Thomas De Decker
commited on
Commit
•
306e18b
1
Parent(s):
0697c55
Add model
Browse files- README.md +256 -0
- config.json +38 -0
- merges.txt +0 -0
- pytorch_model.bin +3 -0
- special_tokens_map.json +1 -0
- tokenizer.json +0 -0
- tokenizer_config.json +1 -0
- training_args.bin +3 -0
- vocab.json +0 -0
README.md
CHANGED
@@ -1,3 +1,259 @@
|
|
1 |
---
|
|
|
|
|
2 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
|
3 |
+
language: en
|
4 |
license: mit
|
5 |
+
tags:
|
6 |
+
- keyphrase-extraction
|
7 |
+
datasets:
|
8 |
+
- midas/kptimes
|
9 |
+
metrics:
|
10 |
+
- seqeval
|
11 |
+
widget:
|
12 |
+
- text: "Keyphrase extraction is a technique in text analysis where you extract the important keyphrases from a document.
|
13 |
+
Thanks to these keyphrases humans can understand the content of a text very quickly and easily without reading
|
14 |
+
it completely. Keyphrase extraction was first done primarily by human annotators, who read the text in detail
|
15 |
+
and then wrote down the most important keyphrases. The disadvantage is that if you work with a lot of documents,
|
16 |
+
this process can take a lot of time.
|
17 |
+
|
18 |
+
Here is where Artificial Intelligence comes in. Currently, classical machine learning methods, that use statistical
|
19 |
+
and linguistic features, are widely used for the extraction process. Now with deep learning, it is possible to capture
|
20 |
+
the semantic meaning of a text even better than these classical methods. Classical methods look at the frequency,
|
21 |
+
occurrence and order of words in the text, whereas these neural approaches can capture long-term semantic dependencies
|
22 |
+
and context of words in a text."
|
23 |
+
example_title: "Example 1"
|
24 |
+
- text: "FoodEx is the largest trade exhibition for food and drinks in Asia, with about 70,000 visitors checking out the products presented by hundreds of participating companies. I was lucky to enter as press; otherwise, visitors must be affiliated with the food industry— and pay ¥5,000 — to enter. The FoodEx menu is global, including everything from cherry beer from Germany and premium Mexican tequila to top-class French and Chinese dumplings. The event was a rare chance to try out both well-known and exotic foods and even see professionals making them. In addition to booths offering traditional Japanese favorites such as udon and maguro sashimi, there were plenty of innovative twists, such as dorayaki , a sweet snack made of two pancakes and a red-bean filling, that came in coffee and tomato flavors. While I was there I was lucky to catch the World Sushi Cup Japan 2013, where top chefs from around the world were competing … and presenting a wide range of styles that you would not normally see in Japan, like the flower makizushi above."
|
25 |
+
example_title: "Example 2"
|
26 |
+
model-index:
|
27 |
+
- name: ml6team/keyphrase-extraction-distilbert-kptimes
|
28 |
+
results:
|
29 |
+
- task:
|
30 |
+
type: keyphrase-extraction
|
31 |
+
name: Keyphrase Extraction
|
32 |
+
dataset:
|
33 |
+
type: midas/kptimes
|
34 |
+
name: kptimes
|
35 |
+
metrics:
|
36 |
+
- type: F1 (Seqeval)
|
37 |
+
value: 0.000
|
38 |
+
name: F1 (Seqeval)
|
39 |
+
- type: F1@M
|
40 |
+
value: 0.331
|
41 |
+
name: F1@M
|
42 |
---
|
43 |
+
# 🔑 Keyphrase Extraction Model: KBIR-KPTimes
|
44 |
+
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 ⏳.
|
45 |
+
|
46 |
+
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.
|
47 |
+
|
48 |
+
|
49 |
+
## 📓 Model Description
|
50 |
+
This model uses [KBIR](https://huggingface.co/distilbert-base-uncased) as its base model and fine-tunes it on the [KPTimes dataset](https://huggingface.co/datasets/midas/kptimes). 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).
|
51 |
+
You can find more information about the architecture in this [paper](https://arxiv.org/abs/2112.08547).
|
52 |
+
|
53 |
+
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.
