# Russian sarcastic comments detection

## About the project

Sarcasm recognition topic is popular nowadays. A lot of articles devoted
to the sarcasm recognition theme have been published since 2010.
Although there was actually no one of them to be about Russian sarcastic
comments recognition.

This project proposes a solution to the problem of Russian sarcastic
comments detection. A corpus of Russian-language comments was formed
with binary markup: whether the comment is sarcastic or not. The problem
was solved using three natural language processing approaches: TF-IDF
and Logistic Regression, Recurrent Neural Networks, Transformers. A
comparative analysis of the results of these approaches was carried out.

The link to the github repository is [here](https://github.com/123-39/NLP_Project).

## Researchers
The project was prepared by [Ivan Trenev](https://ods.ai/users/f3487c760b79) and [Nadezhda Chaplinskaia](https://ods.ai/users/affffadfbb1d).

## Dataset

To build the corpus of Russian sarcastic and non-sarcastic comments we
decided to use the text-based English news headlines corpus ([here](https://github.com/sankalpg10/NLP-Sarcasm-Detection/tree/main/data/) is a link to the website, where the text-based English news headlines corpus could be downloaded). We have
translated to Russian and then we have cleared all the samples. As the
result we have got 9554 sarcastic and 10479 non-sarcastic translated
samples.

But we had the idea that the corpus we want has to be more
"Russian", has to have some exclusively Russian sarcastic
statements that appears due to Russian mentality. That's why we decided
to use also another corpus &ndash; the corpus of Russian jokes ([here](https://www.kaggle.com/datasets/konstantinalbul/russian-jokes/) is a link to the website, where the corpus of Russian jokes could be downloaded). The corpus is huge: 130204 samples. These jokes we would like to consider as Russian sarcastic statements, because every Russian joke has Russian-sarcastic emotional background.

Finally, the corpus we made consisted of 5000 translated sarcastic
samples and 5000 Russian jokes &ndash; as sarcastic comments (labelled 1),
and 10000 translated non-sarcastic samples as non-sarcastic comments
(labelled 0). Thus we have get the corpus of 20000 balanced comments.

The split of the dataset to the train, valid and test samples is
presented on the following table. 

|               | Train | Valid |  Test  |
| ------------- |:-----:| ----- |:------:|
| **Comments**  | 11200 | 2800  | 6000   |
| **Tokens**    | 190400| 36400 | 120000 |


## Methods and approaches

We decided to use three train models: TF-IDF with Logistic Regression,
Recurrent Neural Network and Transformer.

As for RNN, it was attempted to use a relatively good quality, and most
importantly, lightweight glove model "navec" ([here](https://github.com/natasha/navec) you can find navec glove model). Using the glove
model, we will build a dictionary and tokenize (converting text strings
into sequences of numbers) every comment. In this case, the size of the
vocabulary was limited to 30,000 words.

As for transformer, we used pre-trained BERT type transformer. The
rubert-tiny2 (sentence encoder model) model was chosen for
classification. As a result, the pre-trained tokenizer and model were
loaded from huggingface ([here](https://huggingface.co/cointegrated/rubert-tiny2) you can find rubert-tiny2 model). For classification, it was necessary to add
a fully connected layer.

## Results

The evaluations (F1-score) of presented methods on the testing set are
showed in a folowing table. 

|    Method    | F1-score |
| ------------ |:--------:|
| Log-Reg      | 0.7868   |
| RNN          | 0.7940   |
| Transformer  | 0.8478   |

The third approach &ndash; Transformers &ndash; has the higher score of detection,
than Recurrent Neural Networks which showed the result close to the
score of the baseline Logistic Regression.

In respect that the presented model is the first model capable to detect
Russian sarcasm, the received results are quite encouraging for the
further researches connected to this field.


## References

1. Dmitry Davidov, Oren Tsur, Ari Rappoport *Semi-supervised recognition of
sarcastic sentences in twitter and amazon* // Proceedings of the
Fourteenth Conference on Computational Natural Language Learning,
Association for Computational Linguistics, 2010.

2. Roberto González-Ibánez, Smaranda Muresan, Nina Wacholder *Identifying
sarcasm in Twitter: a closer look* // Proceedings of the 49th Annual
Meeting of the Association for Computational Linguistics: Human Language
Technologies: short papers-Volume 2, Association for Computational
Linguistics, 2011.

3. Stephanie Lukin, Marilyn Walker *Really? well. apparently bootstrapping
improves the performance of sarcasm and nastiness classifiers for online
dialogue* // Proceedings of the Workshop on Language Analysis in Social
Media, 2013.

4. Peng Liu, et al. *Sarcasm detection in social media based on imbalanced classification* // International Conference on Web-Age Information Management, Springer International Publishing, 2014.

5. Francesco Barbieri, Horacio Saggion, Francesco Ronzano *Modelling
sarcasm in twitter, a novel approach* ACL 2014, 2014.

6. Elisabetta Fersini, Federico Alberto Pozzi, Enza Messina *Detecting
irony and sarcasm in microblogs: The role of expressive signals and
ensemble classifiers* // Data Science and Advanced Analytics (DSAA), 2015.

7. Ashwin Rajadesingan, Reza Zafarani, Huan Liu *Sarcasm detection on
twitter: A behavioral modeling approach* // Proceedings of the Eighth
ACM International Conference on Web Search and Data Mining, ACM, 2015.

8. Santosh Kumar Bharti, Korra Sathya Babu, Sanjay Kumar Jena *Parsing-based sarcasm sentiment recognition in Twitter data* // 2015
IEEE/ACM International Conference on Advances in Social Networks
Analysis and Mining (ASONAM). IEEE, 2015.

9. David Bamman, Noah A. Smith *Contextualized sarcasm detection on
twitter* // Ninth International AAAI Conference on Web and Social Media, 2015. 

10. Nabeela Altrabsheh, Mihaela Cocea, Sanaz Fallahkhair *Detecting sarcasm from students' feedback in Twitter* // Design for teaching and learning in a networked world, Springer International Publishing, 2015.

11. Debanjan Ghosh, Weiwei Guo, Smaranda Muresan *Sarcastic or not: word embeddings to predict the literal or sarcastic meaning of words* // Proccedings of the 2015 Conference on Empirical Methods in Natural
Language Processing, Accociation for Computational Linguistics, 2015.

