!7

cmodule TheoryMaker > DynConvo {
  /*
  1. measurable features (fields of object)
  2. labels (words the user throws in)
  3. make theories (random connectors between features and labels)
  4. check theories
  
  1. show a random line
  2. user types keyword
  3. assign keyword to line
  4. check if prediction weas correct

  Basic theory making
  -------------------
  
  For any label X:
    test theory (for every M: M has label X)
    test theory (for every M: M doesn't have label X)
  
  For any feature F:
    for every seen value V of F:
      for every label X:
        test theory (for every M: msg M's feature F has value V => msg has label x))
        test theory (for every M: msg M's feature F has value V => msg doesn't have label x))

  */
  
  srecord Theory(BasicLogicRule statement) {
    new PosNeg<Msg> examples;
    //bool iff; // <=> instead of only =>
    toString { ret str(statement.lhs instanceof MPTrue ? "Every message is " + statement.rhs
      : bidiMode ? statement.lhs + " <=> " + statement.rhs : statement); }
  }

  // propositions about a message. check returns null if unknown
  asclass MsgProp { abstract Bool check(Msg msg); }

  srecord MPTrue() > MsgProp {
    Bool check(Msg msg) { true; }
    toString { ret "always"; }
  }
  
  record HasLabel(S label) > MsgProp {
    Bool check(Msg msg) { ret msg2label_new.get(msg, label); }
    toString { ret label; }
  }
  
  record DoesntHaveLabel(S label) > MsgProp {
    Bool check(Msg msg) { ret not(msg2label_new.get(msg, label)); }
    toString { ret "not " + label; }
  }

  record FeatureValueIs(S feature, O value) > MsgProp {
    Bool check(Msg msg) { ret eq(getMsgFeature(msg, feature), value); }
    toString { ret feature + "=" + value; }
  }

  class Label {
    S name;

    *() {}
    *(S *name) {}
    
    TreeSetWithDuplicates<Theory> bestTheories = new(reverseComparatorFromCalculatedField theoryScore());

    int score() { ret theoryScore(first(bestTheories)); }
    Theory bestTheory() { ret first(bestTheories); }
  }
  
  switchable double minAdjustedScoreToDisplay = 50;
  switchable bool autoNext = false;
  static bool bidiMode = true; // treat all theories as bidirectional
  
  L<Msg> msgs; // full dialog
  L<Msg> shownMsgs;
  transient Map<Msg, Map<S, O>> msg2features = AutoMap<>(lambda1 calcMsgFeatures);
  new LinkedHashSet<Theory> theories;
  S analysisText;
  transient JTable theoryTable, labelsTable, trainedExamplesTable;
  transient JTabbedPane tabs;
  transient SingleComponentPanel scpPredictions;
  transient new Map<S, Label> labelsByName;

  new Set<S> allLabels;
  transient new L<IVF1<S>> onNewLabel;
  new DoubleKeyedMap<Msg, S, Bool> msg2label_new;
  transient new Map<S, FeatureExtractor<Msg>> featureExtractors;
  transient Q thinkQ;
  
  sinterface FeatureEnv<A> {
    A mainObject();
    O getFeature(S name);
  }

  sinterface FeatureExtractor<A> {
    O get(FeatureEnv<A> env);
  }

  start {
    thinkQ = dm_startQ("Thought Queue");
    thinkQ.add(r {
      // legacy + after deletion cleaning
      setField(allLabels := asTreeSet(msg2label_new.bKeys()));
      updateLabelsByName();
      
      onNewLabel.add(lbl -> change());
  
      makeTheoriesAboutLabels();
      makeTheoriesAboutFeaturesAndLabels();
      
      for (S field : fields(Msg))
        featureExtractors.put(field, env -> getOpt(env.mainObject(), field));
  
      makeTextExtractors("text");
  
      callFAllOnAll(onNewLabel, allLabels);
      
      msg2labelUpdated();
      if (empty(msgs))
        setField(msgs := mainCruddieLog());
      showRandomMsg();
    });
  }

