There have been a lot of interest in the scaling properties of Transformer models. However, not much has been done on the front of investigating the effect of scaling properties of different inductive biases and model architectures. Do model architectures scale differently? If so, how does inductive bias affect scaling behaviour? How does this influence upstream (pretraining) and downstream (transfer)? This paper conducts a systematic study of scaling behaviour of ten diverse model architectures such as Transformers, Switch Transformers, Universal Transformers, Dynamic convolutions, Performers, and recently proposed MLP-Mixers. Via extensive experiments, we show that (1) architecture is an indeed an important consideration when performing scaling and (2) the best performing model can fluctuate at different scales. We believe that the findings outlined in this work has significant implications to how model architectures are currently evaluated in the community.

Scaling language models improves performance but comes with significant computational costs. This paper proposes UL2R, a method that substantially improves existing language models and their scaling curves with a relatively tiny amount of extra compute. The key idea is to continue training a state-of-the-art large language model on a few more steps with UL2’s mixture-of-denoiser objective. We show that, with almost negligible extra computational costs and no new sources of data, we are able to substantially improve the scaling properties of large language models on downstream metrics. In this paper, we continue training a baseline language model, PaLM, with ULR2, introducing a new set of models at 8B, 62B, and 540B scale which we call U-PaLM. Impressively, at 540B scale, we show an approximately 2x computational savings rate where U-PaLM achieves the same performance as the final PaLM 540B model at around half its computational budget (i.e., saving ~4.4 million TPUv4 hours). We further show that this improved scaling curve leads to “emergent abilities” on challenging BIG-Bench tasks—for instance, U-PaLM does much better on some tasks or demonstrates better quality at much smaller scale (62B as opposed to 540B). Overall, we show that U-PaLM outperforms PaLM on many few-shot setups, including reasoning tasks with chain-of-thought (e.g., GSM8K), multilingual tasks (MGSM, TydiQA), MMLU and challenging BIG-Bench tasks.

Differentiable Search Indices (DSIs) encode a corpus of documents in the parameters of a model and use the same model to map queries directly to relevant document identifiers. Despite the solid performance of DSI models, successfully deploying them in scenarios where document corpora change with time is an open problem. In this work, we introduce DSI++, a continual learning challenge for DSI with the goal of continuously indexing new documents while being able to answer queries related to both previously and newly indexed documents. Across different model scales and document identifier representations, we show that continual indexing of new documents leads to considerable forgetting of previously indexed documents. We also hypothesize and verify that the model experiences forgetting events during training, leading to unstable learning. To mitigate these issues, we investigate two approaches. The first focuses on modifying the training dynamics. Flatter minima implicitly alleviates forgetting, so we explicitly optimize for flatter loss basins and show that the model stably memorizes more documents (+12%). Next, we introduce a parametric memory to generate pseudo-queries for documents and supplement them during incremental indexing to prevent forgetting for the retrieval task. Extensive experiments on a novel continual indexing benchmark based on Natural Questions demonstrate that our proposed solution mitigates the forgetting in DSI++ by a significant margin and improves the average Hits@10 by +21.1% over competitive baselines.

Data-to-text generation focuses on generating fluent natural language responses from structured meaning representations (MRs). Such representations are compositional and it is costly to collect responses for all possible combinations of atomic meaning schemata, thereby necessitating few-shot generalization to novel MRs. In this work, we systematically study the compositional generalization of the state-of-the-art T5 models in few-shot data-to-text tasks. We show that T5 models fail to generalize to unseen MRs, and we propose a template-based input representation that considerably improves the model’s generalization capability. To further improve the model’s performance, we propose an approach based on self-training using fine-tuned BLEURT for pseudo-response selection. On the commonly-used SGD and Weather benchmarks, the proposed self-training approach improves tree accuracy by 46%+ and reduces the slot error rates by 73%+ over the strong T5 baselines in few-shot settings.

Generating fluent natural language responses from structured semantic representations is a critical step in task-oriented conversational systems. Sequence-to-sequence models on flat meaning representations (MR) have been dominant in this task, for example in the E2E NLG Challenge. Previous work has shown that a tree-structured MR can improve the model for better discourse-level structuring and sentence-level planning. In this work, we propose a tree-to-sequence model that uses a tree-LSTM encoder to leverage the tree structures in the input MR, and further enhance the decoding by a structure-enhanced attention mechanism. In addition, we explore combining these enhancements with constrained decoding to improve semantic correctness. Our experiments not only show significant improvements over standard seq2seq baselines, but also is more data-efficient and generalizes better to hard scenarios.

