NER with NLTK in NLP - Model Metrics & Evaluation

Metrics & Evaluation - NER with NLTK

Which metric matters for NER with NLTK and WHY

Named Entity Recognition (NER) finds names like people, places, or organizations in text. The main metrics to check are Precision, Recall, and F1-score.

Precision tells us how many of the entities the model found are actually correct. This matters because we want to avoid wrong names.

Recall tells us how many of the real entities the model found. This matters because missing important names can be bad.

F1-score balances precision and recall to give one clear number showing overall quality.

Confusion matrix for NER (simplified)

          | Predicted Entity | Predicted Non-Entity
    ------|------------------|--------------------
    True  |        TP        |         FN
    Entity|                  |                    
    ------|------------------|--------------------
    True  |        FP        |         TN
    Non-  |                  |                    
    Entity|                  |                    

    TP = Correctly found entities
    FP = Wrongly found entities (false alarms)
    FN = Missed entities
    TN = Correctly ignored non-entities
    

Precision vs Recall tradeoff in NER

If you want to be very sure about the entities you find, focus on high precision. For example, a legal document analyzer should not mark wrong names.

If you want to find as many entities as possible, even if some are wrong, focus on high recall. For example, a news aggregator might want to catch all possible names.

Usually, improving one lowers the other. The F1-score helps find a good balance.

Good vs Bad metric values for NER

Good: Precision and recall above 0.8 means the model finds most names correctly and misses few.

Bad: Precision below 0.5 means many wrong names are found. Recall below 0.5 means many names are missed.

F1-score below 0.6 usually means the model needs improvement.

Common pitfalls in NER metrics

• Accuracy paradox: Most words are not entities, so accuracy can be high even if the model never finds entities.
• Data leakage: Testing on data the model saw during training inflates metrics falsely.
• Overfitting: Very high training scores but low test scores mean the model memorizes instead of learning.
• Ignoring entity types: Treating all entities the same can hide poor performance on important types.

Self-check question

Your NER model has 98% accuracy but only 12% recall on person names. Is it good for production?

Answer: No. The high accuracy is misleading because most words are not person names. The very low recall means the model misses almost all person names, which is bad if you need to find them.

Key Result

For NER with NLTK, F1-score balancing precision and recall best shows model quality.

Practice

1. What is the main purpose of Named Entity Recognition (NER) in Natural Language Processing?

easy

A. To count the number of words in a sentence

B. To translate text from one language to another

C. To find names of people, places, and organizations in text

D. To correct spelling mistakes in text

Solution

1. Step 1: Understand NER's role

   NER is designed to identify and classify named entities like people, places, and organizations in text.

2. Step 2: Compare with other NLP tasks

   Translation, word counting, and spell checking are different tasks unrelated to NER.

3. Final Answer:

   To find names of people, places, and organizations in text -> Option C

4. Quick Check:

   NER = Find names [OK]

Hint: NER extracts names and places from text quickly [OK]

Common Mistakes:

• Confusing NER with translation
• Thinking NER counts words
• Mixing NER with spell checking

2. Which NLTK function is used to perform Named Entity Recognition after POS tagging?

easy

A. ne_chunk()

B. word_tokenize()

C. pos_tag()

D. sent_tokenize()

Solution

1. Step 1: Identify NLTK functions for NER

   NLTK uses ne_chunk() to recognize named entities from POS-tagged tokens.

2. Step 2: Differentiate from other functions

   word_tokenize() splits text into words, pos_tag() tags parts of speech, and sent_tokenize() splits text into sentences.

3. Final Answer:

   ne_chunk() -> Option A

4. Quick Check:

   NER uses ne_chunk() [OK]

Hint: Use ne_chunk() after pos_tag() for NER in NLTK [OK]

Common Mistakes:

• Using word_tokenize() for NER
• Confusing pos_tag() with NER
• Trying sent_tokenize() for entity recognition

3. What will be the output type of ne_chunk(pos_tag(word_tokenize(text))) in NLTK?

medium

A. A plain string with entity labels

B. A list of strings

C. A dictionary mapping words to entity types

D. A tree structure with named entities as subtrees

Solution

1. Step 1: Understand ne_chunk output

   The ne_chunk() function returns a tree structure where named entities are subtrees labeled with entity types.

2. Step 2: Compare output types

   It is not a list, dictionary, or plain string but a hierarchical tree that can be traversed.

3. Final Answer:

   A tree structure with named entities as subtrees -> Option D

4. Quick Check:

   ne_chunk output = tree structure [OK]

Hint: ne_chunk returns a tree, not a list or dict [OK]

Common Mistakes:

• Expecting a list of strings
• Thinking output is a dictionary
• Assuming output is a plain string

4. Given the code snippet:

import nltk
text = "Apple is looking at buying U.K. startup"
tokens = nltk.word_tokenize(text)
pos_tags = nltk.pos_tag(tokens)
entities = nltk.ne_chunk(pos_tags, binary=True)
print(entities)

What is the likely error in this code?

medium

A. Missing import for ne_chunk

B. Incorrect argument 'binary=True' in ne_chunk

C. pos_tag requires a list of sentences, not tokens

D. word_tokenize should be called after ne_chunk

Solution

1. Step 1: Check ne_chunk parameters

   The ne_chunk() function's binary=True limits it to binary NER (labels entities simply as NE, typically focusing on PERSON), which is incorrect for standard NER requiring specific types like PERSON, ORGANIZATION, GPE.

2. Step 2: Verify other parts

   Imports are correct with import nltk, pos_tag() accepts tokenized words, and preprocessing order is proper.

3. Final Answer:

   Incorrect argument 'binary=True' in ne_chunk -> Option B

4. Quick Check:

   binary=True limits to binary NER [OK]

Hint: Use binary=False for detailed entity types in ne_chunk [OK]

Common Mistakes:

• Using binary=True for detailed NER
• Calling word_tokenize after ne_chunk
• Misunderstanding pos_tag input

5. You want to extract only PERSON entities from a text using NLTK's ne_chunk. Which approach correctly filters PERSON entities from the chunked tree?

hard

A. Traverse the tree and select subtrees with label 'PERSON'

B. Use pos_tag to find tokens tagged as 'PERSON'

C. Filter tokens containing capital letters only

D. Use word_tokenize and select words starting with 'P'

Solution

1. Step 1: Understand ne_chunk output structure

   Named entities are subtrees labeled with entity types like 'PERSON', so we must traverse the tree to find these subtrees.

2. Step 2: Evaluate filtering methods

   pos_tag does not label entities, only parts of speech. Capital letters or starting with 'P' are unreliable heuristics.

3. Final Answer:

   Traverse the tree and select subtrees with label 'PERSON' -> Option A

4. Quick Check:

   Filter PERSON by subtree label [OK]

Hint: Filter PERSON entities by subtree label in ne_chunk tree [OK]

Common Mistakes:

• Using pos_tag to find entities
• Filtering by capitalization only
• Selecting words by first letter