Transformers generalize to novel compositions of structures and entities after being trained on a complex dataset, but easily overfit on datasets of insufficient complexity. We observe that when the training set is sufficiently complex, the model encodes structurally equivalent sentences using a systematic attention pattern. Inspired by this observation, we propose SQ-Transformer (Structurally Quantized) that explicitly encourages systematicity in the embeddings and attention layers even with low-complexity data. At the embedding level, we introduce Structure-oriented Vector Quantization (SoVQ) to cluster word embeddings into several classes of structurally equivalent entities. At the attention level, we devise the Systematic Attention Layer (SAL) and an alternative, Systematically Regularized Layer (SRL) that operate on the quantized word embeddings so that sentences of the same structure are encoded with invariant or similar attention patterns. Empirically, we show SQ-Transformer achieves stronger compositional generalization than the vanilla Transformer on multiple low-complexity semantic parsing and machine translation datasets. In our analysis, we show SoVQ indeed learns a syntactically clustered embedding space, and SAL/SRL induces generalizable attention patterns, altogether leading to improved systematicity.

Recent diagnostic datasets on compositional generalization, such as SCAN (Lake and Baroni, 2018) and COGS (Kim and Linzen, 2020), expose severe problems in models trained from scratch on these datasets. However, in contrast to this poor performance, state-of-the-art models trained on larger and more general datasets show better generalization ability. In this work, to reconcile this inconsistency, we conduct an empirical analysis by training Transformer models on a variety of training sets with different data factors, including dataset scale, pattern complexity, example difficulty, etc. First, we show that increased dataset complexity can lead to better generalization behavior on multiple different generalization challenges. To further understand this improvement, we show two axes of the benefit from more complex datasets: they provide more diverse examples so compositional understanding becomes more effective, and they also prevent ungeneralizable memorization of the examples due to reduced example repetition frequency. Finally, we explore how training examples of different difficulty levels influence generalization differently. On synthetic datasets, simple examples invoke stronger compositionality than hard examples do. On larger-scale real language datasets, while hard examples become more important potentially to ensure decent data coverage, a balanced mixture of simple and hard examples manages to induce the strongest generalizability.

Recent datasets expose the lack of the systematic generalization ability in standard sequence-to-sequence models. In this work, we analyze this behavior of seq2seq models and identify two contributing factors: a lack of mutual exclusivity bias (one target sequence can only be mapped to one source sequence), and the tendency to memorize whole examples rather than separating structures from contents. We propose two techniques to address these two issues respectively: Mutual Exclusivity Training that prevents the model from producing seen generations when facing novel examples via an unlikelihood-based loss, and prim2primX data augmentation that automatically diversifies the arguments of every syntactic function to prevent memorizing and provide a compositional inductive bias without exposing test-set data. Combining these two techniques, we show substantial empirical improvements using standard sequence-to-sequence models (LSTMs and Transformers) on two widely-used compositionality datasets: SCAN and COGS. Finally, we provide analysis characterizing the improvements as well as the remaining challenges, and provide detailed ablations of our method.

Machine translation has seen rapid progress with the advent of Transformer-based models. These models have no explicit linguistic structure built into them, yet they may still implicitly learn structured relationships by attending to relevant tokens. We hypothesize that this structural learning could be made more robust by explicitly endowing Transformers with a structural bias, and we investigate two methods for building in such a bias. One method, the TP-Transformer, augments the traditional Transformer architecture to include an additional component to represent structure. The second method imbues structure at the data level by segmenting the data with morphological tokenization. We test these methods on translating from English into morphologically rich languages, Turkish and Inuktitut, and consider both automatic metrics and human evaluations. We find that each of these two approaches allows the network to achieve better performance, but this improvement is dependent on the size of the dataset. In sum, structural encoding methods make Transformers more sample-efficient, enabling them to perform better from smaller amounts of data.

