Effidit: Your AI Writing Assistant

Shuming Shi, Enbo Zhao, Duyu Tang, Yan Wang, Piji Li, Wei Bi, Haiyun Jiang, Guoping Huang, Leyang Cui, Xinting Huang, Cong Zhou, Yong Dai, Dongyang MaarXiv 2022 · DOI · arXiv

Introduction

Effidit (Efficient and Intelligent Editing) is a comprehensive AI writing assistant developed by Tencent AI Lab. Designed to support both Chinese and English, it moves beyond basic grammar checking to offer a suite of advanced features including context-aware polishing, keyword-to-sentence generation, and bidirectional phrase completion. While digital writing assistants like Grammarly or Wordtune are common, Effidit addresses specific limitations in current technology, such as model degeneration, limited error type handling, and rigid completion mechanisms.

Addressing the Gaps in AI Writing

Current AI writing tools often face three major challenges:

  1. Model Degeneration: Large language models (LLMs) can produce repetitive or "dull" text when trained with standard objectives.
  2. Limited Error Types: Many systems excel at fixing typos (substitution) but struggle with insertion or deletion errors.
  3. Rigid Completion: Most tools only consider the text preceding the cursor (the prefix), ignoring any existing text that follows (the suffix).

Effidit tackles these issues through a modular architecture and several core technical innovations.

Technical Innovations

To prevent the repetitive "looping" common in AI-generated text, Effidit utilizes the Simple Contrastive Framework for Neural Text Generation (SimCTG). This framework forces the model to learn more discriminative token representations during training. During decoding, a Contrastive Search algorithm selects words that are both probable and distinct from the previous context, effectively maintaining creativity and coherence.

InsertBERT for Error Correction

For more robust error correction, the team developed InsertBERT. This model introduces a [null] token, allowing it to predict whether a word should exist at a specific position. This makes it significantly more effective at identifying where words need to be deleted or inserted, rather than just substituted.

Keywords to Sentences (K2S) as Infilling

The K2S feature is formulated as a text-infilling task. Instead of simply generating text after a prompt, the model treats user-provided keywords as anchors and "fills in the blanks" between them.

The Effidit interface demonstrating the Keywords to Sentences (K2S) feature with AI-generated suggestions.

Hybrid Retrieval-Generation

Effidit provides two types of results for many tasks: Web Sentences, retrieved from a massive high-quality corpus for factual accuracy, and AI Generation, for creative flexibility.

Key Features and Insights

Bidirectional Completion

Unlike traditional assistants, Effidit considers both the prefix and the suffix. If a user types in the middle of a sentence, the system analyzes the surrounding context to suggest the most appropriate fit.

Effidit's sentence completion interface showing bilingual support for English and Chinese.

Context-Aware Polishing

Effidit’s phrase polishing goes beyond a simple thesaurus. It uses a similarity graph combined with a coherence score to find words that fit the specific "vibe" and syntax of the sentence.

The [null] Token Strategy

Using a [null] token to detect "non-existence" is a clever adaptation of BERT-style models for editing tasks. This approach avoids the need for computationally expensive sequence-to-sequence architectures for minor fixes.

Evaluation and Results

Effidit was evaluated using benchmarks like Wikitext-103 and massive internal datasets, including a general web corpus and academic data from sources like arXiv.

  • Human Evaluation: In head-to-head tests on Wikitext-103, the SimCTG + Contrastive Search approach significantly outperformed standard nucleus sampling in coherence, fluency, and informativeness.
  • Error Correction: Effidit’s Chinese error correction is ranked as "first tier" for word substitution and is notably ahead of competitors in handling insertion and deletion errors.
  • K2S Performance: The generation-based K2S model achieved significantly higher "Novelty" scores (5.42 for Chinese) compared to retrieval-based baselines (1.99), indicating true synthesis rather than simple retrieval.

Future Directions

While Effidit is already a powerful tool, future work includes scaling contrastive techniques to models with billions of parameters and expanding advanced features like sentence expansion to English. The team also plans to introduce specialized modes for domains such as legal and medical writing.

Summary updated

Abstract

In this technical report, we introduce Effidit (Efficient and Intelligent Editing), a digital writing assistant that facilitates users to write higher-quality text more efficiently by using artificial intelligence (AI) technologies. Previous writing assistants typically provide the function of error checking (to detect and correct spelling and grammatical errors) and limited text-rewriting functionality. With the emergence of large-scale neural language models, some systems support automatically completing a sentence or a paragraph. In Effidit, we significantly expand the capacities of a writing assistant by providing functions in five categories: text completion, error checking, text polishing, keywords to sentences (K2S), and cloud input methods (cloud IME). In the text completion category, Effidit supports generation-based sentence completion, retrieval-based sentence completion, and phrase completion. In contrast, many other writing assistants so far only provide one or two of the three functions. For text polishing, we have three functions: (context-aware) phrase polishing, sentence paraphrasing, and sentence expansion, whereas many other writing assistants often support one or two functions in this category. The main contents of this report include major modules of Effidit, methods for implementing these modules, and evaluation results of some key methods.
BibTeX
@misc{shi2022effidit,
  author = {Shuming Shi and Enbo Zhao and Duyu Tang and Yan Wang and Piji Li and Wei Bi and Haiyun Jiang and Guoping Huang and Leyang Cui and Xinting Huang and Cong Zhou and Yong Dai and Dongyang Ma},
  title = {{Effidit: Your AI Writing Assistant}},
  year = {2022},
  eprint = {2208.01815},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.01815},
}
Figure