Revert "[omnibox][ml] Remove scoring signals for history fuzzy provid…

…er matches."

This reverts commit 6300f06.

Reason for revert: match.scoring_signals is null, causing crash

Original change's description:
> [omnibox][ml] Remove scoring signals for history fuzzy provider matches.
>
> The fuzzy provider works by running the bookmark and history quick
> providers with a "corrected" input and applying a penalty proportional
> to the correction amount to the results.  The scoring signals are
> populated according to the correct input and leads to an artificially
> inflated model output.
>
> We previously attempted to correct for this by re-applying the penalty
> after re-scoring.  This was not sufficient in demoting the fuzzy
> suggestions.
>
> This CL attempts to correct for this by clearing the
> scoring signals in the fuzzy provider, as these are not accurate for
> the actual input, resulting in fuzzy matches no longer being re-scored
> by the model.
>
> Bug: 1472374, 1405555
> Change-Id: I10bd5036179513a156c6e1ba15713ff1b12eb8fa
> Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/4775323
> Code-Coverage: findit-for-me@appspot.gserviceaccount.com <findit-for-me@appspot.gserviceaccount.com>
> Reviewed-by: manuk hovanesian <manukh@chromium.org>
> Commit-Queue: Angela Yoeurng <yoangela@chromium.org>
> Cr-Commit-Position: refs/heads/main@{#1184412}

Bug: 1472374, 1405555
Change-Id: Ib94480da1bc69bfce486aebec7bf041f02ff9323
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/4794136
Auto-Submit: Daniel Yip <danielyip@google.com>
Owners-Override: Daniel Yip <danielyip@google.com>
Commit-Queue: Daniel Yip <danielyip@google.com>
Reviewed-by: Srinivas Sista <srinivassista@chromium.org>
Cr-Commit-Position: refs/branch-heads/5956@{#7}
Cr-Branched-From: c219341-refs/heads/main@{#1185115}

components/omnibox/browser/autocomplete_controller.cc

@@ -1807,6 +1807,18 @@ void AutocompleteController::OnUrlScoringModelDone(
       match_itr->RecordAdditionalInfo("ml_legacy_relevance",
                                       match_itr->relevance);
       match_itr->relevance = relevance_heap.top();
+
+      // Fuzzy matches use the scoring signals for history and bookmark
+      // providers with the "corrected" input. This leads to an artificially
+      // high confidence from the model. Correct for this by re-applying the
+      // penalty from the history fuzzy provider.
+      if (match_itr->provider && match_itr->provider->type() ==
+                                     AutocompleteProvider::TYPE_HISTORY_FUZZY) {
+        match_itr->RecordAdditionalInfo("ml_relevance_before_penalty",
+                                        match_itr->relevance);
+        HistoryFuzzyProvider::ApplyRelevancePenalty(
+            *match_itr, match_itr->fuzzy_match_penalty);
+      }
     }
     relevance_heap.pop();
     prediction_and_match_itr_heap.pop();

components/omnibox/browser/autocomplete_match.cc

@@ -258,6 +258,7 @@ AutocompleteMatch::AutocompleteMatch(const AutocompleteMatch& match)
     : provider(match.provider),
       relevance(match.relevance),
       typed_count(match.typed_count),
+      fuzzy_match_penalty(match.fuzzy_match_penalty),
       deletable(match.deletable),
       fill_into_edit(match.fill_into_edit),
       additional_text(match.additional_text),
@@ -318,6 +319,7 @@ AutocompleteMatch& AutocompleteMatch::operator=(
   provider = std::move(match.provider);
   relevance = std::move(match.relevance);
   typed_count = std::move(match.typed_count);
+  fuzzy_match_penalty = std::move(match.fuzzy_match_penalty);
   deletable = std::move(match.deletable);
   fill_into_edit = std::move(match.fill_into_edit);
   additional_text = std::move(match.additional_text);
@@ -387,6 +389,7 @@ AutocompleteMatch& AutocompleteMatch::operator=(
   provider = match.provider;
   relevance = match.relevance;
   typed_count = match.typed_count;
+  fuzzy_match_penalty = match.fuzzy_match_penalty;
   deletable = match.deletable;
   fill_into_edit = match.fill_into_edit;
   additional_text = match.additional_text;

components/omnibox/browser/autocomplete_match.h

@@ -641,6 +641,10 @@ struct AutocompleteMatch {
   // set for matches from HistoryURL and HistoryQuickProvider.
   int typed_count = -1;
 
+  // The percentage deducted from the relevance score by the history fuzzy
+  // provider.  This is currently used to re-apply the penalty after ML scoring.
+  int fuzzy_match_penalty = 0;
+
   // True if the user should be able to delete this match.
   bool deletable = false;
 

components/omnibox/browser/history_fuzzy_provider.cc

@@ -567,6 +567,15 @@ size_t HistoryFuzzyProvider::EstimateMemoryUsage() const {
   return res;
 }
 
+// static
+void HistoryFuzzyProvider::ApplyRelevancePenalty(AutocompleteMatch& match,
+                                                 int penalty) {
+  DCHECK_GE(penalty, 0);
+  DCHECK_LE(penalty, 100);
+  match.relevance -= match.relevance * penalty / 100;
+  match.fuzzy_match_penalty = penalty;
+}
+
 HistoryFuzzyProvider::~HistoryFuzzyProvider() = default;
 
 void HistoryFuzzyProvider::DoAutocomplete() {
@@ -698,15 +707,7 @@ int HistoryFuzzyProvider::AddConvertedMatches(const ACMatches& matches,
 
   // Apply relevance penalty; all corrections are equal and we only apply this
   // to the most relevant result, so edit distance isn't needed.
-  DCHECK_GE(penalty, 0);
-  DCHECK_LE(penalty, 100);
-  match.relevance -= match.relevance * penalty / 100;
-
-  // Scoring signals are calculated in the history and bookmark providers using
-  // the corrected input. These scoring signals are inaccurate for the true
-  // input, so clear them to prevent the ml model assigning an
-  // artificially high confidence to this suggestion.
-  match.scoring_signals->Clear();
+  ApplyRelevancePenalty(match, penalty);
 
   return 1;
 }

components/omnibox/browser/history_fuzzy_provider.h

@@ -207,6 +207,11 @@ class HistoryFuzzyProvider : public HistoryProvider,
   // See base/trace_event/memory_usage_estimator.h for more info.
   size_t EstimateMemoryUsage() const override;
 
+  // Reduces a match's relevance score according to the penalty and records the
+  // penalty amount.  `penalty` is the percentage that will be deducted from the
+  // match's relevance.
+  static void ApplyRelevancePenalty(AutocompleteMatch& match, int penalty);
+
  private:
   ~HistoryFuzzyProvider() override;