1 files changed, 14 insertions, 9 deletions
diff --git a/paper/method.tex b/paper/method.tex
index 1cd3497..a4f7ae7 100644
--- a/paper/method.tex
+++ b/paper/method.tex
@@ -1,11 +1,13 @@
 \section{Method}
 
-This section explains the three main components of our approach: automatically extracting patterns from student submissions, learning classification rules for correct and incorrect programs, and using those rules to generate hints.
+This section explains the three steps in our approach: discovering AST patterns, learning classification rules for correct and incorrect programs, and using those rules to generate hints.
 
 \subsection{Extracting patterns}
 \label{sec:extracting-patterns}
 
-We construct patterns by connecting pairs of leaf nodes in a program’s AST. For this paper we always select a pair of nodes from the same clause: either two nodes referring to the same variable (like the examples in Fig.~\ref{fig:sister}), or a value (such as the empty list \code{[]} or the number \code{0}) and another variable or value in the same \textsf{compound} or \textsf{binop} (like the blue dotted pattern in Fig.~\ref{fig:sum}). For example, in the clause (the second occurrence of each variable -- \code{A}, \code{B} and \code{C} -- is marked with~’ for disambiguation)
+We extract patterns from student submissions. As described above, we are only interested in patterns connecting pairs of leaf nodes in an AST: either two nodes referring to the same variable (like the examples in Fig.~\ref{fig:sister}), or a value (such as the empty list \code{[]} or the number \code{0}) and another variable/value occurring within the same \textsf{compound} or \textsf{binop} (like the blue dotted pattern in Fig.~\ref{fig:sum}).
+
+We induce patterns from such node pairs. Given the clause (the second occurrence of each variable -- \code{A}, \code{B} and \code{C} -- is marked with ’ for disambiguation)
 
 \begin{Verbatim}
 a(A,\textsf{ }B):-
@@ -16,31 +18,34 @@ a(A,\textsf{ }B):-
 \noindent
 we select the following pairs of nodes: \{\code{A},\,\code{A\textsf{'}}\}, \{\code{B},\,\code{B\textsf{'}}\}, \{\code{C},\,\code{C\textsf{'}}\}, \{\code{B\textsf{'}},\,\code{1}\} and \{\code{C\textsf{'}},\,\code{1}\}.
 
-For each selected pair of leaf nodes $(a,b)$ we build a pattern by walking the AST in depth-first order, and recording nodes that lie on the paths to $a$ and $b$. We omit \textsf{and} nodes, as explained in the previous section. We also include certain nodes that do not lie on a path to any selected leaf. Specifically, we include the functor or operator of all \textsf{compound}, \textsf{binop} or \textsf{unop} nodes containing $a$ or $b$.
+For each selected pair of leaf nodes $(a,b)$ we construct a pattern by walking the AST in depth-first order and recording nodes that lie on the paths to $a$ and $b$. We omit \textsf{and} nodes, as explained in the previous section. We also include certain nodes that lie near the paths to selected leaves. Specifically, we include the functor/operator of all \textsf{compound}, \textsf{binop} and \textsf{unop} nodes containing $a$ or $b$.
+
+Patterns are extracted automatically given above constraints (each pattern connecting a pair of variables or values). We find that such patterns work well for Prolog. Other languages, however, will likely require different kinds of patterns to achieve good performance.
 
-Patterns constructed in this way form the set of features for rule learning. To keep this set at a reasonable size, we only use patterns that have occurred at least five times in submitted programs.
+Finally, to avoid learning rules specific to a particular program (covering typos and other idiosyncratic mistakes), we ignore rare patterns. In this study we used patterns that occurred in at least five submissions. These patterns form the feature space for rule learning.
 
 \subsection{Learning rules}
 
-We represent students’ programs in the feature space of AST patterns described above. Each pattern corresponds to one binary feature with value $true$ when the pattern is present and $false$ when it is absent. We classify programs as correct or incorrect based on predefined unit tests for each problem, and use these labels for machine learning.
+We represent students’ programs in the feature space of AST patterns described above. Each pattern corresponds to one binary feature with value \textsf{true} when the pattern is present and \textsf{false} when it is absent. We use unit testing to classify each program as correct if it passes all test cases, and incorrect otherwise. We use these labels for machine learning.
 
