Ana(stasiia) Iurshina

The NIL-linking task in Entity Linking deals with cases where the text mentions do not have a corresponding entity in the associated knowledge base. We created a dataset for this task.
To achieve this, we compared two snapshots of WikiData (from different timestamps), found entities that are in the later snapshot but absent from the older one. For this entities we extracted corresponding text from Wikipedia.
More information can be found in the paper.
The dataset is available at zenodo

A prototype of a pipeline for processing of .tmx files (xml-like files used for translation annotation). Relies on the producer-consumer architecture (implemented using Kafka).
Stack: Python, Kafka, Docker
Code on github

In this work, we explored multilingual methods for the extraction of temporal expressions from text and investigated adversarial training for aligning embedding spaces to one common space.
The work resulted in the publication

During my internship at Amazon (2020) I worked on automatic summarization of user reviews.
Stack: Python, Tensorflow, USE (universal sentence encoder), BERT

My master thesis (2020) addressed the problem of fiction summarization using transformers.
I tried different training and pre-training settings, and different architectures (vanilla transformer, GPT-2), looked into the attention distribution and analyzed errors. The best solution (though still very prone to hallucinations) consisted of a pre-processing step with extraction of "main" sentences followed by summarization by GPT-2.
The pre-processing step relied on k-means clustering of sentences encoded with universal sentence encoder. I took 2-3 sentences closest to the centroid of a cluster for each paragraph.
Stack: Python, Pytorch, USE (universal sentence encoder), scikit-learn