Tolkien character or prescription drug name?

Classification using character-level Long Short-Term Memory (LSTM) neural networks

Table of Contents

After trying to read J.R.R. Tolkien’s The Silmarillion again for the millionth time, I remembered a funny tweet that has been around for a while:

Even though I was a casual fan of The Lord of the Rings and having already taken two pharmacology courses in college, I had no idea who or what a Narmacil was. Should we fear him/her by its sword skills or by its dangerous side effects?

This little trivia prompted me to ask if an artificial neural network (ANN) could succeed where I and many more have failed. Here, I show you how to build a special type of ANN called Long Short-Term Memory (LSTM) to classify Tolkien characters and prescription drug names using Keras.



The first step was to build from scratch a combined dataset with names of Tolkien characters and prescription drugs (a bunch of them, not just antidepressants).

Tolkien characters

Lucky for us, the Behind the Name website has a database of the first names of Tolkien characters that we can directly read from the page’s HTML using pandas.

import pandas as pd

raw_tolkien_chars = pd.read_html('')

0Adalbertm1 character1
1Adaldridaf1 character1
2Adalgarm1 character1
3Adalgrimm1 character1
4Adamantaf1 character1
tolkien_names = raw_tolkien_chars[2]['Name']
350           Gethron
351    Ghân-buri-Ghân
352            Gildis
353            Gildor
354         Gil-galad
Name: Name, dtype: object

We can see that some names are hyphenated and have accented letters. To simplify the analysis I transformed unicode characters to ASCII, removed punctuation marks, transformed them to lowercase, and removed any possible duplicates.

import unidecode

processed_tolkien_names = tolkien_names.apply(unidecode.unidecode).str.lower().str.replace('-', ' ')
processed_tolkien_names = [name[0] for name in processed_tolkien_names.str.split()]
processed_tolkien_names = pd.DataFrame(processed_tolkien_names, columns=['name']).sort_values('name').drop_duplicates()
processed_tolkien_names['tolkien'] = 1
473    gethron
439       ghan
109        gil
341     gildis
324     gildor
Name: name, dtype: object

Done! Now we have 746 different character names.

Prescription drugs

To get a comprehensive list of drug names, I downloaded the medication guide of the U.S. Food & Drug Administration (FDA).

raw_medication_guide = pd.read_csv('data/raw/medication_guides.csv')

Drug NameActive IngredientForm;RouteAppl. No.CompanyDateLink
drug_names = raw_medication_guide['Drug Name']
160                                             Chantix
161         Children's Cetirizine Hydrochloride Allergy
162    Chlordiazepoxide and Amitriptyline Hydrochloride
163                                              Cimzia
164                                              Cimzia
Name: Drug Name, dtype: object

A similar preprocessing step was repeated for this dataset too:

processed_drug_names = drug_names.str.lower().str.replace('.', '').str.replace( '-', ' ').str.replace('/', ' ').str.replace("'", ' ').str.replace(",", ' ')
processed_drug_names = [name[0] for name in processed_drug_names.str.split()]
processed_drug_names = pd.DataFrame(processed_drug_names, columns=['name']).sort_values('name').drop_duplicates()
processed_drug_names['tolkien'] = 0
373             chantix
448            children
395    chlordiazepoxide
185              cimzia
292               cipro
Name: name, dtype: object

Done, 611 different drug names!

We can finally combine the two datasets and move on.

dataset = pd.concat([processed_tolkien_names, processed_drug_names], ignore_index=True)

Data transformation

So now we have a bunch of names, but machine learning models don’t work with raw characters. We need to convert them into a numerical format that can be processed by our soon-to-be-built model.

Using the Tokenizer class from Keras, we set char_level=True to process each word at character-level. The fit_on_texts() method will update the tokenizer internal vocabulary based on our dataset names and then texts_to_sequences() will transform each name into a sequence of integers.

from keras.preprocessing.text import Tokenizer
tokenizer = Tokenizer(char_level=True)
char_index = tokenizer.texts_to_sequences(dataset['name'])

Look how our beloved Bilbo is now:

[16, 3, 6, 16, 5]

Yet, this representation is not ideal. Having integers to represent letters could lead the ANN to assume that the characters have an ordinal scale. To solve this problem we have to:

  1. Set all names to have the length of the longest name (17 characters here). We use pad_sequences to add 0’s to the end of names shorter than 17 letters.

