---

language: en
license: mit
tags:
- keyphrase-extraction
datasets:
- midas/kpcrowd
metrics:
- seqeval
widget:
- 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."
  example_title: "Example 1"
- text: "In this work, we explore how to learn task specific language models aimed towards learning rich representation of keyphrases from text documents. We experiment with different masking strategies for pre-training transformer language models (LMs) in discriminative as well as generative settings. In the discriminative setting, we introduce a new pre-training objective - Keyphrase Boundary Infilling with Replacement (KBIR), showing large gains in performance (up to 9.26 points in F1) over SOTA, when LM pre-trained using KBIR is fine-tuned for the task of keyphrase extraction. In the generative setting, we introduce a new pre-training setup for BART - KeyBART, that reproduces the keyphrases related to the input text in the CatSeq format, instead of the denoised original input. This also led to gains in performance (up to 4.33 points inF1@M) over SOTA for keyphrase generation. Additionally, we also fine-tune the pre-trained language models on named entity recognition(NER), question answering (QA), relation extraction (RE), abstractive summarization and achieve comparable performance with that of the SOTA, showing that learning rich representation of keyphrases is indeed beneficial for many other fundamental NLP tasks."
  example_title: "Example 2"
model-index:
- name: DeDeckerThomas/keyphrase-extraction-kbir-kpcrowd
  results:
  - task: 
      type: keyphrase-extraction
      name: Keyphrase Extraction
    dataset:
      type: midas/kpcrowd
      name: kpcrowd
    metrics:
      - type: F1 (Seqeval)
        value: 0.427
        name: F1 (Seqeval)
      - type: F1@M
        value: 0.335
        name: F1@M        
---
# 🔑 Keyphrase Extraction Model: KBIR-KPCrowd
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 [KBIR](https://huggingface.co/bloomberg/KBIR) as its base model and fine-tunes it on the [KPCrowd dataset](https://huggingface.co/datasets/midas/kpcrowd). KBIR or Keyphrase Boundary Infilling with Replacement is a pre-trained model which utilizes a multi-task learning setup for optimizing a combined loss of Masked Language Modeling (MLM), Keyphrase Boundary Infilling (KBI) and Keyphrase Replacement Classification (KRC).
You can find more information about the architecture in this [paper](https://arxiv.org/abs/2112.08547).

Keyphrase extraction models are transformer models fine-tuned as a token classification problem where each word in the document is classified as being part of a keyphrase or not.

| Label | Description                     |
| ----- | ------------------------------- |
| B-KEY | At the beginning of a keyphrase |
| I-KEY | Inside a keyphrase              |
| O     | Outside a keyphrase             |

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

Sahrawat, Dhruva, Debanjan Mahata, Haimin Zhang, Mayank Kulkarni, Agniv Sharma, Rakesh Gosangi, Amanda Stent, Yaman Kumar, Rajiv Ratn Shah, and Roger Zimmermann. "Keyphrase extraction as sequence labeling using contextualized embeddings." In European Conference on Information Retrieval, pp. 328-335. Springer, Cham, 2020.

## ✋ Intended Uses & Limitations
### 🛑 Limitations
* This keyphrase extraction model is very dataset-specific. It's not recommended to use this model for other domains, but you are free to test it out.
* Only works for English documents.
* Large number of annotated keyphrases.

### ❓ How To Use
```python
from transformers import (
    TokenClassificationPipeline,
    AutoModelForTokenClassification,
    AutoTokenizer,
)
from transformers.pipelines import AggregationStrategy
import numpy as np

# Define keyphrase extraction pipeline
class KeyphraseExtractionPipeline(TokenClassificationPipeline):
    def __init__(self, model, *args, **kwargs):
        super().__init__(
            model=AutoModelForTokenClassification.from_pretrained(model),
            tokenizer=AutoTokenizer.from_pretrained(model),
            *args,
            **kwargs
        )

    def postprocess(self, all_outputs):
        results = super().postprocess(
            all_outputs=all_outputs,
            aggregation_strategy=AggregationStrategy.SIMPLE,
        )
        return np.unique([result.get("word").strip() for result in results])

```

```python
# Load pipeline
model_name = "ml6team/keyphrase-extraction-kbir-kpcrowd"
extractor = KeyphraseExtractionPipeline(model=model_name)

```
```python
# 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 = extractor(text)

print(keyphrases)

```

```
# Output
['Artificial Intelligence' 'Classical' 'Keyphrase' 'Keyphrase extraction'
 'classical' 'content' 'context' 'disadvantage' 'document' 'documents'
 'extract' 'extraction' 'extraction process' 'frequency' 'human' 'humans'
 'important' 'keyphrases' 'learning' 'linguistic' 'long-term'
 'machine learning' 'meaning' 'methods' 'neural approaches' 'occurrence'
 'process' 'quickly' 'semantic' 'statistical' 'technique' 'text'
 'text analysis' 'understand' 'widely' 'words' 'work']
```

## 📚 Training Dataset
[KPCrowd](https://huggingface.co/datasets/midas/kpcrowd) is a broadcast news transcription dataset consisting of 500 English broadcast news stories from 10 different categories (art and culture, business, crime, fashion, health, politics us, politics world, science, sports, technology) with 50 docs per category. This dataset is annotated by multiple annotators that were required to look at the same news story and assign a set of keyphrases from the text itself.

