Yoshimasa Tsuruoka

Sosuke Nishikawa, Ryokan Ri, Ikuya Yamada et al.

We present EASE, a novel method for learning sentence embeddings via contrastive learning between sentences and their related entities. The advantage of using entity supervision is twofold: (1) entities have been shown to be a strong indicator of text semantics and thus should provide rich training signals for sentence embeddings; (2) entities are defined independently of languages and thus offer useful cross-lingual alignment supervision. We evaluate EASE against other unsupervised models both in monolingual and multilingual settings. We show that EASE exhibits competitive or better performance in English semantic textual similarity (STS) and short text clustering (STC) tasks and it significantly outperforms baseline methods in multilingual settings on a variety of tasks. Our source code, pre-trained models, and newly constructed multilingual STC dataset are available at https://github.com/studio-ousia/ease.

Ryokan Ri, Yoshimasa Tsuruoka

We investigate what kind of structural knowledge learned in neural network encoders is transferable to processing natural language. We design artificial languages with structural properties that mimic natural language, pretrain encoders on the data, and see how much performance the encoder exhibits on downstream tasks in natural language. Our experimental results show that pretraining with an artificial language with a nesting dependency structure provides some knowledge transferable to natural language. A follow-up probing analysis indicates that its success in the transfer is related to the amount of encoded contextual information and what is transferred is the knowledge of position-aware context dependence of language. Our results provide insights into how neural network encoders process human languages and the source of cross-lingual transferability of recent multilingual language models.

Kaiyan Zhao, Qiyu Wu, Xin-Qiang Cai et al.

Learning multi-lingual sentence embeddings is a fundamental task in natural language processing. Recent trends in learning both mono-lingual and multi-lingual sentence embeddings are mainly based on contrastive learning (CL) among an anchor, one positive, and multiple negative instances. In this work, we argue that leveraging multiple positives should be considered for multi-lingual sentence embeddings because (1) positives in a diverse set of languages can benefit cross-lingual learning, and (2) transitive similarity across multiple positives can provide reliable structural information for learning. In order to investigate the impact of multiple positives in CL, we propose a novel approach, named MPCL, to effectively utilize multiple positive instances to improve the learning of multi-lingual sentence embeddings. Experimental results on various backbone models and downstream tasks demonstrate that MPCL leads to better retrieval, semantic similarity, and classification performances compared to conventional CL. We also observe that in unseen languages, sentence embedding models trained on multiple positives show better cross-lingual transfer performance than models trained on a single positive instance.

Qiyu Wu, Masaaki Nagata, Yoshimasa Tsuruoka

Most existing word alignment methods rely on manual alignment datasets or parallel corpora, which limits their usefulness. Here, to mitigate the dependence on manual data, we broaden the source of supervision by relaxing the requirement for correct, fully-aligned, and parallel sentences. Specifically, we make noisy, partially aligned, and non-parallel paragraphs. We then use such a large-scale weakly-supervised dataset for word alignment pre-training via span prediction. Extensive experiments with various settings empirically demonstrate that our approach, which is named WSPAlign, is an effective and scalable way to pre-train word aligners without manual data. When fine-tuned on standard benchmarks, WSPAlign has set a new state-of-the-art by improving upon the best-supervised baseline by 3.3~6.1 points in F1 and 1.5~6.1 points in AER. Furthermore, WSPAlign also achieves competitive performance compared with the corresponding baselines in few-shot, zero-shot and cross-lingual tests, which demonstrates that WSPAlign is potentially more practical for low-resource languages than existing methods.

Hayato Watahiki, Ryo Iwase, Ryosuke Unno et al.

Transferring learned skills across diverse situations remains a fundamental challenge for autonomous agents, particularly when agents are not allowed to interact with an exact target setup. While prior approaches have predominantly focused on learning domain translation, they often struggle with handling significant domain gaps or out-of-distribution tasks. In this paper, we present a simple approach for cross-domain policy transfer that learns a shared latent representation across domains and a common abstract policy on top of it. Our approach leverages multi-domain behavioral cloning on unaligned trajectories of proxy tasks and employs maximum mean discrepancy (MMD) as a regularization term to encourage cross-domain alignment. The MMD regularization better preserves structures of latent state distributions than commonly used domain-discriminative distribution matching, leading to higher transfer performance. Moreover, our approach involves training only one multi-domain policy, which makes extension easier than existing methods. Empirical evaluations demonstrate the efficacy of our method across various domain shifts, especially in scenarios where exact domain translation is challenging, such as cross-morphology or cross-viewpoint settings. Our ablation studies further reveal that multi-domain behavioral cloning implicitly contributes to representation alignment alongside domain-adversarial regularization.

Takuya Hiraoka, Takashi Onishi, Yoshimasa Tsuruoka

In reinforcement learning (RL) with experience replay, experiences stored in a replay buffer influence the RL agent's performance. Information about the influence is valuable for various purposes, including experience cleansing and analysis. One method for estimating the influence of individual experiences is agent comparison, but it is prohibitively expensive when there is a large number of experiences. In this paper, we present PI+ToD as a method for efficiently estimating the influence of experiences. PI+ToD is a policy iteration that efficiently estimates the influence of experiences by utilizing turn-over dropout. We demonstrate the efficiency of PI+ToD with experiments in MuJoCo environments.

