---
language:
- en
license: cc-by-4.0
task_categories:
- text-classification
task_ids:
- multi-class-classification
pretty_name: Complex Word Identification
size_categories:
- 1K<n<10K
tags:
- text-simplification
- lexical-complexity
- cwi
- nlp-course
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.csv
      - split: validation
        path: validation.csv
      - split: test
        path: test.csv
  - config_name: biomedical
    data_files:
      - split: test
        path: biomedical_test.csv
  - config_name: news
    data_files:
      - split: test
        path: news_test.csv
dataset_info:
  - config_name: default
    features:
      - name: word
        dtype: string
      - name: label
        dtype:
          class_label:
            names:
              '0': simple
              '1': complex
      - name: annotators
        dtype: int64
      - name: sentence
        dtype: string
      - name: sentence_index
        dtype: int64
    splits:
      - name: train
        num_examples: 4000
      - name: validation
        num_examples: 1000
      - name: test
        num_examples: 922
  - config_name: biomedical
    features:
      - name: word
        dtype: string
      - name: complexity_score
        dtype: float64
      - name: label
        dtype:
          class_label:
            names:
              '0': simple
              '1': complex
    splits:
      - name: test
        num_examples: 289
  - config_name: news
    features:
      - name: word
        dtype: string
      - name: complexity_score
        dtype: float64
      - name: label
        dtype:
          class_label:
            names:
              '0': simple
              '1': complex
    splits:
      - name: test
        num_examples: 1813
---

# Complex Word Identification (CIS 5300)

## Dataset Description

This dataset supports the **Complex Word Identification (CWI)** task: given a word in context, predict whether it is *complex* (likely to be difficult for non-native speakers, children, or people with reading disabilities) or *simple*.

CWI is the first step in **lexical simplification** — the task of rewriting text to make it more accessible. Before you can simplify a word, you need to identify which words need simplification.

### Task

- **Input**: A word and the sentence it appears in
- **Output**: Binary label — `0` (simple) or `1` (complex)
- **Example**: In the sentence *"The beleaguered director resigned Wednesday"*, the word *beleaguered* is complex (label=1) while *director* is simple (label=0).

## Dataset Structure

### Default Configuration (Main Task)

```python
from datasets import load_dataset

dataset = load_dataset("CCB/cis5300-text-classification")

# Splits: train (4,000), validation (1,000), test (922)
print(dataset["train"][0])
# {'word': 'string', 'label': 0, 'annotators': 0, 'sentence': 'WASHINGTON -- The beleaguered...', 'sentence_index': 27}
```

| Split | Examples | Labels |
|-------|----------|--------|
| train | 4,000 | Yes |
| validation | 1,000 | Yes |
| test | 922 | No (held out for evaluation) |

**Fields:**

| Field | Type | Description |
|-------|------|-------------|
| `word` | string | The target word to classify |
| `label` | ClassLabel (simple/complex) | 0 = simple, 1 = complex |
| `annotators` | int | Number of annotators who labeled the word as complex |
| `sentence` | string | The full sentence containing the target word |
| `sentence_index` | int | Token position of the target word in the sentence |

### Domain Generalization Configs

Two additional test sets from different domains for evaluating how well models generalize:

```python
# Biomedical domain (PubMed abstracts)
bio = load_dataset("CCB/cis5300-text-classification", "biomedical")

# News domain (from CWI 2018 shared task)
news = load_dataset("CCB/cis5300-text-classification", "news")
```

| Config | Domain | Examples | Source |
|--------|--------|----------|--------|
| `biomedical` | PubMed abstracts | 289 | CompLex (Shardlow et al., 2020) |
| `news` | News articles | 1,813 | CWI 2018 Shared Task (Yimam et al., 2018) |

### Supplementary Data

The file `ngram_counts.txt.gz` contains word frequency counts from Google Books N-grams (~8.7M entries). This is useful as a feature for classification but is not a structured dataset. Download it separately:

```python
from huggingface_hub import hf_hub_download
path = hf_hub_download("CCB/cis5300-text-classification", "ngram_counts.txt.gz", repo_type="dataset")

import gzip
ngram_counts = {}
with gzip.open(path, 'rt', encoding='utf-8') as f:
    for line in f:
        parts = line.strip().split('\t')
        if len(parts) == 2:
            ngram_counts[parts[0]] = int(parts[1])
```

## Source

The core dataset was introduced in:

> **Simplification Using Paraphrases and Context-Based Lexical Substitution**
> Reno Kriz, Eleni Miltsakaki, Marianna Apidianaki, Chris Callison-Burch
> *Proceedings of NAACL-HLT 2018*, pages 207–217
> [https://aclanthology.org/N18-1019/](https://aclanthology.org/N18-1019/)

The words are drawn from news articles and annotated by both native and non-native English speakers who indicated which words could be difficult for non-native speakers, children, or people with reading disabilities.

### Annotation scheme and binarization

Each word was independently labeled by approximately 10 annotators. The `annotators` column records how many annotators marked the word as complex. Binary labels were derived using a threshold:

- **Simple (0)**: No annotator marked the word as complex (`annotators = 0`)
- **Complex (1)**: Three or more annotators marked it as complex (`annotators >= 3`)
- **Excluded**: Words in the ambiguous zone (1–2 annotators) were removed from the dataset

This means the dataset contains only clear cases — words that are unambiguously simple or where a meaningful fraction of annotators agreed on complexity.

### Domain generalization sources

- **Biomedical**: From the CompLex dataset (Shardlow et al., 2020) — words from biomedical abstracts with continuous complexity scores binarized at threshold 0.5.
- **News**: From the CWI 2018 Shared Task (Yimam et al., 2018) — words from news articles annotated for complexity.

## Intended Use

This dataset is used for **Homework 1** in [CIS 5300: Natural Language Processing](https://www.seas.upenn.edu/~cis5300/) at the University of Pennsylvania. Students build progressively more sophisticated classifiers:

1. **Baselines**: Word length and word frequency thresholds
2. **Machine learning**: Naive Bayes and Logistic Regression with engineered features
3. **Custom models**: Additional features with error analysis
4. **Domain generalization**: Evaluate on biomedical and news text

## Citation

```bibtex
@inproceedings{kriz-etal-2018-simplification,
    title = "Simplification Using Paraphrases and Context-Based Lexical Substitution",
    author = "Kriz, Reno and Miltsakaki, Eleni and Apidianaki, Marianna and Callison-Burch, Chris",
    booktitle = "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    year = "2018",
    address = "New Orleans, Louisiana",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/N18-1019",
    doi = "10.18653/v1/N18-1019",
    pages = "207--217",
}
```