|
54 |
+
|
55 |
+
| Label | Description |
|
56 |
+
| ----- | ------------------------------- |
|
57 |
+
| B-KEY | At the beginning of a keyphrase |
|
58 |
+
| I-KEY | Inside a keyphrase |
|
59 |
+
| O | Outside a keyphrase |
|
60 |
+
|
61 |
+
## ✋ Intended Uses & Limitations
|
62 |
+
### 🛑 Limitations
|
63 |
+
* This keyphrase extraction model is very domain-specific and will perform very well on news articles from NY Times. It's not recommended to use this model for other domains, but you are free to test it out.
|
64 |
+
* Limited amount of predicted keyphrases.
|
65 |
+
* Only works for English documents.
|
66 |
+
* For a custom model, please consult the [training notebook]() for more information.
|
67 |
+
|
68 |
+
### ❓ How To Use
|
69 |
+
```python
|
70 |
+
from transformers import (
|
71 |
+
TokenClassificationPipeline,
|
72 |
+
AutoModelForTokenClassification,
|
73 |
+
AutoTokenizer,
|
74 |
+
)
|
75 |
+
from transformers.pipelines import AggregationStrategy
|
76 |
+
import numpy as np
|
77 |
+
|
78 |
+
# Define keyphrase extraction pipeline
|
79 |
+
class KeyphraseExtractionPipeline(TokenClassificationPipeline):
|
80 |
+
def __init__(self, model, *args, **kwargs):
|
81 |
+
super().__init__(
|
82 |
+
model=AutoModelForTokenClassification.from_pretrained(model),
|
83 |
+
tokenizer=AutoTokenizer.from_pretrained(model),
|
84 |
+
*args,
|
85 |
+
**kwargs
|
86 |
+
)
|
87 |
+
|
88 |
+
def postprocess(self, model_outputs):
|
89 |
+
results = super().postprocess(
|
90 |
+
model_outputs=model_outputs,
|
91 |
+
aggregation_strategy=AggregationStrategy.SIMPLE,
|
92 |
+
)
|
93 |
+
return np.unique([result.get("word").strip() for result in results])
|
94 |
+
|
95 |
+
```
|
96 |
+
|
97 |
+
```python
|
98 |
+
# Load pipeline
|
99 |
+
model_name = "ml6team/keyphrase-extraction-kbir-kptimes"
|
100 |
+
extractor = KeyphraseExtractionPipeline(model=model_name)
|
101 |
+
```
|
102 |
+
```python
|
103 |
+
# Inference
|
104 |
+
text = """
|
105 |
+
Keyphrase extraction is a technique in text analysis where you extract the
|
106 |
+
important keyphrases from a document. Thanks to these keyphrases humans can
|
107 |
+
understand the content of a text very quickly and easily without reading it
|
108 |
+
completely. Keyphrase extraction was first done primarily by human annotators,
|
109 |
+
who read the text in detail and then wrote down the most important keyphrases.
|
110 |
+
The disadvantage is that if you work with a lot of documents, this process
|
111 |
+
can take a lot of time.
|
112 |
+
|
113 |
+
Here is where Artificial Intelligence comes in. Currently, classical machine
|
114 |
+
learning methods, that use statistical and linguistic features, are widely used
|
115 |
+
for the extraction process. Now with deep learning, it is possible to capture
|
116 |
+
the semantic meaning of a text even better than these classical methods.
|
117 |
+
Classical methods look at the frequency, occurrence and order of words
|
118 |
+
in the text, whereas these neural approaches can capture long-term
|
119 |
+
semantic dependencies and context of words in a text.
|
120 |
+
""".replace("\n", " ")
|
121 |
+
|
122 |
+
keyphrases = extractor(text)
|
123 |
+
|
124 |
+
print(keyphrases)
|
125 |
+
|
126 |
+
```
|
127 |
+
|
128 |
+
```
|
129 |
+
# Output
|
130 |
+
['artificial intelligence']
|
131 |
+
```
|
132 |
+
|
133 |
+
## 📚 Training Dataset
|
134 |
+
[KPTimes](https://huggingface.co/datasets/midas/kptimes) is a keyphrase extraction/generation dataset consisting of 279,923 news articles from NY Times and 10K from JPTimes and annotated by professional indexers or editors.
|
135 |
+
|
136 |
+
You can find more information in the [paper](https://arxiv.org/abs/1911.12559).
|
137 |
+
|
138 |
+
## 👷♂️ Training procedure
|
139 |
+
For more in detail information, you can take a look at the [training notebook]().
