notxia/unibo-ai-notes
1
0
Fork 0
mirror of https://github.com/NotXia/unibo-ai-notes.git synced 2025-12-15 02:52:22 +01:00
Code Issues Packages Projects Releases Wiki Activity

Add FAIKR2 description logic

Browse Source
This commit is contained in:
Tian Cheng (Riccardo) Xia
2023-11-26 16:43:21 +01:00
parent 098bb90f07
commit 3763967104
2 changed files with 352 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
src/fundamentals-of-ai-and-kr/module2/main.tex
View File

@ -8,5 +8,6 @@
\makenotesfront
\input{sections/_logic.tex}
\input{sections/_ontoligies.tex}
\input{sections/_descriptive_logic.tex}
\end{document}

351
src/fundamentals-of-ai-and-kr/module2/sections/_descriptive_logic.tex Normal file
View File

@ -0,0 +1,351 @@
\chapter{Description logic}
\section{Syntax}
\begin{description}
\item[Logical symbols] \marginnote{Logical symbols}
Symbols with fixed meaning.
\begin{description}
\item[Punctuation] ( ) [ ]
\item[Positive integers]
\item[Concept-forming operators] \texttt{ALL}, \texttt{EXISTS}, \texttt{FILLS}, \texttt{AND}
\item[Connectives] $\sqsubseteq$, $\doteq$, $\rightarrow$
\end{description}
\item[Non-logical symbols] \marginnote{Non-logical symbols}
Domain-dependant symbols.
\begin{description}
\item[Atomic concepts] Categories (CamelCase, e.g. \texttt{Person}).
\item[Roles] Used to describe objects (:CamelCase, e.g. \texttt{:Height}).
\item[Constants] (camelCase, e.g. \texttt{johnDoe}).
\end{description}
\item[Complex concept] \marginnote{Complex concept}
Concept-forming operators can be used to combine atomic concepts and form complex concepts.
A well-formed concept follows the conditions:
\begin{itemize}
\item An atomic concept is a concept.
\item If \texttt{r} is a role and \texttt{d} is a concept, then \texttt{[ALL r d]} is a concept.
\item If \texttt{r} is a role and $n$ is a positive integer, then \texttt{[EXISTS $n$ r]} is a concept.
\item If \texttt{r} is a role and \texttt{c} is a constant, then \texttt{[FILLS r c]} is a concept.
\item If $\texttt{d}_1 \dots \texttt{d}_n$ are concepts,
then \texttt{[AND $\texttt{d}_1 \dots \texttt{d}_n$]} is a concept.
\end{itemize}
\item[Sentence] \marginnote{Sentence}
Connectives can be used to combine concepts and form sentences.
A well-formed sentence follows the conditions:
\begin{itemize}
\item If $\texttt{d}_1$ and $\texttt{d}_2$ are concepts,
then $(\texttt{d}_1 \sqsubseteq \texttt{d}_2)$ is a sentence.
\item If $\texttt{d}_1$ and $\texttt{d}_2$ are concepts,
then $(\texttt{d}_1 \doteq \texttt{d}_2)$ is a sentence.
\item If $\texttt{c}$ is a constant and $\texttt{d}$ is a concept,
then $(\texttt{c} \rightarrow \texttt{d})$ is a sentence.
\end{itemize}
\item[Knowledge base] \marginnote{Knowledge base}
Collection of sentences.
\begin{description}
\item[Constants] are individuals of the domain.
\item[Concepts] are categories of individuals.
\item[Roles] are binary relations between individuals.
\end{description}
\item[Assetion box (A-box)] \marginnote{Assetion box (A-box)}
List of facts about individuals.
\item[Terminological box (T-box)] \marginnote{Terminological box (T-box)}
List of sentences (axioms) about concepts.
\end{description}
\section{Semantics}
\subsection{Concept-forming operators}
\marginnote{Concept-forming operators}
Let \texttt{r} be a role, \texttt{d} be a concept, \texttt{c} be a constant and $n$ a positive integer.
The semantics of concept-forming operators are:
\begin{descriptionlist}
\item[\texttt{[ALL r d]}]
Individuals \texttt{r}-related to the individuals of the category \texttt{d}.
\begin{example}
\texttt{[ALL :HasChild Male]} individuals that have zero children or only male children.
\end{example}
\item[\texttt{[EXISTS $n$ r]}]
Individuals \texttt{r}-related to at least $n$ other individuals.
\begin{example}
\texttt{[EXISTS 1 :Child]} individuals with at least one child.
