What to Know:
– Microsoft Research has developed a new conversational question answering model.
– The model uses a generative retrieval approach to retrieve and rank relevant passages.
– The model outperforms existing retrieval models on several benchmark datasets.
– It achieves state-of-the-art performance on the TREC CAR dataset.
The Full Story:
Microsoft Research has introduced a new conversational question answering model that uses a generative retrieval approach to retrieve and rank relevant passages. The model, called Generative Retrieval for Ranking Answers (GRRA), aims to improve the accuracy and relevance of answers provided by conversational question answering systems.
Conversational question answering systems are designed to provide answers to user queries in a conversational manner, similar to how a human would respond in a conversation. These systems are widely used in various applications, including virtual assistants and chatbots.
The GRRA model consists of two main components: a retriever and a ranker. The retriever is responsible for retrieving relevant passages from a large collection of documents, while the ranker ranks these passages based on their relevance to the user query.
To train the model, the researchers used a large-scale dataset called TREC CAR, which contains over 300,000 documents and 1.2 million queries. They also used a technique called contrastive learning, which helps the model learn to differentiate between relevant and irrelevant passages.
The researchers evaluated the performance of the GRRA model on several benchmark datasets, including MS MARCO, Natural Questions, and TREC CAR. The results showed that the model outperformed existing retrieval models on all three datasets.
On the MS MARCO dataset, the GRRA model achieved a Mean Reciprocal Rank (MRR) of 0.384, compared to the previous state-of-the-art model’s MRR of 0.374. On the Natural Questions dataset, the GRRA model achieved an MRR of 0.571, compared to the previous state-of-the-art model’s MRR of 0.563. On the TREC CAR dataset, the GRRA model achieved an MRR of 0.384, which is the highest reported performance on this dataset.
The researchers also conducted an ablation study to analyze the impact of different components of the GRRA model. They found that both the retriever and the ranker contribute significantly to the overall performance of the model. They also found that contrastive learning plays a crucial role in improving the model’s ability to retrieve and rank relevant passages.
In conclusion, Microsoft Research has developed a new conversational question answering model, GRRA, that uses a generative retrieval approach to retrieve and rank relevant passages. The model outperforms existing retrieval models on several benchmark datasets and achieves state-of-the-art performance on the TREC CAR dataset. This research could potentially improve the accuracy and relevance of answers provided by conversational question answering systems, leading to better user experiences in applications such as virtual assistants and chatbots.
Original article: https://www.searchenginejournal.com/generative-retrieval-for-conversational-question-answering/496373/