Part-of-speech tagging in NLP - Model Metrics & Evaluation

Metrics & Evaluation - Part-of-speech tagging

Which metric matters for Part-of-speech tagging and WHY

For part-of-speech (POS) tagging, accuracy is the main metric. This is because POS tagging is a classification task where each word is assigned one correct tag. Accuracy tells us the percentage of words tagged correctly out of all words. Since every word must have exactly one tag, accuracy directly shows how well the model is doing overall.

Sometimes, per-tag precision and recall are also useful to understand how well the model predicts specific tags, especially if some tags are rare or more important.

Confusion matrix example for POS tagging

Imagine a simple POS tagger that predicts three tags: Noun (N), Verb (V), and Adjective (Adj). Here is a confusion matrix for 100 words:

      | Predicted N | Predicted V | Predicted Adj |
    ---------------------------------------------
    N |     40      |      5      |       5       |
    V |      3      |     30      |       2       |
    Adj|     2      |      3      |      10       |
    ---------------------------------------------
    Total words = 100
    

From this matrix:

• True Positives for Noun = 40
• False Positives for Noun = 3 + 2 = 5 (words wrongly predicted as Noun)
• False Negatives for Noun = 5 + 5 = 10 (Nouns predicted as other tags)

Accuracy = (40 + 30 + 10) / 100 = 80%

Precision vs Recall tradeoff with examples

In POS tagging, precision and recall per tag help understand errors:

• Precision for a tag means: Of all words predicted as that tag, how many were correct?
• Recall for a tag means: Of all words that truly have that tag, how many did the model find?

Example: For the Verb tag, if precision is high but recall is low, the model is very sure when it says a word is a verb but misses many verbs. This might happen if the model is cautious and only tags clear verbs.

For POS tagging, a balance is important because missing tags (low recall) or wrongly tagging words (low precision) both reduce usefulness.

What "good" vs "bad" metric values look like for POS tagging

Good metrics:

• Accuracy above 90% on a balanced dataset means most words are tagged correctly.
• Precision and recall above 85% for common tags like Noun and Verb.
• Consistent performance across tags, not just on frequent ones.

Bad metrics:

• Accuracy below 70% means many words are tagged wrong.
• Very low recall for some tags means the model misses many words of that type.
• High precision but very low recall or vice versa indicates imbalance and poor tagging quality.

Common pitfalls in POS tagging metrics

• Ignoring rare tags: Some tags appear rarely but are important. Ignoring their performance hides problems.
• Accuracy paradox: If the dataset has many nouns, a model tagging everything as noun can get high accuracy but is useless.
• Data leakage: Using test sentences seen during training inflates accuracy falsely.
• Overfitting: Very high training accuracy but low test accuracy means the model memorizes training data, not generalizing.

Self-check question

Your POS tagging model has 98% accuracy but only 12% recall on the "Verb" tag. Is this good for production? Why or why not?

Answer: No, it is not good. The model misses most verbs (low recall), which means many verbs are tagged incorrectly or missed. Even though overall accuracy is high, the poor recall on verbs can cause serious problems in understanding sentences. The model needs improvement to better detect verbs.

Key Result

Accuracy is key for POS tagging, but per-tag precision and recall reveal detailed strengths and weaknesses.