100 NLP Questions
100 NLP Questions
Steve Nouri
Q1. Which of the following techniques can be used for keyword normalization in
NLP, the process of converting a keyword into its base form?
a. Lemmatization
b. Soundex
c. Cosine Similarity
d. N-grams
Answer : a) Lemmatization helps to get to the base form of a word, e.g. are playing -> play, eating
-> eat, etc.Other options are meant for different purposes.
Q2. Which of the following techniques can be used to compute the distance
between two word vectors in NLP?
a. Lemmatization
b. Euclidean distance
c. Cosine Similarity
d. N-grams
Answer : b) and c)
Distance between two word vectors can be computed using Cosine similarity and Euclidean
Distance. Cosine Similarity establishes a cosine angle between the vector of two words. A cosine
angle close to each other between two word vectors indicates the words are similar and vice a
versa.
E.g. cosine angle between two words “Football” and “Cricket” will be closer to 1 as compared to
angle between the words “Football” and “New Delhi”
Answer : e)All of the above can be used as features of the text corpus.
Q4. You created a document term matrix on the input data of 20K documents for a
Machine learning model. Which of the following can be used to reduce the
dimensions of data?
a. only 1
b. 2, 3
c. 1, 3
d. 1, 2, 3
Answer : d)
Q5. Which of the text parsing techniques can be used for noun phrase detection,
verb phrase detection, subject detection, and object detection in NLP.
a. Part of speech tagging
b. Skip Gram and N-Gram extraction
c. Continuous Bag of Words
d. Dependency Parsing and Constituency Parsing
Answer : d)
Q6. Dissimilarity between words expressed using cosine similarity will have values
significantly higher than 0.5
a. True
b. False
Answer : a)
Q7. Which one of the following are keyword Normalization techniques in NLP
a. Stemming
b. Part of Speech
c. Named entity recognition
d. Lemmatization
Answer : a) and d)
Part of Speech (POS) and Named Entity Recognition(NER) are not keyword Normalization
techniques. Named Entity help you extract Organization, Time, Date, City, etc..type of entities
from the given sentence, whereas Part of Speech helps you extract Noun, Verb, Pronoun,
adjective, etc..from the given sentence tokens.
Answer : (c)
formula for TF is K/T
formula for IDF is log(total docs / no of docs containing “data”)
= log(1 / (⅓))
= log (3)
Hence correct choice is Klog(3)/T
Q10. In NLP, The algorithm decreases the weight for commonly used words and
increases the weight for words that are not used very much in a collection of
documents
a. Term Frequency (TF)
b. Inverse Document Frequency (IDF)
c. Word2Vec
d. Latent Dirichlet Allocation (LDA)
Answer : b)
Q11. In NLP, The process of removing words like “and”, “is”, “a”, “an”, “the” from
a sentence is called as
a. Stemming
b. Lemmatization
c. Stop word
d. All of the above
Answer : c) In Lemmatization, all the stop words such as a, an, the, etc.. are removed. One can
also define custom stop words for removal.
Answer : b) The statement describes the process of tokenization and not stemming, hence it is
False.
Q13. In NLP, Tokens are converted into numbers before giving to any Neural
Network
a. True
b. False
Answer : a) In NLP, all words are converted into a number before feeding to a Neural Network.
Answer : d) All the ones mentioned are NLP libraries except BERT, which is a word embedding
Answer : b) TF-IDF helps to establish how important a particular word is in the context of the
document corpus. TF-IDF takes into account the number of times the word appears in the
document and offset by the number of documents that appear in the corpus.
● TF is the frequency of term divided by a total number of terms in the document.
● IDF is obtained by dividing the total number of documents by the number of documents
containing the term and then taking the logarithm of that quotient.
