Posterior Control of Blackbox Generation

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2731–2743 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 2731 Posterior Control of Blackbox Generation Xiang Lisa Li Department of Computer Science Johns Hopkins University xli150@jhu.edu Alexander M. Rush Department of Computer Science Cornell Tech arush@cornell.edu Abstract Text generation often requires high-precision output that obeys task-specific rules. This fine-grained control is difficult to enforce with off-the-shelf deep learning models. In this work, we consider augmenting neural gen- eration models with discrete control states learned through a structured latent-variable ap- proach. Under this formulation, task-specific knowledge can be encoded through a range of rich, posterior constraints that are effec- tively trained into the model. This approach allows users to ground internal model deci- sions based on prior knowledge, without sacri- ficing the representational power of neural gen- erative models. Experiments consider applica- tions of this approach for text generation. We find that this method improves over standard benchmarks, while also providing fine-grained control. 1 Introduction A core challenge in using deep learning for NLP is developing methods that allow for controlled output while maintaining the broad coverage of data-driven methods. While this issue is less prob- lematic in classification tasks, it has hampered the deployment of systems for conditional natural lan- guage generation (NLG), where users often need to control output through task-specific knowledge or plans. While there have been significant improve- ments in generation quality from automatic systems (Mei et al., 2016; Dusek and Jurcicek, 2016; Le- bret et al., 2016b), these methods are still far from being able to produce controlled output (Wiseman et al., 2017). Recent state-of-the-art system have even begun to utilize manual control through rule- based planning modules (Moryossef et al., 2019; Puduppully et al., 2019). Consider the case of encoder-decoder models for generation, built with RNNs or transformers. These models generate fluent output and provide flexible representations of their conditioning. Un- fortunately, auto-regressive decoders are also glob- ally dependent, which makes it challenging to in- corporate domain constraints. Research into controllable deep models aims to circumvent the all-or-nothing dependency trade- off of encoder-decoder systems and expose ex- plicit higher-level decisions. One line of research has looked at global control states that represent sentence-level properties for the full decoder. For example, Hu et al. (2017) uses generative adversar- ial networks where the attributes of the text (e.g., sentiment, tense) are exposed. Another line of re- search exposes fine-level properties, such as phrase type, but requires factoring the decoder to expose local decisions, e.g. Wiseman et al. (2018). This work proposes a method for augmenting any neural decoder architecture to incorporate fine- grained control states. The approach first modi- fies training to incorporate structured latent control variables. Then, training constraints are added to anchor the state values to problem-specific knowl- edge. At test time, the control states can be ignored or utilized as grounding for test-time constraints. Technically, the approach builds on recent advances in structured amortized variational inference to en- force …

Content truncated to 500 words