Cloze Test – Advanced Study Notes
Program: Insurance (LIC / GIC)
Subject: English Language
Level: MIT‑style, first‑principles treatment
A Cloze test is a reading‑comprehension assessment in which words (or other linguistic units) are systematically omitted from a passage and the test taker must supply the missing items. It is widely used in language proficiency testing, educational research, and increasingly in industry‑specific contexts such as insurance documentation comprehension.
1. Overview
| Feature | Description |
|---|---|
| Goal | Measure predictive and contextual linguistic competence. |
| Typical format | 5–20% of words removed, often 1–2 words per sentence. |
| Scoring | Correct fill‑in → +1; wrong or blank → 0. |
| Applications | Language exams, curriculum assessment, training modules, policy‑document readability checks. |
2. Theoretical Foundations
2.1 Linguistic Theory
| Aspect | Key Idea |
|---|---|
| Lexical Semantics | Predictability of a word depends on semantic fit with surrounding context. |
| Syntactic Structure | The grammatical role (subject, object, complement) constrains possible fillers. |
| Morphology | Inflectional endings (tense, number, case) provide strong cues. |
| Pragmatics | Discourse coherence and world knowledge influence the choice. |
Example:
“The ___ of the policy was amended.”
Possible fillers: terms, coverage, premium. The syntactic structure (noun phrase following the) and the context (policy) narrow the set.
2.2 Psycholinguistic Theory
- Predictive Coding – the brain continuously generates predictions about upcoming words (Hagoort, 2014).
- Lexical Retrieval – the lexical access process is facilitated when contextual constraints are strong.
- Entropy Reduction – the test measures the extent to which the reader reduces uncertainty about the missing word.
Mathematically, for a token w in context C, the conditional probability is
[
P(w \mid C) = \frac{P(C \mid w),P(w)}{P(C)} \quad \text{(Bayes’ theorem)}.
]
A high P(w|C) indicates a strong, unambiguous prediction.
2.3 Computational Modeling
| Model | Relevance |
|---|---|
| n‑gram Language Models | Estimate *P(w |
| Transformer‑based Models (BERT, GPT) | Capture long‑range dependencies; can generate candidate fillers. |
| Item Response Theory (IRT) | Models test taker ability θ and item parameters (difficulty b, discrimination a) for scoring. |
IRT Equation (2PL model):
[
P(\text{correct} \mid \theta) = \frac{1}{1 + e^{-a(\theta - b)}}.
]
This allows us to calibrate cloze items and compute standardized scores.
3. Construction of Cloze Tests
3.1 Item Selection
- Corpus‑driven
- Use a representative insurance corpus (e.g., LIC policy documents, GIC terms).
- Extract high‑frequency