Abstractive summarization, the task of generating a concise summary of input documents, requires: (1) reasoning over the source document to determine the salient pieces of information scattered across the long document, and (2) composing a cohesive text by reconstructing these salient facts into a shorter summary that faithfully reflects the complex relations connecting these facts. In this paper, we adapt TP-Transformer (Schlag et al., 2019), an architecture that enriches the original Transformer (Vaswani et al., 2017) with the explicitly compositional Tensor Product Representation (TPR), for the task of abstractive summarization. The key feature of our model is a structural bias that we introduce by encoding two separate representations for each token to represent the syntactic structure (with role vectors) and semantic content (with filler vectors) separately. The model then binds the role and filler vectors into the TPR as the layer output. We argue that the structured intermediate representations enable the model to take better control of the contents (salient facts) and structures (the syntax that connects the facts) when generating the summary. Empirically, we show that our TP-Transformer outperforms the Transformer and the original TP-Transformer significantly on several abstractive summarization datasets based on both automatic and human evaluations. On several syntactic and semantic probing tasks, we demonstrate the emergent structural information in the role vectors and the performance gain by information specificity of the role vectors and improved syntactic interpretability in the TPR layer outputs.(Code and models are available at https://github.com/jiangycTarheel/TPT-Summ)

Undirected neural sequence models have achieved performance competitive with the state-of-the-art directed sequence models that generate monotonically from left to right in machine translation tasks. In this work, we train a policy that learns the generation order for a pre-trained, undirected translation model via reinforcement learning. We show that the translations decoded by our learned orders achieve higher BLEU scores than the outputs decoded from left to right or decoded by the learned order from Mansimov et al. (2019) on the WMT’14 German-English translation task. On examples with a maximum source and target length of 30 from De-En and WMT’16 English-Romanian tasks, our learned order outperforms all heuristic generation orders on three out of four language pairs. We next carefully analyze the learned order patterns via qualitative and quantitative analysis. We show that our policy generally follows an outer-to-inner order, predicting the left-most and right-most positions first, and then moving toward the middle while skipping less important words at the beginning. Furthermore, the policy usually predicts positions for a single syntactic constituent structure in consecutive steps. We believe our findings could provide more insights on the mechanism of undirected generation models and encourage further research in this direction.

Systematic compositionality is an essential mechanism in human language, allowing the recombination of known parts to create novel expressions. However, existing neural models have been shown to lack this basic ability in learning symbolic structures. Motivated by the failure of a Transformer model on the SCAN compositionality challenge (Lake and Baroni, 2018), which requires parsing a command into actions, we propose two auxiliary sequence prediction tasks as additional training supervision. These automatically-generated sequences are more representative of the underlying compositional symbolic structures of the input data. During inference, the model jointly predicts the next action and the next tokens in the auxiliary sequences at each step. Experiments on the SCAN dataset show that our method encourages the Transformer to understand compositional structures of the command, improving its accuracy on multiple challenging splits from ≤ 10% to 100%. With only 418 (5%) training instances, our approach still achieves 97.8% accuracy on the MCD1 split. Therefore, we argue that compositionality can be induced in Transformers given minimal but proper guidance. We also show that a better result is achieved using less contextualized vectors as the attention’s query, providing insights into architecture choices in achieving systematic compositionality. Finally, we show positive generalization results on the grounded-SCAN task (Ruis et al., 2020).

We introduce HoVer (HOppy VERification), a dataset for many-hop evidence extraction and fact verification. It challenges models to extract facts from several Wikipedia articles that are relevant to a claim and classify whether the claim is supported or not-supported by the facts. In HoVer, the claims require evidence to be extracted from as many as four English Wikipedia articles and embody reasoning graphs of diverse shapes. Moreover, most of the 3/4-hop claims are written in multiple sentences, which adds to the complexity of understanding long-range dependency relations such as coreference. We show that the performance of an existing state-of-the-art semantic-matching model degrades significantly on our dataset as the number of reasoning hops increases, hence demonstrating the necessity of many-hop reasoning to achieve strong results. We hope that the introduction of this challenging dataset and the accompanying evaluation task will encourage research in many-hop fact retrieval and information verification.