Shumpei Kubosawa, Takashi Onishi, Yoshimasa Tsuruoka

During the operation of a chemical plant, product quality must be consistently maintained, and the production of off-specification products should be minimized. Accordingly, process variables related to the product quality, such as the temperature and composition of materials at various parts of the plant must be measured, and appropriate operations (that is, control) must be performed based on the measurements. Some process variables, such as temperature and flow rate, can be measured continuously and instantaneously. However, other variables, such as composition and viscosity, can only be obtained through time-consuming analysis after sampling substances from the plant. Soft sensors have been proposed for estimating process variables that cannot be obtained in real time from easily measurable variables. However, the estimation accuracy of conventional statistical soft sensors, which are constructed from recorded measurements, can be very poor in unrecorded situations (extrapolation). In this study, we estimate the internal state variables of a plant by using a dynamic simulator that can estimate and predict even unrecorded situations on the basis of chemical engineering knowledge and an artificial intelligence (AI) technology called reinforcement learning, and propose to use the estimated internal state variables of a plant as soft sensors. In addition, we describe the prospects for plant operation and control using such soft sensors and the methodology to obtain the necessary prediction models (i.e., simulators) for the proposed system.

Zhongtao Miao, Kaiyan Zhao, Yoshimasa Tsuruoka

Current representations used in reasoning steps of large language models can mostly be categorized into two main types: (1) natural language, which is difficult to verify; and (2) non-natural language, usually programming code, which is difficult for people who are unfamiliar with coding to read. In this paper, we propose to use a semi-structured form to represent reasoning steps of large language models. Specifically, we use relation tuples, which are not only human-readable but also machine-friendly and easier to verify than natural language. We implement a framework that includes three main components: (1) introducing relation tuples into the reasoning steps of large language models; (2) implementing an automatic verification process of reasoning steps with a local code interpreter based on relation tuples; and (3) integrating a simple and effective dynamic feedback mechanism, which we found helpful for self-improvement of large language models. The experimental results on various arithmetic datasets demonstrate the effectiveness of our method in improving the arithmetic reasoning ability of large language models. The source code is available at https://github.com/gpgg/art.

Ryokan Ri, Ikuya Yamada, Yoshimasa Tsuruoka

Recent studies have shown that multilingual pretrained language models can be effectively improved with cross-lingual alignment information from Wikipedia entities. However, existing methods only exploit entity information in pretraining and do not explicitly use entities in downstream tasks. In this study, we explore the effectiveness of leveraging entity representations for downstream cross-lingual tasks. We train a multilingual language model with 24 languages with entity representations and show the model consistently outperforms word-based pretrained models in various cross-lingual transfer tasks. We also analyze the model and the key insight is that incorporating entity representations into the input allows us to extract more language-agnostic features. We also evaluate the model with a multilingual cloze prompt task with the mLAMA dataset. We show that entity-based prompt elicits correct factual knowledge more likely than using only word representations. Our source code and pretrained models are available at https://github.com/studio-ousia/luke.

Zhongtao Miao, Qiyu Wu, Yoshimasa Tsuruoka

Text embedding and generative tasks are usually trained separately based on large language models (LLMs) nowadays. This causes a large amount of training cost and deployment effort. Context compression is also a challenging and pressing task, which is vital to reasoning-driven generation, and agentic tasks requiring long context and continual learning. In this paper, we explore how to unify reasoning-driven generation, reasoning-enhanced text representation and context compression tasks in one forward pass for LLMs. Through meta latent tokens and a unified generative, representative and compressive tuning approach, we propose a training framework named GRC that bridges the three tasks. The trained models can accomplish three objectives in a single forward pass while maintaining modular, LEGO-style flexibility during inference. This design greatly reduces the deployment effort for retrieval-augmented generation (RAG) and achieves efficient inference and three times data utilization during training. Furthermore, this framework design enables a new paradigm for text embedding: self-reason-latent embeds, and a new generation paradigm, latent memory-augmented generation, where compressed and internalized KV cache with O(1) length is used as the updatable memory. We also propose hybrid paged attention to speed up the inference of our models. Extensive experiments on reasoning-intensive retrieval benchmarks, generative tasks, document compression, latency evaluation, and RAG settings demonstrate the effectiveness of our method and may shed light on the truly unified model that can handle reasoning-driven generation, embedding and compression tasks seamlessly.

Zhongtao Miao, Kaiyan Zhao, Masaaki Nagata et al.

Neologism-aware machine translation aims to translate source sentences containing neologisms into target languages. This field remains underexplored compared with general machine translation (MT). In this paper, we propose an agentic framework, NeoAMT, for neologism-aware machine translation using a Wiktionary search tool. Specifically, we first create a new dataset for neologism-aware machine translation and develop a search tool based on Wiktionary. The new dataset covers 16 languages and 75 translation directions and is derived from approximately 10 million records of an English Wiktionary dump. The retrieval corpus of the search tool is also constructed from around 3 million cleaned records of the Wiktionary dump. We then use it for training the translation agent with reinforcement learning (RL) and evaluating the accuracy of neologism-aware machine translation. Based on this, we also propose an RL training framework that contains a novel reward design and an adaptive rollout generation approach by leveraging "translation difficulty" to further improve the translation quality of translation agents using our search tool.

Zhongtao Miao, Qiyu Wu, Kaiyan Zhao et al.