  void makeTheoriesAboutLabels {
    // For any label X:
    onNewLabel.add(lbl -> {
      // test theory (for every M: M has label X)
      addTheory(new Theory(BasicLogicRule(new MPTrue, new HasLabel(lbl))));
      // test theory (for every M: M doesn't have label X)
      addTheory(new Theory(BasicLogicRule(new MPTrue, new DoesntHaveLabel(lbl))));
    });
  }

  void makeTheoriesAboutFeaturesAndLabels {
    // for every label X:
    onNewLabel.add(lbl -> {
      // For any feature F:
      for (S feature : keys(featureExtractors))
        // for every seen value V of F:
        for (O value : possibleValuesOfFeatureRelatedToLabel(feature, lbl))
          for (O rhs : ll(new HasLabel(lbl), new DoesntHaveLabel(lbl)))
            // test theory (for every M: msg M's feature F has value V => msg has/doesn't have label x))
            addTheory(new Theory(BasicLogicRule(
              new FeatureValueIs(feature, value), rhs)));
    });
  }

  Set possibleValuesOfFeature(S feature) {
    if (isBoolField(Msg, feature))
      ret litset(false, true);
    ret litset();
  }

  Set possibleValuesOfFeatureRelatedToLabel(S feature, S label) {
    Set set = possibleValuesOfFeature(feature);
    fOr (Msg msg : getMsgsRelatedToLabel(label))
      set.add(getMsgFeature(msg, feature));
    ret set;
  }

  // returns AutoMap with no realized entries
  Map<S, O> calcMsgFeatures(Msg msg) {
    new Var<FeatureEnv<Msg>> env;
    AutoMap<S, O> map = new(feature -> featureExtractors.get(feature).get(env!));
    env.set(new FeatureEnv<Msg> {
      Msg mainObject() { ret msg; }
      O getFeature(S feature) { ret map.get(feature); }
    });
    ret map;    
  }

  void showMsgs(L<Msg> l) {
    setField(shownMsgs := l);
    setMsgs(l);
    if (l(shownMsgs) == 1) {
      Msg msg = first(shownMsgs);
      setField(analysisText := joinWithEmptyLines(
        "Trained Labels: " + or2(renderBoolMap(getMsgLabels(msg)), "-"),
        "Features:\n" + formatColonProperties_quoteStringValues(
msg2features.get(msg))
      ));
      setSCPComponent(scpPredictions,
        scrollableStackWithSpacing(map(predictionsForMsg(msg), p ->
          withSideMargin(jLabelWithButtons(iround(p.adjustedConfidence) + "%: " + p.predictedLabel(), 
            "Right", rThread { acceptPrediction(p) },
            "Wrong", rThread { rejectPrediction(p) })))));
    } else setField(analysisText := "");
  }

  void updatePredictions() {
    showMsgs(shownMsgs);
  }

  srecord Prediction(S label, bool plus, double adjustedConfidence) {
    toString {
      ret predictedLabel() + " (confidence: " + iround(adjustedConfidence) + "%)";
    }

    S predictedLabel() {
      ret (plus ? "" : "not ") + label;
    }
  }

  L<Prediction> predictionsForMsg(Msg msg) {
    // positive labels first, then "not"s. sort by score in each group
    new L<Prediction> out;
    for (Label label : values(labelsByName)) {
      Theory t = label.bestTheory(), continue if null;
      Bool lhs = evalTheoryLHS(t, msg), continue if null;
      bool prediction = t.statement.rhs instanceof DoesntHaveLabel ? !lhs : lhs;
      double conf = threeB1BScore(t.examples), adjusted = adjustConfidence(conf);
      if (adjusted < minAdjustedScoreToDisplay) continue;
      out.add(new Prediction(label.name, prediction, adjusted));
    }
    ret sortedByCalculatedFieldDesc(out, p -> pair(p.plus, p.adjustedConfidence));
  }

  // go from range 50-100 to 0-100 (might look better)
  double adjustConfidence(double x) {
    ret max(0, (x-50)*2);
  }

  void showRandomMsg {
    showMsgs(randomElementAsList(msgs));
  }

  void acceptPrediction(Prediction p) {
    if (p != null) sendInput2(p.predictedLabel());
  }

  void rejectPrediction(Prediction p) {
    if (p != null) sendInput2(cloneWithFlippedBoolField plus(p).predictedLabel());
  }