12. Mondher Bouazizi, Tomoaki Otsuki *A pattern-based approach for sarcasm detection on twitter* // IEEE Transl., 2016.

13. S.K. Bharti, et al. *Sarcastic sentiment detection in tweets streamed in real time: a big data approach* // Digital Communications and Networks, 2016.

This project proposes a solution to the problem of Russian sarcastic comments detection. A corpus of Russian-language comments was formed with binary markup: whether the comment is sarcastic or not. The problem was solved using three natural language processing approaches: TF- IDF and Logistic Regression, Recurrent Neural Networks, Transformers. A comparative analysis of the results of these approaches was carried out.

Russian sarcastic comments detection

<h2>Deteсt trolls and bots in YouTube comments</h2><h3>Research project</h3><p>For detection noise information (bot and troll messages) we need to make following:</p><ul><li>Because we don't have labeled data we have to analyze data in order to be able to find method for noise message and usefully &nbsp;information separating. A small study was done to solve this problem. (See <a target="_blank" rel="noopener noreferrer" href="https://github.com/mvlebedev/ods_nlp_huawei_fall_2022/blob/main/Report/report_eng.pdf">report</a> and <a target="_blank" rel="noopener noreferrer" href="https://github.com/mvlebedev/ods_nlp_huawei_fall_2022">repository</a> with project).</li><li>In the second part of project we need to develop machine learning model for troll and bots detection.</li></ul><h3>Team and contribution</h3><ul><li>At the moment, the project has only one main developer (email mixam.maxim@gmail.com), you are free to contact and ask any questions. We may discuss further development.</li></ul><h3><strong>Related Work</strong></h3><p>A short list of related works and their descriptions can be found <a target="_blank" rel="noopener noreferrer" href="https://github.com/mvlebedev/ods_nlp_huawei_fall_2022/blob/main/Report/report_eng.pdf">here</a>.</p><h3>Experiment design</h3><figure class="image"><img src="https://storage.yandexcloud.net/ds-ods/files/content/2023/02/05/abb3b99e/diagram.png"></figure><h3>Dataset</h3><p>The data will be comments from YouTube channel "вДудь". On these videos, the blogger interviews famous actors, politicians, etc. The program for uploading comments to a file is in the <a target="_blank" rel="noopener noreferrer" href="https://github.com/mvlebedev/ods_nlp_huawei_fall_2022">repository.</a></p><h3>Results</h3><p>Three approaches were used to identify troll and bot comments (See pictures below too). The first is topic modeling with <strong>LDA</strong>. The second method of parsing messages is based on the use of embeddings obtained from <strong>RuBert</strong>. Further, for this vector representation, the <strong>GMM</strong> and <strong>K-means</strong> clustering methods were used. Based on the hand review of the data for the corresponding clusters, a conclusion was made about the class number to which the noise messages correspond. The third approach is {t-distributed} stochastic neighbor nesting (<strong>t-SNE</strong>) applied to <strong>RuBert embeddings</strong>. Noise comments were considered points lying on the periphery of a circle of a given radius.As a result, the analysis of texts that were marked as noise (with bots and trolls), we can conclude that thematic modeling showed the worst. There is a lot of information in his texts that does not look like the work of trolls and bots. Methods based on <strong>RuBert embeddings</strong> showed very good results. Approach based on the identification of noise points at the <strong>t-SNE</strong> boundary performance showed good results. This method does not require additional analysis of classes like all other methods. The only thing he selected is less than all the points (objects) for the noise class. This is due to the choice of a large circle radius.&nbsp;</p><h4><strong>Latent Dirichlet Allocation</strong></h4><p>On the following picture is demonstrated distribution of words in topics. We may guess that topic 6 has spam information.</p><figure class="image image_resized" style="width:96.49%;"><img src="https://storage.yandexcloud.net/ds-ods/files/content/2023/02/28/cd8dc79f/themes.png"></figure><h4><strong>t-SNE и GMM (Noise comments: class 3)</strong></h4><figure class="image"><img src="https://storage.yandexcloud.net/ds-ods/files/content/2023/02/28/e2b65574/tsne_gmm.png"></figure><h4><strong>t-SNE и K-means (Noise comments: class 3)</strong></h4><figure class="image"><img src="https://storage.yandexcloud.net/ds-ods/files/content/2023/02/28/20a98479/tsne_kmeans.png"></figure><p>&nbsp;</p><h4><strong>t-SNE and selection of points on the periphery of a circle}</strong></h4><figure class="image"><img src="https://storage.yandexcloud.net/ds-ods/files/content/2023/02/28/2257b77d/tsne_outside_circle.png"></figure>

Max Lebedev

The project devoted to the development of an approach for
definitions of bots and trolls in Russian-language comments on the YouTube service.
It was created as part of a project for the «Natural Language Processing course» by HUAWEI

Deteсt trolls and bots in YouTube comments

# Practical Assignment 2: Text multiclass classification: movie's genre

 The design of an upgraded version of the second practice assignment for the **«Natural Language Processing course» by HUAWEI** company is presented in the current survey. The contribution to this course includes an expanded dataset, four baseline pipelines and a flexible point distribution for student assessment. It is believed that a new assignment will be a valuable opportunity for students to practise their skills in pipeline creation, model fine-tuning and comparison analysis.

[**Link to the repository.**](https://github.com/e0xextazy/nlp_huawei_new2_task)

_This project appeared during the Natural Language Processing course (stream 3, autumn 2022)._

## Motivation
The first practical assignment of the NLP course is related to classical methods and its implementation from scratch, while the second practical task aims at maximizing metrics using arbitrary tools. Almost all real-world projects require developers to fine-tune previously implemented and trained models for specific cases. The skills of choosing parameters, analysing appropriate solutions and fine-tuning them for use cases are essential for implementing comprehensive architectures. Therefore, the idea of the discussed task is indeed created to practise these skills. We implement four baselines based on both state-of-the-art solutions and base algorithms for further detailed implementation by the students. 

## Dataset
The **RT-movie** dataset was compiled for the purpose of facilitating the application of various methods and tools for problem-solving. This dataset was obtained from the [RottenTomatoes](https://www.rottentomatoes.com) movie review website, which features over 300 pages. It includes six classes, as depicted in Figure or in Table below, and is licensed under the [MIT license](https://github.com/e0xextazy/nlp_huawei_new2_task/blob/master/LICENSE), allowing for both commercial and research use.