The field of cross-lingual sentence embeddings has recently experienced significant advancements, but research concerning low-resource languages has lagged due to the scarcity of parallel corpora. This paper shows that cross-lingual word representation in low-resource languages is notably under-aligned with that in high-resource languages in current models. To address this, we introduce a novel framework that explicitly aligns words between English and eight low-resource languages, utilizing off-the-shelf word alignment models. This framework incorporates three primary training objectives: aligned word prediction and word translation ranking, along with the widely used translation ranking. We evaluate our approach through experiments on the bitext retrieval task, which demonstrate substantial improvements on sentence embeddings in low-resource languages. In addition, the competitive performance of the proposed model across a broader range of tasks in high-resource languages underscores its practicality.

Qiyu Wu, Masaaki Nagata, Zhongtao Miao et al.

The problem of hallucination and omission, a long-standing problem in machine translation (MT), is more pronounced when a large language model (LLM) is used in MT because an LLM itself is susceptible to these phenomena. In this work, we mitigate the problem in an LLM-based MT model by guiding it to better word alignment. We first study the correlation between word alignment and the phenomena of hallucination and omission in MT. Then we propose to utilize word alignment as preference to optimize the LLM-based MT model. The preference data are constructed by selecting chosen and rejected translations from multiple MT tools. Subsequently, direct preference optimization is used to optimize the LLM-based model towards the preference signal. Given the absence of evaluators specifically designed for hallucination and omission in MT, we further propose selecting hard instances and utilizing GPT-4 to directly evaluate the performance of the models in mitigating these issues. We verify the rationality of these designed evaluation methods by experiments, followed by extensive results demonstrating the effectiveness of word alignment-based preference optimization to mitigate hallucination and omission. On the other hand, although it shows promise in mitigating hallucination and omission, the overall performance of MT in different language directions remains mixed, with slight increases in BLEU and decreases in COMET.

Megha Sharma, Muhammad Taimoor Haseeb, Gus Xia et al.

This paper introduces M2M Gen, a multi modal framework for generating background music tailored to Japanese manga. The key challenges in this task are the lack of an available dataset or a baseline. To address these challenges, we propose an automated music generation pipeline that produces background music for an input manga book. Initially, we use the dialogues in a manga to detect scene boundaries and perform emotion classification using the characters faces within a scene. Then, we use GPT4o to translate this low level scene information into a high level music directive. Conditioned on the scene information and the music directive, another instance of GPT 4o generates page level music captions to guide a text to music model. This produces music that is aligned with the mangas evolving narrative. The effectiveness of M2M Gen is confirmed through extensive subjective evaluations, showcasing its capability to generate higher quality, more relevant and consistent music that complements specific scenes when compared to our baselines.

Kaiyan Zhao, Tsuguchika Tabaru, Kenichi Kobayashi et al.

Although recent quantized Large Language Models (LLMs), such as BitNet, have paved the way for significant reduction in memory usage during deployment with binary or ternary weights, training these models still demands substantial memory footprints. This is partly because high-precision (i.e., unquantized) weights required for straight-through estimation must be maintained throughout the whole training process. To address this, we explore directly updating the quantized low-precision weights without relying on straight-through estimation during backpropagation, aiming to save memory usage during training. Specifically, we employ a stochastic rounding technique to minimize the information loss caused by the use of low-bit weights throughout training. Experimental results on our LLaMA-structured models of various sizes indicate that (1) training with only low-precision weights is feasible even when they are constrained to ternary values; (2) extending the bit width to 8 bits achieves performance on par with BitNet b1.58; (3) our models remain robust to precision scaling and memory reduction, showing minimal performance degradation when moving from FP32 to lower-memory environments (BF16/FP8); and (4) our models also support inference using ternary weights, showcasing their flexibility in deployment.

Kaiyan Zhao, Zhongtao Miao, Yoshimasa Tsuruoka

Multimodal sentence embedding models typically leverage image-caption pairs in addition to textual data during training. However, such pairs often contain noise, including redundant or irrelevant information on either the image or caption side. To mitigate this issue, we propose MCSEO, a method that enhances multimodal sentence embeddings by incorporating fine-grained object-phrase alignment alongside traditional image-caption alignment. Specifically, MCSEO utilizes existing segmentation and object detection models to extract accurate object-phrase pairs, which are then used to optimize a contrastive learning objective tailored to object-phrase correspondence. Experimental results on semantic textual similarity (STS) tasks across different backbone models demonstrate that MCSEO consistently outperforms strong baselines, highlighting the significance of precise object-phrase alignment in multimodal representation learning.

Guanquan Wang, Takuya Hiraoka, Yoshimasa Tsuruoka

This paper introduces Consistency Trajectory Planning (CTP), a novel offline model-based reinforcement learning method that leverages the recently proposed Consistency Trajectory Model (CTM) for efficient trajectory optimization. While prior work applying diffusion models to planning has demonstrated strong performance, it often suffers from high computational costs due to iterative sampling procedures. CTP supports fast, single-step trajectory generation without significant degradation in policy quality. We evaluate CTP on the D4RL benchmark and show that it consistently outperforms existing diffusion-based planning methods in long-horizon, goal-conditioned tasks. Notably, CTP achieves higher normalized returns while using significantly fewer denoising steps. In particular, CTP achieves comparable performance with over $120\times$ speedup in inference time, demonstrating its practicality and effectiveness for high-performance, low-latency offline planning.

