Facebook's AI technology can simplify sentences while preserving their meaning.

2019-10-15

Simplifying the syntax and structure of text is a useful skill that most of us learn in school. Over time, it becomes almost instinctive, simple and easy. However, for artificial intelligence, mastering this skill is difficult due to a lack of language knowledge. To address this challenge, scientists from Facebook and Inria are researching a simplification model called ACCESS. They claim that this model can customize text length, amount of paraphrasing, lexical complexity, syntactic complexity, and other parameters while preserving the meaning of sentences. In a detailed article describing their work, the researchers wrote, "Text simplification is beneficial not only for individuals with cognitive impairments such as aphasia, dyslexia, and autism, but also for second language learners and individuals with lower literacy skills. Previous research on text simplification has mainly focused on developing models that generate a single, universally simplified output for a given source text, without the ability to adjust the output based on the needs of different target audiences. We propose a controllable simplification model that provides users with an explicit way to manipulate and update the simplified output according to their needs." The research team developed a general encoder-decoder framework called seq2seq, which takes data as context input. Additionally, the researchers set a special token value at the beginning of the source sentence to calculate the ratio of target sentence parameters (such as length) relative to the source sentence. They then set the model for four selected parameters: length, paraphrasing, lexical complexity, and syntactic complexity. Finally, a series of tests were conducted. In the experiment, the research group trained a transformer model on the WikiLarge dataset, which consists of 296,402 automatically aligned complex-simple sentence pairs from English Wikipedia and Simple English Wikipedia. They evaluated the model using validation and test sets from a Turkish corpus, where each complex sentence had eight "paraphrased sentences" created by Amazon Mechanical Turk workers (without splitting, overly simplified structure, or reduced content). On the SARI metric, ACCESS achieved a score of 41.87, which is a "significant" improvement compared to the previous level of 40.45. SARI is a popular benchmark that compares predicted simplifications with source and target references. In terms of readability measures that do not consider grammar and meaning preservation, ACCESS ranked third with a score of 7.22. The researchers wrote in their text report, "Through analysis, we confirmed that each parameter has the expected effect on the generated simplifications. Explicitly adjusting the model on parameters such as length, paraphrasing, lexical complexity, or syntactic complexity can significantly improve its performance in sentence simplification. Our method and framework can help text simplification adapt to the needs of different audiences."

RECENT AI NEWS

RECENT AI TOOLS