Controlled Paraphrase Generation with Linguistic Feature Control

This system utilizes an encoder-decoder model to generate text with controlled complexity, guided by 40 linguistic complexity indices. The model can generate diverse paraphrases of a given sentence, each adjusted to maintain consistent meaning while varying in linguistic complexity according to the desired level.

It is important to note that not all index combinations are feasible (e.g., a sentence of "length" 5 with 10 "unique words"). To ensure high-quality outputs, our approach compares the initial generation with the target linguistic indices, and performs iterative refinement to match the closest, yet coherent achievable set of indices for the given target.
  1. Select Operation Mode: Choose from the available modes:
    • Linguistically-diverse Paraphrase Generation: Generate diverse paraphrases.
      • Steps:
        1. Enter the source text in the provided textbox.
        2. Specify the number of paraphrases you want.
        3. Click "Generate" to produce paraphrases with varying linguistic complexity.
    • Complexity-Matched Paraphrasing: Match the complexity of the input text.
      • Steps:
        1. Enter the source text in the provided textbox.
        2. Provide another sentence to extract linguistic indices.
        3. Click "Generate" to produce a paraphrase matching the complexity of the given sentence.
    • Manual Linguistic Control: Manually adjust linguistic features using sliders.
      • Steps:
        1. Enter the source text in the provided textbox.
        2. Activate or deactivate features of interest using the checkboxes.
        3. Use the sliders to adjust linguistic features.
        4. Use Tools: Access additional tools under "Tools to assist in setting linguistic indices" for advanced control.
        • Impute Missing Values: Automatically fill inactive features.
        • Random Target: Generate a random set of linguistic indices.
        • Copy Source to Target: Copy linguistic indices from the source to the target.
        • Add/Subtract Complexity: Adjust the complexity of the target indices.
        1. Click "Generate" to produce the output text based on the adjusted features.
Operation Modes

Advanced Options:

  • Approximate vs. Exact Computation: Choose between faster approximate computation or more precise exact computation of linguistic indices.
  • View Intermediate Generations: Enable this option to see the intermediate sentences generated during the quality control process.
Radio
Examples