  2. Convert each integer representation to its one-hot encoded vector representation. The vector consists of 0s in all cells except for a single 1 in a cell to identify the letter.

from keras.preprocessing.sequence import pad_sequences
from keras.utils import to_categorical
import numpy as np

char_index = pad_sequences(char_index, maxlen=dataset['name'].apply(len).max(), padding="post")
x = to_categorical(char_index)  # onehot encoding
y = np.array(dataset['tolkien'])
(1357, 17, 27)

We have 1357 names. Each name has 17 letters and each letter is a one-hot encoded vector of size 27 (26 letters of the Latin alphabet + padding character).

Data split

I split the data into train, validation, and test sets with a 60/20/20 ratio using a custom function since sklearn train_test_split only outputs two sets.

from sklearn.model_selection import train_test_split

def data_split(data, labels, train_ratio=0.5, rand_seed=42):

    x_train, x_temp, y_train, y_temp = train_test_split(data,

    x_val, x_test, y_val, y_test = train_test_split(x_temp,

    return x_train, x_val, x_test, y_train, y_val, y_test

x_train, x_val, x_test, y_train, y_val, y_test = data_split(x, y, train_ratio=0.6)

Let’s take a look at the splits:

from collections import Counter
import matplotlib.pyplot as plt

dataset_count = pd.DataFrame([Counter(y_train), Counter(y_val), Counter(y_test)],
                                index=["train", "val", "test"])

print(f"Total number of samples: \n{dataset_count.sum(axis=0).sum()}")
print(f"Class/Samples: \n{dataset_count.sum(axis=0)}")
print(f"Split/Class/Samples: \n{dataset_count}")


Total number of samples: 
1    746
0    611
dtype: int64
        1    0
train  451  363
val    149  122
test   146  126

There are more Tolkien characters than drug names, but it seems like a decent balance.

LSTM model

Long Short-Term Memory is a type of Recurrent Neural Network proposed by Hochreiter S. & Schmidhuber J. (1997) to store information over extended time intervals. Names are just sequences of characters in which the order is important, so LSTM networks are a great choice for our name prediction task. You can read more about LSTMs in this awesome illustrated guide written by Michael Phi.


Keras was used to build this simple LSTM model after some tests and hyperparameter tuning. It is just a hidden layer with 8 LSTM blocks, one dropout layer to prevent overfitting, and one output neuron with a sigmoid activation function to make a binary classification.

from tensorflow.keras import Sequential
from tensorflow.keras.layers import Dense, Activation, Dropout, LSTM
from tensorflow.random import set_seed

model = Sequential()
model.add(LSTM(8, return_sequences=False,
               input_shape=(x.shape[1], x.shape[2])))
Model: "sequential"
Layer (type)                 Output Shape              Param #   
lstm (LSTM)                  (None, 8)                 1152      
dropout (Dropout)            (None, 8)                 0         
dense (Dense)                (None, 1)                 9         
activation (Activation)      (None, 1)                 0         

Total params: 1,161
Trainable params: 1,161
Non-trainable params: 0

Adam is a good default optimizer and produces great results in deep learning applications. Binary cross-entropy is the default loss function to binary classification problems and it is compatible with our single neuron output architecture.

from tensorflow.keras.optimizers import Adam

              optimizer=Adam(learning_rate=1e-3), metrics=['accuracy'])

Two callbacks were implemented. EarlyStopping to stop the training process after 20 epochs without reducing the validation loss and ModelCheckpoint to always save the model when the validation loss drops.

from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint

es = EarlyStopping(monitor='val_loss', verbose=1, patience=20)
mc = ModelCheckpoint("best_model.h5", monitor='val_loss',
                     verbose=1, save_best_only=True)

history =, y_train, batch_size=32, epochs=100,
                    validation_data=(x_val, y_val), callbacks=[es, mc])
Epoch 00071: val_loss did not improve from 0.34949
Epoch 72/100
26/26 [==============================] - 1s 24ms/step - loss: 0.3085 - accuracy: 0.8836 - val_loss: 0.3861 - val_accuracy: 0.8487

Epoch 00072: val_loss did not improve from 0.34949
Epoch 00072: early stopping
val_loss_per_epoch = history.history['val_loss']
best_epoch = val_loss_per_epoch.index(min(val_loss_per_epoch)) + 1
print(f"Best epoch: {best_epoch}")
Best epoch: 52

