djackson-proofpoint/uplimit-synthetic-data-week-2-filtered-customemails

Custom Email categorization with readability score

The dataset is a "distilabeled" version of the dataset: snoop2head/enron_aeslc_emails. The dataset was filtered to smaller sized messages(less than 256 chars).

The dataset was analyzed by textdescriptives(flesch_reading_ease) and then categorized by gliclass against these labels:
"business and industrial",
"books and literature",
"home and garden",
"adult",
"science",
"food and drink",
"real estate",
"news",
"jobs and education",
"health",
"hobbies and leisure",
"games",
"beauty and fitness",
"computers and electronics",
"arts and entertainment",
"travel and transportation",
"finance",
"law and government",
"sensitive subjects",
"autos and vehicles",
"people and society",
"sports",
"shopping",
"online communities",
"pets and animals",
"internet and telecom"

Using this dataset

This dataset is useful for labelling and qualifying example email interactions to help train LLMs for this purpose.

Observations:

Narrowing this set down to very short messages seems to have been detrimental to the classification.
Most of the content that is available is the email headers only, and maybe one or two lines of text.
The readability score is directly related to the body content. If a URL is present in the body, it will have a low(<50) readability score.


Due to the majority of the data being email headers, it is categorizing most as 'online communities', and only occasionally finding other topics.

Reproducing the dataset

from datasets import load_dataset
import textdescriptives as td
from gliclass import GLiClassModel, ZeroShotClassificationPipeline
from transformers import AutoTokenizer

#Start with the source dataset
dataset = load_dataset("snoop2head/enron_aeslc_emails")

#Filter by length
dataset_quality = dataset.filter(
    lambda r:
        not r["text"] is None and
        len(r["text"]) <= 256
)

#Function to work in batches for readability of text
def process_quality(batch):
    df = td.extract_metrics(
        text=batch["text"],
        lang="en",
        metrics=["readability"]
    )
    batch["flesch_reading_ease"] = df["flesch_reading_ease"].values
    return batch


dataset_quality_reading = dataset_quality.map(
    process_quality, batched=True
)

#Now process for classification of data
repo_id = "knowledgator/gliclass-modern-base-v2.0-init"
model = GLiClassModel.from_pretrained(repo_id)
tokenizer = AutoTokenizer.from_pretrained(repo_id)

pipeline = ZeroShotClassificationPipeline(
    model,
    tokenizer,
    device="cuda"
)

labels = [
    "business and industrial",
    "books and literature",
    "home and garden",
    "adult",
    "science",
    "food and drink",
    "real estate",
    "news",
    "jobs and education",
    "health",
    "hobbies and leisure",
    "games",
    "beauty and fitness",
    "computers and electronics",
    "arts and entertainment",
    "travel and transportation",
    "finance",
    "law and government",
    "sensitive subjects",
    "autos and vehicles",
    "people and society",
    "sports",
    "shopping",
    "online communities",
    "pets and animals",
    "internet and telecom",
]

def process_domains(batch):
    results = pipeline(batch["text"], labels, threshold=0)
    batch["domain"] = [max(res, key=lambda x: x['score'])["label"]
 if res else None for res in results]
    return batch

dataset_quality_reading_domain = dataset_quality_reading.map(
    process_domains, batched=True
)