Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ReALM: Reference Resolution As Language Modeling

Joel Ruben Antony Moniz, Soundarya Krishnan, Melis Ozyildirim, Prathamesh Saraf, Halim Cagri Ates, Yuan Zhang, Hong Yu

TL;DR

ReALM addresses reference resolution across conversational, on-screen, and background contexts using a text-based encoding of screen content to enable on-device language modeling. The approach fine-tunes FLAN-T5 on reformatted inputs and compares against MARRS and ChatGPT baselines, achieving strong results, including parity with GPT-4 on small models and surpassing it with larger variants. Key contributions include a novel onscreen encoding algorithm that preserves spatial layout and the demonstration that small models can rival large LLMs on this task. This work advances natural, hands-free user interactions by bridging textual and non-textual context in a modular, privacy-friendly pipeline.

Abstract

Reference resolution is an important problem, one that is essential to understand and successfully handle context of different kinds. This context includes both previous turns and context that pertains to non-conversational entities, such as entities on the user's screen or those running in the background. While LLMs have been shown to be extremely powerful for a variety of tasks, their use in reference resolution, particularly for non-conversational entities, remains underutilized. This paper demonstrates how LLMs can be used to create an extremely effective system to resolve references of various types, by showing how reference resolution can be converted into a language modeling problem, despite involving forms of entities like those on screen that are not traditionally conducive to being reduced to a text-only modality. We demonstrate large improvements over an existing system with similar functionality across different types of references, with our smallest model obtaining absolute gains of over 5% for on-screen references. We also benchmark against GPT-3.5 and GPT-4, with our smallest model achieving performance comparable to that of GPT-4, and our larger models substantially outperforming it.

ReALM: Reference Resolution As Language Modeling

TL;DR

ReALM addresses reference resolution across conversational, on-screen, and background contexts using a text-based encoding of screen content to enable on-device language modeling. The approach fine-tunes FLAN-T5 on reformatted inputs and compares against MARRS and ChatGPT baselines, achieving strong results, including parity with GPT-4 on small models and surpassing it with larger variants. Key contributions include a novel onscreen encoding algorithm that preserves spatial layout and the demonstration that small models can rival large LLMs on this task. This work advances natural, hands-free user interactions by bridging textual and non-textual context in a modular, privacy-friendly pipeline.

Abstract

Reference resolution is an important problem, one that is essential to understand and successfully handle context of different kinds. This context includes both previous turns and context that pertains to non-conversational entities, such as entities on the user's screen or those running in the background. While LLMs have been shown to be extremely powerful for a variety of tasks, their use in reference resolution, particularly for non-conversational entities, remains underutilized. This paper demonstrates how LLMs can be used to create an extremely effective system to resolve references of various types, by showing how reference resolution can be converted into a language modeling problem, despite involving forms of entities like those on screen that are not traditionally conducive to being reduced to a text-only modality. We demonstrate large improvements over an existing system with similar functionality across different types of references, with our smallest model obtaining absolute gains of over 5% for on-screen references. We also benchmark against GPT-3.5 and GPT-4, with our smallest model achieving performance comparable to that of GPT-4, and our larger models substantially outperforming it.
Paper Structure (19 sections, 4 figures, 12 tables, 2 algorithms)

This paper contains 19 sections, 4 figures, 12 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Realated Work and Motivation
  3. Task
  4. Datasets
  5. Conversational Data
  6. Synthetic Data
  7. On-screen Data
  8. Models
  9. MARRS
  10. ChatGPT
  11. Our Approach
  12. Conversational References
  13. Onscreen References
  14. Results
  15. Analysis
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Sample screenshots used in the annotation of on-screen data. The data was annotated in a two-step process, as described in Section \ref{['sec:onscreen_data']}.
  • Figure 2: Technical diagrams representing user turns with a conversational assistant in (a), and a user screen in (b). Shaded rectangles represent various elements shown on the screen detectable by screen parser-extractors.
  • Figure 3: Technical diagrams representing user screens. Shaded rectangles represent various elements shown on the screen detectable by screen parser-extractors.
  • Figure 4: Performance improvements with each experiment -- (a) Baseline Finetuned LLM, (b) Obtaining screen elements through OCR, (c) Obtaining screen elements through UI elements and Clustering (d) Adding an extra newline between the instruction and user request, (e) Onscreen Grab, (f) Onscreen Grab with injected turn objects, (g) Onscreen Grab with injected turn object + needing lines to be separated by at least Margin, (h) Separating elements in the same line by a tab