Google : Rethinking Search: Making Experts out of Dilettantes

[engine]

Google has published a paper in which it argues that search as it stands today may not be the search of the future. The current model requires ranking to play a part in providing an answer, and then places "a rather significant cognitive burden on the user." Today's search services, including Google are wanting to provide answers instead of purely ranked results.Google goes on to say that this has had limited success.
In addition, is explains that there has been much progress into natural language understanding, including "word-embeddings", "sequence modelling, large pre-trained language models which capture relationships between entities.

This paper envisions a unified model-based approach to building IR systems that eliminates the need for indexes as we know them today by encoding all of the knowledge for a given corpus in a model that can be used for a wide range of tasks. As the remainder of this paper shows, once everything is viewed through a model-centric lens instead of an index-centric one, many new and interesting opportunities emerge to significantly advance IR systems. If successful, IR models that synthesize elements of classical IR systems and modern large-scale NLP models have the potential to yield a transformational shift in thinking and a significant leap in capabilities across a wide range of IR tasks, such as document retrieval, question answering, summarization, classification,recommendation, etc.

When you read this paper, it's quite clearly proposing the idea of an alternative way of searching and answering. Instead of document retrieval from an index, it's discussing pre-trained language models, model-based information retrieval, and even beyond language models.

If all of these research ambitions were to come to fruition, the resulting system would be a very early version of the system that we envisioned in the introduction. That is, the resulting system would be able to provide expert answers to a wide range of information needs in a way that neither modern IR systems, question answering systems, or pre-trained LMs can do today.Some of the key benefits of the model-based IR paradigm de-scribed herein include:
•It abstracts away the long-lived, and possibly unnecessary,distinction between “retrieval” and “scoring”.
•It results in a unified model that encodes all of the knowledge contained in a corpus, eliminating the need for traditional indexes.
•It allows for dozens of new tasks to easily be handled by the model, either via multi-task learning or via few-shot learning, with minimal amounts of labelled training data.
•It allows seamless integration of multiple modalities and languages within a unified model.

Here's the paper (PDF) [arxiv.org...]

[brotherhood of LAN]

>wrong answers

A trending one
Google turned me into a serial killer [hristo-georgiev.com]
[news.ycombinator.com...]