Classification¶
Table of contents¶
-
3.1. Predict using Python
3.2. Predict using CLI
-
4.1. from Python
4.2. from CLI
-
5.1. from Python
5.2. from CLI
-
7.1. Few-shot setting
7.2. Multiple languages support
7.3. Dataset and Scores
1. Introduction to the task¶
This section describes a family of BERT-based models that solve a variety of different classification tasks.
Insults detection is a binary classification task of identying wether a given sequence is an insult of another participant of communication.
Sentiment analysis is a task of classifying the polarity of the the given sequence. The number of classes may vary depending on the data: positive/negative binary classification, multiclass classification with a neutral class added or with a number of different emotions.
The models trained for the paraphrase detection task identify whether two sentences expressed with different words convey the same meaning.
Topic classification refers to the task of classifying an utterance by the topic which belongs to the conversational domain.
2. Get started with the model¶
First make sure you have the DeepPavlov Library installed. More info about the first installation.
[ ]:
!pip install -q deeppavlov
Then make sure that all the required packages for the model are installed.
[ ]:
!python -m deeppavlov install insults_kaggle_bert
insults_kaggle_bert
is the name of the model’s config_file. What is a Config File?
Configuration file defines the model and describes its hyperparameters. To use another model, change the name of the config_file here and further. The full list of NER models with their config names can be found in the table.
3. Use the model for prediction¶
3.1 Predict using Python¶
After model installation build it from the config and use it for prediction.
[ ]:
from deeppavlov import build_model
model = build_model('insults_kaggle_bert', download=True)
Input format: List[sentences]
Output format: List[labels]
[ ]:
model(['You are kind of stupid', 'You are a wonderful person!'])
['Insult', 'Not Insult']
3.2 Predict using CLI¶
You can also get predictions in an interactive mode through CLI (Command Line Interface).
[ ]:
!python deeppavlov interact insults_kaggle_bert -d
-d
is an optional download key (alternative to download=True
in Python code). The key -d
is used to download the pre-trained model along with embeddings and all other files needed to run the model.
Or make predictions for samples from stdin.
[ ]:
!python deeppavlov predict insults_kaggle_bert -f <file-name>
4. Evaluation¶
4.1 Evaluate from Python¶
[ ]:
from deeppavlov import evaluate_model
model = evaluate_model('insults_kaggle_bert', download=True)
4.2 Evaluate from CLI¶
[ ]:
!python -m deeppavlov evaluate insults_kaggle_bert -d
5. Train the model on your data¶
5.1 Train your model from Python¶
Provide your data path¶
To train the model on your data, you need to change the path to the training data in the config_file.
Parse the config_file and change the path to your data from Python.
[ ]:
from deeppavlov import train_model
from deeppavlov.core.commands.utils import parse_config
model_config = parse_config('insults_kaggle_bert')
# dataset that the model was trained on
print(model_config['dataset_reader']['data_path'])
~/.deeppavlov/downloads/insults_data
Provide a data_path to your own dataset. You can also change any of the hyperparameters of the model.
[ ]:
# download and unzip a new example dataset
!wget http://files.deeppavlov.ai/datasets/insults_data.tar.gz
!tar -xzvf "insults_data.tar.gz"
[ ]:
# provide a path to the directory with your train, valid and test files
model_config['dataset_reader']['data_path'] = "./contents/"
Train dataset format¶
Train the model using new config¶
[ ]:
model = train_model(model_config)
Use your model for prediction.
[ ]:
model(['You are kind of stupid', 'You are a wonderful person!'])
['Insult', 'Not Insult']
5.2 Train your model from CLI¶
To train the model on your data, create a copy of a config file and change the data_path variable in it. After that, train the model using your new config_file. You can also change any of the hyperparameters of the model.
[ ]:
!python -m deeppavlov train model_config.json
6. Models list¶
The table presents a list of all of the classification models available in DeepPavlov Library.
Config name |
Language |
Task |
Dataset |
Model Size |
Metric |
Score |
---|---|---|---|---|---|---|
insults_kaggle_bert |
En |
Insults |
1.1 GB |
ROC-AUC |
0.8770 |
|
paraphraser_rubert |
Ru |
Paraphrase |
2.0 GB |
F1 |
0.8738 |
|
paraphraser_convers_distilrubert _2L |
Ru |
Paraphrase |
1.2 GB |
F1 |
0.7396 |
|
paraphraser_convers_distilrubert _6L |
Ru |
Paraphrase |
1.6 GB |
F1 |
0.8354 |
|
sentiment_sst_conv_bert |
En |
Sentiment |
1.1 GB |
Accuracy |
0.6626 |
|
sentiment_twitter |
Ru |
Sentiment |
6.2 GB |
F1-macro |
0.9961 |
|
rusentiment_bert |
Ru |
Sentiment |
1.3 GB |
F1-weighted |
0.7005 |
|
rusentiment_convers_bert |
Ru |
Sentiment |
1.5 GB |
F1-weighted |
0.7724 |
|
topics_distilbert_base_uncased |
En |
Topics |
6.2 GB |
F1-macro |
0.9961 |
7. Simple few-shot classifiers¶
Additionally, in the faq section you can find a config for a fast and simple pre-BERT model, which consists of a fasttext vectorizer and a simple logistic regression classifier.
7.1 Few-shot setting¶
In the current setting the config can be used for few-shot classification - a task, in which only a few training examples are available for each class (usually from 5 to 10). Note that the config takes the full version of the dataset as the input and samples N examples for each class of the train data in the iterator.
The sampling is done within the basic_classification_iterator
component of the pipeline and the shot
parameter defines the number of examples to be sampled. By default the shot
parameter is set to None
(no sampling applied).
7.2 Multiple languages support¶
By default fasttext_logreg
supports classification in English, but can be modified for classification in Russian.
In order to change fasttext_logreg
language to Russian, change LANGUAGE
variable in the metadata.variables
section from en
to ru
and change the Spacy model by changing SPACY_MODEL
variable from en_core_web_sm
to ru_core_news_sm
.
You can do that by directly editing the config file through an editor or change it through Python (example below). N.B. read_json
and find_config
combination is intentionally used instead of parse_config
to read config in the example, because parse_config
will replace all LANGUAGE
and SPACY_MODEL
usages in the config with the default values from metadata.variables
.
[ ]:
from deeppavlov import build_model
from deeppavlov.core.common.file import read_json, find_config
model_config = read_json(find_config('fasttext_logreg'))
model_config['metadata']['variables']['LANGUAGE'] = 'ru'
model_config['metadata']['variables']['SPACY_MODEL'] = 'ru_core_news_sm'
model = build_model(model_config, install=True, download=True)
7.3 Dataset and Scores¶
To demonstrate the performance of the model in two languages, we use the English and Russian subsets of the MASSIVE dataset.
MASSIVE is a parallel dataset of utterrances in 52 languages with annotations for the Natural Language Understanding tasks of intent prediction and slot annotation. We only employ the intent classification data. You can see the results of the given configs in 5-shot classification setting in the table below.
Config name |
Language |
Train accuracy |
Validation accuracy |
Test accuracy |
---|---|---|---|---|
fasttext_logreg |
en |
0.9632 |
0.5239 |
0.5155 |
fasttext_logreg |
ru |
0.9231 |
0.4565 |
0.4304 |