Takuya Hiraoka, Guanquan Wang, Takashi Onishi et al.

In reinforcement learning (RL) with experience replay, experiences stored in a replay buffer influence the RL agent's performance. Information about how these experiences influence the agent's performance is valuable for various purposes, such as identifying experiences that negatively influence underperforming agents. One method for estimating the influence of experiences is the leave-one-out (LOO) method. However, this method is usually computationally prohibitive. In this paper, we present Policy Iteration with Turn-over Dropout (PIToD), which efficiently estimates the influence of experiences. We evaluate how correctly PIToD estimates the influence of experiences and its efficiency compared to LOO. We then apply PIToD to amend underperforming RL agents, i.e., we use PIToD to estimate negatively influential experiences for the RL agents and to delete the influence of these experiences. We show that RL agents' performance is significantly improved via amendments with PIToD.

Takahisa Imagawa, Takuya Hiraoka, Yoshimasa Tsuruoka

Recently, methods for learning diverse skills to generate various behaviors without external rewards have been actively studied as a form of unsupervised reinforcement learning. However, most of the existing methods learn a finite number of discrete skills, and thus the variety of behaviors that can be exhibited with the learned skills is limited. In this paper, we propose a novel method for learning potentially an infinite number of different skills, which is named discovery of continuous skills on a sphere (DISCS). In DISCS, skills are learned by maximizing mutual information between skills and states, and each skill corresponds to a continuous value on a sphere. Because the representations of skills in DISCS are continuous, infinitely diverse skills could be learned. We examine existing methods and DISCS in the MuJoCo Ant robot control environments and show that DISCS can learn much more diverse skills than the other methods.

Shumpei Kubosawa, Takashi Onishi, Makoto Sakahara et al.

The number of railway service disruptions has been increasing owing to intensification of natural disasters. In addition, abrupt changes in social situations such as the COVID-19 pandemic require railway companies to modify the traffic schedule frequently. Therefore, automatic support for optimal scheduling is anticipated. In this study, an automatic railway scheduling system is presented. The system leverages reinforcement learning and a dynamic simulator that can simulate the railway traffic and passenger flow of a whole line. The proposed system enables rapid generation of the traffic schedule of a whole line because the optimization process is conducted in advance as the training. The system is evaluated using an interruption scenario, and the results demonstrate that the system can generate optimized schedules of the whole line in a few minutes.

Sosuke Nishikawa, Ikuya Yamada, Yoshimasa Tsuruoka et al.

We present a multilingual bag-of-entities model that effectively boosts the performance of zero-shot cross-lingual text classification by extending a multilingual pre-trained language model (e.g., M-BERT). It leverages the multilingual nature of Wikidata: entities in multiple languages representing the same concept are defined with a unique identifier. This enables entities described in multiple languages to be represented using shared embeddings. A model trained on entity features in a resource-rich language can thus be directly applied to other languages. Our experimental results on cross-lingual topic classification (using the MLDoc and TED-CLDC datasets) and entity typing (using the SHINRA2020-ML dataset) show that the proposed model consistently outperforms state-of-the-art models.

Takuya Hiraoka, Takahisa Imagawa, Taisei Hashimoto et al.

Randomized ensembled double Q-learning (REDQ) (Chen et al., 2021b) has recently achieved state-of-the-art sample efficiency on continuous-action reinforcement learning benchmarks. This superior sample efficiency is made possible by using a large Q-function ensemble. However, REDQ is much less computationally efficient than non-ensemble counterparts such as Soft Actor-Critic (SAC) (Haarnoja et al., 2018a). To make REDQ more computationally efficient, we propose a method of improving computational efficiency called DroQ, which is a variant of REDQ that uses a small ensemble of dropout Q-functions. Our dropout Q-functions are simple Q-functions equipped with dropout connection and layer normalization. Despite its simplicity of implementation, our experimental results indicate that DroQ is doubly (sample and computationally) efficient. It achieved comparable sample efficiency with REDQ, much better computational efficiency than REDQ, and comparable computational efficiency with that of SAC.

Ryokan Ri, Toshiaki Nakazawa, Yoshimasa Tsuruoka

Placeholder translation systems enable the users to specify how a specific phrase is translated in the output sentence. The system is trained to output special placeholder tokens, and the user-specified term is injected into the output through the context-free replacement of the placeholder token. However, this approach could result in ungrammatical sentences because it is often the case that the specified term needs to be inflected according to the context of the output, which is unknown before the translation. To address this problem, we propose a novel method of placeholder translation that can inflect specified terms according to the grammatical construction of the output sentence. We extend the sequence-to-sequence architecture with a character-level decoder that takes the lemma of a user-specified term and the words generated from the word-level decoder to output the correct inflected form of the lemma. We evaluate our approach with a Japanese-to-English translation task in the scientific writing domain, and show that our model can incorporate specified terms in the correct form more successfully than other comparable models.

Ryokan Ri, Toshiaki Nakazawa, Yoshimasa Tsuruoka

For Japanese-to-English translation, zero pronouns in Japanese pose a challenge, since the model needs to infer and produce the corresponding pronoun in the target side of the English sentence. However, although fully resolving zero pronouns often needs discourse context, in some cases, the local context within a sentence gives clues to the inference of the zero pronoun. In this study, we propose a data augmentation method that provides additional training signals for the translation model to learn correlations between local context and zero pronouns. We show that the proposed method significantly improves the accuracy of zero pronoun translation with machine translation experiments in the conversational domain.

