Natural Language Processing - Text Summarisation
Lead Research Organisation:
University of Southampton
Department Name: Sch of Electronics and Computer Sci
Abstract
We live in a world where textual content is abundant. In recent years machine techniques to understand textual information have started to become useful (particularly with respect to language translation, and now towards question answering). Being able to consume entire texts in short periods of time is unfeasible for humans. However, it is possible to condense the information into a shorter form, while still containing the key pieces of content. This research project will look at issues related to learned representations of information, and explore how these can be used to automatically produce meaningful summaries that can be adapted to specific human consumers (for example by varying levels of reading difficulty, or the depth of information contained).
Organisations
People |
ORCID iD |
| Katayoun Farrahi (Primary Supervisor) | |
| Bhumika Mistry (Student) |
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/R513325/1 | 01/10/2018 | 30/09/2023 | |||
| 2280528 | Studentship | EP/R513325/1 | 01/10/2019 | 30/09/2022 | Bhumika Mistry |
| Description | In the area of neural networks for deep learning, we present findings to show the advantages of developing networks that are able to learn interpretable functions to do specialist tasks. In particular, we show: - how existing models in the area (of neural arithmetic logic modules) are unstable and lack robustness which makes them unable to be utilised as intended in larger (applicational) systems. - we identify sources of weakness in specialised modules which lead to a lack of robustness in training them and propose a solution using stochasticity by taking advantage of the training paradigm they learn in. - we introduce a publicly available resource (Github repository) giving researchers access to a benchmark for testing and developing neural arithmetic logic modules. |
| Exploitation Route | Academics: The findings of the work around the neural arithmetic logic modules offers a strong foundation for this new niche field making it easier for researchers to become involved. In particular, the primer paper offers a comprehensive guide for those with no knowledge of the research field and is complemented by the work open-sourced in the Github repository allowing researchers access to the specialist modules in a reproducible environment. Non-academics: Will eventually be able to use the developed architectures and put them in larger end-to-end deep neural networks which can be used in application-based tasks requiring the functionality of the specialist. For example, learning networks that can schedule views for crowdsourced live-streaming (CLS). |
| Sectors | Digital/Communication/Information Technologies (including Software) |
| Title | Neural Arithmetic Logic Modules: Arithmetic Benchmarking Tasks |
| Description | This is a MIT Licensed github repository containing code for being able to benchmark various types of Neural Arithmetic Logic Modules on synthetic arithmetic tasks. Details are found in the readme (accessed by the provided link). |
| Type Of Material | Computer model/algorithm |
| Year Produced | 2021 |
| Provided To Others? | Yes |
| Impact | Provides a benchmark tool for any researchers wanting to compare their new modules against existing ones. |
| URL | https://github.com/bmistry4/nalm-benchmark |