-Since programs can be validated with appropriate unit tests, our goal is not classifying new submissions, but rather to discover patterns associated with program correctness. This approach to machine learning is called \emph{descriptive induction} -- the automatic discovery of patterns describing regularities in data. We use rule learning for this task, because conditions of rules can be easily translated to hints.
+Since we can establish program correctness using appropriate unit tests, our goal here is not classifying new submissions. Instead, we wish to discover patterns associated with program correctness. This approach to machine learning is called \emph{descriptive induction} -- the automatic discovery of patterns describing regularities in data. We use rule learning for this task, because conditions of rules can be easily translated to hints.
 
 Before explaining the algorithm, let us discuss the reasons why a program can be incorrect. Our experience indicates that bugs in student programs can often be described by 1) some incorrect or \emph{buggy} pattern, which needs to be removed, or 2) some missing relation (pattern) that should be included before the program can be correct. We shall now explain how both types of errors can be identified with rules.
 
 To discover buggy patterns, the algorithm first learns rules that describe incorrect programs (I-rules). We use a variant of the CN2 algorithm~\cite{clark1991rule} implemented within the Orange data-mining toolbox~\cite{demsar2013orange}. Since we use rules to generate hints, and since hints should not be presented to students unless they are likely to be correct, we impose additional constraints on the rule learner:
+
 \begin{enumerate}
  \item The classification accuracy of each learned rule must exceed a threshold (we selected 90\%, as 10\% error  seems acceptable for our application).
  \item Each conjunct in a condition must be significant with respect to the likelihood-ratio test (in our experiments significance threshold was set to $p=0.05$).
- \item A conjunct can only specify the presence of a pattern: we allow feature-value pairs with only $true$ as value. 
+ \item A conjunct can only specify the presence of a pattern: we allow feature-value pairs with only \textsf{true} as value.
 \end{enumerate}
 
 \noindent The former two constraints are needed to induce good rules with significant patterns, while the latter constraint assures that rules mention only presence (and not absence) of patterns as reasons for a program to be incorrect. This is important, since conditions of I-rules ought to contain patterns symptomatic of incorrect programs.
 
 With respect to the second type of error, we could try the same approach and learn rules using the above algorithm for the class of correct programs (C-rules). Having accurate rules for correct programs, the conditional part of these rules would define sufficient combinations of patterns that render a program correct.
-It turns out that it is difficult to learn accurate rules for correct programs, because there are many programs that are incorrect despite having all important patterns, because they include also incorrect patterns. 
+It turns out that it is difficult to learn accurate rules for correct programs, because there are many programs that are incorrect despite having all important patterns, because they include also incorrect patterns.
 
-A possible way to solve this problem is to remove programs that are covered by rules for incorrect class. This way all known buggy patterns are removed from the data and will not be included in C-rules. However, removing incorrect patterns also removes the need for relevant patterns. For example, if all incorrect programs were removed, the single C-rule “$\mathsf{true} ⇒ \mathsf{correct}$” would suffice, which cannot be used to generate hints. We achieved the best results by learning from the complete data set, whereas the accuracy of rules was estimated on data without programs covered by I-rules. 
+A possible way to solve this problem is to remove programs that are covered by rules for incorrect class. This way all known buggy patterns are removed from the data and will not be included in C-rules. However, removing incorrect patterns also removes the need for relevant patterns. For example, if all incorrect programs were removed, the single C-rule “$\mathsf{true} ⇒ \mathsf{correct}$” would suffice, which cannot be used to generate hints. We achieved the best results by learning from the complete data set, whereas the accuracy of rules was estimated on data without programs covered by I-rules.
 
 Even though our main interest is discovery of patterns, we can still use induced rules to classify new programs, for example to evaluate the quality of rules. The classification procedure has three steps: 1) if an I-rule covers the program, classify it as incorrect; 2) else if a C-rule covers the program, classify it as correct; 3) otherwise, if no rule covers the program, classify it as incorrect.