![](https://raw.githubusercontent.com/e0xextazy/nlp_huawei_new2_task/master/images/class_distribution.png) 

| Label |Percentage, in \%|
|:-:|:-:|
|Comedy|23|
|Drama|21.8|
|Horror|17.7|
|Mystery|14.4|
|Action|12.8|
|Kids|10.3|

The **general statistics** of datasets are in Table below. The distribution of the **number of characters**, the **number of words**, the **most common stop words** and the **most common words** can be seen in the Figures below.

|   |Train|Test|
|:-:|:-:|:-:|
|Articles|5495|788|
|Words|327357|47711|
|Vocabulary size|48721|48721|

![](https://raw.githubusercontent.com/e0xextazy/nlp_huawei_new2_task/master/images/num_chars.png)
![](https://raw.githubusercontent.com/e0xextazy/nlp_huawei_new2_task/master/images/num_words.png)
![](https://raw.githubusercontent.com/e0xextazy/nlp_huawei_new2_task/master/images/common_stopwords.png) 
![](https://raw.githubusercontent.com/e0xextazy/nlp_huawei_new2_task/master/images/word_cloud.png)

## Metrics

The main metric by which students will be evaluated is **Accuracy**:
![](https://raw.githubusercontent.com/e0xextazy/nlp_huawei_new2_task/master/images/accuracy.png)
![](https://raw.githubusercontent.com/e0xextazy/nlp_huawei_new2_task/master/images/exact_match.png)
where _N_ is the number of samples in the test dataset.

## Baselines

In all implementations, we used the `LabelEncoder` from the [scikit-learn](https://scikit-learn.org/) library to translate a categorical target variable into a numeric data type.

### Logistic Regression

To get a vector representation of the text, we trained `TfidfVectorizer` from the scikit-learn library with parameters:

```Python
vect_word = TfidfVectorizer(
    max_features=100,
    lowercase=True,
    analyzer="word",
    stop_words="english",
    ngram_range=(1, 3),
    dtype=np.float32
)
```

Next, using vector representations of the text, we trained Logistic regression with parameters:

```Python
logreg = LogisticRegression(
    C=2,
    random_state=42,
    max_iter=100
)
```

We did not use validation set and early stopping of training, as we limited the maximum number of training steps.

### Catboost

We trained Catboost Classifier with parameters:

```Python
cbc = CatBoostClassifier(
    iterations=100,
    depth=5,
    random_seed=42
)
```

We used the description of the films as a text feature into the model.

### LSTM

For the validation set, we took 15% of all data and used the whole word as a token.  Our dictionary consists of 30,000 of the most popular tokens in our corpus.  And we get vector representations from a pre-trained Glove model: `"glove-wiki-gigaword-100"`, where each vector has a dimension of 100. 

Below are the configurations for the model and training, respectively:

```Python
config = {
    "freeze": False,
    "cell_type": "LSTM",
    "cell_dropout": 0.2,
    "num_layers": 2,
    "hidden_size": 128,
    "out_activation": "relu",
    "bidirectional": True,
    "out_dropout": 0.2,
    "out_sizes": [200],
}
```
 
```Python
trainer_config = {
    "lr": 3e-3,
    "n_epochs": 10,
    "weight_decay": 1e-6,
    "batch_size": 128,
    "device": "cuda" if torch.cuda.is_available() else "cpu"
}
```

### DistilBERT

We split the training set into train and val in the ratio of 85\% and 15\%, respectively, and for each text in both sets we received tokens by using a pre-trained tokenizer `distilbert-base-uncased` with parameters: `truncation=True, do_lower_case=True`. For the input sequence, we used a maximum length equal to 128 tokens and combined the examples into a batch of 64 elements each. The model had 6 classes and a dropout probability of 0.3. For training, we used the following configuration:

```Python
trainer_config = {
    "lr": 3e-4,
    "n_epochs": 2,
    "weight_decay": 1e-6,
    "batch_size": 64,
    "device": "cuda" if torch.cuda.is_available() else "cpu"
}
```

## Results

For each of the baseline models, we were able to obtain a target metric — Accuracy, as shown in the table below.

|Baseline|Accuracy|
|---|:-:|
|LogReg|0.36294|
|CatBoost|0.59137|
|LSTM|0.52284|
|DistilBERT|0.61294|