Taisei Hashimoto, Yoshimasa Tsuruoka

In many deep reinforcement learning settings, when an agent takes an action, it repeats the same action a predefined number of times without observing the states until the next action-decision point. This technique of action repetition has several merits in training the agent, but the data between action-decision points (i.e., intermediate frames) are, in effect, discarded. Since the amount of training data is inversely proportional to the interval of action repeats, they can have a negative impact on the sample efficiency of training. In this paper, we propose a simple but effective approach to alleviate to this problem by introducing the concept of pseudo-actions. The key idea of our method is making the transition between action-decision points usable as training data by considering pseudo-actions. Pseudo-actions for continuous control tasks are obtained as the average of the action sequence straddling an action-decision point. For discrete control tasks, pseudo-actions are computed from learned action embeddings. This method can be combined with any model-free reinforcement learning algorithm that involves the learning of Q-functions. We demonstrate the effectiveness of our approach on both continuous and discrete control tasks in OpenAI Gym.

Takahisa Imagawa, Takuya Hiraoka, Yoshimasa Tsuruoka

Meta-reinforcement learning (RL) addresses the problem of sample inefficiency in deep RL by using experience obtained in past tasks for a new task to be solved. However, most meta-RL methods require partially or fully on-policy data, i.e., they cannot reuse the data collected by past policies, which hinders the improvement of sample efficiency. To alleviate this problem, we propose a novel off-policy meta-RL method, embedding learning and evaluation of uncertainty (ELUE). An ELUE agent is characterized by the learning of a feature embedding space shared among tasks. It learns a belief model over the embedding space and a belief-conditional policy and Q-function. Then, for a new task, it collects data by the pretrained policy, and updates its belief based on the belief model. Thanks to the belief update, the performance can be improved with a small amount of data. In addition, it updates the parameters of the neural networks to adjust the pretrained relationships when there are enough data. We demonstrate that ELUE outperforms state-of-the-art meta RL methods through experiments on meta-RL benchmarks.

Takuya Hiraoka, Takahisa Imagawa, Voot Tangkaratt et al.

Model-based meta-reinforcement learning (RL) methods have recently been shown to be a promising approach to improving the sample efficiency of RL in multi-task settings. However, the theoretical understanding of those methods is yet to be established, and there is currently no theoretical guarantee of their performance in a real-world environment. In this paper, we analyze the performance guarantee of model-based meta-RL methods by extending the theorems proposed by Janner et al. (2019). On the basis of our theoretical results, we propose Meta-Model-Based Meta-Policy Optimization (M3PO), a model-based meta-RL method with a performance guarantee. We demonstrate that M3PO outperforms existing meta-RL methods in continuous-control benchmarks.

Sosuke Nishikawa, Ryokan Ri, Yoshimasa Tsuruoka

Unsupervised cross-lingual word embedding (CLWE) methods learn a linear transformation matrix that maps two monolingual embedding spaces that are separately trained with monolingual corpora. This method relies on the assumption that the two embedding spaces are structurally similar, which does not necessarily hold true in general. In this paper, we argue that using a pseudo-parallel corpus generated by an unsupervised machine translation model facilitates the structural similarity of the two embedding spaces and improves the quality of CLWEs in the unsupervised mapping method. We show that our approach outperforms other alternative approaches given the same amount of data, and, through detailed analysis, we show that data augmentation with the pseudo data from unsupervised machine translation is especially effective for mapping-based CLWEs because (1) the pseudo data makes the source and target corpora (partially) parallel; (2) the pseudo data contains information on the original language that helps to learn similar embedding spaces between the source and target languages.

Ryokan Ri, Yoshimasa Tsuruoka

Existing approaches to mapping-based cross-lingual word embeddings are based on the assumption that the source and target embedding spaces are structurally similar. The structures of embedding spaces largely depend on the co-occurrence statistics of each word, which the choice of context window determines. Despite this obvious connection between the context window and mapping-based cross-lingual embeddings, their relationship has been underexplored in prior work. In this work, we provide a thorough evaluation, in various languages, domains, and tasks, of bilingual embeddings trained with different context windows. The highlight of our findings is that increasing the size of both the source and target window sizes improves the performance of bilingual lexicon induction, especially the performance on frequent nouns.

Takahisa Imagawa, Takuya Hiraoka, Yoshimasa Tsuruoka

Reinforcement Learning, a machine learning framework for training an autonomous agent based on rewards, has shown outstanding results in various domains. However, it is known that learning a good policy is difficult in a domain where rewards are rare. We propose a method, optimistic proximal policy optimization (OPPO) to alleviate this difficulty. OPPO considers the uncertainty of the estimated total return and optimistically evaluates the policy based on that amount. We show that OPPO outperforms the existing methods in a tabular task.

Shiqi Gao, Fuminori Okuya, Yoshihiro Kawahara et al.

