Lambda(LambdaVar("P"), Application(LambdaVar("P"), Variable("x"))),
#             Atom("Q", []))

Free variables

term = parser.parse("λP. P(x)")
free_variables(term)
# {Variable("x")}  — x is free; P is lambda-bound and does not appear

The result is a mixed set that may contain both Variable (logical) and LambdaVar (lambda-bound) objects.

Beta-reduction

beta_reduce reduces to beta-normal form using a normal-order (leftmost-outermost) strategy with full capture-avoiding substitution. It raises ReductionLimitError after 10 000 steps if the term does not normalise.

# (λP. λx. P(x))(Q) → λx. Application(Atom("Q",[]), LambdaVar("x"))
result = beta_reduce(parser.parse("(λP. λx. P(x))(Q)"))

# Full pipeline: parse → resolve → reduce
reduced = beta_reduce(parser.parse("(λP. P(x))(λy. Q(y))"))
# Atom("Q", [Variable("x")])

Eta-reduction

eta_reduce performs a single bottom-up pass contracting all eta-redexes: λp. f(p) → f when p is not free in f. Quantifiers are recursed into but never treated as eta-redexes.

from unicode_fol_kit import LambdaVar, Lambda, Application, Atom, Variable

f = Atom("P", [Variable("x")])                  # some formula
term = Lambda(LambdaVar("p"), Application(f, LambdaVar("p")))  # λp. f(p)
eta_reduce(term)   # → f  (the Atom node, not the Lambda)

Beta-eta normalisation

beta_eta_normalize alternates beta_reduce and eta_reduce to fixpoint (up to 100 rounds). The alternation loop is a genuine necessity: eta-reduction can expose fresh beta-redexes, requiring another beta pass.

normal = beta_eta_normalize(parser.parse("(λP. P(x))(Q)"))

ReductionLimitError is raised if the inner beta-reduction limit or the outer round limit is exceeded.

Scope resolution (standalone)

resolve_lambda_scope is also available as a standalone function for hand-built ASTs:

from unicode_fol_kit import resolve_lambda_scope, Lambda, LambdaVar, Atom, Variable

raw = Lambda(LambdaVar("x"), Atom("P", [Variable("x")]))
resolved = resolve_lambda_scope(raw)
# Lambda(LambdaVar("x"), Atom("P", [LambdaVar("x")]))

Reducing MSFL formulas to classical FOL

to_fol() performs a two-phase reduction: it first lowers Łukasiewicz operators to classical ones (to_msfol()), then eliminates sort annotations via relativisation (_relativize()):

from unicode_fol_kit import MSFLParser, to_fol

parser = MSFLParser(many_sorted=True, fuzzy=True)
formula = parser.parse("∀x:Human (P(x) ∧ ¬Q(x))")

classical = to_fol(formula)
# Quantifier(∀, x, Implies(Atom(Human, [x]), And(Atom(P,[x]), Not(Atom(Q,[x])))))

# Optionally, conjoin sort-membership facts for constants at the top level:
classical_with_facts = to_fol(formula, include_sort_facts=True)

Equivalence checking (Z3)

from unicode_fol_kit import MSFLParser, formulas_are_equivalent

parser = MSFLParser()
f1 = parser.parse("¬(P(x) ∧ Q(x))")
f2 = parser.parse("¬P(x) ∨ ¬Q(x)")

formulas_are_equivalent(f1, f2)  # True

Entailment checking (Prover9)

from unicode_fol_kit import MSFLParser, check_logical_entailment

parser = MSFLParser()
premises = [
    parser.parse("∀x (Human(x) → Mortal(x))"),
    parser.parse("Human(socrates)"),
]
conclusion = parser.parse("Mortal(socrates)")

check_logical_entailment(premises, conclusion, prover9_path="/usr/bin/prover9")  # True

Syntax reference

This section describes the full surface syntax accepted by the parser. Because the three modes share the same term and atom layer, most of the syntax is identical across modes; differences are called out explicitly.