Mark Baushenko

Artyom Boldinov

Milana Shkhanukova

Мы подготовили новое практические задание №2 с учетом новых архитектур

Новое практическое задание для курса HUAWEI NLP

<h2>Stocks Tweets Project</h2><p>This project is devoted to the study of the relationship between publications in social networks and the corresponding stock prices.&nbsp;</p><p>All code is open-source, here is the link to GitHub: <a target="_blank" rel="noopener noreferrer" href="https://github.com/fin-algo-lake-ai/stocks-tweets-project">https://github.com/fin-algo-lake-ai/stocks-tweets-project</a></p><h3>Dataset</h3><p>At the first stage of the project, the original dataset from the article [Jaggi M. at al., 2021] was taken for work. It contains 6.4 million tweets for 25 stocks.&nbsp;</p><p>After examining the dataset a number of issues had been found and fixed:&nbsp;</p><ul><li>added stop words preprocessing (initially stop words were just removed)</li><li>corrected filling missing data algorithm to more stable one</li><li>data labels were recalculated in a different way (with isolation in time of the message release date and the price change date)<br>&nbsp;</li></ul><h3>Experiment setup</h3><p>After loading the dataset, it is divided into train and test sets in a ratio of 0.85 to 0.15. To check the stability and variability of each model, the split is carried out several times with a different seed value. Each partition is tested independently of the others.</p><p>Of all the dataset fields, only one is used - the preprocessed message text, the rest of the features (date, time, author, etc.) are intentionally not used in the pipeline of this work to avoid distortion in the search for the correlation of the entities under study.</p><p>To estimate model quality, we used the Accuracy metric. As the dataset is perfectly balanced, this metric is very intuitive, as the value 0.5 corresponds to random noise.</p><p>After training and saving all the models, an ensemble of models is created to evaluate the cross-correlation of the models and obtain the final maximum achievable accuracy value.<br>&nbsp;</p><h3>Results</h3><p>Below is a final table and graph with results of individual models and their ensembles.</p><figure class="table" style="float:left;"><table><tbody><tr><td style="background-color:#9fc5e8;border:1pt solid #000000;padding:5pt;"><p style="text-align:center;"><strong>Model Type</strong></p></td><td style="background-color:#9fc5e8;border:1pt solid #000000;padding:5pt;"><p style="text-align:center;"><strong>Accuracy</strong></p></td><td style="background-color:#9fc5e8;border:1pt solid #000000;padding:5pt;"><p style="text-align:center;"><strong>Model</strong></p><p style="text-align:center;"><strong>File Size</strong></p></td><td style="background-color:#9fc5e8;border:1pt solid #000000;padding:5pt;"><p style="text-align:center;"><strong>Comments</strong></p></td></tr><tr><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;"><p>Baseline</p><p>(Random Guess)</p></td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">0.500 +- 0.003</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">-</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">It’s not an actual model (DummyClassifier with strategy="uniform")</td></tr><tr><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">Transformer (Roberta)</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">0.545 +- 0.002</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">350 MB</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">Roberta and DistilRoberta had the same accuracy in our experiments</td></tr><tr><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">GBDT (Catboost)</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">0.560 +- 0.003</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">5 MB</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;"><p>Model details: CatBoostClassifier</p><p>n_estimators=300, max_depth = 8</p></td></tr><tr><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">Naive Bayes</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">0.568 +- 0.002</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">58 MB + 49 MB</td><td style="border:1pt solid #000000;padding:5pt;vertical-align:top;">Model details: MultinomialNB on TF-IDF features for word n-grams with length from 1 to&nbsp; 3, alpha = 0.1</td></tr></tbody></table></figure><p><img src="https://lh4.googleusercontent.com/-JpcdTjg_XplXbavyguxB2ZMhK4kylOX7HrA9o1mXiwmQ5XB9O7ECZChm9XB7GdVh15UWKtQUkwCDubqSQbSC-SQ78q6mGoPH2--QjUBKYMcjQhnqLp1uXxf1voqmALWDt4c7hCd-eG0CLK-Abq9BuOw2NZVAo8urLxVR-ntDrbbg4FnVlQgFTZ-4DjcMA"></p><p style="text-align:center;">Boxplot graphs for pairwise and triple ensembles.<br><i>(BL - baseline model (random uniform), CB - CatBoost model</i><br><i>NB - Naive Bayes model, TR - Transformer model)</i></p><p>&nbsp;</p><h3>Conclusion and future work</h3><p>The results turned out to be promising - despite the complexity of the subject area (predicting future price behavior), we managed to get statistically significant accuracy values (0.582) with a random baseline of 0.500 + -0.003. A separate interesting conclusion was that the size and algorithmic complexity of machine learning models do not always determine the result of their application in each specific case, requiring researchers to conduct more versatile experiments at the modeling stage.</p><p>Possible next steps:</p><ul><li>Calculate business-oriented metrics (closer to the real potential usage of the idea)</li><li>Labels: Take not just the next calendar day, but the next N days (up to a week). Also it's possible to experiment with the 0.5% threshold that is used for class label calculation.</li><li>Models: Complete results with RNN approaches (LSTM / GRU) - the main runs have already been carried out, they need to be added to the final ensemble.</li></ul><p>&nbsp;</p><h3>References</h3><ol><li>Yilmaz, E. S., Ozpolat, A., &amp; Destek, M. A. (2022). Do Twitter sentiments really effective on energy stocks? Evidence from the intercompany dependency. Environmental Science and Pollution Research, 29(52), 78757-78767.<br><a target="_blank" rel="noopener noreferrer" href="https://www.academia.edu/download/89681397/s11356-022-21269-9.pdf">https://www.academia.edu/download/89681397/s11356-022-21269-9.pdf</a></li><li>Jaggi, M., Mandal, P., Narang, S., Naseem, U., &amp; Khushi, M. (2021). Text Mining of Stocktwits Data for Predicting Stock Prices. arXiv preprint arXiv:2103.16388.<br><a target="_blank" rel="noopener noreferrer" href="https://arxiv.org/abs/2103.16388">https://arxiv.org/abs/2103.16388</a></li><li>Rahul Pandey R. (2021) Common Pitfalls to Avoid in Forecasting Models for Stock Price Prediction<br><a target="_blank" rel="noopener noreferrer" href="https://medium.com/geekculture/common-pitfalls-to-avoid-in-forecasting-models-for-stock-price-prediction-3a7c3ff8b80">https://medium.com/geekculture/common-pitfalls-to-avoid-in-forecasting-models-for-stock-price-prediction-3a7c3ff8b80</a></li><li>Adusumilli, R. (2020). NLP in the Stock Market. Leveraging sentiment analysis on 10-k fillings as an edge<br><a target="_blank" rel="noopener noreferrer" href="https://towardsdatascience.com/nlp-in-the-stock-market-8760d062eb92">https://towardsdatascience.com/nlp-in-the-stock-market-8760d062eb92</a></li><li>Sun, Y., Liu, X., Chen, G., Hao, Y., &amp; Zhang, Z. J. (2020). How mood affects the stock market: Empirical evidence from microblogs. Information &amp; Management, 57(5), 103181.<br><a target="_blank" rel="noopener noreferrer" href="https://www.sciencedirect.com/science/article/pii/S0378720618307183">https://www.sciencedirect.com/science/article/pii/S0378720618307183</a></li><li>Yuz, T. (2018). A Sentiment Analysis Approach to Predicting Stock Returns<br><a target="_blank" rel="noopener noreferrer" href="https://medium.com/@tomyuz/a-sentiment-analysis-approach-to-predicting-stock-returns-d5ca8b75a42">https://medium.com/@tomyuz/a-sentiment-analysis-approach-to-predicting-stock-returns-d5ca8b75a42</a></li></ol><h3><br>Contacts and feedback</h3><p>Please, feel free to contact us at this Telegram group: <a target="_blank" rel="noopener noreferrer" href="https://t.me/tweets_for_stocks_feed">https://t.me/tweets_for_stocks_feed</a></p><p>&nbsp;</p>

Finding correlation between tweets and future stock prices using ML/NLP approaches
[this is a graduation project of the course "Natural Language Processing course, stream 3, autumn 2022"]

Tweets for Stocks

# Main goals:
- To create a sign language search engine that enables users to easily find sign language content.(MVP - Minimal Viable Product)
- To create an interactive sign language tutoring system that effectively teaches users sign language.(Medium Viable Product)
- To create a machine translation system that accurately and efficiently translates between sign languages and spoken or written language, making communication more accessible for users.(Maximal Viable Product)
### Terms & Definitions
- Minimal Viable Product - The minimum set of features necessary to test the basic functionality of the sign language interface.
- Medium Viable Product - An expanded version of the MVP with additional features to improve usability and functionality.
- Maximal Viable Product - The final version of the product, with all the features and capabilities required for full functionality.


