\section{Background}

\section{Problem Formulation} \label{sec3:problem}


\section{Model Description}


\subsection{Input Representation}

\begin{algorithm}
	\caption{Sentence Grouping via Locality-Sensitive Hashing (LSH)}
	\label{alg3:lsh}
	\begin{algorithmic}[1]
		\REQUIRE shingle size $n$, similarity threshold $t$, minimum group size $g$
		\ENSURE explanation set $\mathcal{E}$, groups of sentences $\mathcal{M}$
		\STATE Pre-process textual data to obtain the sentence collection $\mathcal{S}$
		\STATE $lsh \gets MinHashLSH(t)$, $\mathcal{C} \gets \varnothing$
		\FOR{sentence $s$ in $\mathcal{S}$}
		\STATE $m \gets MinHash()$  // create MinHash for $s$
		\FOR{$n$-shingle $h$ in $s$}
		\STATE $m.update(h)$  // convert $s$ into $m$ by encoding its $n$-shingles
		\ENDFOR
		\STATE $lsh.insert(m)$, $\mathcal{C}.add(m)$  // $\mathcal{C}$: set of all sentences' MinHash
		\ENDFOR
		\STATE $\mathcal{M} \gets \varnothing$, $\mathcal{Q} \gets \varnothing$  // $\mathcal{Q}$: set of queried sentences
		\FOR{$m$ in $\mathcal{C}$}
		\IF{$m$ not in $\mathcal{Q}$}
		\STATE $\mathcal{G} \gets lsh.query(m)$  // $\mathcal{G}$: ID set of duplicate sentences
		\IF{$\mathcal{G}.size > g$}
		\STATE $\mathcal{M}.add(\mathcal{G})$  // only keep groups with enough sentences
		\STATE $\mathcal{E}.add(\mathcal{G}.get())$  // keep one explanation in each group
		\ENDIF
		\FOR{$m'$ in $\mathcal{G}$}
		\STATE $lsh.remove(m')$, $\mathcal{Q}.add(m')$  // for efficiency
		\ENDFOR
		\ENDIF
		\ENDFOR
	\end{algorithmic}
\end{algorithm}

\subsection{Transformer and Attention Masking}



\subsection{Explanation and Recommendation}


\paragraph{Explanation Generation:}


\paragraph{Context Prediction:}


\paragraph{Rating Prediction:}


\paragraph{Multi-task Learning:}


\section{Experimental Setup}


\subsection{Datasets}

\begin{table}
	\caption{Statistics of the three datasets.}
	\begin{center}	
		\begin{tabular}{l|r|r|r}
			\hline
			& Yelp & Amazon & TripAdvisor \\
			\hline
			\#users & 27,147 & 7,506 & 9,765 \\
			\#items & 20,266 & 7,360 & 6,280 \\
			\#records & 1,293,247 & 441,783 & 320,023 \\
			\#features & 7,340 & 5,399 & 5,069 \\
			\#records / user & 47.64 &58.86&32.77 \\
			\#records / item & 63.81 &60.02&50.96 \\
			\#words / explanation & 12.32 &14.14&13.01 \\
			\hline
		\end{tabular}
	\end{center}	
	\label{table3:dataset}
\end{table}

\subsection{Evaluation Metrics}


\subsection{Compared Methods}


\subsection{Implementation Details}





\section{Results and Analysis}

\subsection{Quantitative Analysis on Explanations}


\subsection{Qualitative Case Study on Explanations}


\subsection{Recommendation Performance}


\subsection{Ablation Study}

\begin{sidewaystable}[!tbp]
	\caption[Ablation study on the smallest dataset TripAdvisor.]{Ablation study on the smallest dataset TripAdvisor. Arrows $\uparrow$ and $\downarrow$ respectively denote the performance increase and decrease compared with PETER.}
	\begin{center}
		\begin{tabular}{l|lll|lll|ll}
			\hline
			\multirow{2}{*}{} & \multicolumn{3}{c|}{Explainability} & \multicolumn{3}{c|}{Text Quality}  & \multicolumn{2}{c}{Recommendation} \\ \cline{2-9}
			& FMR & FCR & DIV & USR & BLEU-1 & BLEU-4 & RMSE & MAE \\ \hline		
			Disable $\mathcal{L}_c$ & 0.06 $\downarrow$ & 0.03 $\downarrow$ & 5.75 $\downarrow$ & 0.01 $\downarrow$ & 15.37 $\downarrow$ & 0.86 $\downarrow$ & 0.80 $\uparrow$ & 0.61 $\uparrow$ \\
			Disable $\mathcal{L}_r$ & 0.07 & 0.14 $\uparrow$ & 2.90 $\uparrow$ & 0.10 $\uparrow$ & 16.16 $\uparrow$ & 1.15 $\uparrow$ & 3.23 $\downarrow$ & 3.10 $\downarrow$ \\
			Left-to-Right Masking & 0.07 & 0.15 $\uparrow$ & 2.68 $\uparrow$ & 0.12 $\uparrow$ & 15.73 $\downarrow$ & 1.11 & 0.87 $\downarrow$ & 0.68 $\downarrow$ \\ \hline
			PETER & 0.07 & 0.13 & 2.95 & 0.08 & 15.96 & 1.11 & 0.81 & 0.63 \\ \hline
		\end{tabular}
	\end{center}
	\label{table3:ablation}
\end{sidewaystable}

\section{Summary}