Let’s plot the accuracy and loss values per epoch to see the progression of these metrics.

def plot_metrics(history):
    plt.plot(history.history['accuracy'], label='Training')
    plt.plot(history.history['val_accuracy'], label='Validation')
    plt.legend(loc='lower right')

    plt.plot(history.history['loss'], label='Training')
    plt.plot(history.history['val_loss'], label='Validation')
    plt.legend(loc='upper right')



We can see that the accuracy quickly reaches a good plateau around 80%. Visually the model appears to start overfitting after epoch 50. It shouldn’t be a problem to use the version saved by ModelCheckpoint at epoch 52.

Performance evaluation

Finally, let’s see how our model does with the test dataset.

from tensorflow.keras.models import load_model

model = load_model("best_model.h5")
metrics = model.evaluate(x=x_test, y=y_test)
9/9 [==============================] - 1s 7ms/step - loss: 0.4595 - accuracy: 0.8125
print("Accuracy: {0:.2f} %".format(metrics[1]*100))
Accuracy: 81.25 %

81.25 %

Not bad!


We can explore the results a little more with the confusion matrix and classification report:

from sklearn.metrics import confusion_matrix
from seaborn import heatmap

def plot_confusion_matrix(y_true, y_pred, labels):
    cm = confusion_matrix(y_true, y_pred)
    heatmap(cm, annot=True, fmt="d", cmap="rocket_r", xticklabels=labels, yticklabels=labels)
    plt.ylabel('True label')
    plt.xlabel('Predicted label')

predictions = model.predict(x_test)
threshold = 0.5
y_pred = predictions > threshold
plot_confusion_matrix(y_test, y_pred, labels=['Drug','Tolkien'])


from sklearn.metrics import classification_report

print(classification_report(y_test, y_pred, target_names=['Drug', 'Tolkien']))
              precision    recall  f1-score   support

        Drug       0.80      0.80      0.80       126
     Tolkien       0.83      0.82      0.82       146

    accuracy                           0.81       272
   macro avg       0.81      0.81      0.81       272
weighted avg       0.81      0.81      0.81       272

Besides the good accuracy, the model has almost the same number of false positives and false negatives. We can see this reflecting in a balanced precision and recall.

So, just out of curiosity, which Tolkien characters could pass as prescription drugs?

def onehot_to_text(onehot_word):
    """Reverse one-hot encoded words to strings"""

    char_index = [[np.argmax(char) for char in onehot_word]]
    word = tokenizer.sequences_to_texts(char_index)
    return ''.join(word[0].split())
test_result = pd.DataFrame()
test_result['true'] = y_test
test_result['prediction'] = y_pred.astype(int)
test_result['name'] = [onehot_to_text(name) for name in x_test]

test_result['name'].loc[(test_result['true']==1) & (test_result['prediction']==0)]
13              ivy
17         camellia
44      celebrindor
47         meriadoc
63        vanimelde
64        finduilas
75        eglantine
84             ruby
87            poppy
89             otto
100           tanta
102          myrtle
108          prisca
132          cottar
151          stybba
171            este
175           daisy
189          tulkas
195        arciryas
205        odovacar
206          tarcil
207    hyarmendacil
229            jago
230            tata
240           ponto
271       landroval
Name: name, dtype: object


So, here we covered how to work with character embeddings and build a simple LSTM model capable of telling apart Tolkien character names from prescription drug names. Full code, including requirements, dataset, a Jupyter Notebook code version, and a script version, can be found at my GitHub repo.

You can also play around with this popular interactive quiz found on the web: Antidepressant or Tolkien?. I only got 70.8% right! Can you guess better than the LSTM network?



Hu, Y., Hu, C., Tran, T., Kasturi, T., Joseph, E., & Gillingham, M. (2021). What’s in a Name?–Gender Classification of Names with Character Based Machine Learning Models. arXiv preprint arXiv:2102.03692.

Bhagvati, C. (2018). Word representations for gender classification using deep learning. Procedia computer science, 132, 614-622.

Liang, X. (2018). How to Preprocess Character Level Text with Keras.

Guilherme Bauer Negrini
Guilherme Bauer Negrini
Biomedical Scientist and Researcher | Software Developer

Neuroscience, bioinformatics and biomedical data analysis.