You can find more information in the [paper](https://arxiv.org/abs/1306.4606).

## 👷‍♂️ Training Procedure
### Training Parameters

| Parameter | Value |
| --------- | ------|
| Learning Rate | 1e-4 |
| Epochs | 50 |
| Early Stopping Patience | 3 |

### Preprocessing
The documents in the dataset are already preprocessed into list of words with the corresponding labels. The only thing that must be done is tokenization and the realignment of the labels so that they correspond with the right subword tokens.

```python
from datasets import load_dataset
from transformers import AutoTokenizer

# Labels
label_list = ["B", "I", "O"]
lbl2idx = {"B": 0, "I": 1, "O": 2}
idx2label = {0: "B", 1: "I", 2: "O"}

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained("bloomberg/KBIR", add_prefix_space=True)
max_length = 512

# Dataset parameters
dataset_full_name = "midas/kpcrowd"
dataset_subset = "raw"
dataset_document_column = "document"
dataset_biotags_column = "doc_bio_tags"

def preprocess_fuction(all_samples_per_split):
    tokenized_samples = tokenizer.batch_encode_plus(
        all_samples_per_split[dataset_document_column],
        padding="max_length",
        truncation=True,
        is_split_into_words=True,
        max_length=max_length,
    )
    total_adjusted_labels = []
    for k in range(0, len(tokenized_samples["input_ids"])):
        prev_wid = -1
        word_ids_list = tokenized_samples.word_ids(batch_index=k)
        existing_label_ids = all_samples_per_split[dataset_biotags_column][k]
        i = -1
        adjusted_label_ids = []

        for wid in word_ids_list:
            if wid is None:
                adjusted_label_ids.append(lbl2idx["O"])
            elif wid != prev_wid:
                i = i + 1
                adjusted_label_ids.append(lbl2idx[existing_label_ids[i]])
                prev_wid = wid
            else:
                adjusted_label_ids.append(
                    lbl2idx[
                        f"{'I' if existing_label_ids[i] == 'B' else existing_label_ids[i]}"
                    ]
                )

        total_adjusted_labels.append(adjusted_label_ids)
    tokenized_samples["labels"] = total_adjusted_labels
    return tokenized_samples

# Load dataset
dataset = load_dataset(dataset_full_name, dataset_subset)

# Preprocess dataset
tokenized_dataset = dataset.map(preprocess_fuction, batched=True)
    
```

### Postprocessing (Without Pipeline Function)
If you do not use the pipeline function, you must filter out the B and I labeled tokens. Each B and I will then be merged into a keyphrase. Finally, you need to strip the keyphrases to make sure all unnecessary spaces have been removed.
```python
# Define post_process functions
def concat_tokens_by_tag(keyphrases):
    keyphrase_tokens = []
    for id, label in keyphrases:
        if label == "B":
            keyphrase_tokens.append([id])
        elif label == "I":
            if len(keyphrase_tokens) > 0:
                keyphrase_tokens[len(keyphrase_tokens) - 1].append(id)
    return keyphrase_tokens


def extract_keyphrases(example, predictions, tokenizer, index=0):
    keyphrases_list = [
        (id, idx2label[label])
        for id, label in zip(
            np.array(example["input_ids"]).squeeze().tolist(), predictions[index]
        )
        if idx2label[label] in ["B", "I"]
    ]

    processed_keyphrases = concat_tokens_by_tag(keyphrases_list)
    extracted_kps = tokenizer.batch_decode(
        processed_keyphrases,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=True,
    )
    return np.unique([kp.strip() for kp in extracted_kps])

```

## 📝 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.
The model achieves the following results on the Inspec test set:

| Dataset           | P@5  | R@5  | F1@5 | P@10 | R@10 | F1@10 | P@M  | R@M  | F1@M |
|:-----------------:|:----:|:----:|:----:|:----:|:----:|:-----:|:----:|:----:|:----:|
| Inspec Test Set   | 0.47 | 0.07 | 0.12 | 0.46 | 0.13 | 0.20  | 0.37 | 0.33 | 0.33 |

## 🚨 Issues
Please feel free to start discussions in the Community Tab.