Bag-of-Vectors Autoencoders for Unsupervised Conditional Text Generation

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 468–488 November 20–23, 2022. ©2022 Association for Computational Linguistics 468 Bag-of-Vectors Autoencoders for Unsupervised Conditional Text Generation Florian Mai Idiap Research Institute / EPFL Rue Marconi 19, 1920 Martigny Switzerland florian.mai@idiap.ch James Henderson Idiap Research Institute Rue Marconi 19, 1920 Martigny Switzerland james.henderson@idiap.ch Abstract Text autoencoders are often used for unsu- pervised conditional text generation by apply- ing mappings in the latent space to change at- tributes to the desired values. Recently, Mai et al. (2020) proposed Emb2Emb, a method to learn these mappings in the embedding space of an autoencoder. However, their method is restricted to autoencoders with a single-vector embedding, which limits how much informa- tion can be retained. We address this issue by extending their method to Bag-of-Vectors Au- toencoders (BoV-AEs), which encode the text into a variable-size bag of vectors that grows with the size of the text, as in attention-based models. This allows to encode and reconstruct much longer texts than standard autoencoders. Analogous to conventional autoencoders, we propose regularization techniques that facilitate learning meaningful operations in the latent space. Finally, we adapt Emb2Emb for a train- ing scheme that learns to map an input bag to an output bag, including a novel loss function and neural architecture. Our empirical evalua- tions on unsupervised sentiment transfer show that our method performs substantially better than a standard autoencoder. 1 Introduction In conditional text generation, we would like to produce an output text given an input text. Hence, parallel input-output pairs are required to train a good supervised machine learning model on this type of task. Large-scale pretraining (Peters et al., 2018; Devlin et al., 2019; Lewis et al., 2020) can al- leviate the necessity for training examples to some extent, but even this requires a substantial number of annotations (Yogatama et al., 2019). This is an expensive process and can introduce unwanted ar- tifacts itself, which are henceforth learned by the model (Gururangan et al., 2018). For these rea- sons, there is substantial interest in unsupervised solutions. Text autoencoders (AEs) don’t require la- beled data for training, and are therefore a popular model for unsupervised approaches to many tasks, such as machine translation (Artetxe et al., 2018), sentence compression (Févry and Phang, 2018) and sentiment transfer (Shen et al., 2017). The classical text AE (Bowman et al., 2016) embeds the input text into a single fixed-size vector via the encoder, and then tries to reconstruct the input text from the single vector via the decoder. Single-vector embeddings are very useful, because they allow to perform conditional text generation through simple mappings in the embedding space, e.g. by adding a constant offset vector to change attributes such as sentiment (Shen et al., 2020). Recently, Mai et al. (2020) proposed Emb2Emb, a method that can learn these mappings directly in the embedding space of any pretrained single-vector AE. …

Content truncated to 500 words