Tree-Planted Transformers: Large Language Models with Implicit Syntactic Supervision

Large Language Models (LLMs) have achieved remarkable success thanks to scalability on large text corpora, but have some drawback in training efficiency. In contrast, Syntactic Language Models (SLMs) can be trained efficiently to reach relatively high performance thanks to syntactic supervision, but have trouble with scalability. Thus, given these complementary advantages of LLMs and SLMs, it is necessary to develop an architecture that integrates the scalability of LLMs with the training efficiency of SLMs, namely Syntactic Large Language Models (SLLM). In this paper, we propose a novel method dubbed tree-planting: implicitly "plant" trees into attention weights of Transformer LMs to reflect syntactic structures of natural language. Specifically, Transformer LMs trained with tree-planting will be called Tree-Planted Transformers (TPT), which learn syntax on small treebanks via tree-planting and then scale on large text corpora via continual learning with syntactic scaffolding. Targeted syntactic evaluations on the SyntaxGym benchmark demonstrated that TPTs, despite the lack of explicit syntactic supervision, significantly outperformed various SLMs with explicit syntactic supervision that generate hundreds of syntactic structures in parallel, suggesting that tree-planting and TPTs are the promising foundation for SLLMs.

Emergent Word Order Universals from Cognitively-Motivated Language Models

The world's languages exhibit certain so-called typological or implicational universals; for example, Subject-Object-Verb (SOV) word order typically employs postpositions. Explaining the source of such biases is a key goal in linguistics. We study the word-order universals through a computational simulation with language models (LMs). Our experiments show that typologically typical word orders tend to have lower perplexity estimated by LMs with cognitively plausible biases: syntactic biases, specific parsing strategies, and memory limitations. This suggests that the interplay of these cognitive biases and predictability (perplexity) can explain many aspects of word-order universals. This also showcases the advantage of cognitively-motivated LMs, which are typically employed in cognitive modeling, in the computational simulation of language universals.

Shotgun crystal structure prediction using machine-learned formation energies

Stable or metastable crystal structures of assembled atoms can be predicted by finding the global or local minima of the energy surface with respect to the atomic configurations. Generally, this requires repeated first-principles energy calculations that are impractical for large systems, such as those containing more than 30 atoms in the unit cell. Here, we have made significant progress in solving the crystal structure prediction problem with a simple but powerful machine-learning workflow; using a machine-learning surrogate for first-principles energy calculations, we performed non-iterative, single-shot screening using a large library of virtually created crystal structures. The present method relies on two key technical components: transfer learning, which enables a highly accurate energy prediction of pre-relaxed crystalline states given only a small set of training samples from first-principles calculations, and generative models to create promising and diverse crystal structures for screening. Here, first-principles calculations were performed only to generate the training samples, and for the optimization of a dozen or fewer finally narrowed-down crystal structures. Our shotgun method was more than 5--10 times less computationally demanding and achieved an outstanding prediction accuracy that was 2--6 times higher than that of the conventional methods that rely heavily on iterative first-principles calculations.

Composition, Attention, or Both?

In this paper, we propose a novel architecture called Composition Attention Grammars (CAGs) that recursively compose subtrees into a single vector representation with a composition function, and selectively attend to previous structural information with a self-attention mechanism. We investigate whether these components -- the composition function and the self-attention mechanism -- can both induce human-like syntactic generalization. Specifically, we train language models (LMs) with and without these two components with the model sizes carefully controlled, and evaluate their syntactic generalization performance against six test circuits on the SyntaxGym benchmark. The results demonstrated that the composition function and the self-attention mechanism both play an important role to make LMs more human-like, and closer inspection of linguistic phenomenon implied that the composition function allowed syntactic features, but not semantic features, to percolate into subtree representations.

Transfer learning with affine model transformation

Supervised transfer learning (TL) has received considerable attention because of its potential to boost the predictive power of machine learning in cases with limited data. In a conventional scenario, cross-domain differences are modeled and estimated using a given set of source models and samples from a target domain. For example, if there is a functional relationship between source and target domains, only domain-specific factors are additionally learned using target samples to shift the source models to the target. However, the general methodology for modeling and estimating such cross-domain shifts has been less studied. This study presents a TL framework that simultaneously and separately estimates domain shifts and domain-specific factors using given target samples. Assuming consistency and invertibility of the domain transformation functions, we derive an optimal family of functions to represent the cross-domain shift. The newly derived class of transformation functions takes the same form as invertible neural networks using affine coupling layers, which are widely used in generative deep learning. We show that the proposed method encompasses a wide range of existing methods, including the most common TL procedure based on feature extraction using neural networks. We also clarify the theoretical properties of the proposed method, such as the convergence rate of the generalization error, and demonstrate the practical benefits of separately modeling and estimating domain-specific factors through several case studies.

