January 04–07, 2024, Bangalore, India | Ashish Mittal, Rudra Murthy V, Vishwajeet Kumar, and Riyaz Bhat
The tutorial "Towards Understanding and Mitigating the Hallucinations in NLP and Speech" by Ashish Mittal, Vishwajeet Kumar, Rudra Murthy V, and Riyaz Bhat aims to address the issue of hallucinations in natural language processing (NLP) and speech recognition systems. Despite significant advancements in deep learning architectures like the transformer, these models often generate text that is unfaithful to the input, containing irrelevant or unsupported information. Similar issues are observed in speech recognition, where the output text may differ from the original speech signal.
The tutorial covers the problem of hallucinations in various tasks such as machine translation, summarization, and speech recognition. It categorizes the types of hallucinations, describes techniques to quantify them, and presents recent methods to mitigate these issues. The authors highlight the root causes of hallucinations in ASR models and provide a generic framework applicable to multiple NLP tasks. The tutorial is designed to be 1.5 hours long, with breaks and discussion sessions, and is intended for researchers in dialog generation, speech-to-text generation, and machine translation, as well as academicians and industry professionals involved in natural language generation.
The tutorial's goals include raising awareness about the prevalence and impact of hallucinations in neural generation models and encouraging the research community to expand beyond English in NLP and speech research. The target audience is expected to have a basic understanding of deep-learning-based solutions like sequence-to-sequence models and Transformers. The proposers will be available to present the tutorial in person, and the session will be structured to cover the problem, causes, and mitigation strategies for hallucinations in NLP and speech models.The tutorial "Towards Understanding and Mitigating the Hallucinations in NLP and Speech" by Ashish Mittal, Vishwajeet Kumar, Rudra Murthy V, and Riyaz Bhat aims to address the issue of hallucinations in natural language processing (NLP) and speech recognition systems. Despite significant advancements in deep learning architectures like the transformer, these models often generate text that is unfaithful to the input, containing irrelevant or unsupported information. Similar issues are observed in speech recognition, where the output text may differ from the original speech signal.
The tutorial covers the problem of hallucinations in various tasks such as machine translation, summarization, and speech recognition. It categorizes the types of hallucinations, describes techniques to quantify them, and presents recent methods to mitigate these issues. The authors highlight the root causes of hallucinations in ASR models and provide a generic framework applicable to multiple NLP tasks. The tutorial is designed to be 1.5 hours long, with breaks and discussion sessions, and is intended for researchers in dialog generation, speech-to-text generation, and machine translation, as well as academicians and industry professionals involved in natural language generation.
The tutorial's goals include raising awareness about the prevalence and impact of hallucinations in neural generation models and encouraging the research community to expand beyond English in NLP and speech research. The target audience is expected to have a basic understanding of deep-learning-based solutions like sequence-to-sequence models and Transformers. The proposers will be available to present the tutorial in person, and the session will be structured to cover the problem, causes, and mitigation strategies for hallucinations in NLP and speech models.