# The tasks can include

- Develop sign language generation and recognition system

- Develop a tool for collecting and labeling sign language data, and create a dataset using this tool

- Use the tool to mark up sign language data, such as videos or images, with linguistic and grammatical information, such as signs, movements, handshape, hand orientation, location, and facial expression

- Develop a sign language search engine that allows users to search for sign language content.

- Develop an interactive sign language tutoring system

- Continuously monitor and analyse the performance of the system to identify areas for improvement.

- Design a UI/UX interface that is easy to navigate and understand for sign language users.

### App task examples:

Data labeling
- Sign language data labeler uses the tool [youtube](https://youtu.be/NIa3q5zdT-U)
- Collection of sign language sentence at NGTU NETI [youtube](https://youtu.be/oLZsv8j_nJ4)

Sign language recognition
- Recognition of Russian fingerspelling [youtube](https://youtu.be/NxdVHfo_ZLo)
- Recognize hand configurations-orientations to present examples of words [youtube](https://youtu.be/VRlgWQFEUzA)
- Static one-handed gestures to present examples of words [youtube](https://youtu.be/qQwLTXNp9DY)
- In the context of the course 'Designing a Machine Learning System', a Telegram bot is demonstrated that is able to recognize the shape of the hand and then match it to the corresponding word in Russian Sign Language. [youtube](https://www.youtube.com/live/3pJSXQuBd-s?feature=share) (see 4:42:37-5:03:16)

UI/UX layout
- Random gesture selection [youtube](https://youtu.be/oOWatIspYwE)
- Search for gesture by word [youtube](https://youtu.be/V1XYH2L7Gfo)
- Sequential display of selected gesture parameters for gesture search [youtube](https://youtu.be/lryv_njvHtE)
- Search for gesture by photo with its performance [youtube](https://youtu.be/jcAkCfhPKls)
- AdaptisCon#2 in Novosibirsk - Presentation of the Russian sign language dictionary [youtube](https://www.youtube.com/live/oKyXMrA6ghs?feature=share) (see 1:48:10-2:22:14 of our speaker)

### Site, bot & mobile apps
- https://signlanguages.ru/
- https://t.me/slsru_bot (In developing)
- iOS app "Signlanguage" [AppStore](https://apps.apple.com/ru/app/signlanguages/id1602656863)
-  Android app "Signlanguage" [PlayMarket](https://play.google.com/store/apps/details?id=com.nstu.rsl)

# Social media:
- [github](https://github.com/orgs/sign-language-interface)
- [vk](https://vk.com/signlanguages)
- [telegram](https://t.me/slsru)
- [instagram](https://www.instagram.com/signlanguages.ru)

# About
Email: alexeyayaya@gmail.com - Alexey Prikhodko

- Alexey Prikhodko - Domain expert, full-stack developer  
- Julia Prikhodko - Sign language data labeler and informer
- Mark Belousov - Frontend-developer, system analyst
- Yana Kondratenko - ML developer

Sign Language Interface project aims to improve accessibility of technology for individuals who use sign languages by incorporating sign language recognition, generation, translation, and tutoring functionality into UI/UX design and AI systems.

Sign language interface

# Medical transcriptions classification
[github repository](https://github.com/evgainy/HMLP)

## About 
In this project, we have implemented a model to correctly classify the medical diagnosis based on the given medical transcriptions. The main goal is to correctly classify the medical specialties based on the transcription text and suggest our approach to solving this problem.

## Researchers
[Anastasia](https://github.com/apiskunova1): created a text corpus, developed the programm code, prepared the report.  
[Arina](https://github.com/Arina3103): developed the programm code and prepared the report.  
[Evgeniy](https://github.com/evgainy): developed the programm code and prepared the report.

## Dataset
Medical data is extremely hard to find due to HIPAA privacy regulations. But https://www.mtsamples.com/ is contained sample transcription reports for many specialties and different work types. This webcite is designed to give access to a large collection of transcribed medical reports. These reports can be used by learning, as well as working medical transcriptionists for their daily transcription needs.
Data collection was carried out manually by copying data from the site. Thus, 4314 medical transcriptions from various fields of medicine were obtained. This dataset contains six columns: ’description’, ’medical specialty’, ’sample
name’, ’transcription’, and ’keywords’.

As part of pre-processing, we have filtered out the categories which have more than 50 samples, so the number of categories got reduced from 33 to 17.
There are a huge number of records belonging to the class ’Consult - History and Phy.’, which is almost thrice when compared with some of the other classes in the dataset. Since we are trying to classify the medical specialities based on medical transcriptions, we need only the ’transcription’ and ’medical specialty’ columns in the dataset.

Then we transformed all the texts to lower case, deleted punctuations, removed stop words, performed lemmatization and tokenization.

## Results
After applying the methods of SVM and multiclass logistic regression, we get the following results. As we can see in the table, the best classification result is achieved by improving the dataset.

|Model |Precision |Recall |F1 score| Accuracy(%)|
|-------|-------------|------|--|--|
|SVC+SMOTE| 0.56| 0.45                       |0.47|0.63|
|LR+SMOTE  | 0.54   | 0.42                    |0.45|0.62|
| model-impr.+SVC| 0.79| 0.75              |0.76|0.81|
| model-impr.+LR |0.78| 0.67                 |0.71|0.78|
| model-impr.+SMOTE+SVC| 0.82| 0.76|0.79|0.85|
| model-impr.+SMOTE+LR| 0.81| 0.72.  |0.75|0.84|

There is a study that used data similar to ours. Most likely, the dataset was collected from the same site as ours. In any case, the result of this study is much worse than ours (the best result is 0.65). In addition, they used completely different classification methods. Thus, we can claim that our model shows the best training outcomes on medical transcription data.