Bayesian Sequential Stacking Algorithm for Concurrently Designing Molecules and Synthetic Reaction Networks

In the last few years, de novo molecular design using machine learning has made great technical progress but its practical deployment has not been as successful. This is mostly owing to the cost and technical difficulty of synthesizing such computationally designed molecules. To overcome such barriers, various methods for synthetic route design using deep neural networks have been studied intensively in recent years. However, little progress has been made in designing molecules and their synthetic routes simultaneously. Here, we formulate the problem of simultaneously designing molecules with the desired set of properties and their synthetic routes within the framework of Bayesian inference. The design variables consist of a set of reactants in a reaction network and its network topology. The design space is extremely large because it consists of all combinations of purchasable reactants, often in the order of millions or more. In addition, the designed reaction networks can adopt any topology beyond simple multistep linear reaction routes. To solve this hard combinatorial problem, we present a powerful sequential Monte Carlo algorithm that recursively designs a synthetic reaction network by sequentially building up single-step reactions. In a case study of designing drug-like molecules based on commercially available compounds, compared with heuristic combinatorial search methods, the proposed method shows overwhelming performance in terms of computational efficiency and coverage and novelty with respect to existing compounds.

Crystal structure prediction with machine learning-based element substitution

The prediction of energetically stable crystal structures formed by a given chemical composition is a central problem in solid-state physics. In principle, the crystalline state of assembled atoms can be determined by optimizing the energy surface, which in turn can be evaluated using first-principles calculations. However, performing the iterative gradient descent on the potential energy surface using first-principles calculations is prohibitively expensive for complex systems, such as those with many atoms per unit cell. Here, we present a unique methodology for crystal structure prediction (CSP) that relies on a machine learning algorithm called metric learning. It is shown that a binary classifier, trained on a large number of already identified crystal structures, can determine the isomorphism of crystal structures formed by two given chemical compositions with an accuracy of approximately 96.4\%. For a given query composition with an unknown crystal structure, the model is used to automatically select from a crystal structure database a set of template crystals with nearly identical stable structures to which element substitution is to be applied. Apart from the local relaxation calculation of the identified templates, the proposed method does not use ab initio calculations. The potential of this substation-based CSP is demonstrated for a wide variety of crystal systems.

Modeling Human Sentence Processing with Left-Corner Recurrent Neural Network Grammars

In computational linguistics, it has been shown that hierarchical structures make language models (LMs) more human-like. However, the previous literature has been agnostic about a parsing strategy of the hierarchical models. In this paper, we investigated whether hierarchical structures make LMs more human-like, and if so, which parsing strategy is most cognitively plausible. In order to address this question, we evaluated three LMs against human reading times in Japanese with head-final left-branching structures: Long Short-Term Memory (LSTM) as a sequential model and Recurrent Neural Network Grammars (RNNGs) with top-down and left-corner parsing strategies as hierarchical models. Our computational modeling demonstrated that left-corner RNNGs outperformed top-down RNNGs and LSTM, suggesting that hierarchical and left-corner architectures are more cognitively plausible than top-down or sequential architectures. In addition, the relationships between the cognitive plausibility and (i) perplexity, (ii) parsing, and (iii) beam size will also be discussed.

Lower Perplexity is Not Always Human-Like

In computational psycholinguistics, various language models have been evaluated against human reading behavior (e.g., eye movement) to build human-like computational models. However, most previous efforts have focused almost exclusively on English, despite the recent trend towards linguistic universal within the general community. In order to fill the gap, this paper investigates whether the established results in computational psycholinguistics can be generalized across languages. Specifically, we re-examine an established generalization -- the lower perplexity a language model has, the more human-like the language model is -- in Japanese with typologically different structures from English. Our experiments demonstrate that this established generalization exhibits a surprising lack of universality; namely, lower perplexity is not always human-like. Moreover, this discrepancy between English and Japanese is further explored from the perspective of (non-)uniform information density. Overall, our results suggest that a cross-lingual evaluation will be necessary to construct human-like computational models.