\end{example}
\item[\texttt{[FILLS r c]}]
Individuals \texttt{r}-related to the individual \texttt{c}.
\begin{example}
\texttt{[FILLS :Child john]} individuals with child \texttt{john}.
\end{example}
\item[\texttt{[AND $\texttt{d}_1 \dots \texttt{d}_n$]}]
Individuals belonging to all the categories $\texttt{d}_1 \dots \texttt{d}_n$.
\end{descriptionlist}
\subsection{Sentences}
\marginnote{Sentences}
Sentences are expressions with truth values in the domain.
Let \texttt{d} be a concept and \texttt{c} be a constant.
The semantics of sentences are:
\begin{descriptionlist}
\item[$\texttt{d}_1 \sqsubseteq \texttt{d}_2$]
Concept $\texttt{d}_1$ is subsumed by $\texttt{d}_2$.
\begin{example}
$\texttt{PhDStudent} \sqsubseteq \texttt{Student}$ as every PhD is also a student.
\end{example}
\item[$\texttt{d}_1 \doteq \texttt{d}_2$]
Concept $\texttt{d}_1$ is equivalent to $\texttt{d}_2$.
\begin{example}
$\texttt{PhDStudent} \doteq \texttt{[AND Student :Graduated :HasFunding]}$
\end{example}
\item[$\texttt{c} \rightarrow \texttt{d}$]
The individual \texttt{c} satisfies the description of the concept \texttt{d}.
\begin{example}
$\texttt{federico} \rightarrow \texttt{Professor}$
\end{example}
\end{descriptionlist}
\subsection{Interpretation}
\begin{description}
\item[Interpretation] \marginnote{Interpretation}
An interpretation $\mathfrak{I}$ in description logic is a pair ($\mathcal{D}, \mathcal{I}$) where:
\begin{itemize}
\item $\mathcal{D}$ is the domain.
\item $\mathcal{I}$ is the interpretation mapping.
\begin{description}
\item[Constant]
Let \texttt{c} be a constant, $\mathcal{I}[\texttt{c}] \in \mathcal{D}$.
\item[Atomic concept]
Let \texttt{a} be an atomic concept, $\mathcal{I}[\texttt{a}] \subseteq \mathcal{D}$.
\item[Role]
Let \texttt{r} be a role, $\mathcal{I}[\texttt{r}] \subseteq \mathcal{D} \times \mathcal{D}$.
\item[\texttt{Thing}]
The concept \texttt{Thing} corresponds to the domain: $\mathcal{I}[\texttt{Thing}] = \mathcal{D}$.
\item[\texttt{[ALL r d]}]
\[
\mathcal{I}[\texttt{[ALL r d]}] =
\{ \texttt{x} \in \mathcal{D} \mid \forall \texttt{y}:
\langle \texttt{x}, \texttt{y} \rangle \in \mathcal{I}[r] \text{ then } \texttt{y} \in \mathcal{I}[d] \}
\]
\item[\texttt{[EXISTS $n$ r]}]
\[
\mathcal{I}[\texttt{[EXISTS $n$ r]}] =
\{ \texttt{x} \in \mathcal{D} \mid \text{ exists at least $n$ distinct } \texttt{y}:
\langle \texttt{x}, \texttt{y} \rangle \in \mathcal{I}[r] \}
\]
\item[\texttt{[FILLS r c]}]
\[
\mathcal{I}[\texttt{[FILLS r c]}] = \{ \texttt{x} \in \mathcal{D} \mid
\langle \texttt{x}, \mathcal{I}[\texttt{c}] \rangle \in \mathcal{I}[\texttt{r}] \}
\]
\item[\texttt{[AND $\texttt{d}_1 \dots \texttt{d}_n$]}]
\[
\mathcal{I}[\texttt{[AND $\texttt{d}_1 \dots \texttt{d}_n$]}] =
\mathcal{I}[\texttt{d}_1] \cap \dots \cap \mathcal{I}[\texttt{d}_n]
\]
\end{description}
\end{itemize}
\item[Model] \marginnote{Model}
Given an interpretation $\mathfrak{I} = (\mathcal{D}, \mathcal{I})$,
a sentence is true under $\mathfrak{I}$ ($\mathfrak{I} \models \text{sentence}$) if:
\begin{itemize}
\item $\mathfrak{I} \models (\texttt{c} \rightarrow \texttt{d})$ iff $\mathcal{I}[\texttt{c}] \in \mathcal{I}[\texttt{d}]$.
