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:
- Model Degeneration: Large language models (LLMs) can produce repetitive or "dull" text when trained with standard objectives.
- Limited Error Types: Many systems excel at fixing typos (substitution) but struggle with insertion or deletion errors.
- 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
SimCTG and Contrastive Search
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.
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.
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