Mention Flags (MF): Constraining Transformer-based Text Generators

Abstract

This paper focuses on Seq2Seq (S2S) constrained text generation where the text generator is constrained to mention specific words, which are inputs to the encoder, in the generated outputs. Pre-trained S2S models such as T5 or a Copy Mechanism can be trained to copy the surface tokens from encoders to decoders, but they cannot guarantee constraint satisfaction. Constrained decoding algorithms always produce hypotheses satisfying all constraints. However, they are computationally expensive and can lower the generated text quality. In this paper, we propose Mention Flags (MF), which trace whether lexical constraints are satisfied in the generated outputs of an S2S decoder. The MF models are trained to generate tokens until all constraints are satisfied, guaranteeing high constraint satisfaction. Our experiments on the Common Sense Generation task (CommonGen) (Lin et al., 2020), End2end Data-to-Text task (E2ENLG) (Dušek et al., 2020) and Novel Object Captioning task (nocaps) (Agrawal et al., 2019) show that the MF models maintain higher constraint satisfaction and text quality than the baseline models and other constrained text generation algorithms, achieving state-of-the-art performance on all three tasks. These results are achieved with a much lower run-time than constrained decoding algorithms. We also show that the MF models work well in the low-resource setting. 1

Knowledge Is Flat: A Seq2Seq Generative Framework for Various Knowledge Graph Completion

“…In the future, we would explore extending KG-S2S to other Seq2Seq PLMs, such as BART (Lewis et al, 2020) and MASS (Song et al, 2019). In addition, it is interesting to combine KG-S2S with other knowledge-intensive NLP tasks, such as conversation recommendation (Li et al, 2018b) and commonsense generation (Wang et al, 2021b) in the Seq2Seq framework, and see if the KG knowledge could benefit these downstream tasks.…”

Guided Beam Search to Improve Generalization in Low-Resource Data-to-Text Generation

“…where score(b) is the score of the current beam state, logGen(w) is the output logit of the generator, f ( * ) are functions that scores word w weighted by α i , and V suc is a predefined vocabulary. Similarily, Mention Flags (Wang et al, 2021) tries to identify the presence of tokens in the hypothesis given a set of flags. Both methods face the same problem since they operate on surface tokens.…”

KC-GEE: Knowledge-based Conditioning for Generative Event Extraction

“…These decomposition strategies showed high performance while introducing more detailed annotation to model training [5,7]. Inspired by the success of pretrained language models and the corresponding natural language generation-based paradigm for various NLP tasks [4,[21][22][23] tackle event extraction as controlled event generation. [6] is an end-to-end conditional generation method with manually designed discrete prompts for each event type, which needs more human effort to find the Fig.…”