Answer : d)
Q18 In text mining, converting text into tokens and then converting them into an
integer or floating-point vectors can be done using
a. CountVectorizer
b. TF-IDF
c. Bag of Words
d. NERs
Answer : a) CountVectorizer helps do the above, while others are not applicable.
text =[“Rahul is an avid writer, he enjoys studying understanding and presenting. He loves to
play”]
vectorizer = CountVectorizer()
vectorizer.fit(text)
vector = vectorizer.transform(text)
Q19. In NLP, Words represented as vectors are called as Neural Word Embeddings
a. True
b. False
Answer : a) Word2Vec, GloVe based models build word embedding vectors that are
multidimensional.
Q20. In NLP, Context modeling is supported with which one of the following word
embeddings
1. a. Word2Vec
2. b) GloVe
3. c) BERT
4. d) All of the above
Answer : c) Only BERT (Bidirectional Encoder Representations from Transformer) supports
context modelling where the previous and next sentence context is taken into consideration. In
Word2Vec, GloVe only word embeddings are considered and previous and next sentence context
is not considered.
Q22. Which one of the following Word embeddings can be custom trained for a
specific subject in NLP
a. Word2Vec
b. BERT
Answer : b) BERT allows Transform Learning on the existing pre-trained models and hence can
be custom trained for the given specific subject, unlike Word2Vec and GloVe where existing word
embeddings can be used, no transfer learning on text is possible.
Q23. Word embeddings capture multiple dimensions of data and are represented
as vectors
a. True
b. False
Answer : a)
Q24. In NLP, Word embedding vectors help establish distance between two tokens
a. True
b. False
Answer : a) One can use Cosine similarity to establish distance between two vectors represented
through Word Embeddings
Q25. Language Biases are introduced due to historical data used during training of
word embeddings, which one amongst the below is not an example of bias
a. New Delhi is to India, Beijing is to China
b. Man is to Computer, Woman is to Homemaker
Answer : a)
Statement b) is a bias as it buckets Woman into Homemaker, whereas statement a) is not a
biased statement.
Q26. Which of the following will be a better choice to address NLP use cases such
as semantic similarity, reading comprehension, and common sense reasoning
a. ELMo
b. Open AI’s GPT
c. ULMFit
Answer : b) Open AI’s GPT is able to learn complex pattern in data by using the Transformer
models Attention mechanism and hence is more suited for complex use cases such as semantic
similarity, reading comprehensions, and common sense reasoning.
Answer : c) ULMFit has an LSTM based Language modeling architecture. This got replaced into
Transformer architecture with Open AI’s GPT
Q28. Which of the following architecture can be trained faster and needs less
amount of training data
a. LSTM based Language Modelling
b. Transformer architecture
Answer : b) Transformer architectures were supported from GPT onwards and were faster to
train and needed less amount of data for training too.
Q29. Same word can have multiple word embeddings possible with ____________?
a. GloVe
b. Word2Vec
c. ELMo
d. nltk
Answer : c) EMLo word embeddings supports same word with multiple embeddings, this helps
in using the same word in a different context and thus captures the context than just meaning of
the word unlike in GloVe and Word2Vec. Nltk is not a word embedding.
Q30 For a given token, its input representation is the sum of embedding from the
token, segment and position embedding
a. ELMo
b. GPT
c. BERT
d. ULMFit
Q31. Trains two independent LSTM language model left to right and right to left and
shallowly concatenates them
a. GPT
b. BERT
c. ULMFit
d. ELMo
Answer : b) GPT is a unidirectional model and word embedding are produced by training on
information flow from left to right. ELMo is bidirectional but shallow. Word2Vec provides simple
word embedding.
Answer : c)BERT Transformer architecture models the relationship between each word and all
other words in the sentence to generate attention scores. These attention scores are later used
as weights for a weighted average of all words’ representations which is fed into a fully-connected
network to generate a new representation.
Q35. Transformer model pays attention to the most important word in Sentence
a. True
b. False
Q36. Which NLP model gives the best accuracy amongst the following?
a. BERT
b. XLNET
c. GPT-2
d. ELMo
Answer : b) XLNET has given best accuracy amongst all the models. It has outperformed BERT
on 20 tasks and achieves state of art results on 18 tasks including sentiment analysis, question
answering, natural language inference, etc.
