Natural Language Processing Seminar 2023–2024
The NLP Seminar is organised by the Linguistic Engineering Group at the Institute of Computer Science, Polish Academy of Sciences (ICS PAS). It takes place on (some) Mondays, usually at 10:15 am, often online – please use the link next to the presentation title. All recorded talks are available on YouTube. |
|
9 October 2023 |
Agnieszka Mikołajczyk-Bareła, Wojciech Janowski (VoiceLab), Piotr Pęzik (University of Łódź / VoiceLab), Filip Żarnecki, Alicja Golisowicz (VoiceLab) |
|
This talk will summarize our recent work on fine-tuning a large generative language model on bilingual instruction datasets, which resulted in the release of an open version of Trurl (trurl.ai). The motivation behind creating this model was to improve the performance of the original Llama 2 7B- and 13B-parameter models (Touvron et al. 2023), from which it was derived in a number of areas such as information extraction from customer-agent interactions and data labeling with a special focus on processing texts and instructions written in Polish. We discuss the process of optimizing the instruction datasets and the effect of the fine-tuning process on a number of selected downstream tasks. |
16 October 2023 |
Konrad Wojtasik, Vadim Shishkin, Kacper Wołowiec, Arkadiusz Janz, Maciej Piasecki (Wrocław University of Science and Technology) |
|
Information Retrieval over large collections of documents is an extremely important research direction in the field of natural language processing. It is a key component in question-answering systems, where the answering model often relies on information contained in a database with up-to-date knowledge. This not only allows for updating the knowledge upon which the system responds to user queries but also limits its hallucinations. Currently, information retrieval models are neural networks and require significant training resources. For many years, lexical matching methods like BM25 outperformed trained neural models in Open Domain setting, but current architectures and extensive datasets allow surpassing lexical solutions. In the presentation, I will introduce available datasets for the evaluation and training of modern information retrieval architectures in document collections from various domains, as well as future development directions. |
30 October 2023 |
Agnieszka Faleńska (University of Stuttgart) |
|
The summary will be available soon. |
13 November 2023 |
Piotr Rybak (Institute of Computer Science, Polish Academy of Sciences) |
Advancing Polish Question Answering: Datasets and Models |
Although question answering (QA) is one of the most popular topics in natural language processing, until recently it was virtually absent in the Polish scientific community. However, the last few years have seen a significant increase in work related to this topic. In this talk, I will discuss what question answering is, how current QA systems work, and what datasets and models are available for Polish QA. In particular, I will discuss the resources created at IPI PAN, namely the PolQA and MAUPQA datasets and the Silver Retriever model. Finally, I will point out further directions of work that are still open when it comes to Polish question answering. |
Please see also the talks given in 2000–2015 and 2015–2023. |
2 April 2020
Stan Matwin (Dalhousie University)
Efficient training of word embeddings with a focus on negative examples
This presentation is based on our AAAI 2018 and AAAI 2019 papers on English word embeddings. In particular, we examine the notion of “negative examples”, the unobserved or insignificant word-context co-occurrences, in spectral methods. we provide a new formulation for the word embedding problem by proposing a new intuitive objective function that perfectly justifies the use of negative examples. With the goal of efficient learning of embeddings, we propose a kernel similarity measure for the latent space that can effectively calculate the similarities in high dimensions. Moreover, we propose an approximate alternative to our algorithm using a modified Vantage Point tree and reduce the computational complexity of the algorithm with respect to the number of words in the vocabulary. We have trained various word embedding algorithms on articles of Wikipedia with 2.3 billion tokens and show that our method outperforms the state-of-the-art in most word similarity tasks by a good margin. We will round up our discussion with some general thought s about the use of embeddings in modern NLP.