Tokens

The lexer distinguishes the following token kinds. Because the patterns are mutually exclusive, a given identifier is unambiguously a variable, a constant, a function/predicate name, a number, or a sort annotation.

Token	Pattern	Examples	Meaning
Variable	one lowercase letter, optional trailing digits	x, y, x1, z42	a (possibly bound) logical variable
Name	lowercase, at least two letters, may contain digits and uppercase after the first letter	socrates, distance, centerOf, foo1	a bare constant or a function symbol
Constant (c_)	c_ followed by letters/digits	c_a, c_zero, c_42	an explicitly marked constant
Predicate	one uppercase letter, then letters/digits	P, Human, OnSurfaceOf	a predicate symbol
Number	digits, optional decimal part	0, 42, 3.14	a numeric literal
Sort annotation	: followed by an uppercase letter and letters/digits	:Human, :Sort1	a sort tag (MSFOL and MSFL modes only)

The c_ form exists so that single-letter constants can be written without colliding with variables. A bare a is always a variable; if you need the constant a, write c_a.

A function or predicate is recognised by being immediately followed by a parenthesised argument list, e.g. distance(x, y) or Human(socrates). The same token class (Name) serves both as a bare constant and, when applied, as a function symbol.

The sort annotation token always begins with :, which makes it lexically disjoint from all other tokens. Whitespace before the colon is optional: ∀x:Human P(x) and ∀x :Human P(x) are both valid and produce identical parse trees.

Terms

A term is one of:

• a variable (x, x1)
• a constant (socrates, c_a) or number (42, 3.14)
• in MSFOL / MSFL modes: a sort-annotated constant (alice:Human, c_a:Sort1)
• a function application (f(t1, …, tn), e.g. centerOf(x))
• an arithmetic combination of terms using +, -, *, /
• a parenthesised term ((t))

Arithmetic follows the usual precedence: * and / bind tighter than + and -, and both groups are left-associative. For example x + y * z parses as x + (y * z).

Sort rules in MSFOL / MSFL modes: variables are sorted implicitly by the quantifier that binds them; ground constants must carry an explicit sort annotation. An unsorted constant (e.g. bare alice) is a syntax error in sorted modes.

Atomic formulas

An atomic formula is either:

• a predicate applied to terms: P, Human(socrates), OnSurfaceOf(y, x) (a predicate may be nullary, i.e. used without arguments)
• an infix comparison between two terms: =, ≠, <, >, ≤, ≥, e.g. x1 + 1 = y1 or distance(y, c) > distance(z, c)

Compound formulas

Atomic formulas are combined with connectives and quantifiers. The available connectives and their interpretations depend on the mode:

FOL mode

Syntax	Operator	Interpretation
¬φ	negation	classical
φ ∧ ψ	conjunction	classical
φ ∨ ψ	disjunction	classical
φ ⊕ ψ	exclusive or	classical
φ → ψ	implication	classical
φ ↔ ψ	biconditional	classical
∀x φ	universal	unsorted
∃x φ	existential	unsorted

MSFOL mode

Same connectives as FOL except ⊕ (exclusive or) is not available. Quantifiers require a sort annotation:

Syntax	Operator
¬φ, φ ∧ ψ, φ ∨ ψ, φ → ψ, φ ↔ ψ	classical (as FOL)
∀x:Sort φ, ∃x:Sort φ	sorted quantifiers

MSFL mode

Connectives are reinterpreted as Łukasiewicz operators:

Syntax	Operator	Semantics
¬φ	Łuk. negation	1 − φ
φ ∧ ψ	weak conjunction	min(φ, ψ)
φ ∨ ψ	weak disjunction	max(φ, ψ)
φ ⊗ ψ	strong conjunction	max(0, φ + ψ − 1)
φ ⊕ ψ	strong disjunction	min(1, φ + ψ)
φ → ψ	Łuk. implication	min(1, 1 − φ + ψ)
φ ↔ ψ	Łuk. equivalence	1 − |φ − ψ|
∀x:Sort φ, ∃x:Sort φ	sorted quantifiers

A formula may be wrapped in parentheses ( … ) or square brackets [ … ]; the two are interchangeable for grouping.