|
140 |
+
|
141 |
+
### Training parameters
|
142 |
+
|
143 |
+
| Parameter | Value |
|
144 |
+
| --------- | ------|
|
145 |
+
| Learning Rate | 1e-4 |
|
146 |
+
| Epochs | 50 |
|
147 |
+
| Early Stopping Patience | 3 |
|
148 |
+
|
149 |
+
### Preprocessing
|
150 |
+
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.
|
151 |
+
|
152 |
+
```python
|
153 |
+
from datasets import load_dataset
|
154 |
+
from transformers import AutoTokenizer
|
155 |
+
|
156 |
+
# Labels
|
157 |
+
label_list = ["B", "I", "O"]
|
158 |
+
lbl2idx = {"B": 0, "I": 1, "O": 2}
|
159 |
+
idx2label = {0: "B", 1: "I", 2: "O"}
|
160 |
+
|
161 |
+
# Tokenizer
|
162 |
+
tokenizer = AutoTokenizer.from_pretrained("bloomberg/KBIR")
|
163 |
+
max_length = 512
|
164 |
+
|
165 |
+
# Dataset parameters
|
166 |
+
dataset_full_name = "midas/kptimes"
|
167 |
+
dataset_subset = "raw"
|
168 |
+
dataset_document_column = "document"
|
169 |
+
dataset_biotags_column = "doc_bio_tags"
|
170 |
+
|
171 |
+
def preprocess_fuction(all_samples_per_split):
|
172 |
+
tokenized_samples = tokenizer.batch_encode_plus(
|
173 |
+
all_samples_per_split[dataset_document_column],
|
174 |
+
padding="max_length",
|
175 |
+
truncation=True,
|
176 |
+
is_split_into_words=True,
|
177 |
+
max_length=max_length,
|
178 |
+
)
|
179 |
+
total_adjusted_labels = []
|
180 |
+
for k in range(0, len(tokenized_samples["input_ids"])):
|
181 |
+
prev_wid = -1
|
182 |
+
word_ids_list = tokenized_samples.word_ids(batch_index=k)
|
183 |
+
existing_label_ids = all_samples_per_split[dataset_biotags_column][k]
|
184 |
+
i = -1
|
185 |
+
adjusted_label_ids = []
|
186 |
+
|
187 |
+
for wid in word_ids_list:
|
188 |
+
if wid is None:
|
189 |
+
adjusted_label_ids.append(lbl2idx["O"])
|
190 |
+
elif wid != prev_wid:
|
191 |
+
i = i + 1
|
192 |
+
adjusted_label_ids.append(lbl2idx[existing_label_ids[i]])
|
193 |
+
prev_wid = wid
|
194 |
+
else:
|
195 |
+
adjusted_label_ids.append(
|
196 |
+
lbl2idx[
|
197 |
+
f"{'I' if existing_label_ids[i] == 'B' else existing_label_ids[i]}"
|
198 |
+
]
|
199 |
+
)
|
200 |
+
|
201 |
+
total_adjusted_labels.append(adjusted_label_ids)
|
202 |
+
tokenized_samples["labels"] = total_adjusted_labels
|
203 |
+
return tokenized_samples
|
204 |
+
|
205 |
+
# Load dataset
|
206 |
+
dataset = load_dataset(dataset_full_name, dataset_subset)
|
207 |
+
|
208 |
+
# Preprocess dataset
|
209 |
+
tokenized_dataset = dataset.map(preprocess_fuction, batched=True)
|
210 |
+
|
211 |
+
```
|
212 |
+
|
213 |
+
### Postprocessing (Without Pipeline Function)
|
214 |
+
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.
|
215 |
+
```python
|
216 |
+
# Define post_process functions
|
217 |
+
def concat_tokens_by_tag(keyphrases):
|
218 |
+
keyphrase_tokens = []
|
219 |
+
for id, label in keyphrases:
|
220 |
+
if label == "B":
|
221 |
+
keyphrase_tokens.append([id])
|
222 |
+
elif label == "I":
|
223 |
+
if len(keyphrase_tokens) > 0:
|
224 |
+
keyphrase_tokens[len(keyphrase_tokens) - 1].append(id)
|
225 |
+
return keyphrase_tokens
|
226 |
+
|
227 |
+
|
228 |
+
def extract_keyphrases(example, predictions, tokenizer, index=0):
|
229 |
+
keyphrases_list = [
|
230 |
+
(id, idx2label[label])
|
231 |
+
for id, label in zip(
|
232 |
+
np.array(example["input_ids"]).squeeze().tolist(), predictions[index]
|
233 |
+
)
|
234 |
+
if idx2label[label] in ["B", "I"]
|
235 |
+
]
|
236 |
+
|
237 |
+
processed_keyphrases = concat_tokens_by_tag(keyphrases_list)
|
238 |
+
extracted_kps = tokenizer.batch_decode(
|
239 |
+
processed_keyphrases,
|
240 |
+
skip_special_tokens=True,
|
241 |
+
clean_up_tokenization_spaces=True,
|
242 |
+
)
|
243 |
+
return np.unique([kp.strip() for kp in extracted_kps])
|
244 |
+
|
245 |
+
```
|
246 |
+
|
247 |
+
## 📝 Evaluation Results
|
248 |
+
|
249 |
+
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.
