MR-to-Text |
Robocup and WeatherGov | [210] Coarse-to-fine aligner and penalty based on learned priors | LSTM \(\rightarrow\) LSTM + regularization | N |
Recipe and SF H&R | [160] Neural agenda-checklist modeling | GRU \(\rightarrow\) GRU + agenda encoders | Y |
BAGEL | [79] Reranking beam outputs w/ RNN-based reranker | LSTM \(\rightarrow\) LSTM + reranker | N |
Restaurant Ratings | [227] Nondelexicalized inputs w/ data augmentation | LSTM \(\rightarrow\) LSTM | Y |
WikiData | [52] Complementary text-to-data translation | GRU \(\rightarrow\) GRU | Y |
E2E | [81] Comparative evaluation of 62 systems | Seq2Seq + Data-driven + Templated | Y |
[256] MLP encoder attuned to the dataset | MLP \(\rightarrow\) GRU | Y |
[360] Two-level hierarchical encoder | CAEncoder \(\rightarrow\) GRU | N |
[150] Ensemble w/ heuristic reranking | Ensemble w/ LSTM + CNN | N |
[310] Hierarchical decoding with POS tags | GRU \(\rightarrow\) GRU | N |
[95] Unsupervised DTG with DAEs | LSTM \(\rightarrow\) LSTM + DAEs | Y |
[103] Comparative evaluations w/ ensembling & penalties | Ensemble w/ LSTM + T | N |
[62] Syntactic controls with SC-LSTM | SC-LSTM | Y |
[297] Computational pragmatics based DTG | GRU \(\rightarrow\) GRU | N |
[59] Extensive anonymization | LSTM \(\rightarrow\) LSTM | Y |
[78] Semantic correctness in neural DTG | LSTM \(\rightarrow\) LSTM | Y |
[159] Self-training w/ noise injection sampling | GRU \(\rightarrow\) GRU | Y |
[277] Char-level GRU w/ input reconstruction | GRU \(\rightarrow\) GRU | N |
[97] CRFs w/ Gumbel categorical sampling | CRF | N |
Graph-to-Text |
AGENDA | [168] Graph-centric Transformer and AGENDA dataset | T \(\rightarrow\) LSTM + LSTM encoding | Y |
LDC2015E25 | [88] Phrase vs Neural MR-text w/ preprocessing analysis | LSTM \(\rightarrow\) LSTM + Phrase-based | N |
LDC2015E86 | [169] Unlabeled pre-training and linearization agnosticism | LSTM \(\rightarrow\) LSTM | N |
[273] Dual encoding for hybrid traversal | GNN \(\rightarrow\) LSTM | Y |
LDC2017T10 | [12] Graph reconstruction w/ node and edge projection | T \(\rightarrow\) T + reconstruction loss | Y |
[203] Fine-tuning GPT-2 on AMR-text joint distribution | GPT-2 | Y |
WebNLG | [207] Graph encoding with GCNs | GCN \(\rightarrow\) LSTM | N |
[68] LSTM based triple encoder | LSTM \(\rightarrow\) LSTM | Y |
[91] Discrete neural pipelines and comparisons to end-to-end | GRU \(\rightarrow\) GRU + T | Y |
[219] Sentence planning with ordered trees | LSTM \(\rightarrow\) LSTM | Y |
[275] Complementary graph contextualization | GAT \(\rightarrow\) T | Y |
[306] Detachable multi-view reconstruction | T \(\rightarrow\) T | N |
[364] Dual encoder for structure and planning | GCN \(\rightarrow\) LSTM | Y |
[274] Task-adaptive pretraining for PLMs | BART + T5 | Y |
[2] Knowledge enhanced language models and KeLM dataset | T5 | Y |
[158] Graph-text joint representations and pretraining strategies | BART + T5 | Y |
Record-to-Text (Table-to-text) |
WikiBio | [174] Tabular positional embeddings and WikiBio dataset | LSTM \(\rightarrow\) LSTM + Kneser-Ney | N |
[14] Encoding tabular attributes and WikiTableText dataset | GRU \(\rightarrow\) GRU | N |
[193] Field information through modified LSTM gating | LSTM \(\rightarrow\) LSTM | N |
[292] Link-based and content-based attention | LSTM \(\rightarrow\) LSTM | N |
[244] Multi-instance learning w/ alignment-based rewards | LSTM \(\rightarrow\) LSTM | Y |
[202] Key fact identification and data augment for few shot | LSTM + T | N |
[191] Hierarchical encoding w/ supervised auxiliary learning | LSTM \(\rightarrow\) LSTM | Y |