Semantic similarity modeling is central to many NLP problems such as natural language inference and question answering. Syntactic structures interact closely with semantics in learning compositional representations and alleviating long-range dependency issues. How-ever, such structure priors have not been well exploited in previous work for semantic mod-eling. To examine their effectiveness, we start with the Pairwise Word Interaction Model, one of the best models according to a recent reproducibility study, then introduce components for modeling context and structure using multi-layer BiLSTMs and TreeLSTMs. In addition, we introduce residual connections to the deep convolutional neural network component of the model. Extensive evaluations on eight benchmark datasets show that incorporating structural information contributes to consistent improvements over strong baselines.

A core problem of information retrieval (IR) is relevance matching, which is to rank documents by relevance to a user’s query. On the other hand, many NLP problems, such as question answering and paraphrase identification, can be considered variants of semantic matching, which is to measure the semantic distance between two pieces of short texts. While at a high level both relevance and semantic matching require modeling textual similarity, many existing techniques for one cannot be easily adapted to the other. To bridge this gap, we propose a novel model, HCAN (Hybrid Co-Attention Network), that comprises (1) a hybrid encoder module that includes ConvNet-based and LSTM-based encoders, (2) a relevance matching module that measures soft term matches with importance weighting at multiple granularities, and (3) a semantic matching module with co-attention mechanisms that capture context-aware semantic relatedness. Evaluations on multiple IR and NLP benchmarks demonstrate state-of-the-art effectiveness compared to approaches that do not exploit pretraining on external data. Extensive ablation studies suggest that relevance and semantic matching signals are complementary across many problem settings, regardless of the choice of underlying encoders.

We describe our exploratory system for the shallow surface realization task, which combines morphological inflection using character sequence-to-sequence models with a baseline linearizer that implements a tree-to-tree model using sequence-to-sequence models on serialized trees. Results for morphological inflection were competitive across languages. Due to time constraints, we could only submit complete results (including linearization) for English. Preliminary linearization results were decent, with a small benefit from reranking to prefer valid output trees, but inadequate control over the words in the output led to poor quality on longer sentences.

This paper explores the problem of ranking short social media posts with respect to user queries using neural networks. Instead of starting with a complex architecture, we proceed from the bottom up and examine the effectiveness of a simple, word-level Siamese architecture augmented with attention-based mechanisms for capturing semantic “soft” matches between query and post tokens. Extensive experiments on datasets from the TREC Microblog Tracks show that our simple models not only achieve better effectiveness than existing approaches that are far more complex or exploit a more diverse set of relevance signals, but are also much faster.

Generating fluent natural language responses from structured semantic representations is a critical step in task-oriented conversational systems. Avenues like the E2E NLG Challenge have encouraged the development of neural approaches, particularly sequence-to-sequence (Seq2Seq) models for this problem. The semantic representations used, however, are often underspecified, which places a higher burden on the generation model for sentence planning, and also limits the extent to which generated responses can be controlled in a live system. In this paper, we (1) propose using tree-structured semantic representations, like those used in traditional rule-based NLG systems, for better discourse-level structuring and sentence-level planning; (2) introduce a challenging dataset using this representation for the weather domain; (3) introduce a constrained decoding approach for Seq2Seq models that leverages this representation to improve semantic correctness; and (4) demonstrate promising results on our dataset and the E2E dataset.

Many state-of-the-art neural models for NLP are heavily parameterized and thus memory inefficient. This paper proposes a series of lightweight and memory efficient neural architectures for a potpourri of natural language processing (NLP) tasks. To this end, our models exploit computation using Quaternion algebra and hypercomplex spaces, enabling not only expressive inter-component interactions but also significantly (75%) reduced parameter size due to lesser degrees of freedom in the Hamilton product. We propose Quaternion variants of models, giving rise to new architectures such as the Quaternion attention Model and Quaternion Transformer. Extensive experiments on a battery of NLP tasks demonstrates the utility of proposed Quaternion-inspired models, enabling up to 75% reduction in parameter size without significant loss in performance.

This paper tackles the problem of reading comprehension over long narratives where documents easily span over thousands of tokens. We propose a curriculum learning (CL) based Pointer-Generator framework for reading/sampling over large documents, enabling diverse training of the neural model based on the notion of alternating contextual difficulty. This can be interpreted as a form of domain randomization and/or generative pretraining during training. To this end, the usage of the Pointer-Generator softens the requirement of having the answer within the context, enabling us to construct diverse training samples for learning. Additionally, we propose a new Introspective Alignment Layer (IAL), which reasons over decomposed alignments using block-based self-attention. We evaluate our proposed method on the NarrativeQA reading comprehension benchmark, achieving state-of-the-art performance, improving existing baselines by 51% relative improvement on BLEU-4 and 17% relative improvement on Rouge-L. Extensive ablations confirm the effectiveness of our proposed IAL and CL components.