Google Open Sources LaserTagger, an AI Model for Accelerating Text Generation

2020-02-03

The sequence-to-sequence AI model was introduced by Google in 2014 with the aim of mapping inputs (usually text) to outputs, where the lengths of the inputs and outputs may vary. They are used for text generation tasks, including summarization, grammar error correction, and sentence fusion, and recent architectural breakthroughs have made them more effective than before. However, they are not perfect as they require a large amount of training data to achieve acceptable performance levels, and they typically generate output character by character (which inherently slows them down).


This is why Google researchers developed LaserTagger, an open-source text editing model that can predict a series of editing operations that transform source text into target text. They claim that LaserTagger handles text generation in a less error-prone way and is easier to train and execute.LaserTagger's release marks an important step for Google in the field of natural language processing and understanding. This week, they completed the summary of Meena, a neural network with 2.6 billion parameters capable of handling multi-turn conversations.


The working principle of LaserTagger is that there is often overlap between the input and output for many text generation tasks. For example, in detecting and correcting grammar errors or fusing multiple sentences, most of the input text can remain unchanged, with only a small portion of words needing modification. LaserTagger then generates a series of editing operations instead of actual words, such as keep (copying the word to the output), delete (deleting the word), and keep-addx or delete-addx (adding the phrase X before or after the marked word, with the option to delete the marked word).


The added phrases come from a restricted vocabulary that has been optimized to minimize the vocabulary size and increase the number of training examples. The only necessary words added to the target text come from the vocabulary, avoiding the model from adding arbitrary words and alleviating the problem of ambiguity (i.e., generating output that is not supported by the input text). Additionally, LaserTagger can predict editing operations with high precision in parallel, achieving end-to-end acceleration compared to models that predict sequentially.LaserTagger has been evaluated on multiple text generation tasks and has shown "considerable strength" in performance compared to baseline models trained with a large number of training examples, and it is 100 times faster. Even with only a few hundred or thousand training examples, it produces "reasonable" results that can be manually edited or curated.


The team writes, "The advantages of LaserTagger become more apparent in large-scale applications, such as improving the format of voice responses in certain services by reducing response length and redundancy. The high inference speed allows the model to be inserted into existing technology stacks without adding any noticeable latency on the user side, and the improved data efficiency enables the collection of training data in multiple languages, benefiting users from different language backgrounds."


GitHub link: https://github.com/google-research/lasertagger

RECENT AI NEWS

RECENT AI TOOLS