[192] Forced attention for omission control | LSTM \(\rightarrow\) LSTM | Y |
[46] External contextual information w/ knowledge graphs | GRU \(\rightarrow\) GRU | Y |
[318] Confidence priors for hallucination control | BERT + Pointer Networks | Y |
[51] Soft copy switching policy for few-shot learning | GPT-2 | Y |
[356] Variational auto-encoders for template induction | VAE modified to VTM | Y |
[337] Autoregressive modeling with iterative text-editing | Pointer networks + Text editing | Y |
[365] Reinforcement learning with adversarial networks | GAN | N |
[105] Linearly combined multi-reward policy | Pointer networks | N |
[354] Source-target disagreement auxiliary loss | T | N |
[311] BERT-based IR system for contextual examples | T5 + BERT | Y |
Rotowire | [347] Classification-based metrics and RotoWire dataset | LSTM \(\rightarrow\) LSTM + Templated | N |
[229] Numeric operations and operation-result encoding | GRU \(\rightarrow\) GRU + operation encoders | Y |
[253] Dynamic hierarchical entity-modeling and MLB dataset | LSTM \(\rightarrow\) LSTM | Y |
[252] Content selection and planning w/ gating and IE | LSTM \(\rightarrow\) LSTM | Y |
[107] Contextualized numeric representations | LSTM \(\rightarrow\) LSTM | Y |
[139] Dynamic salient record tracking w/ stylized generation | GRU \(\rightarrow\) GRU | N |
[261] Two-tier hierarchical input encoding | T \(\rightarrow\) LSTM | N |
[180] Auxiliary supervision w/ reasoning over entity graphs | LSTM + GAT | Y |
[255] Paragraph-centric macro planning | LSTM \(\rightarrow\) LSTM | Y |
[254] Interweaved plan and generation w/ variational models | LSTM \(\rightarrow\) LSTM | Y |
TabFact | [47] Coarse-to-fine two-stage generation | LSTM + T + GPT-2 + BERT | Y |
WikiPerson | [339] Disagreement loss w/ optimal transport matching loss | T | Y |
Humans, Books and Songs | [109] Attribute prediction-based reconstruction loss | GPT-2 | Y |
ToTTo LogicNLG NumericNLG | [189] Contextual examples through k nearest neighbors | GPT-3 | Y |
[312] Targeted table cell representation | GPT-2 | Y |
[49] Semantic confounders w/ Pearl’s do-calculus | DCVED + GPT | Y |
[187] Table-to-logic pretraining for logic text generation | T5 + BART | Y |
[223] Faithful generation with unlikelihood and replacement detection | T5 | Y |
[44] Table serialization and structural encoding | T \(\rightarrow\) GPT-2 | Y |
[7] T5 infused with tabular embeddings | T5 | N |
Cross-domain |
E2E WebNLG DART WikiBio RotoWire WITA | [140] Char-based vs word-based seq2seq | GRU \(\rightarrow\) GRU | Y |
[348] Template induction w/ neural HSMM decoder | HSMM | N |
[98] Training w/ partially aligned dataset | T \(\rightarrow\) T + supportiveness | Y |
[157] Iteratively editing templated text | GPT-2 + LaserTagger | N |
[127] RoBERTa-based semantic fidelity classifier | GPT-2 + RoBERTa | Y |
[48] Knowledge-grounded pre-training and KGTEXT dataset | T \(\rightarrow\) T + GAT | N |
[186] Hybrid attention-copy for stylistic imitation | LSTM + T | Y |
[351] Disambiguation and stitching with PLMs | GPT3 + T5 | N |
[77] Unified learning of D2T and T2D | T5 + VAE | N |
[144] Search and learn in a few-shot setting | T5 + Search and Learn | Y |
Time series-to-text |
WebNLG and DART | [294] Open-domain transfer learning for time-series narration | BART + T5 + Time series analysis | Y |
Chart-to-text |
Chart2Text | [233] Preprocessing w/ variable substitution | T \(\rightarrow\) T | Y |
Chart-to-text | [153] Neural baselines for Chart-to-text dataset | LSTM + T + BART + T5 | Y |