ml6team/keyphrase-generation-keybart-inspec

🔑 Keyphrase Generation Model: KeyBART-inspec

Keyphrase extraction is a technique in text analysis where you extract the important keyphrases from a document. Thanks to these keyphrases humans can understand the content of a text very quickly and easily without reading it completely. Keyphrase extraction was first done primarily by human annotators, who read the text in detail and then wrote down the most important keyphrases. The disadvantage is that if you work with a lot of documents, this process can take a lot of time ⏳.

Here is where Artificial Intelligence 🤖 comes in. Currently, classical machine learning methods, that use statistical and linguistic features, are widely used for the extraction process. Now with deep learning, it is possible to capture the semantic meaning of a text even better than these classical methods. Classical methods look at the frequency, occurrence and order of words in the text, whereas these neural approaches can capture long-term semantic dependencies and context of words in a text.

📓 Model Description

This model uses KeyBART as its base model and fine-tunes it on the Inspec dataset. KeyBART focuses on learning a better representation of keyphrases in a generative setting. It produces the keyphrases associated with the input document from a corrupted input. The input is changed by token masking, keyphrase masking and keyphrase replacement. This model can already be used without any fine-tuning, but can be fine-tuned if needed. You can find more information about the architecture in this paper.

Kulkarni, Mayank, Debanjan Mahata, Ravneet Arora, and Rajarshi Bhowmik. "Learning Rich Representation of Keyphrases from Text." arXiv preprint arXiv:2112.08547 (2021).

✋ Intended Uses & Limitations

🛑 Limitations

• This keyphrase generation model is very domain-specific and will perform very well on abstracts of scientific papers. It's not recommended to use this model for other domains, but you are free to test it out.
• Only works for English documents.

❓ How To Use

# Model parameters
from transformers import (
    Text2TextGenerationPipeline,
    AutoModelForSeq2SeqLM,
    AutoTokenizer,
)


class KeyphraseGenerationPipeline(Text2TextGenerationPipeline):
    def __init__(self, model, keyphrase_sep_token=";", *args, **kwargs):
        super().__init__(
            model=AutoModelForSeq2SeqLM.from_pretrained(model),
            tokenizer=AutoTokenizer.from_pretrained(model),
            *args,
            **kwargs
        )
        self.keyphrase_sep_token = keyphrase_sep_token

    def postprocess(self, model_outputs):
        results = super().postprocess(
            model_outputs=model_outputs
        )
        return [[keyphrase.strip() for keyphrase in result.get("generated_text").split(self.keyphrase_sep_token) if keyphrase != ""] for result in results]

# Load pipeline
model_name = "ml6team/keyphrase-generation-keybart-inspec"
generator = KeyphraseGenerationPipeline(model=model_name)

# Inference
text = """
Keyphrase extraction is a technique in text analysis where you extract the
important keyphrases from a document. Thanks to these keyphrases humans can
understand the content of a text very quickly and easily without reading it
completely. Keyphrase extraction was first done primarily by human annotators,
who read the text in detail and then wrote down the most important keyphrases.
The disadvantage is that if you work with a lot of documents, this process
can take a lot of time. 

Here is where Artificial Intelligence comes in. Currently, classical machine
learning methods, that use statistical and linguistic features, are widely used
for the extraction process. Now with deep learning, it is possible to capture
the semantic meaning of a text even better than these classical methods.
Classical methods look at the frequency, occurrence and order of words
in the text, whereas these neural approaches can capture long-term
semantic dependencies and context of words in a text.
""".replace("\n", " ")

keyphrases = generator(text)

print(keyphrases)

# Output
[['keyphrase extraction', 'text analysis', 'keyphrases', 'human annotators', 'artificial']]

📚 Training Dataset

Inspec is a keyphrase extraction/generation dataset consisting of 2000 English scientific papers from the scientific domains of Computers and Control and Information Technology published between 1998 to 2002. The keyphrases are annotated by professional indexers or editors.

You can find more information in the paper.

👷‍♂️ Training Procedure

Training Parameters

Parameter	Value
Learning Rate	5e-5
Epochs	15
Early Stopping Patience	1

Preprocessing

The documents in the dataset are already preprocessed into list of words with the corresponding keyphrases. The only thing that must be done is tokenization and joining all keyphrases into one string with a certain seperator of choice( ; ).

from datasets import load_dataset
from transformers import AutoTokenizer

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained("bloomberg/KeyBART", add_prefix_space=True)

# Dataset parameters
dataset_full_name = "midas/inspec"
dataset_subset = "raw"
dataset_document_column = "document"

keyphrase_sep_token = ";"

def preprocess_keyphrases(text_ids, kp_list):
    kp_order_list = []
    kp_set = set(kp_list)
    text = tokenizer.decode(
        text_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
    )
    text = text.lower()
    for kp in kp_set:
        kp = kp.strip()
        kp_index = text.find(kp.lower())
        kp_order_list.append((kp_index, kp))

    kp_order_list.sort()
    present_kp, absent_kp = [], []

    for kp_index, kp in kp_order_list:
        if kp_index < 0:
            absent_kp.append(kp)
        else:
            present_kp.append(kp)
    return present_kp, absent_kp


def preprocess_fuction(samples):
    processed_samples = {"input_ids": [], "attention_mask": [], "labels": []}
    for i, sample in enumerate(samples[dataset_document_column]):
        input_text = " ".join(sample)
        inputs = tokenizer(
            input_text,
            padding="max_length",
            truncation=True,
        )
        present_kp, absent_kp = preprocess_keyphrases(
            text_ids=inputs["input_ids"],
            kp_list=samples["extractive_keyphrases"][i]
            + samples["abstractive_keyphrases"][i],
        )
        keyphrases = present_kp
        keyphrases += absent_kp

        target_text = f" {keyphrase_sep_token} ".join(keyphrases)

        with tokenizer.as_target_tokenizer():
            targets = tokenizer(
                target_text, max_length=40, padding="max_length", truncation=True
            )
            targets["input_ids"] = [
                (t if t != tokenizer.pad_token_id else -100)
                for t in targets["input_ids"]
            ]
        for key in inputs.keys():
            processed_samples[key].append(inputs[key])
        processed_samples["labels"].append(targets["input_ids"])
    return processed_samples

# Load dataset
dataset = load_dataset(dataset_full_name, dataset_subset)
# Preprocess dataset
tokenized_dataset = dataset.map(preprocess_fuction, batched=True)
    

Postprocessing

For the post-processing, you will need to split the string based on the keyphrase separator.

def extract_keyphrases(examples):
    return [example.split(keyphrase_sep_token) for example in examples]

📝 Evaluation results

Traditional evaluation methods are the precision, recall and F1-score @k,m where k is the number that stands for the first k predicted keyphrases and m for the average amount of predicted keyphrases. In keyphrase generation you also look at F1@O where O stands for the number of ground truth keyphrases.

The model achieves the following results on the Inspec test set:

Extractive Keyphrases

Dataset	P@5	R@5	F1@5	P@10	R@10	F1@10	P@M	R@M	F1@M	P@O	R@O	F1@O
Inspec Test Set	0.40	0.37	0.35	0.20	0.37	0.24	0.42	0.37	0.36	0.33	0.33	0.33

Abstractive Keyphrases

Dataset	P@5	R@5	F1@5	P@10	R@10	F1@10	P@M	R@M	F1@M	P@O	R@O	F1@O
Inspec Test Set	0.07	0.12	0.08	0.03	0.12	0.05	0.08	0.12	0.08	0.08	0.12	0.08

🚨 Issues

Please feel free to start discussions in the Community Tab.