Operator precedence

The precedence levels are the same across all three modes (MSFL uses the same syntactic structure with Łukasiewicz semantics):

Precedence	Operators	Associativity
1 (highest)	¬, quantifiers ∀ / ∃	prefix
2	∧ ∨ ⊕ (FOL) / ∧ ∨ (MSFOL) / ∧ ∨ ⊗ ⊕ (MSFL)	left
3	→	right
4 (lowest)	↔	right

Worked examples (parenthesised to show how the parser groups them):

• ¬P(x) ∧ Q(x) → (¬P(x)) ∧ Q(x) — negation binds tighter than conjunction
• P(x) ∧ Q(x) → R(x) → (P(x) ∧ Q(x)) → R(x) — conjunction binds tighter than implication
• P(x) → Q(x) ↔ R(x) → (P(x) → Q(x)) ↔ R(x) — implication binds tighter than biconditional
• P(x) → Q(x) → R(x) → P(x) → (Q(x) → R(x)) — implication is right-associative
• P(x) ∧ Q(x) ∧ R(x) → (P(x) ∧ Q(x)) ∧ R(x) — conjunction is left-associative

Mixing same-level operators

The same-level connectives (level 2 above) cannot be mixed without explicit parentheses. This is deliberate: it avoids the silent, easy-to-misread grouping that a default precedence would impose.

FOL mode — ∧, ∨, ⊕ cannot be mixed:

P(x) ∧ Q(x) ∨ R(x)      # rejected
(P(x) ∧ Q(x)) ∨ R(x)    # accepted

MSFOL mode — ∧ and ∨ cannot be mixed:

P(x) ∧ Q(x) ∨ R(x)      # rejected
(P(x) ∧ Q(x)) ∨ R(x)    # accepted

MSFL mode — ∧, ∨, ⊗, ⊕ cannot be mixed:

P(x) ∧ Q(x) ⊗ R(x)        # rejected
(P(x) ∧ Q(x)) ⊗ R(x)      # accepted

A chain of the same operator is always fine: P ∧ Q ∧ R, P ⊗ Q ⊗ R, etc.

Quantifier scope

A quantifier binds only the immediately following (tightly bound) formula, not the rest of the line:

∀x P(x) ∧ Q(x)      # parses as (∀x P(x)) ∧ Q(x)
∀x P(x) → Q(x)      # parses as (∀x P(x)) → Q(x)

If you intend the quantifier to range over the whole formula — which is usually what is meant — add parentheses:

∀x (P(x) → Q(x))    # quantifier ranges over the implication
∀x (P(x) ∧ Q(x))    # quantifier ranges over the conjunction

Quantifiers can be stacked directly: ∀x:H ∀y:H ∃z:A φ.

Supported symbols

Category	FOL	MSFOL	MSFL
Quantifiers	∀ ∃ (unsorted)	∀ ∃ (sorted :Sort)	∀ ∃ (sorted :Sort)
Connectives	∧ ∨ ⊕ ¬ → ↔	∧ ∨ ¬ → ↔	∧ ∨ ⊗ ⊕ ¬ → ↔
Lambda	λ	λ	λ
Sort annotations	—	:Sort	:Sort
Equality / comparison	= ≠ < > ≤ ≥	same	same
Arithmetic	+ - * /	same	same
Grouping	( ) [ ]	same	same
Argument separator	,	same	same

Whitespace is insignificant and may be used freely between tokens — including before sort annotation colons.

Lambda abstraction and application (all modes)

A lambda abstraction is written λ followed by a parameter name, a literal ., and a body formula. All three parser modes support identical lambda surface notation.

