Tuesday, 7th February 2023 - 12.00 Marzieh Saeidi : Seminar Title: Challenges of Using Modern Language Models in Real World Applications Abstract: In recent years, transformer based language models have shown impressive performance gains on long-standing NLP problems, such as language understanding, answering queries from text and machine translation. In this talk, I explore the challenges of using such models in two real world scenarios and ways to overcome them. In the first part, I will discuss whether we can extend these models to database systems. As we know, databases are an essential part of any system for storing and querying data and they work with predefined schemas and structured data. However, there is a huge amount of unstructured data such as text that one may need to query over. For instance, you may want to know the answer to “, such as “List/Count all female athletes who were born in the 20th century in Africa''. I will walk you through the challenges of using NLP to do database like reasoning over text and ways to overcome it. In the second part of the talk, I will look at using language models to detect policy compliance. Policy compliance detection can be used with respect to governmental rules or in online social media platforms to identify whether a post is in accordance with community standards. In such cases, policies are very long and complex and hard to grasp entirely by a model at a time. I propose an approach that is modular, interpretable, adaptable and requires annotations in a way that is less subjective for annotators to label. I will finish my talk by proposing ways in which language models can be reduced in size by distilling the reasoning and leaving out the knowledge which can be added later explicitly. Bio: Marzieh Saeidi is a research scientist at Shiftlab AI, working on generating synthetic textual content using Large Language Models. Previously, she worked as an Applied Research Scientist at Meta AI (formerly known as Facebook AI), focusing on different areas such as automation of detection of policy violations, improving customer service experience and recommendation of short videos . She obtained her PhD from University College London in Natural Language processing and Machine Learning under the supervision of Sebastian Riedel Add to your calendar vCal iCal Feb 07 2023 12.00 - 13.00 Tuesday, 7th February 2023 - 12.00 Marzieh Saeidi : Seminar This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk. Informatics Forum, G.03 Contact
Tuesday, 7th February 2023 - 12.00 Marzieh Saeidi : Seminar Title: Challenges of Using Modern Language Models in Real World Applications Abstract: In recent years, transformer based language models have shown impressive performance gains on long-standing NLP problems, such as language understanding, answering queries from text and machine translation. In this talk, I explore the challenges of using such models in two real world scenarios and ways to overcome them. In the first part, I will discuss whether we can extend these models to database systems. As we know, databases are an essential part of any system for storing and querying data and they work with predefined schemas and structured data. However, there is a huge amount of unstructured data such as text that one may need to query over. For instance, you may want to know the answer to “, such as “List/Count all female athletes who were born in the 20th century in Africa''. I will walk you through the challenges of using NLP to do database like reasoning over text and ways to overcome it. In the second part of the talk, I will look at using language models to detect policy compliance. Policy compliance detection can be used with respect to governmental rules or in online social media platforms to identify whether a post is in accordance with community standards. In such cases, policies are very long and complex and hard to grasp entirely by a model at a time. I propose an approach that is modular, interpretable, adaptable and requires annotations in a way that is less subjective for annotators to label. I will finish my talk by proposing ways in which language models can be reduced in size by distilling the reasoning and leaving out the knowledge which can be added later explicitly. Bio: Marzieh Saeidi is a research scientist at Shiftlab AI, working on generating synthetic textual content using Large Language Models. Previously, she worked as an Applied Research Scientist at Meta AI (formerly known as Facebook AI), focusing on different areas such as automation of detection of policy violations, improving customer service experience and recommendation of short videos . She obtained her PhD from University College London in Natural Language processing and Machine Learning under the supervision of Sebastian Riedel Add to your calendar vCal iCal Feb 07 2023 12.00 - 13.00 Tuesday, 7th February 2023 - 12.00 Marzieh Saeidi : Seminar This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk. Informatics Forum, G.03 Contact
Feb 07 2023 12.00 - 13.00 Tuesday, 7th February 2023 - 12.00 Marzieh Saeidi : Seminar This event is co-organised by ILCC and by the UKRI Centre for Doctoral Training in Natural Language Processing, https://nlp-cdt.ac.uk.