The evaluation function for imperfect information games is always hard to define but owns a significant impact on the playing strength of a program. Deep learning has made great achievements these years, and already exceeded the top human players' level even in the game of Go. In this paper, we introduce a new data model to represent the available imperfect information on the game table, and construct a well-designed convolutional neural network for game record training. We choose the accuracy of tile discarding which is also called as the agreement rate as the benchmark for this study. Our accuracy on test data reaches 70.44%, while the state-of-art baseline is 62.1% reported by Mizukami and Tsuruoka (2015), and is significantly higher than previous trials using deep learning, which shows the promising potential of our new model. For the AI program building, besides the tile discarding strategy, we adopt similar predicting strategies for other actions such as stealing (pon, chi, and kan) and riichi. With the simple combination of these several predicting networks and without any knowledge about the concrete rules of the game, a strength evaluation is made for the resulting program on the largest Japanese Mahjong site `Tenhou'. The program has achieved a rating of around 1850, which is significantly higher than that of an average human player and of programs among past studies.

Takuya Hiraoka, Takahisa Imagawa, Tatsuya Mori et al.

Options are generally learned by using an inaccurate environment model (or simulator), which contains uncertain model parameters. While there are several methods to learn options that are robust against the uncertainty of model parameters, these methods only consider either the worst case or the average (ordinary) case for learning options. This limited consideration of the cases often produces options that do not work well in the unconsidered case. In this paper, we propose a conditional value at risk (CVaR)-based method to learn options that work well in both the average and worst cases. We extend the CVaR-based policy gradient method proposed by Chow and Ghavamzadeh (2014) to deal with robust Markov decision processes and then apply the extended method to learning robust options. We conduct experiments to evaluate our method in multi-joint robot control tasks (HopperIceBlock, Half-Cheetah, and Walker2D). Experimental results show that our method produces options that 1) give better worst-case performance than the options learned only to minimize the average-case loss, and 2) give better average-case performance than the options learned only to minimize the worst-case loss.

Shumpei Kubosawa, Takashi Onishi, Yoshimasa Tsuruoka

Chemical plants are complex and dynamical systems consisting of many components for manipulation and sensing, whose state transitions depend on various factors such as time, disturbance, and operation procedures. For the purpose of supporting human operators of chemical plants, we are developing an AI system that can semi-automatically synthesize operation procedures for efficient and stable operation. Our system can provide not only appropriate operation procedures but also reasons why the procedures are considered to be valid. This is achieved by integrating automated reasoning and deep reinforcement learning technologies with a chemical plant simulator and external knowledge. Our preliminary experimental results demonstrate that it can synthesize a procedure that achieves a much faster recovery from a malfunction compared to standard PID control.

Keigo Kawamura, Yoshimasa Tsuruoka

Despite the notable successes in video games such as Atari 2600, current AI is yet to defeat human champions in the domain of real-time strategy (RTS) games. One of the reasons is that an RTS game is a multi-agent game, in which single-agent reinforcement learning methods cannot simply be applied because the environment is not a stationary Markov Decision Process. In this paper, we present a first step toward finding a game-theoretic solution to RTS games by applying Neural Fictitious Self-Play (NFSP), a game-theoretic approach for finding Nash equilibria, to Mini-RTS, a small but nontrivial RTS game provided on the ELF platform. More specifically, we show that NFSP can be effectively combined with policy gradient reinforcement learning and be applied to Mini-RTS. Experimental results also show that the scalability of NFSP can be substantially improved by pretraining the models with simple self-play using policy gradients, which by itself gives a strong strategy despite its lack of theoretical guarantee of convergence.

Naoya Taguchi, Yoshimasa Tsuruoka

Memory-Augmented Neural Networks (MANNs) are a class of neural networks equipped with an external memory, and are reported to be effective for tasks requiring a large long-term memory and its selective use. The core module of a MANN is called a controller, which is usually implemented as a recurrent neural network (RNN) (e.g., LSTM) to enable the use of contextual information in controlling the other modules. However, such an RNN-based controller often allows a MANN to directly solve the given task by using the (small) internal memory of the controller, and prevents the MANN from making the best use of the external memory, thereby resulting in a suboptimally trained model. To address this problem, we present a novel type of RNN-based controller that is partially non-recurrent and avoids the direct use of its internal memory for solving the task, while keeping the ability of using contextual information in controlling the other modules. Our empirical experiments using Neural Turing Machines and Differentiable Neural Computers on the Toy and bAbI tasks demonstrate that the proposed controllers give substantially better results than standard RNN-based controllers.

Takuya Hiraoka, Takashi Onishi, Takahisa Imagawa et al.

Hierarchical planners that produce interpretable and appropriate plans are desired, especially in its application to supporting human decision making. In the typical development of the hierarchical planners, higher-level planners and symbol grounding functions are manually created, and this manual creation requires much human effort. In this paper, we propose a framework that can automatically refine symbol grounding functions and a high-level planner to reduce human effort for designing these modules. In our framework, symbol grounding and high-level planning, which are based on manually-designed knowledge bases, are modeled with semi-Markov decision processes. A policy gradient method is then applied to refine the modules, in which two terms for updating the modules are considered. The first term, called a reinforcement term, contributes to updating the modules to improve the overall performance of a hierarchical planner to produce appropriate plans. The second term, called a penalty term, contributes to keeping refined modules consistent with the manually-designed original modules. Namely, it keeps the planner, which uses the refined modules, producing interpretable plans. We perform preliminary experiments to solve the Mountain car problem, and its results show that a manually-designed high-level planner and symbol grounding function were successfully refined by our framework.