Multi-hop reading comprehension requires the model to explore and connect relevant information from multiple sentences/documents in order to answer the question about the context. To achieve this, we propose an interpretable 3-module system called Explore-Propose-Assemble reader (EPAr). First, the Document Explorer iteratively selects relevant documents and represents divergent reasoning chains in a tree structure so as to allow assimilating information from all chains. The Answer Proposer then proposes an answer from every root-to-leaf path in the reasoning tree. Finally, the Evidence Assembler extracts a key sentence containing the proposed answer from every path and combines them to predict the final answer. Intuitively, EPAr approximates the coarse-to-fine-grained comprehension behavior of human readers when facing multiple long documents. We jointly optimize our 3 modules by minimizing the sum of losses from each stage conditioned on the previous stage’s output. On two multi-hop reading comprehension datasets WikiHop and MedHop, our EPAr model achieves significant improvements over the baseline and competitive results compared to the state-of-the-art model. We also present multiple reasoning-chain-recovery tests and ablation studies to demonstrate our system’s ability to perform interpretable and accurate reasoning.

Multi-hop question answering requires a model to connect multiple pieces of evidence scattered in a long context to answer the question. In this paper, we show that in the multi-hop HotpotQA (Yang et al., 2018) dataset, the examples often contain reasoning shortcuts through which models can directly locate the answer by word-matching the question with a sentence in the context. We demonstrate this issue by constructing adversarial documents that create contradicting answers to the shortcut but do not affect the validity of the original answer. The performance of strong baseline models drops significantly on our adversarial test, indicating that they are indeed exploiting the shortcuts rather than performing multi-hop reasoning. After adversarial training, the baseline’s performance improves but is still limited on the adversarial test. Hence, we use a control unit that dynamically attends to the question at different reasoning hops to guide the model’s multi-hop reasoning. We show that our 2-hop model trained on the regular data is more robust to the adversaries than the baseline. After adversarial training, it not only achieves significant improvements over its counterpart trained on regular data, but also outperforms the adversarially-trained baseline significantly. Finally, we sanity-check that these improvements are not obtained by exploiting potential new shortcuts in the adversarial data, but indeed due to robust multi-hop reasoning skills of the models.

Multi-hop QA requires a model to connect multiple pieces of evidence scattered in a long context to answer the question. The recently proposed HotpotQA (Yang et al., 2018) dataset is comprised of questions embodying four different multi-hop reasoning paradigms (two bridge entity setups, checking multiple properties, and comparing two entities), making it challenging for a single neural network to handle all four. In this work, we present an interpretable, controller-based Self-Assembling Neural Modular Network (Hu et al., 2017, 2018) for multi-hop reasoning, where we design four novel modules (Find, Relocate, Compare, NoOp) to perform unique types of language reasoning. Based on a question, our layout controller RNN dynamically infers a series of reasoning modules to construct the entire network. Empirically, we show that our dynamic, multi-hop modular network achieves significant improvements over the static, single-hop baseline (on both regular and adversarial evaluation). We further demonstrate the interpretability of our model via three analyses. First, the controller can softly decompose the multi-hop question into multiple single-hop sub-questions to promote compositional reasoning behavior of the main network. Second, the controller can predict layouts that conform to the layouts designed by human experts. Finally, the intermediate module can infer the entity that connects two distantly-located supporting facts by addressing the sub-question from the controller.

A good neural sequence-to-sequence summarization model should have a strong encoder that can distill and memorize the important information from long input texts so that the decoder can generate salient summaries based on the encoder’s memory. In this paper, we aim to improve the memorization capabilities of the encoder of a pointer-generator model by adding an additional ‘closed-book’ decoder without attention and pointer mechanisms. Such a decoder forces the encoder to be more selective in the information encoded in its memory state because the decoder can’t rely on the extra information provided by the attention and possibly copy modules, and hence improves the entire model. On the CNN/Daily Mail dataset, our 2-decoder model outperforms the baseline significantly in terms of ROUGE and METEOR metrics, for both cross-entropy and reinforced setups (and on human evaluation). Moreover, our model also achieves higher scores in a test-only DUC-2002 generalizability setup. We further present a memory ability test, two saliency metrics, as well as several sanity-check ablations (based on fixed-encoder, gradient-flow cut, and model capacity) to prove that the encoder of our 2-decoder model does in fact learn stronger memory representations than the baseline encoder.