Parameter types

Parameter form	Example	Typical use
Single lowercase letter	λx. P(x)	value variable
Multi-letter lowercase name	λfoo. P(foo(x))	named-constant parameter
Uppercase predicate symbol	λP. P(x)	predicate / higher-order parameter

All three token classes become a LambdaVar in the AST. Scope resolution (applied automatically by parse()) then rewrites body occurrences of the lambda-bound name:

• Variable occurrence — λx. P(x): the x in P(x) becomes LambdaVar("x").
• Predicate-application occurrence — λP. P(x): the P(x) in the body becomes Application(LambdaVar("P"), Variable("x")). Multi-argument atoms curry left: P(x, y) → Application(Application(LambdaVar("P"), x), y).
• Named-function occurrence — λfoo. P(foo(x)): the foo(x) in P's argument list (a term-level function call) becomes Application(LambdaVar("foo"), Variable("x")).

The scope obeys the innermost-binder rule: a quantifier removes the quantified name from the lambda-bound set. Inside λx. ∀x P(x), the x in P(x) is logical (stays Variable).

Body scope

The body extends rightward through all connectives — lambda has lower precedence than every binary operator:

λx. P(x) ∧ Q(x)      # body is the And node P(x) ∧ Q(x)
λx. P(x) → Q(x)      # body is the Implies node P(x) → Q(x)

Multi-parameter lambdas are written by nesting: λP. λx. P(x).

Application syntax

A lambda application requires both sides to be parenthesised: (func)(arg).

(λx. P(x))(a)         # arg is variable a
(λx. P(x))(alice)     # arg is constant alice
(λP. P(x))(Q)         # arg is the zero-arity atom Q
(λP. P(x))(Q(y))      # arg is the atom Q(y)

Higher-order application inside the body — a predicate parameter applied to arguments — is written in the natural P(x) notation, not as (P)(x). Scope resolution handles the rewrite automatically.

Parse examples

parser = MSFLParser()

parser.parse("λx. P(x)")
# Lambda(LambdaVar("x"), Atom("P", [LambdaVar("x")]))

parser.parse("λP. P(x)")
# Lambda(LambdaVar("P"), Application(LambdaVar("P"), Variable("x")))

parser.parse("λP. λx. P(x)")
# Lambda(LambdaVar("P"), Lambda(LambdaVar("x"), Application(LambdaVar("P"), LambdaVar("x"))))

parser.parse("λx. ∀x P(x)")
# Lambda(LambdaVar("x"), Quantifier("∀", Variable("x"), Atom("P", [Variable("x")])))
# x inside ∀ is quantifier-bound — NOT rewritten to LambdaVar

parser.parse("(λP. P(x))(Q)")
# Application(Lambda(LambdaVar("P"), Application(LambdaVar("P"), Variable("x"))), Atom("Q", []))

A complete FOL example

∀x ((Object(x) ∧ HasThreeDimensionalShape(x) ∧
     ∀y ∀z ((Point(y) ∧ OnSurfaceOf(y, x) ∧ Point(z) ∧ OnSurfaceOf(z, x))
            → distance(y, centerOf(x)) = distance(z, centerOf(x))))
    → Sphere(x))

A complete MSFOL example

∀x:Person ∀y:Person (Knows(x, y) ∧ Trusted(y:Person)) → Shares(x, y)

A complete MSFL example

∀x:Patient ∀y:Treatment
    (Effective(y:Treatment) ⊗ Tolerable(x:Patient, y:Treatment))
    → Recommended(x:Patient, y:Treatment)

AST nodes

All nodes are Python dataclasses and can be imported from unicode_fol_kit.

