
Dialogue systems aim to build an agent that mimics human conversation. These models have a higher requirement in consistency, semantics, and interactiveness and constraints are implemented on emotion, personal style, dialogue intent. 

Dialogue Generation


Storytelling requires the the generation of text with complete narrative logic, which scopes over longer strings. Control conditions include story lines and ending. 

Storytelling 

The goal of data-to-text generation is to convert structured data (tables, graphs) into natural language tasks such as a weather forecast. Control conditions ensure the text contains information reflected in the structured data.  

Data to Text



Neural networks are used to expand existing datasets and create new data since data collection and cleaning comes at a high cost. CTG can be used to identify and replace entities in given text or generating new sentences according to given attributes. 

Data Augmentation


Biased training data causes models to output inaccurate results. Rewriting biased to unbiased text or changing the data distribution of CTG-generated text is feasible, and this task is controlled on gender, race, and toxicity. 

Debiasing 


Formatting the generated text (length, rhythm in poems) control on the number of words, structure, etc.. 

Format Control

A focus of attention in controllable text generation (CTG), attribute-based generation aims to produce sentences that satisfy specific attributes such as topic, emotion, and keywords. Combining multiple control attributes allows the system to create interpretable and controllable paragraphs or articles.

Attribute-Based Controllable Text Generation

List of CTG-related tasks with explicit controlled conditions and goals. 

University of California, Berkeley

Control conditions range from text attribution (sentiment, topic, keywords) to author style (gender, age), text genre and format (poems, couplets), event ordering (story lines), or structured data description.

Learn Before

Related

Learn After