svenstahlmann
commited on
Commit
•
092176d
1
Parent(s):
59164b1
Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,66 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language: en
|
3 |
+
tags:
|
4 |
+
- distilbert
|
5 |
+
- needmining
|
6 |
+
license: apache-2.0
|
7 |
+
metric:
|
8 |
+
- f1
|
9 |
+
---
|
10 |
+
|
11 |
+
# Finetuned-Distilbert-needmining (uncased)
|
12 |
+
|
13 |
+
This model is a finetuned version of the [Distilbert base model](https://huggingface.co/distilbert-base-uncased). It was
|
14 |
+
trained to predict need-containing sentences from amazon product reviews.
|
15 |
+
|
16 |
+
## Model description
|
17 |
+
|
18 |
+
This mode is part of ongoing research, after the publication of the research more information will be added.
|
19 |
+
|
20 |
+
## Intended uses & limitations
|
21 |
+
|
22 |
+
You can use this model to identify sentences that contain customer needs in user-generated content. This can act as a filtering process to remove uninformative content for market research.
|
23 |
+
|
24 |
+
### How to use
|
25 |
+
|
26 |
+
You can use this model directly with a pipeline for text classification:
|
27 |
+
|
28 |
+
```python
|
29 |
+
>>> from transformers import pipeline
|
30 |
+
>>> classifier = pipeline("text-classification", model="svenstahlmann/finetuned-distilbert-needmining")
|
31 |
+
>>> classifier("the plasic feels super cheap.")
|
32 |
+
[{'label': 'contains need', 'score': 0.9397542476654053}]
|
33 |
+
```
|
34 |
+
|
35 |
+
### Limitations and bias
|
36 |
+
|
37 |
+
We are not aware of any bias in the training data.
|
38 |
+
|
39 |
+
## Training data
|
40 |
+
|
41 |
+
The training was done on a dataset of 6400 sentences. The sentences were taken from product reviews off amazon and coded if they express customer needs.
|
42 |
+
## Training procedure
|
43 |
+
For the training, we used [Population Based Training (PBT)](https://www.deepmind.com/blog/population-based-training-of-neural-networks) and optimized for f1 score on a validation set of 1600 sentences.
|
44 |
+
|
45 |
+
|
46 |
+
### Preprocessing
|
47 |
+
|
48 |
+
The preprocessing follows the [Distilbert base model](https://huggingface.co/distilbert-base-uncased).
|
49 |
+
|
50 |
+
|
51 |
+
### Pretraining
|
52 |
+
|
53 |
+
The model was trained on a titan RTX for 1 hour.
|
54 |
+
|
55 |
+
## Evaluation results
|
56 |
+
|
57 |
+
Results on the validation set:
|
58 |
+
|
59 |
+
| F1 |
|
60 |
+
|:----:|
|
61 |
+
| 76.0 |
|
62 |
+
|
63 |
+
|
64 |
+
### BibTeX entry and citation info
|
65 |
+
|
66 |
+
coming soon
|