Answer : d) XLNET provides permutation-based language modelling and is a key difference from
BERT. In permutation language modeling, tokens are predicted in a random manner and not
sequential. The order of prediction is not necessarily left to right and can be right to left. The
original order of words is not changed but a prediction can be random.
The conceptual difference between BERT and XLNET can be seen from the following diagram.
Q39. What is Naive Bayes algorithm, When we can use this algorithm in NLP?
Naive Bayes algorithm is a collection of classifiers which works on the principles of the Bayes’
theorem. This series of NLP model forms a family of algorithms that can be used for a wide range
of classification tasks including sentiment prediction, filtering of spam, classifying documents and
more.
Naive Bayes algorithm converges faster and requires less training data. Compared to other
discriminative models like logistic regression, Naive Bayes model it takes lesser time to train. This
algorithm is perfect for use while working with multiple classes and text classification where the
data is dynamic and changes frequently.
Q63. There have some various common elements of natural language processing.
Those elements are very important for understanding NLP properly, can you please
explain the same in details with an example?
Answer:
There have a lot of components normally using by natural language processing (NLP). Some of
the major components are explained below:
● Extraction of Entity: It actually identifying and extracting some critical data from the
available information which help to segmentation of provided sentence on identifying each
entity. It can help in identifying one human that it’s fictional or real, same kind of reality
identification for any organization, events or any geographic location etc.
● The analysis in a syntactic way: it mainly helps for maintaining ordering properly of the
available words.
Q64 In the case of processing natural language, we normally mentioned one
common terminology NLP and binding every language with the same terminology
properly. Please explain in details about this NLP terminology with an example?
Answer:
This is the basic NLP Interview Questions asked in an interview. There have some several factors
available in case of explaining natural language processing. Some of the key factors are given
below:
● Vectors and Weights: Google Word vectors, length of TF-IDF, varieties documents, word
vectors, TF-IDF.
● Structure of Text: Named Entities, tagging of part of speech, identifying the head of the
sentence.
● Analysis of sentiment: Know about the features of sentiment, entities available for the
sentiment, sentiment common dictionary.
● Classification of Text: Learning supervising, set off a train, set of validation in Dev, Set of
define test, a feature of the individual text, LDA.
● Reading of Machine Language: Extraction of the possible entity, linking with an individual
entity, DBpedia, some libraries like Pikes or FRED.
Q68 How does the encoder-decoder structure work for language modelling?
The encoder-decoder structure is a deep learning model architecture responsible for several state
of the art solutions, including Machine Translation.
The input sequence is passed to the encoder where it is transformed to a fixed-dimensional vector
representation using a neural network. The transformed input is then decoded using another
neural network. Then, these outputs undergo another transformation and a softmax layer. The
final output is a vector of probabilities over the vocabularies. Meaningful information is extracted
based on these probabilities.
Q70 How would you implement an NLP system as a service, and what are some
pitfalls you might face in production?
This is less of a NLP question than a question for productionizing machine learning models. There
are however certain intricacies to NLP models.
Q72 Which of the following models can perform tweet classification with regards
to context mentioned above?
A) Naive Bayes
B) SVM
C) None of the above
Solution: (C)
Since, you are given only the data of tweets and no other information, which means there is no
target variable present. One cannot train a supervised learning model, both svm and naive bayes
are supervised learning techniques.
Q73 You have created a document term matrix of the data, treating every tweet as
one document. Which of the following is correct, in regards to document term
matrix?
1. Removal of stopwords from the data will affect the dimensionality of data
2. Normalization of words in the data will reduce the dimensionality of data
Q74 Which of the following features can be used for accuracy improvement of a
classification model?
