Description
For Project-2, you will implement a feature-driven model that extracts named entities (NEs) from a document. The goal is
to extract relevant concepts from a document such as names of persons, locations, geo-political entities, etc. To motivate the
idea, consider the document given below:
To train your NER model, you are provided with the CoNLL-2003 English dataset [1]. This corpus describes four kinds
of NE’s: Locations (LOC), Organizations (ORG), Persons (PER) and Miscellaneous (MISC). The figure given below is a
snapshot of how data is arranged in the corpus:
1
As you can see in the figure, each new line in the corpus represents a token and its associated label. This corpus follows the
BIO notation for representing labels where B marks the beginning of a tag, I marks the continuation of a previous tag and
O marks that the given token is a named entity. Also note that a new line marks the beginning of a new sentence in the
document.
Task – 1: Data preprocessing (5 points)
As an initial pre-processing step, carry out the following:
• Extract all sentences, tokens and the associated NER tags from the corpus. Identify how many sentences, unique tokens
and tags are present in this corpus.
• Convert all tokens to lowercase.
• Replace each tag by a unique identifier integer.
Task – 2: Feature engineering (15 points)
Additionally, for each token, extract its lemma and POS tag by consulting an API like NLTK [2]. Note that to extract these
features, you must pass the entire sentence as an argument to the associated functions as the POS tag of a word depends on
its context. For example, the POS tag of the word ‘race’ depends on the context in which it is used. In the sentence “The
horse is expected to race tomorrow”, its POS tag is VB and in the sentence “The race for outer space”, its POS tag is NN.
Represent the lemma and POS tag of each token in your corpus as two one-hot vectors. Recall that a one-hot vector is a
binary vector which is used to distinguish each word or tag in a vocabulary or tagset from every other word in the vocabulary
or tagset. The vector consists of 0s in all cells with the exception of a single 1 in a cell used uniquely to identify that word.
Thus if there are V words in your vocabulary and t tags in your tagset, you will create a V – dimensional vector for the
lemma of your token and a t – dimensional vector for the token’s POS tag. Concatenate the two representations to create a
(V + t) – dimensional vector for each token.
Task – 3: Learning (10 points)
Use your favourite machine learning (not deep learning) algorithm to train a system that recognizes NER tags for a given
word. Specifically, you will use the feature vectors created in Task-2 as inputs to the machine learning model. The output of
the model will be the tag associated with that token.
Note that you may find some words in the test set which were not present in the training set. This is also known as the OOV
(out-of-vocabulary) problem in NLP. To handle OOV words, you must create a special ‘UNK’ token in the vocabulary you
created in Task-1. The lemmas of all OOV words must be replaced by this UNK token.
Task – 4: Model performance (10 points)
Report the model’s performance on the supplied test set. Specifically, you will report the accuracy, precision, recall and
F-score on the test set. Additionally, report the throughput of your system at inference time (you must report both the time
taken as well as the throughput in kbps)
2
To Submit
Submit the following:
1. Your source code (either as a notebook or regular code file)
2. Instructions on how to run the code
3. A text file outlining your model’s performance on the test set
External Links
1. NLTK: An easy-to-use API for NLP
2. SpaCy: An industrial-strength NLP tool
3. CoreNLP: Java library released by Stanford for common NLP tasks
4. MITIE: A C++ library for Information Extraction
5. Scikit-learn: Machine Learning with Python
6. OpenML with Java
7. A very simple tutorial for NER with scikit-learn
References
[1] Erik F Sang and Fien De Meulder. Introduction to the conll-2003 shared task: Language-independent named entity
recognition. arXiv preprint cs/0306050, 2003.
[2] Edward Loper Bird, Steven and Ewan Klein. Natural language processing with python. O’Reilly Media Inc., 2009.
3