Nowadays, the capabilities of language models are getting stronger and stronger, and they play a large role in various tasks. Open-Domain Dialogs are probably the most difficult type of task to complete, requiring the ability to have conversations on any topic.
In conversational tasks, in addition to generating responses that humans perceive to be reasonable, interesting, and context-specific, language models must work according to a Responsible AI framework so as not to generate content that is not proven by an information source.
Recently, Google introduced the latest progress in the language model LaMDA in a paper entitled "LaMDA: Language Models for Dialog Applications" in terms of safe, reliable and high-quality dialogue.
The LaMDA model, with 137 billion parameters, is built by fine-tuning a dedicated conversational neural language model using the Transformer architecture, allowing conversations to be conducted using external knowledge sources.
Defining goals and metrics is critical for training a conversational model. The LaMDA model has three key objectives of quality, safety and solidity, each with its own metrics.
Qualitatively, Google splits it into three aspects: sensibleness, specificity, Interestingness , SSI.
Among them, rationality refers to the model's meaningful response in the dialogue, such as no common sense error in the response; specificity refers to the model's response to a specific context in the dialogue, rather than a general response in the general situation; interestingness refers to the model's insightful and intelligent response.
Figure | LaMDA Model Dialogue (Source: Google)
In terms of security, Google has made great strides in developing and deploying responsible AI. To avoid biases and content that could harm users, it uses a set of security goals that limit the model's output dialogue to form security metrics.
Because language models sometimes output content that seems plausible and actual but contradicts known facts, Google has studied the solidity of the LaMDA model.
Solidity is the percentage of external sources of information that can be supported by authoritative external sources of information on the outside world' claims. However, the LaMDA model built in these sources does not fully guarantee the accuracy of the generated response, so Google allows users and the reliability of external system sources to determine whether the response is valid or not.
In addition, Google introduced the pre-training and fine-tuning phases of the LaMDA model in the paper.
During the pre-training phase, Google made a dataset containing 1.56T words, marked the vocabulary in the dataset as a 2.81T SentencePiece token, and then pre-trained the LaMDA model through the GSPMD system.
It is understood that Google will use the pre-trained LaMDA model for its natural language processing research, including program synthesis, zero-sample learning and style migration.
During the fine-tuning phase, Google let the LaMDA model perform two types of tasks, one is to make a hybrid generation task that responds to the natural language of the specified context, and the other is a classification task that responds to safety and high quality, thus becoming a multitasking model.
During a conversation, the LaMDA generator generates several candidate responses to the targeted context, and then the LaMDA classifier predicts the SSI and safe score of each candidate response, and finally selects the best response based on the ranking of the two data.
Humans can clarify facts through existing tools and knowledge bases, while language models can only rely on their internal parameters to get information.
To do this, Google made a dataset of human communication with the LaMDA model, and through this dataset, the generator and classifier of the LaMDA model were fine-tuned, allowing it to invoke an external information retrieval system during the conversation to improve the solidity of the response.
Google said, "After evaluating the LaMDA model, it was significantly better than the pre-trained model in every dimension and at all model sizes." Regardless of whether fine-tuned or not, quality metrics such as reasonableness, specificity, and interest often increase with the amount of model parameters. Security, while not only dependent on model scaling, can be improved by fine-tuning. ”
In addition, the solidity of the LaMDA model will continue to increase due to the increase in the size of the model. The reason for this may be that the larger the model, the stronger its ability to remember uncommon knowledge, and the fine-tuning allows the model to access external knowledge sources and shift the load of remembering knowledge to external knowledge sources.
However, while fine-tuning narrows the gap between the language model and humans, the model's level of security and solidity is still weaker than that of humans.
Figure | Evaluation data on all aspects of the LaMDA model (Source: Google)
The emergence of the LaMDA model opens up new avenues for completing open-domain conversations, while illustrating key challenges facing neural language models, such as the use and improvement of security metrics, and how to fine-tune them with larger models and clearly labeled data.
However, this is still a very early work and has great limitations. In the future, Google will explore new ways to further improve the solidity of safety metrics and LaMDA models and align with its AI principles.
-End-
reference:
https://ai.googleblog.com/2022/01/lamda-towards-safe-grounded-and-high.html