diff --git a/cpp/src/arrow/compute/kernels/scalar_string_test.cc b/cpp/src/arrow/compute/kernels/scalar_string_test.cc
index 4969beada682..c56546f7966b 100644
--- a/cpp/src/arrow/compute/kernels/scalar_string_test.cc
+++ b/cpp/src/arrow/compute/kernels/scalar_string_test.cc
@@ -1245,7 +1245,7 @@ TYPED_TEST(TestStringKernels, Utf8Normalize) {
 
   // decomposed: U+0061(LATIN SMALL LETTER A) + U+0301(COMBINING ACUTE ACCENT)
   // composed: U+00E1(LATIN SMALL LETTER A WITH ACUTE)
-  const char* json_composed = "[\"foo\", \"á\"]";
+  const char* json_composed = "[\"foo\", \"\xc3\xa1\"]";
   const char* json_decomposed = "[\"foo\", \"a\xcc\x81\"]";
   for (const auto& options : compose_options) {
     this->CheckUnary("utf8_normalize", json_decomposed, this->type(), json_composed,
@@ -1260,6 +1260,25 @@ TYPED_TEST(TestStringKernels, Utf8Normalize) {
                      &options);
   }
 
+  // Hangul composes algorithmically in utf8proc, not through the composition table.
+  // decomposed: U+1112(HANGUL CHOSEONG HIEUH) + U+1161(HANGUL JUNGSEONG A) +
+  //             U+11AB(HANGUL JONGSEONG NIEUN)
+  // composed: U+D55C(HANGUL SYLLABLE HAN)
+  json_composed = "[\"\xed\x95\x9c\"]";
+  json_decomposed = "[\"\xe1\x84\x92\xe1\x85\xa1\xe1\x86\xab\"]";
+  for (const auto& options : compose_options) {
+    this->CheckUnary("utf8_normalize", json_decomposed, this->type(), json_composed,
+                     &options);
+    this->CheckUnary("utf8_normalize", json_composed, this->type(), json_composed,
+                     &options);
+  }
+  for (const auto& options : decompose_options) {
+    this->CheckUnary("utf8_normalize", json_composed, this->type(), json_decomposed,
+                     &options);
+    this->CheckUnary("utf8_normalize", json_decomposed, this->type(), json_decomposed,
+                     &options);
+  }
+
   // canonical: U+00B2(Superscript Two)
   // compatibility: "2"
   const char* json_canonical = "[\"01\xc2\xb2!\"]";
diff --git a/cpp/src/arrow/compute/kernels/scalar_string_utf8.cc b/cpp/src/arrow/compute/kernels/scalar_string_utf8.cc
index 582d559dde2e..6d4525c28696 100644
--- a/cpp/src/arrow/compute/kernels/scalar_string_utf8.cc
+++ b/cpp/src/arrow/compute/kernels/scalar_string_utf8.cc
@@ -547,6 +547,16 @@ struct Utf8NormalizeBase {
     if (res < 0) {
       return Status::Invalid("Cannot normalize utf8 string: ", utf8proc_errmsg(res));
     }
+    if (decompose_options_ & UTF8PROC_COMPOSE) {
+      // utf8proc_decompose() only decomposes; the canonical composition step for
+      // NFC and NFKC is done in-place by utf8proc_normalize_utf32().
+      res = utf8proc_normalize_utf32(
+          reinterpret_cast<utf8proc_int32_t*>(codepoints_.data()), res,
+          decompose_options_);
+      if (res < 0) {
+        return Status::Invalid("Cannot normalize utf8 string: ", utf8proc_errmsg(res));
+      }
+    }
     return res;
   }
 
diff --git a/python/pyarrow/tests/test_compute.py b/python/pyarrow/tests/test_compute.py
index 8b2ad2b333fc..0548be7c9740 100644
--- a/python/pyarrow/tests/test_compute.py
+++ b/python/pyarrow/tests/test_compute.py
@@ -3825,6 +3825,15 @@ def test_utf8_normalize():
     assert pc.utf8_normalize(arr, form="NFKC") == pa.array(["0123"])
     assert pc.utf8_normalize(arr, "NFD") == arr
     assert pc.utf8_normalize(arr, "NFKD") == pa.array(["0123"])
+    # GH-51225: composing forms must compose, not only decompose
+    composed = pa.array(["\u00e9", "\ud55c", None])
+    decomposed = pa.array(["e\u0301", "\u1112\u1161\u11ab", None])
+    for form in ("NFC", "NFKC"):
+        assert pc.utf8_normalize(decomposed, form=form) == composed
+        assert pc.utf8_normalize(composed, form=form) == composed
+    for form in ("NFD", "NFKD"):
+        assert pc.utf8_normalize(composed, form=form) == decomposed
+        assert pc.utf8_normalize(decomposed, form=form) == decomposed
     with pytest.raises(
             ValueError,
             match='"NFZ" is not a valid Unicode normalization form'):