Anastasia

Evgenesh

Virginia

Final project for Natural Language Processing course (stream 3, autumn 2022)

Medical transcriptions classification

<p>Этот проект появился при прохождении курса Natural Language Processing course (stream 3, autumn 2022).</p><p>При изучении некоторых тем в начальных классах школ на предмете «Русский язык» случаются проблемы, что нет достаточного количества подготовленных материалов для учителей. Примером такой темы является «функциональные стили речи». Данная работа подразумевает под собой несколько логических этапов: сбор и разметка датасета, очистка и предобработка датасета, выбор на основе сравнительного анализа наилучшей модели решающей данную задачу. &nbsp;Для создания размеченного датасета использовались популярные источники информации: новостные сайты, литературные произведения, архив научных публикация и статей и т.д. В результате обучения моделей на собранном датасете получились итересные результаты. &nbsp;В данной работе было реализовано 2 модели, которые основаны на разных подходах: классическое машинной обучения и методы глубокого обучения.</p><p><a target="_blank" rel="noopener noreferrer" href="https://github.com/d3vyatk4ru/style-sentences-clf">Ссылка на репозиторий с проектом</a></p>

Andrew

Daniil

Predictions of functional styles of sentence speech in Russian language

The training models methodology for predicting the rate of cryptocurrency against the dollar two weeks ahead from the current date, so the models are rebuild on a daily basis. Model training uses the trends data from Yahoo Finance. List of cryptocurrencies: BTC, ETH, LTC, BNB, XMR, ATOM. But this list can be extended.
The training of model takes into account the news sentiment about cryptocurrency. Historical news are taken from the news sites and the actual news from the rss feeds.

The value of the cryptocurrency rate is predicted for the next 14 days. Data discretization - one day. Interval forecast, the range of acceptable values of the forecasted exchange rate is indicated.

<p>The public part of the project is a telegram bot @cryptonitto_bot<strong> (https://t.me/cryptonitto_bot)</strong></p><p>The bot menu is shown on the figure below.</p><figure class="image"><img src="https://storage.yandexcloud.net/ds-ods/files/content/2023/01/13/1eabd43c/bot_menu.jpg"></figure>

<p>We are developing a unique forecasting method using the combination of time series prediction with news sentiment analysis.</p><p>A special service for collecting news of the cryptomarket has been developed.</p><p>The news structure is shown on the picture below.</p><figure class="image"><img src="https://storage.yandexcloud.net/ds-ods/files/content/2023/01/13/7a25dab2/news_infrastructure.jpg"></figure><p>The view of news parser is shown on the picture below.<img src="https://storage.yandexcloud.net/ds-ods/files/content/2023/01/13/7017bed9/feed_parser.png">he picture</p><p>&nbsp;</p>

<p>the Prophet library has been used in the pilot version of the predictive model .</p>

<p>The project appeared on the ML System Design Course by <a href="ODS.ai">ODS.ai</a></p><p>Git repository&nbsp;<a href="github.com/lunaplush/crypto/"> </a><a target="_blank" rel="noopener noreferrer" href="https://github.com/lunaplush/crypto/">https://github.com/lunaplush/crypto/</a></p><p>Team &nbsp;Aleksandra Maksimova and &nbsp;Dmitry Kotsarev</p><p>The Donbass Data Science community &nbsp;<a href="vk.com/DonbassDataScience ">vk.com/DonbassDataScience </a>we are supported by</p><p>Project support on <a href="dondata.ru">dondata.ru</a></p><p>Email: maximova.alexandra@mail.ru</p>

Cryptonitto is an assistance analytic system for a forecasting of crypto currency price movement on a daily time frames

Cryptocurrency rate forecasting

<p>Manifold aware <code>pytorch.optim</code>.</p>

<p>Unofficial implementation for <a href="https://openreview.net/forum?id=r1eiqi09K7" rel="noopener noreferrer" target="_blank">&quot;Riemannian Adaptive Optimization&nbsp;</a><a href="https://openreview.net/forum?id=r1eiqi09K7" rel="noopener noreferrer" target="_blank">Methods&quot;</a>&nbsp;at ICLR2019 and more.</p>

<p>Guthub page: <a href="https://github.com/geoopt/geoopt" rel="noopener noreferrer" target="_blank">https://github.com/geoopt/geoopt</a></p>

<h2>Installation</h2>

<ol>
<li>GitHub (preferred so far) due to active development
<pre>
<code>pip install git+https://github.com/geoopt/geoopt.git</code></pre>
</li>
<li>pypi (this might be significantly behind master branch)
<pre>
<code>pip install geopt</code></pre>

<ol>
</ol>
</li>
</ol>

<p>The preferred way to install geoopt will change once stable project stage is achieved.&nbsp;Now, pypi is behind master as we actively develop and implement new features.</p>

<h2>PyTorch Support</h2>

<p>Geoopt officially supports <strong>2 latest stable versions</strong> of pytorch upstream or the latest major release.</p>

<h2>What is done so far?</h2>

<p>Work is in progress but you can already use this. Note that API might&nbsp;change in future releases.</p>

<h4>Tensors</h4>

<ul>
<li>
<p><code>geoopt.ManifoldTensor</code> -- just as torch.Tensor with additional <code>manifold</code> keyword argument.</p>
</li>
<li>
<p><code>geoopt.ManifoldParameter</code>&nbsp;-- same as above, recognized in torch.nn.Module.parameters as correctly subclassed.</p>
</li>
</ul>

<p>All above containers have special methods to work with them as with&nbsp;points on a certain manifold</p>

<ul>
<li><code>.proj_()</code>&nbsp;--&nbsp;inplace projection on the manifold.</li>
<li><code>.proju(u)</code>&nbsp;--&nbsp;project vector <code>u</code> on the tangent space. You need&nbsp;to project all vectors for all methods below.</li>
<li><code>.egrad2rgrad(u)</code>&nbsp;--&nbsp;project gradient <code>u</code> on Riemannian manifold</li>
<li><code>.inner(u, v=None)</code>--&nbsp;inner product at this point for two&nbsp;<strong>tangent</strong> vectors at this point. The passed vectors are not<br />
&nbsp; projected, they are assumed to be already projected.</li>
<li><code>.retr(u)</code>--&nbsp;retraction map following vector <code>u</code></li>
<li><code>.expmap(u)</code>--&nbsp;exponential map following vector <code>u</code> (if expmap is not available in closed form, best approximation is used)</li>
<li><code>.transp(v,u)</code>--&nbsp;transport vector <code>v</code> &nbsp;with direction <code>u</code></li>
<li><code>.retr_transp(v,u)</code>--&nbsp;transport <code>self</code>, vector <code>v&nbsp;</code>(and possibly more vectors) with direction <code>u</code>&nbsp;(returns are plain tensors)</li>
</ul>