Akari Asai, Akiko Eriguchi, Kazuma Hashimoto et al.

Despite recent work in Reading Comprehension (RC), progress has been mostly limited to English due to the lack of large-scale datasets in other languages. In this work, we introduce the first RC system for languages without RC training data. Given a target language without RC training data and a pivot language with RC training data (e.g. English), our method leverages existing RC resources in the pivot language by combining a competitive RC model in the pivot language with an attentive Neural Machine Translation (NMT) model. We first translate the data from the target to the pivot language, and then obtain an answer using the RC model in the pivot language. Finally, we recover the corresponding answer in the original language using soft-alignment attention scores from the NMT model. We create evaluation sets of RC data in two non-English languages, namely Japanese and French, to evaluate our method. Experimental results on these datasets show that our method significantly outperforms a back-translation baseline of a state-of-the-art product-level machine translation system.

Seydou Ba, Takuya Hiraoka, Takashi Onishi et al.

Monte Carlo Tree Search (MCTS) is particularly adapted to domains where the potential actions can be represented as a tree of sequential decisions. For an effective action selection, MCTS performs many simulations to build a reliable tree representation of the decision space. As such, a bottleneck to MCTS appears when enough simulations cannot be performed between action selections. This is particularly highlighted in continuously running tasks, for which the time available to perform simulations between actions tends to be limited due to the environment's state constantly changing. In this paper, we present an approach that takes advantage of the anytime characteristic of MCTS to increase the simulation time when allowed. Our approach is to effectively balance the prospect of selecting an action with the time that can be spared to perform MCTS simulations before the next action selection. For that, we considered the simulation time as a decision variable to be selected alongside an action. We extended the Hierarchical Optimistic Optimization applied to Tree (HOOT) method to adapt our approach to environments with a continuous decision space. We evaluated our approach for environments with a continuous decision space through OpenAI gym's Pendulum and Continuous Mountain Car environments and for environments with discrete action space through the arcade learning environment (ALE) platform. The evaluation results show that, with variable simulation times, the proposed approach outperforms the conventional MCTS in the evaluated continuous decision space tasks and improves the performance of MCTS in most of the ALE tasks.

Kazuma Hashimoto, Yoshimasa Tsuruoka

A major obstacle in reinforcement learning-based sentence generation is the large action space whose size is equal to the vocabulary size of the target-side language. To improve the efficiency of reinforcement learning, we present a novel approach for reducing the action space based on dynamic vocabulary prediction. Our method first predicts a fixed-size small vocabulary for each input to generate its target sentence. The input-specific vocabularies are then used at supervised and reinforcement learning steps, and also at test time. In our experiments on six machine translation and two image captioning datasets, our method achieves faster reinforcement learning ($\sim$2.7x faster) with less GPU memory ($\sim$2.3x less) than the full-vocabulary counterpart. The reinforcement learning with our method consistently leads to significant improvement of BLEU scores, and the scores are equal to or better than those of baselines using the full vocabularies, with faster decoding time ($\sim$3x faster) on CPUs.

Kazeto Yamamoto, Takashi Onishi, Yoshimasa Tsuruoka

One of the key challenges in applying reinforcement learning to real-life problems is that the amount of train-and-error required to learn a good policy increases drastically as the task becomes complex. One potential solution to this problem is to combine reinforcement learning with automated symbol planning and utilize prior knowledge on the domain. However, existing methods have limitations in their applicability and expressiveness. In this paper we propose a hierarchical reinforcement learning method based on abductive symbolic planning. The planner can deal with user-defined evaluation functions and is not based on the Herbrand theorem. Therefore it can utilize prior knowledge of the rewards and can work in a domain where the state space is unknown. We demonstrate empirically that our architecture significantly improves learning efficiency with respect to the amount of training examples on the evaluation domain, in which the state space is unknown and there exist multiple goals.

Akiko Eriguchi, Yoshimasa Tsuruoka, Kyunghyun Cho

There has been relatively little attention to incorporating linguistic prior to neural machine translation. Much of the previous work was further constrained to considering linguistic prior on the source side. In this paper, we propose a hybrid model, called NMT+RNNG, that learns to parse and translate by combining the recurrent neural network grammar into the attention-based neural machine translation. Our approach encourages the neural machine translation model to incorporate linguistic prior during training, and lets it translate on its own afterward. Extensive experiments with four language pairs show the effectiveness of the proposed NMT+RNNG.

Kazuma Hashimoto, Yoshimasa Tsuruoka

This paper presents a novel neural machine translation model which jointly learns translation and source-side latent graph representations of sentences. Unlike existing pipelined approaches using syntactic parsers, our end-to-end model learns a latent graph parser as part of the encoder of an attention-based neural machine translation model, and thus the parser is optimized according to the translation objective. In experiments, we first show that our model compares favorably with state-of-the-art sequential and pipelined syntax-based NMT models. We also show that the performance of our model can be further improved by pre-training it with a small amount of treebank annotations. Our final ensemble model significantly outperforms the previous best models on the standard English-to-Japanese translation dataset.

Kazuma Hashimoto, Caiming Xiong, Yoshimasa Tsuruoka et al.

