Introduction
"Text Beautifier" represents an early attempt to bridge the gap between grammatical correctness and emotional resonance in digital writing. While contemporary word processors are excellent at identifying typos and syntax errors, they remain largely "emotionally blind." They can tell a user if a sentence is correct, but not if it is cold, overly formal, or insufficiently enthusiastic.
For non-native English speakers, the challenge of writing often extends beyond grammar into the mastery of tone and nuance. This research proposes a system that moves beyond mere "correctness" toward "expressiveness," allowing users to tailor their text to specific emotional goals using a prototype linguistic tool.
System Architecture
The system, developed in 2008, relies on a rule-based NLP pipeline built upon three distinct pillars: Semantic Parsing, Affect Sensing, and Lexical Scoring.
1. Semantic Parsing: The Triplet Model
To manipulate a sentence's tone without losing its core meaning, the system first decomposes the input into a "triplet" structure consisting of a Subject (agent), an Action (verb), and an Object. Using the Machinese Syntax parser, the tool identifies functional dependencies. This structural awareness allows the system to make surgical edits—such as adding an adjective to an object or replacing a verb with a more evocative synonym—rather than simply performing a blanket sentiment swap.
2. Affect Sensing: The OCC Model
To quantify the emotional weight of a sentence, the researchers utilized the OCC (Ortony, Clore, and Collins) theory. This psychological framework categorizes emotions based on how individuals evaluate events, agents, and objects. By mapping linguistic structures to these evaluations, the system can estimate the "affective affinity" of a sentence across various emotional categories like Joy, Hope, or Pride.
3. Lexical Scoring: SenseNet
The final pillar is SenseNet, a custom-built tool designed to calculate the "contextual valence" of individual words. Unlike static sentiment dictionaries, SenseNet attempts to account for how a word's emotional impact changes based on its surroundings. It draws from a proprietary database of valence-assigned words and utilizes online thesauruses and commonsense knowledge bases to score new vocabulary.
The "Beauty-Scale" Interface
The user interacts with the system through a "mood slider" interface known as the Beauty-Scale. This scale is divided into four distinct dimensions:
- Valence: General positive or negative sentiment.
- Affect: Emotional affinity (e.g., Joy or Hope).
- Prospect: The desirability of future events.
- Praise: The praiseworthiness of specific actions.
By adjusting these sliders, users can "scale up" or "scale down" specific emotional qualities. For example, increasing the "Praise" slider might prompt the system to replace a neutral verb with one that implies higher social value.
Results and Examples
The paper demonstrates the system's utility through sentence transformation. In one example, the system processed the input: "I am going to revise the paper according to the feedbacks given by the reviewers." The initial analysis calculated a valence of +7.44, with primary affective affinities for "Joy" and "Happy-for."
By interacting with the sliders to enhance the tone, the system suggested a "beautified" version: "I am aiming to improve the paper according to the optimistic feedback given by the prudent reviewers." Here, the system made several strategic changes: swapping "revise" for "improve," and adding the adjectives "optimistic" and "prudent" to the objects "feedback" and "reviewers," respectively.
Limitations and Future Work
As a 2008 prototype, Text Beautifier relies on a rule-based NLP pipeline rather than the large language models (LLMs) prevalent today. This reliance on manual rules and predefined lexical lists limits its ability to handle idiomatic language or complex, multi-clause prose. Furthermore, the paper lacks a formal user study to quantify how human readers perceive the quality of the "beautified" text.
Despite these limitations, the project remains a significant early exploration of affective computing in assistive writing. It highlights a persistent need in NLP: the transition from tools that merely correct errors to tools that help humans express themselves with greater emotional precision.
Summary updated