<h3>Manifolds</h3>

<ul>
<li><code>geoopt.Euclidean</code> -- unconstrained manifold in <code>R</code> with&nbsp;Euclidean metric</li>
<li><code>geoopt.Stiefel</code> --&nbsp;Stiefel manifold on matrices&nbsp;&nbsp;<code>A in R^{n x p} : A^t A=I, n &gt;= p</code></li>
<li><code>geoopt.Sphere</code> -- Sphere manifold <code>||x||=1</code></li>
<li><code>geoopt.BirkhoffPolytope</code> --&nbsp;manifold of Doubly Stochastic matrices</li>
<li><code>geoopt.Stereographic</code> --&nbsp;Constant curvature stereographic projection model</li>
<li><code>geoopt.SphereProjection</code> --&nbsp;Sphere stereographic projection model</li>
<li><code>geoopt.PoincareBall</code> --&nbsp;<a href="https://en.wikipedia.org/wiki/Poincar%C3%A9_disk_model" rel="noopener noreferrer" target="_blank">Poincare ball model</a></li>
<li><code>geoopt.Lorentz</code> --&nbsp;<a href="https://en.wikipedia.org/wiki/Hyperboloid_model" rel="noopener noreferrer" target="_blank">Hyperboloid model</a></li>
<li><code>geoopt.ProductManifold</code> --&nbsp;Product manifold constructor</li>
<li><code>geoopt.Scaled</code> --&nbsp;Scaled version of the manifold. Similar to <a href="https://openreview.net/forum?id=HJxeWnCcF7" rel="noopener noreferrer" target="_blank">Learning Mixed-Curvature Representations in Product Spaces</a> if combined with <code>ProductManifold</code></li>
<li><code>geoopt.SymmetricPositiveDefinite</code> --&nbsp;SPD matrix manifold</li>
<li><code>geoopt.UpperHalf</code> --&nbsp;Siegel Upper half manifold. Supports Riemannian and Finsler metrics, as in <a href="https://arxiv.org/abs/2106.04941" rel="noopener noreferrer" target="_blank">Symmetric Spaces for Graph Embeddings: A Finsler-Riemannian Approach</a></li>
<li><code>geoopt.BoundedDomain</code> --&nbsp;Siegel Bounded domain manifold. Supports Riemannian and Finsler metrics.</li>
</ul>

<p>All manifolds implement methods necessary to manipulate tensors on manifolds and&nbsp;tangent vectors to be used in general purpose. See more in `documentation`.</p>

<h2>Optimizers</h2>

<ul>
<li><code>geoopt.optim.RiemannianSGD</code> --&nbsp;a subclass of <code>torch.optim.SGD</code>&nbsp;with the same API and a Sparse version</li>
<li><code>geoopt.optim.RiemannianAdam&nbsp;</code>--&nbsp;a subclass of <code>torch.optim.Adam</code> and a Sparse version</li>
</ul>

<h2>Samplers</h2>

<ul>
<li><code>geoopt.samplers.RSGLD</code> -- Riemannian Stochastic Gradient Langevin&nbsp;Dynamics</li>
<li><code>geoopt.samplers.RHMC</code> --&nbsp;Riemannian Hamiltonian Monte-Carlo</li>
<li><code>geoopt.samplers.SGRHMC</code> --&nbsp;Stochastic Gradient Riemannian&nbsp;Hamiltonian Monte-Carlo</li>
</ul>

<h2><br />
Citing Geoopt</h2>

<p>If you find this project useful in your research, please kindly add this bibtex entry in references and cite.</p>

<pre>
<code>@misc{geoopt2020kochurov,</code>
<code>    title={Geoopt: Riemannian Optimization in PyTorch},</code>
<code>    author={Max Kochurov and Rasul Karimov and Serge Kozlukov},</code>
<code>    year={2020},</code>
<code>    eprint={2005.02819},</code>
<code>    archivePrefix={arXiv},</code>
<code>    primaryClass={cs.CG}</code>
<code>}</code></pre>

Maxim Kochurov

PyTorch framework for Riemannian Deep Learning

Geoopt

<h2><code>Optimum Transformers</code></h2><p><code>Accelerated NLP pipelines for fast inference&nbsp;</code></p><figure class="image image_resized" style="width:20px;"><img src="https://emoji.discourse-cdn.com/apple/rocket.png?v=12" alt=":rocket:"></figure><p><code>on CPU and GPU. Built with&nbsp;</code></p><figure class="image image_resized" style="width:20px;"><img src="https://emoji.discourse-cdn.com/apple/hugs.png?v=12" alt=":hugs:"></figure><p><code>Transformers, Optimum and ONNX runtime.</code></p><figure class="image"><img src="https://img.shields.io/github/stars/AlekseyKorshuk/optimum-transformers?style=social" alt="GitHub stars"></figure><h2><code>Disclaimer</code></h2><p><code>This project is my inspiration of </code><a target="_blank" rel="noopener noreferrer" href="https://huggingface.co/infinity"><code>Huggingface Infinity 3</code></a><code>. And first step done by </code><a target="_blank" rel="noopener noreferrer" href="https://twitter.com/psuraj28"><code>Suraj Patil</code></a><code>.</code></p><blockquote><p><a target="_blank" rel="noopener noreferrer" href="https://twitter.com/huggingface"><code>@huggingface</code></a><code>’s pipeline API is awesome!</code></p><figure class="image image_resized" style="width:20px;"><img src="https://emoji.discourse-cdn.com/apple/star_struck.png?v=12" alt=":star_struck:"></figure><p><code>, right? And onnxruntime is super fast !</code></p><figure class="image image_resized" style="width:20px;"><img src="https://emoji.discourse-cdn.com/apple/rocket.png?v=12" alt=":rocket:"></figure><p><code>. Wouldn’t it be great to combine these two?</code><br><code><i>– Tweet by Suraj Patil</i></code></p></blockquote><p><code>It was under this slogan that I started doing this project!</code></p><blockquote><p><code>And the main goal was to show myself to get into <strong>@huggingface</strong> team</code></p><figure class="image image_resized" style="width:20px;"><img src="https://emoji.discourse-cdn.com/apple/hugs.png?v=12" alt=":hugs:"></figure></blockquote><h2><code>How to use</code></h2><h3><code>Quick start</code></h3><p><code>The usage is exactly the same as original pipelines, except minor improves:</code></p><pre><code class="language-bash">from optimum_transformers import pipeline