  @Override
  void sendInput2(S s) {
    // treat input as a label
    if (l(shownMsgs) == 1) {
      Msg shown = first(shownMsgs);
      new Matches m;
      if "not ..." {
        S label = cleanLabel(m.rest());
        doubleKeyedMapPutVerbose(+msg2label_new, shown, label, false);
        msg2labelUpdated(label);
        if (autoNext) showRandomMsg();
      } else {
        S label = cleanLabel(s);
        doubleKeyedMapPutVerbose(+msg2label_new, shown, label, true);
        msg2labelUpdated(label);
        if (autoNext) showRandomMsg();
      }
      change();
    }
  }

  Map<S, Bool> getMsgLabels(Msg msg) {
 ret msg2label_new.getA(msg);
 }
  Set<Msg> getMsgsRelatedToLabel(S label) { ret msg2label_new.asForB(label); }

  void msg2labelUpdated(S label) {
    for (Theory t : cloneList(labelByName(label).bestTheories))
      checkTheory(t);
    msg2labelUpdated();
  }

  void msg2labelUpdated() {
    callFAllOnAll(onNewLabel, addAll_returnNew(allLabels, msg2label_new.bKeys()));
    updateTrainedExamplesTable();
  }

  void updateTrainedExamplesTable {
    dataToTable_uneditable(trainedExamplesTable, map(msg2label_new.map1, (msg, map) ->
      litorderedmap(
        "Message" := (msg.fromUser ? "User" : "Bot") + ": " + msg.text,
        "Labels" := renderBoolMap(map))));
  }

  JComponent mainPart() {
    ret jhsplit(jvsplit(
        jCenteredSection("Focused Message", super.mainPart()),
        jhsplit(
          jCenteredSection("Message Analysis", dm_textArea analysisText()),
          jCenteredSection("Predictions", scpPredictions = singleComponentPanel())
        )),
      with(r updateTabs, tabs = jtabs(
        "", with(r updateLabelsTable, labelsTable = sexyTable()),
        "", with(r updateTheoryTable, tableWithSearcher2_returnPanel(theoryTable = sexyTable())),
        "", with(r updateTrainedExamplesTable, tableWithSearcher2_returnPanel(trainedExamplesTable = sexyTable()))
      )));
  }

  void updateTabs {
    setTabTitles(tabs,
      firstLetterToUpper(nLabels(labelsByName)),
      firstLetterToUpper(nTheories(theories)),
      n2(msg2label_new.aKeys(), "Trained Example"));
  }

  void updateTheoryTable {
    L<Theory> sorted = sortedByCalculatedFieldDesc(theories, t ->
      t.examples == null ? null : t.examples.score());
    dataToTable_uneditable(theoryTable, map(sorted, t -> litorderedmap(
      "Score" := renderTheoryScore(t),
      "Theory" := str(t))));
  }

  Map<S, Theory> labelsToBestTheoryMap() {
    Map<S, L<Theory>> map = multiMapToMap(multiMapIndex targetLabelOfTheory(theories));
    ret mapValues(map,
      theories -> highestBy theoryScore(theories));
  }

  void updateLabelsTable {
    L<Label> sorted = sortedByCalculatedFieldDesc(values(labelsByName), l -> l.score());
    dataToTable_uneditable(labelsTable, map(sorted, label -> {
      Cl<Theory> bestTheories = label.bestTheories.tiedForFirst();
      ret litorderedmap(
        "Label" := label.name,
        "Prediction Confidence" := renderTheoryScore(first(bestTheories)),
        "Best Theory" := empty(bestTheories) ? "" :
          (l(bestTheories) > 1 ? "[+" + (l(bestTheories)-1) + "] " : "") +  first(bestTheories));
    }));
  }
  