A) Frequency count of terms
B) Vector Notation of sentence
C) Part of Speech Tag
D) Dependency Grammar
E) All of these
Solution: (E)
All of the techniques can be used for the purpose of engineering features in a model.
Q75 What percentage of the total statements are correct with regards to Topic
Modeling?
1. It is a supervised learning technique
2. LDA (Linear Discriminant Analysis) can be used to perform topic modeling
3. Selection of number of topics in a model does not depend on the size of data
4. Number of topic terms are directly proportional to size of the data
A) 0
B) 25
C) 50
D) 75
E) 100
Solution: (A)
LDA is unsupervised learning model, LDA is latent Dirichlet allocation, not Linear discriminant
analysis. Selection of the number of topics is directly proportional to the size of the data, while
number of topic terms is not directly proportional to the size of the data. Hence none of the
statements are correct.
Q78 What is the difference between learning latent features using SVD and getting
embedding vectors using deep network?
SVD uses linear combination of inputs while a neural network uses nonlinear combination.
Q79 What is the information in the hidden and cell state of LSTM?
Hidden stores all the information till that time step and cell state stores particular information that
might be needed in the future time step.
Number of parameters in an LSTM model with bias
4(𝑚h+h²+h) where 𝑚 is input vectors size and h is output vectors size a.k.a. hidden
The point to see here is that mh dictates the model size as m>>h. Hence it's important to have a
small vocab.
Time complexity of LSTM
seq_length*hidden²
Time complexity of transfomer
seq_length²*hidden
When hidden size is more than the seq_length(which is normally the case), transfomer is faster
than LSTM.
Q82 What’s the difference between hard and soft parameter sharing in multi-task
learning?
Hard sharing is where we train for all the task at the same time and update our weights using all
the losses whereas soft sharing is where we train for one task at a time.
Q85 Why does the transformer block have LayerNorm instead of BatchNorm?
Looking at the advantages of LayerNorm, it is robust to batch size and works better as it works at
the sample level and not batch level.
Q86 What changes would you make to your deep learning code if you knew there
are errors in your training data?
We can do label smoothening where the smoothening value is based on % error. If any particular
class has known error, we can also use class weights to modify the loss.
Q87 What are the tricks used in ULMFiT? (Not a great questions but checks the
awareness)
● LM tuning with task text
● Weight dropout
● Discriminative learning rates for layers
● Gradual unfreezing of layers
● Slanted triangular learning rate schedule
This can be followed up with a question on explaining how they help.
Q89 What are the differences between GPT and GPT-2? (From Lilian Weng)
Q92 How does parameter sharing in ALBERT affect the training and inference time?
Q93 How would you reduce the inference time of a trained NN model?
● Serve on GPU/TPU/FPGA
● 16 bit quantisation and served on GPU with fp16 support
● Pruning to reduce parameters
● Knowledge distillation (To a smaller transformer model or simple neural network)
● Hierarchical softmax/Adaptive softmax
● You can also cache results as explained here.
Of course! BPE is a smart tokeniser and it can help us get a smaller vocabulary which can help
us find a model with less parameters.
Q95 How would you make an arxiv papers search engine? (I was asked — How
would you make a plagiarism detector?)
Get top k results with TF-IDF similarity and then rank results with
● semantic encoding + cosine similarity
● a model trained for ranking
This is a trick question. The interviewee can say all things such as using transfer learning and
latest models but they need to talk about having a neutral class too otherwise you can have really
good accuracy/f1 and still, the model will classify everything into positive or negative.
The truth is that a lot of news is neutral and so the training needs to have this class. The
interviewee should also talk about how he will create a dataset and his training strategies like the
selection of language model, language model fine-tuning and using various datasets for multi-
task learning.
Q98 What is the difference between regular expression and regular grammar?
A regular expression is the representation of natural language in the form of mathematical
expressions containing a character sequence. On the other hand, regular grammar is the
generator of natural language, defining a set of defined rules and syntax which the strings in the
natural language must follow.