pipe = pipeline("text-classification", use_onnx=True, optimize=True)
pipe("This restaurant is awesome")
# [{'label': 'POSITIVE', 'score': 0.9998743534088135}]
</code></pre><ul><li><code><strong>use_onnx</strong>- converts default model to ONNX graph</code></li><li><code><strong>optimize</strong> - optimizes converted ONNX graph with </code><a target="_blank" rel="noopener noreferrer" href="https://huggingface.co/docs/optimum/index"><code>Optimum 8</code></a></li></ul><h3><code>Optimum config</code></h3><p><code>Read </code><a target="_blank" rel="noopener noreferrer" href="https://huggingface.co/docs/optimum/index"><code>Optimum 8</code></a><code> documentation for more details</code></p><pre><code class="language-bash">from optimum_transformers import pipeline
from optimum.onnxruntime import ORTConfig

ort_config = ORTConfig(quantization_approach="dynamic")
pipe = pipeline("text-classification", use_onnx=True, optimize=True, ort_config=ort_config)
pipe("This restaurant is awesome")
# [{'label': 'POSITIVE', 'score': 0.9998743534088135}]
</code></pre><h2><code>Benchmark</code></h2><h3><code>With notebook</code></h3><p><code>You can benchmark pipelines easier with </code><a target="_blank" rel="noopener noreferrer" href="https://github.com/AlekseyKorshuk/optimum-transformers/blob/master/notebooks/benchmark_pipelines.ipynb"><code>benchmark_pipelines 3</code></a><code> notebook.</code></p><h3><code>With own script</code></h3><pre><code class="language-bash">from optimum_transformers import Benchmark

task = "sentiment-analysis"
model_name = "philschmid/MiniLM-L6-H384-uncased-sst2"
num_tests = 100

benchmark = Benchmark(task, model_name)
results = benchmark(num_tests, plot=True)
</code></pre><h3><code>Results</code></h3><blockquote><p><code>Note: These results were collected on my local machine. So if you have high performance machine to benchmark, please contact me</code></p><figure class="image image_resized" style="width:20px;"><img src="https://emoji.discourse-cdn.com/apple/hugs.png?v=12" alt=":hugs:"></figure></blockquote><h4><code><strong>sentiment-analysis</strong></code></h4><p><code>Almost the same as in </code><a target="_blank" rel="noopener noreferrer" href="https://www.youtube.com/watch?v=jiftCAhOYQA"><code>Inifinity launch</code></a><code> video</code></p><figure class="image image_resized" style="width:20px;"><img src="https://emoji.discourse-cdn.com/apple/hugs.png?v=12" alt=":hugs:"></figure><p><code>AWS VM: g4dn.xlarge</code><br><code>GPU: NVIDIA T4</code><br><code>128 tokens</code><br><code>2.6 ms</code></p><p>&nbsp;</p><figure class="image"><img src="https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/b/b0bbebce9d8186be0d20ecdbbc34ca3e327cb9fb_2_666x500.jpeg" alt="Resulting plot" srcset="https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/b/b0bbebce9d8186be0d20ecdbbc34ca3e327cb9fb_2_666x500.jpeg, https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/b/b0bbebce9d8186be0d20ecdbbc34ca3e327cb9fb_2_999x750.jpeg 1.5x, https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/original/2X/b/b0bbebce9d8186be0d20ecdbbc34ca3e327cb9fb.jpeg 2x" sizes="100vw" width="666"></figure><p>&nbsp;</p><h4><code><strong>zero-shot-classification</strong></code></h4><p><code>With typeform/distilbert-base-uncased-mnli</code></p><p>&nbsp;</p><figure class="image"><img src="https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/d/d673c795ad1e6547cf92c5e4951371d8c61eb8bd_2_666x500.jpeg" alt="Resulting plot" srcset="https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/d/d673c795ad1e6547cf92c5e4951371d8c61eb8bd_2_666x500.jpeg, https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/d/d673c795ad1e6547cf92c5e4951371d8c61eb8bd_2_999x750.jpeg 1.5x, https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/original/2X/d/d673c795ad1e6547cf92c5e4951371d8c61eb8bd.jpeg 2x" sizes="100vw" width="666"></figure><p>&nbsp;</p><h4><code><strong>token-classification</strong></code></h4><p>&nbsp;</p><figure class="image"><img src="https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/e/e66b1027e2c2b9af57be5e376402fb3d37cd3a43_2_666x500.jpeg" alt="Resulting plot" srcset="https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/e/e66b1027e2c2b9af57be5e376402fb3d37cd3a43_2_666x500.jpeg, https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/e/e66b1027e2c2b9af57be5e376402fb3d37cd3a43_2_999x750.jpeg 1.5x, https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/original/2X/e/e66b1027e2c2b9af57be5e376402fb3d37cd3a43.jpeg 2x" sizes="100vw" width="666"></figure><p>&nbsp;</p><p>&nbsp;</p><figure class="image"><img src="https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/0/0c4ee5c9d3d4c65cda80668164f1520548a950d7_2_666x500.jpeg" alt="Resulting plot" srcset="https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/0/0c4ee5c9d3d4c65cda80668164f1520548a950d7_2_666x500.jpeg, https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/optimized/2X/0/0c4ee5c9d3d4c65cda80668164f1520548a950d7_2_999x750.jpeg 1.5x, https://aws1.discourse-cdn.com/standard14/uploads/hellohellohello/original/2X/0/0c4ee5c9d3d4c65cda80668164f1520548a950d7.jpeg 2x" sizes="100vw" width="666"></figure><p>&nbsp;</p><p><code>More results are available in project repository: </code><a target="_blank" rel="noopener noreferrer" href="https://github.com/AlekseyKorshuk/optimum-transformers"><code>GitHub 13</code></a><code>.</code></p><p><br><br>&nbsp;</p><p>&nbsp;</p><p>&nbsp;</p>

Aleksey Korshuk

Accelerated NLP pipelines for fast inference on CPU and GPU

Projects

All Projects