  S renderTheoryScore(Theory t) {
    //ret renderPosNegCounts(t.examples);
    ret t == null || t.examples.isEmpty() ? "" : iround(adjustConfidence(threeB1BScore(t.examples))) + "%"
      + " / " + renderPosNegScoreAndCount(t.examples);
  }

  int theoryScore(Theory t) {
    ret t == null ? -100 : t.examples.score();
  }

  void theoriesChanged {
    updateTheoryTable();
    updateLabelsTable();
    updateTabs();
    updatePredictions();
    change();
  }

  visual
    withCenteredButtons(super,
      "Show random msg", rInThinkQ(r showRandomMsg),
      jPopDownButton_noText(flattenObjectArray(
        "Check theories", rInThinkQ(r checkAllTheories),
        "Clear theories", rInThinkQ(r clearTheories),
        "Update predictions", rInThinkQ(r updatePredictions),
        dm_importAndExportAllDataMenuItems())));

  Runnable rInThinkQ(Runnable r) { ret rInQ(thinkQ, r); }

  void addTheory(Theory theory) {
    if (theories.add(theory)) {
      print("New theory: " + theory);
      addTheoryToCollectors(theory);
      theoriesChanged();
    }
  }

  void clearTheories { theories.clear(); theoriesChanged(); }

  Bool checkMsgProp(O prop, Msg msg) {
    if (prop cast And) ret checkMsgProp(prop.a, msg) && checkMsgProp(prop.b, msg);
    if (prop cast Not) ret not(checkMsgProp(prop.a, msg));
    ret ((MsgProp) prop).check(msg);
  }

  Bool evalTheoryLHS(Theory theory, Msg msg) {
    ret theory == null ? null
      : checkMsgProp(theory.statement.lhs, msg);
  }

  Bool testTheoryOnMsg(Theory theory, Msg msg) {
    Bool lhs = evalTheoryLHS(theory, msg);
    Bool rhs = checkMsgProp(theory.statement.rhs, msg);
    if (lhs == null || rhs == null) null;
    if (bidiMode)
      ret eq(lhs, rhs);
    else
      ret isTrue(rhs) || isFalse(lhs);
  }

  void checkAllTheories {
    for (Theory theory : theories)
      checkTheory_noTrigger(theory);
    theoriesChanged();
  }

  void checkTheory(Theory theory) {
    checkTheory_noTrigger(theory);
    theoriesChanged();
  }

  void checkTheory_noTrigger(Theory theory) {
    new PosNeg<Msg> pn;
    for (Msg msg : msgs)
      pn.add(msg, testTheoryOnMsg(theory, msg));
    if (!eq(theory.examples, pn)) {
      removeTheoryFromCollectors(theory);
      theory.examples = pn;
      addTheoryToCollectors(theory);
      change();
    }
  }

  S cleanLabel(S label) { ret upper(label); }
  
  S targetLabelOfTheory(Theory theory) {
    O o = theory.statement.rhs;
    if (o cast HasLabel) ret o.label;
    if (o cast DoesntHaveLabel) ret o.label;
    null;
  }

  void addTheoryToCollectors(Theory theory) {
    S lbl = targetLabelOfTheory(theory);
    if (lbl != null)
      labelByName(lbl).bestTheories.add(theory);
  }

  void removeTheoryFromCollectors(Theory theory) {
    S lbl = targetLabelOfTheory(theory);
    if (lbl != null)
      labelByName(lbl).bestTheories.remove(theory);
  }

  Label labelByName(S name) {
    ret getOrCreate(labelsByName, name, () -> new Label(name));
  }

  void updateLabelsByName() {
    for (S lbl : allLabels)
      labelByName(lbl);
    for (Theory t : theories)
      addTheoryToCollectors(t);
  }

  O getMsgFeature(Msg msg, S feature) {
    ret msg2features.get(msg).get(feature);
  }

  void makeTextExtractors(S textFeature) {
    for (WithName<IF1<S, O>> f : textExtractors()) {
      IF1<S, O> theFunction = f!;
      featureExtractors.put(f.name, env -> theFunction.get((S) env.getFeature(textFeature)));
    }
  }

  L<WithName<IF1<S, O>>> textExtractors() {
    new L<WithName<IF1<S, O>>> l;
    l.add(WithName<>("number of words", lambda1 numberOfWords));
    l.add(WithName<>("number of characters", lambda1 l));
    for (char c : characters("\"', .-_"))
      l.add(WithName<>("contains " + quote(c), s -> contains(s, c)));
    ret l;
  }
}