\item $\mathfrak{I} \models (\texttt{d}_\texttt{1} \sqsubseteq \texttt{d}_\texttt{2})$ iff
$\mathcal{I}[\texttt{d}_\texttt{1}] \subseteq \mathcal{I}[\texttt{d}_\texttt{2}]$.
\item $\mathfrak{I} \models (\texttt{d}_\texttt{1} \doteq \texttt{d}_\texttt{2})$ iff
$\mathcal{I}[\texttt{d}_\texttt{1}] = \mathcal{I}[\texttt{d}_\texttt{2}]$.
\end{itemize}
Given a set of sentences $S$, $\mathfrak{I}$ models $S$ if $\mathfrak{I} \models S$.
\item[Entailment] \marginnote{Entailment}
A set of sentences $S$ logically entails a sentence $\alpha$ if:
\[ \forall \mathfrak{I}:\, (\mathfrak{I} \models S) \rightarrow (\mathfrak{I} \models \alpha) \]
\end{description}
\section{Reasoning}
\subsection{T-box reasoning}
Given a knowledge base of a set of sentences $S$, we would like to be able to determine the following:
\begin{description}
\item[Satisfiability] \marginnote{Satisfiability}
A concept \texttt{d} is satisfiable w.r.t. $S$ if:
\[ \exists \mathfrak{I}, (\mathfrak{I} \models S): \mathfrak{I}[\texttt{d}] \neq \varnothing \]
\item[Subsumption] \marginnote{Subsumption}
A concept $\texttt{d}_1$ is subsumed by $\texttt{d}_2$ w.r.t. $S$ if:
\[ \forall \mathfrak{I}, (\mathfrak{I} \models S): \mathfrak{I}[\texttt{d}_1] \subseteq \mathfrak{I}[\texttt{d}_2] \]
\item[Equivalence] \marginnote{Equivalence}
A concept $\texttt{d}_1$ is equivalent to $\texttt{d}_2$ w.r.t. $S$ if:
\[ \forall \mathfrak{I}, (\mathfrak{I} \models S): \mathfrak{I}[\texttt{d}_1] = \mathfrak{I}[\texttt{d}_2] \]
\item[Disjointness] \marginnote{Disjointness}
A concept $\texttt{d}_1$ is disjoint to $\texttt{d}_2$ w.r.t. $S$ if:
\[ \forall \mathfrak{I}, (\mathfrak{I} \models S): \mathfrak{I}[\texttt{d}_1] \neq \mathfrak{I}[\texttt{d}_2] \]
\end{description}
\begin{theorem}[Reduction to subsumption]
\marginnote{Reduction to subsumption}
Given the concepts $\texttt{d}_1$ and $\texttt{d}_2$, it holds that:
\begin{itemize}
\item $\texttt{d}_1 \text{ is unsatisfiable } \iff \texttt{d}_1 \sqsubseteq \bot$.
\item $\texttt{d}_1 \doteq \texttt{d}_2 \iff \texttt{d}_1 \sqsubseteq \texttt{d}_2 \land \texttt{d}_2 \sqsubseteq \texttt{d}_1$.
\item $\texttt{d}_1 \text{ and } \texttt{d}_2 \text{ are disjoint} \iff (\texttt{d}_1 \cap \texttt{d}_2) \sqsubseteq \bot$.
\end{itemize}
\end{theorem}
\subsection{A-box reasoning}
Given a constant \texttt{c}, a concept \texttt{d} and a set of sentences $S$, we can determine the following:
\begin{description}
\item[Satisfiability] \marginnote{Satisfiability}
A constant \texttt{c} satisfies the concept \texttt{d} if:
\[ S \models (\texttt{c} \rightarrow \texttt{d}) \]
Note that it can be reduced to subsumption.
\end{description}
\subsection{Computing subsumptions}
Given a knowledge base $KB$ and two concepts \texttt{d} and \texttt{e},
we want to prove:
\[ KB \models (\texttt{d} \sqsubseteq \texttt{e}) \]
The following algorithms can be employed:
\begin{descriptionlist}
\item[Structural matching] \marginnote{Structural matching}
\phantom{}
\begin{enumerate}
\item Normalize \texttt{d} and \texttt{e} into a conjunctive form:
\[ \texttt{d} = \texttt{[AND d$_1$ \dots d$_n$]} \hspace*{1cm} \texttt{e} = \texttt{[AND e$_1$ \dots e$_m$]} \]
\item Check if each part of \texttt{e} is accounted by at least a component of \texttt{d}.