Transfer and multi-task learning have traditionally focused on either a single source-target pair or very few, similar tasks. Ideally, the linguistic levels of morphology, syntax and semantics would benefit each other by being trained in a single model. We introduce a joint many-task model together with a strategy for successively growing its depth to solve increasingly complex tasks. Higher layers include shortcut connections to lower-level task predictions to reflect linguistic hierarchies. We use a simple regularization term to allow for optimizing all model weights to improve one task's loss without exhibiting catastrophic interference of the other tasks. Our single end-to-end model obtains state-of-the-art or competitive results on five different tasks from tagging, parsing, relatedness, and entailment tasks.

Yusuke Watanabe, Kazuma Hashimoto, Yoshimasa Tsuruoka

We propose a simple domain adaptation method for neural networks in a supervised setting. Supervised domain adaptation is a way of improving the generalization performance on the target domain by using the source domain dataset, assuming that both of the datasets are labeled. Recently, recurrent neural networks have been shown to be successful on a variety of NLP tasks such as caption generation; however, the existing domain adaptation techniques are limited to (1) tune the model parameters by the target dataset after the training by the source dataset, or (2) design the network to have dual output, one for the source domain and the other for the target domain. Reformulating the idea of the domain adaptation technique proposed by Daume (2007), we propose a simple domain adaptation method, which can be applied to neural networks trained with a cross-entropy loss. On captioning datasets, we show performance improvements over other domain adaptation methods.

Yun-Ching Liu, Yoshimasa Tsuruoka

The combination of multi-armed bandit (MAB) algorithms with Monte-Carlo tree search (MCTS) has made a significant impact in various research fields. The UCT algorithm, which combines the UCB bandit algorithm with MCTS, is a good example of the success of this combination. The recent breakthrough made by AlphaGo, which incorporates convolutional neural networks with bandit algorithms in MCTS, also highlights the necessity of bandit algorithms in MCTS. However, despite the various investigations carried out on MCTS, nearly all of them still follow the paradigm of treating every node as an independent instance of the MAB problem, and applying the same bandit algorithm and heuristics on every node. As a result, this paradigm may leave some properties of the game tree unexploited. In this work, we propose that max nodes and min nodes have different concerns regarding their value estimation, and different bandit algorithms should be applied accordingly. We develop the Asymmetric-MCTS algorithm, which is an MCTS variant that applies a simple regret algorithm on max nodes, and the UCB algorithm on min nodes. We will demonstrate the performance of the Asymmetric-MCTS algorithm on the game of $9\times 9$ Go, $9\times 9$ NoGo, and Othello.

Akiko Eriguchi, Kazuma Hashimoto, Yoshimasa Tsuruoka

Most of the existing Neural Machine Translation (NMT) models focus on the conversion of sequential data and do not directly use syntactic information. We propose a novel end-to-end syntactic NMT model, extending a sequence-to-sequence model with the source-side phrase structure. Our model has an attention mechanism that enables the decoder to generate a translated word while softly aligning it with phrases as well as words of the source sentence. Experimental results on the WAT'15 English-to-Japanese dataset demonstrate that our proposed model considerably outperforms sequence-to-sequence attentional NMT models and compares favorably with the state-of-the-art tree-to-string SMT system.

Kazuma Hashimoto, Yoshimasa Tsuruoka

We present a novel method for jointly learning compositional and non-compositional phrase embeddings by adaptively weighting both types of embeddings using a compositionality scoring function. The scoring function is used to quantify the level of compositionality of each phrase, and the parameters of the function are jointly optimized with the objective for learning phrase embeddings. In experiments, we apply the adaptive joint learning method to the task of learning embeddings of transitive verb phrases, and show that the compositionality scores have strong correlation with human ratings for verb-object compositionality, substantially outperforming the previous state of the art. Moreover, our embeddings improve upon the previous best model on a transitive verb disambiguation task. We also show that a simple ensemble technique further improves the results for both tasks.

Yun-Ching Liu, Yoshimasa Tsuruoka

The UCT algorithm, which combines the UCB algorithm and Monte-Carlo Tree Search (MCTS), is currently the most widely used variant of MCTS. Recently, a number of investigations into applying other bandit algorithms to MCTS have produced interesting results. In this research, we will investigate the possibility of combining the improved UCB algorithm, proposed by Auer et al. (2010), with MCTS. However, various characteristics and properties of the improved UCB algorithm may not be ideal for a direct application to MCTS. Therefore, some modifications were made to the improved UCB algorithm, making it more suitable for the task of game tree search. The Mi-UCT algorithm is the application of the modified UCB algorithm applied to trees. The performance of Mi-UCT is demonstrated on the games of $9\times 9$ Go and $9\times 9$ NoGo, and has shown to outperform the plain UCT algorithm when only a small number of playouts are given, and rougly on the same level when more playouts are available.

Kazuma Hashimoto, Pontus Stenetorp, Makoto Miwa et al.

We present a novel learning method for word embeddings designed for relation classification. Our word embeddings are trained by predicting words between noun pairs using lexical relation-specific features on a large unlabeled corpus. This allows us to explicitly incorporate relation-specific information into the word embeddings. The learned word embeddings are then used to construct feature vectors for a relation classification model. On a well-established semantic relation classification task, our method significantly outperforms a baseline based on a previously introduced word embedding method, and compares favorably to previous state-of-the-art models that use syntactic information or manually constructed external resources.