Model is stuck at 0.51-0.52 accuracy and not improving

Question

This is for a kaggle competition wherein I have essays which I have to grade/predict on 6 parameters (vocabulary, cohesion, conventions, grammar, phraseology, syntax). I have implemented RandomSearch hypermodel to get better results but the model is peaking at 0.52 accuracy and is giving terrible predictions.

Following is the X and y for the model: (y=1:7, X=7:

Following is the model_builder:

def model_builder(hp):
  model = tf.keras.Sequential()
#   model_input = tf.keras.layers.Input(shape=(6,))
    
  model.add(Dense(6,input_shape=(6,), activation = "relu"))
#   hp_units = random.randrange(32,512,32)

  units1 = random.randrange(32,512,32)  
  model.add(tf.keras.layers.Dense(units=units1, activation='relu'))
  model.add(tf.keras.layers.Dropout(0.3))
  
  units2 = random.randrange(32,512,32)  
  model.add(tf.keras.layers.Dense(units=units2, activation='relu'))
  model.add(tf.keras.layers.Dropout(0.3))

  units3 = random.randrange(32,512,32)  
  model.add(tf.keras.layers.Dense(units=units3, activation='relu'))
  model.add(tf.keras.layers.Dropout(0.4))

#   units4 = random.randrange(32,512,32)  
#   model.add(tf.keras.layers.Dense(units=units4, activation='LeakyReLU'))  
#   model.add(tf.keras.layers.Dropout(0.4))

#   units5 = random.randrange(32,512,32)  
#   model.add(tf.keras.layers.Dense(units=units5, activation='LeakyReLU'))  
#   model.add(tf.keras.layers.Dropout(0.5))

#   units6 = random.randrange(32,512,32)  
#   model.add(tf.keras.layers.Dense(units=units6, activation='LeakyReLU'))  
#   model.add(tf.keras.layers.Dropout(0.4))
    
  units7 = random.randrange(32,512,32)  
  model.add(tf.keras.layers.Dense(units=units7, activation='relu'))  
  model.add(tf.keras.layers.Dropout(0.3))

  units8 = random.randrange(32,512,32)  
  model.add(tf.keras.layers.Dense(units=units8, activation='relu'))  
  model.add(tf.keras.layers.Dropout(0.2))
    
  model.add(tf.keras.layers.Dense(6))

  # Tune the learning rate for the optimizer
  # Choose an optimal value from 0.01, 0.001, or 0.0001
  hp_learning_rate = hp.Choice('learning_rate', values=[1e-2, 1e-3, 1e-4])
  hp_momentum = hp.Choice('momentum', values = [1e-1, 3e-1, 5e-1, 7e-1, 9e-1])

  model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=hp_learning_rate,
                                                   momentum=hp_momentum,
                                                  ),
                loss=tf.keras.losses.MeanSquaredError(),
                metrics=['accuracy'])

  return model


#Following is the tuner and search code:


tuner = kt.RandomSearch(
        hypermodel=model_builder,
        objective='val_accuracy',
        max_trials=10,
    #     seed=None,
    #     hyperparameters=None,
    #     tune_new_entries=True,
    #     allow_new_entries=True,
    #     **kwargs
    )

early_stopping_monitor = EarlyStopping(monitor='val_accuracy', patience = 10, restore_best_weights=True)

tuner.search(X_train,
             y_train, 
             epochs=80, 
             batch_size=16,
             validation_data=(X_test,y_test), 
             callbacks=[early_stopping_monitor])

# Get the optimal hyperparameters
best_hps=tuner.get_best_hyperparameters(num_trials=1)[0]

print(f"""
The hyperparameter search is complete. The optimal learning rate for the optimizer
is {best_hps.get('learning_rate')}.
""")

History Plot:

Test vs train accuracy plot:

I have tried feature extraction for better correlation but it's not great. I tried to use different optimizers, activations, etc but none prevailed.

My last shot at this is probably to scale the data to 0-1 and use sigmoid on it.