\end{enumerate}
\item[Tableaux-based algorithms] \marginnote{Tableaux-based algorithms}
Exploit the following theorem:
\[ (KB \models (C \sqsubseteq D)) \iff (KB \cup (x : C \sqcap \lnot D)) \text{ is inconsistent} \]
Note: similar to refutation.
\end{descriptionlist}
\subsection{Open world assumption}
\begin{description}
\item[Open world assumption] \marginnote{Open world assumption}
If a sentence cannot be inferred, its truth values is unknown.
\end{description}
Description logics are based on the open world assumption.
To reason in open world assumption, all the possible models are split upon encountering an unknown facts
depending on the possible cases (Oedipus example).
\section{Expanding description logic}
It is possible to expand a description logic by:
\begin{descriptionlist}
\item[Adding concept-forming operators] \marginnote{Adding concept-forming operators}
Let \texttt{r} be a role, \texttt{d} be a concept, \texttt{c} be a constant and $n$ a positive integer.
We can extend our description logic with:
\begin{description}
\item[\texttt{[AT-MOST $n$ r]}]
Individuals \texttt{r}-related to at most $n$ other individuals.
\begin{example}
\texttt{[AT-MOST $1$ :Child]} individuals with only a child.
\end{example}
\item[\texttt{[ONE-OF c$_1$ $\dots$ c$_n$]}]
Concept only satisfied by \texttt{c$_1$ $\dots$ c$_n$}.
\begin{example}
$\texttt{Beatles} \doteq \texttt{[ALL :BandMember [ONE-OF john paul george ringo]]}$
\end{example}
\item[\texttt{[EXISTS $n$ r d]}]
Individuals \texttt{r}-related to at least $n$ individuals in the category \texttt{d}.
\begin{example}
\texttt{[EXISTS $2$ :Child Male]} individuals with at least two male children.
\end{example}
Note: this increases the computational complexity of entailment.
\end{description}
\item[Relating roles] \marginnote{Relating roles}
\begin{description}
\item[\texttt{[SAME-AS r$_1$ r$_2$]}]
Equates fillers of the roles r$_1$ and r$_2$
\begin{example}
\texttt{[SAME-AS :CEO :Owner]}
\end{example}
Note: this increases the computational complexity of entailment.
Role chaining also leads to undecidability.
\end{description}
\item[Adding rules] \marginnote{Adding rules}
Rules are useful to add conditions (e.g. \texttt{if d$_1$ then [FILLS r c]}).
\end{descriptionlist}
\section{Description logics family}
Depending on the number of operators, a description logic can be:
\begin{itemize}
\item More expressive.
\item Computationally more expensive.
\item Undecidable.
\end{itemize}
\begin{description}
\item[Attributive language ($\mathcal{AL}$)]
Minimal description logic with:
\begin{itemize}
\item Atomic concepts.
\item Universal concept (\texttt{Thing} or $\top$).
\item Bottom concept (\texttt{Nothing} or $\bot$).
\item Atomic negation (only for atomic concepts).
\item \texttt{AND} operator ($\sqcap$).
\item \texttt{ALL} operator ($\forall$).
\item \texttt{[EXISTS $1$ r]} operator ($\exists$).
\end{itemize}
\item[Attributive language complement ($\mathcal{ALC}$)]
$\mathcal{AL}$ with negation for concepts.
\end{description}
\begin{table}[h]
\centering
\begin{tabular}{|c|c|}
\hline
$\mathcal{F}$ & Functional properties \\ \hline
$\mathcal{E}$ & Full existential quantification \\ \hline
$\mathcal{U}$ & Concept union \\ \hline
$\mathcal{C}$ & Complex concept negation \\ \hline
$\mathcal{S}$ & $\mathcal{ALC}$ with transitive roles \\ \hline
$\mathcal{H}$ & Role hierarchy \\ \hline
$\mathcal{R}$ & \makecell[c]{Limited complex roles axioms\\Reflexivity and irreflexivity\\Roles disjointness} \\ \hline
$\mathcal{O}$ & Nominals \\ \hline
$\mathcal{I}$ & Inverse properties \\ \hline
$\mathcal{N}$ & Cardinality restrictions \\ \hline
$\mathcal{Q}$ & Qualified cardinality restrictions \\ \hline
$\mathcal{(D)}$ & Datatype properties, data values and data types \\ \hline
\end{tabular}
\caption{Name and expressivity of logics}
\end{table}