TL;DR
This paper introduces a browser-based editor that simplifies English text by automatically breaking down complex sentences and suggesting simpler word alternatives. Its key innovation is the ability to accept custom vocabulary lists, allowing the simplification process to be tailored to specific reader populations, such as language learners or individuals with low literacy.
What Problem Does This Paper Address?
Making complex English text accessible to a wider audience, including language learners, children, and adults with low literacy skills, is a significant challenge. Manually simplifying texts is time-consuming and requires specialized linguistic knowledge. While automated text simplification systems exist, they often lack flexibility in adapting to different reader needs. Existing tools may focus solely on lexical or syntactic simplification, or they don't adequately support human post-editing to refine the output. This paper addresses the need for a user-friendly, customizable tool that combines both syntactic and lexical simplification and allows for fine-grained control by the user.
Methods
The editor combines two main simplification techniques: syntactic and lexical.
Syntactic Simplification: The system identifies and simplifies seven types of complex grammatical constructs, including coordinated clauses, adverbial clauses, and participial phrases. It achieves this by parsing sentences using the Stanford dependency parser and then applying manually crafted rules to split complex sentences into simpler ones or transform passive voice to active voice. Users can undo splits if they deem them inappropriate.
Lexical Simplification: This process involves substituting difficult words with simpler synonyms. The system uses a three-step approach:
- Substitution Generation: It leverages the
word2vecmodel, trained on Wikipedia, to find words semantically similar to the target word. - Substitution Selection: A language model (trigram model trained on Wikipedia) is used to re-rank these candidates, selecting the top 10 most probable ones.
- Customization: A novel feature is the ability to upload a custom vocabulary list. The system identifies words not in this list as candidates for simplification and prioritizes substitutions from within the list, effectively tailoring the output to a specific reader's vocabulary level.
Datasets
- Training: The
word2vecmodel and the trigram language model were trained on all sentences from Wikipedia. - Evaluation:
- Syntactic simplification was evaluated on the first 300 sentences from the Mechanical Turk Lexical Simplification Data Set (Horn et al., 2014).
- Lexical simplification was evaluated on a subset of the Mechanical Turk Lexical Simplification Data Set (Horn et al., 2014), specifically focusing on sentences where the target word was not in the custom vocabulary list but human annotations contained at least one word from the list.
- The custom vocabulary list used for evaluation was provided by the Hong Kong Education Bureau (EDB, 2012), comprising approximately 4,000 words expected of primary school graduates.
Key Insights and Ideas
- Customization is Key: The ability to input a custom vocabulary list is a significant advancement, moving beyond generic simplification to reader-specific adaptation. This acknowledges that "simplicity" is relative to the audience.
- Integrated Approach: Combining syntactic and lexical simplification within a single editor provides a more comprehensive solution than tools focusing on only one aspect.
- User-in-the-Loop: The editor empowers users to refine the automated process, offering control over sentence splitting and word substitution, which is crucial for ensuring accuracy and naturalness.
- Leveraging Pre-trained Models: The use of
word2vecand language models demonstrates an effective way to leverage existing NLP resources for simplification tasks.
Key Results
- Syntactic Simplification: Achieved 79% precision and 64% recall when evaluated against a human gold standard for applicable simplification types.
- Lexical Simplification: For 57% of target words, a valid substitution was found within the top ten candidates provided in the drop-down list. Precision for the top candidate was 31%.
Limitations and Future Work
- Regeneration Tasks: The system currently leaves complex regeneration tasks (e.g., generating referring expressions, determiners, conjunctions) to the user for post-editing, as these often require real-world knowledge.
- Empirical Studies: The authors plan to conduct empirical studies to measure the editor's effectiveness in assisting teachers with language lesson planning.
- Further Improvement: Future work aims to further enhance simplification quality and potentially offer annotations for difficult words that cannot be simplified.
Summary updated