|
250 |
+
The model achieves the following results on the KPTimes test set:
|
251 |
+
|
252 |
+
| Dataset | P@5 | R@5 | F1@5 | P@10 | R@10 | F1@10 | P@M | R@M | F1@M |
|
253 |
+
|:-----------------:|:----:|:----:|:----:|:----:|:----:|:-----:|:----:|:----:|:----:|
|
254 |
+
| KPTimes Test Set | 0.19 | 0.35 | 0.23 | 0.10 | 0.36 | 0.15 | 0.36 | 0.36 | 0.33 |
|
255 |
+
|
256 |
+
For more information on the evaluation process, you can take a look at the keyphrase extraction [evaluation notebook]().
|
257 |
+
|
258 |
+
## 🚨 Issues
|
259 |
+
Please feel free to start discussions in the Community Tab.
|
config.json
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "bloomberg/KBIR",
|
3 |
+
"architectures": [
|
4 |
+
"RobertaForTokenClassification"
|
5 |
+
],
|
6 |
+
"attention_probs_dropout_prob": 0.1,
|
7 |
+
"bos_token_id": 0,
|
8 |
+
"classifier_dropout": null,
|
9 |
+
"eos_token_id": 2,
|
10 |
+
"gradient_checkpointing": false,
|
11 |
+
"hidden_act": "gelu",
|
12 |
+
"hidden_dropout_prob": 0.1,
|
13 |
+
"hidden_size": 1024,
|
14 |
+
"id2label": {
|
15 |
+
"0": "B-KEY",
|
16 |
+
"1": "I-KEY",
|
17 |
+
"2": "O"
|
18 |
+
},
|
19 |
+
"initializer_range": 0.02,
|
20 |
+
"intermediate_size": 4096,
|
21 |
+
"label2id": {
|
22 |
+
"B-KEY": 0,
|
23 |
+
"I-KEY": 1,
|
24 |
+
"O": 2
|
25 |
+
},
|
26 |
+
"layer_norm_eps": 1e-05,
|
27 |
+
"max_position_embeddings": 514,
|
28 |
+
"model_type": "roberta",
|
29 |
+
"num_attention_heads": 16,
|
30 |
+
"num_hidden_layers": 24,
|
31 |
+
"pad_token_id": 1,
|
32 |
+
"position_embedding_type": "absolute",
|
33 |
+
"torch_dtype": "float32",
|
34 |
+
"transformers_version": "4.18.0",
|
35 |
+
"type_vocab_size": 1,
|
36 |
+
"use_cache": true,
|
37 |
+
"vocab_size": 50265
|
38 |
+
}
|
merges.txt
ADDED
The diff for this file is too large to render.
See raw diff
|
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:27dff89525c5760ed85b726024e2566f5f775ff0b8eed72389152bb93a71a3ab
|
3 |
+
size 1417389425
|
special_tokens_map.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"bos_token": {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "eos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "unk_token": {"content": "<unk>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "sep_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "pad_token": {"content": "<pad>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "cls_token": {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true}}
|
tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tokenizer_config.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"errors": "replace", "bos_token": {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "eos_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "sep_token": {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "cls_token": {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "unk_token": {"content": "<unk>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "pad_token": {"content": "<pad>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "add_prefix_space": true, "trim_offsets": true, "use_fast": true, "max_length": 512, "special_tokens_map_file": null, "name_or_path": "bloomberg/KBIR", "tokenizer_class": "RobertaTokenizer"}
|
training_args.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:33b9936b9af770f36469bcaa654f94d8ff192f8e1b6582e32bde8468b1db9dc4
|
3 |
+
size 3055
|
vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|