Shared term and atom nodes (all modes)

Class	Fields	Notes
Variable	name: str	bound or free variable
Constant	name: str	bare constant or c_-prefixed
Number	value: int | float	numeric literal
Function	name: str, args: list	function application and arithmetic ops
Atom	predicate: str, args: list	predicate or infix comparison

Classical formula nodes (FOL / MSFOL)

Class	Fields
Not	formula
And	left, right
Or	left, right
Xor	left, right (FOL only)
Implies	left, right
Iff	left, right
Quantifier	type: str, variable, formula (FOL only)

MSFOL / MSFL nodes

Class	Fields	Notes
SortedQuantifier	type: str, variable, sort: str, formula	sort annotation without leading :
SortedConstant	name: str, sort: str	sort annotation without leading :

MSFL Łukasiewicz nodes

Class	Fields	Semantics
LukNegation	formula	1 − φ
WeakConjunction	left, right	min(φ, ψ)
WeakDisjunction	left, right	max(φ, ψ)
StrongConjunction	left, right	max(0, φ + ψ − 1)
StrongDisjunction	left, right	min(1, φ + ψ)
LukImplication	left, right	min(1, 1 − φ + ψ)
LukEquivalence	left, right	1 − |φ − ψ|

Lambda-calculus nodes (all modes)

Class	Fields	Notes
LambdaVar	name: str	lambda-bound variable; frozen and hashable — distinct from Variable
Lambda	param: LambdaVar, body: Node	lambda abstraction λparam. body
Application	func: Node, arg: Node	lambda application func(arg)

LambdaVar is kept separate from Variable so that logical binding (by quantifiers) and lambda binding never get confused. free_variables() returns a mixed set that may contain both.

Reductions

Every MSFL node implements two reduction steps:

• to_msfol() — lowers Łukasiewicz connectives to classical nodes while preserving sort annotations (SortedQuantifier and SortedConstant survive unchanged).
• _relativize(facts) — eliminates sort annotations by replacing ∀x:S φ with ∀x (S(x) → φ) and ∃x:S φ with ∃x (S(x) ∧ φ), and replacing SortedConstant(name, sort) with a plain Constant(name).

The top-level helper to_fol(node, include_sort_facts=False) chains both steps and optionally conjoins sort-membership atoms for all ground constants at the top level.

Error handling

Parse errors are reported with human-readable messages rather than raw parser internals. Lexer-level problems (an invalid character, a malformed name or number) raise NamingError; structural problems (an incomplete formula, a misplaced operator, or an attempt to mix same-level connectives without parentheses) raise ParsingError. Both report the offending position and, where useful, a hint. The hint text is mode-aware:

from unicode_fol_kit import MSFLParser

# FOL mode — hint names ∧, ∨, and ⊕
MSFLParser().parse("P(x) ∧ Q(x) ∨ R(x)")
# SYNTAX_ERROR: … Hint: Cannot mix conjunction (∧), disjunction (∨), and exclusive or (⊕) without parentheses

# MSFOL mode — hint names only ∧ and ∨
MSFLParser(many_sorted=True).parse("P(x) ∧ Q(x) ∨ R(x)")
# SYNTAX_ERROR: … Hint: Cannot mix conjunction (∧) and disjunction (∨) without parentheses

# MSFL mode — hint names all four Łukasiewicz connectives
MSFLParser(many_sorted=True, fuzzy=True).parse("P(x) ∧ Q(x) ⊗ R(x)")
# SYNTAX_ERROR: … Hint: Cannot mix weak conjunction (∧), weak disjunction (∨),
#                        strong conjunction (⊗), and strong disjunction (⊕) without parentheses

Citation

If you use this toolkit in academic work, please cite the accompanying preprint:

@misc{vossel2025advancingnaturallanguageformalization,
      title={Advancing Natural Language Formalization to First Order Logic with Fine-tuned LLMs},
      author={Felix Vossel and Till Mossakowski and Björn Gehrke},
      year={2025},
      eprint={2509.22338},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.22338},
}

Vossel, F., Mossakowski, T., & Gehrke, B. (2025). Advancing Natural Language Formalization to First Order Logic with Fine-tuned LLMs. arXiv preprint arXiv:2509.22338.

License

MIT