Yang Zhao

Yang Zhao, Li Du, Xiao Ding et al.

LLMs' performance on complex tasks is still unsatisfactory. A key issue is that presently LLMs learn in a data-driven schema, while the instructions about these complex tasks are both scarce and hard to collect or construct. On the contrary, a prominent phenomenon is that LLMs can learn rather fast on simpler tasks with adequate prior knowledge captured during pretraining stage. Thus, if the prerequisite and mechanism of such rapid generalization could be elucidated, it could enhance the efficiency and effectiveness of the LLM's ability to learn complex tasks. Thus, in this paper, we employ a gradient-based method, to dissect the process that the SFT process adapts LLMs to downstream tasks via the perspective of attention patterns. We find that: (1) LLMs selectively activate task-specific attention heads during SFT; (2) activation patterns for complex tasks are combinations of basic task patterns; and (3) changes in a few parameters can significantly impact activation patterns after SFT on a small number of samples.Based on these insights, experiments are conducted to actually enhance the efficiency and effectiveness of SFT.

Donglei Yu, Yang Zhao, Jie Zhu et al.

Simultaneous Machine Translation (SiMT) generates translations while receiving streaming source inputs. This requires the SiMT model to learn a read/write policy, deciding when to translate and when to wait for more source input. Numerous linguistic studies indicate that audiences in SiMT scenarios have distinct preferences, such as accurate translations, simpler syntax, and no unnecessary latency. Aligning SiMT models with these human preferences is crucial to improve their performances. However, this issue still remains unexplored. Additionally, preference optimization for SiMT task is also challenging. Existing methods focus solely on optimizing the generated responses, ignoring human preferences related to latency and the optimization of read/write policy during the preference optimization phase. To address these challenges, we propose Simultaneous Preference Learning (SimulPL), a preference learning framework tailored for the SiMT task. In the SimulPL framework, we categorize SiMT human preferences into five aspects: \textbf{translation quality preference}, \textbf{monotonicity preference}, \textbf{key point preference}, \textbf{simplicity preference}, and \textbf{latency preference}. By leveraging the first four preferences, we construct human preference prompts to efficiently guide GPT-4/4o in generating preference data for the SiMT task. In the preference optimization phase, SimulPL integrates \textbf{latency preference} into the optimization objective and enables SiMT models to improve the read/write policy, thereby aligning with human preferences more effectively. Experimental results indicate that SimulPL exhibits better alignment with human preferences across all latency levels in Zh$\rightarrow$En, De$\rightarrow$En and En$\rightarrow$Zh SiMT tasks. Our data and code will be available at https://github.com/EurekaForNLP/SimulPL.

Yang Zhao, Li Du, Xiao Ding et al.

Through pretraining on a corpus with various sources, Large Language Models (LLMs) have gained impressive performance. However, the impact of each component of the pretraining corpus remains opaque. As a result, the organization of the pretraining corpus is still empirical and may deviate from the optimal. To address this issue, we systematically analyze the impact of 48 datasets from 5 major categories of pretraining data of LLMs and measure their impacts on LLMs using benchmarks about nine major categories of model capabilities. Our analyses provide empirical results about the contribution of multiple corpora on the performances of LLMs, along with their joint impact patterns, including complementary, orthogonal, and correlational relationships. We also identify a set of ``high-impact data'' such as Books that is significantly related to a set of model capabilities. These findings provide insights into the organization of data to support more efficient pretraining of LLMs.

Zhao Yang, Junhong Lian, Xiang Ao

Personalized news headline generation, aiming at generating user-specific headlines based on readers' preferences, burgeons a recent flourishing research direction. Existing studies generally inject a user interest embedding into an encoderdecoder headline generator to make the output personalized, while the factual consistency of headlines is inadequate to be verified. In this paper, we propose a framework Fact-Preserved Personalized News Headline Generation (short for FPG), to prompt a tradeoff between personalization and consistency. In FPG, the similarity between the candidate news to be exposed and the historical clicked news is used to give different levels of attention to key facts in the candidate news, and the similarity scores help to learn a fact-aware global user embedding. Besides, an additional training procedure based on contrastive learning is devised to further enhance the factual consistency of generated headlines. Extensive experiments conducted on a real-world benchmark PENS validate the superiority of FPG, especially on the tradeoff between personalization and factual consistency.

Deyuan Liu, Zhanyue Qin, Hairu Wang et al.

While large language models (LLMs) excel in many domains, their complexity and scale challenge deployment in resource-limited environments. Current compression techniques, such as parameter pruning, often fail to effectively utilize the knowledge from pruned parameters. To address these challenges, we propose Manifold-Based Knowledge Alignment and Layer Merging Compression (MKA), a novel approach that uses manifold learning and the Normalized Pairwise Information Bottleneck (NPIB) measure to merge similar layers, reducing model size while preserving essential performance. We evaluate MKA on multiple benchmark datasets and various LLMs. Our findings show that MKA not only preserves model performance but also achieves substantial compression ratios, outperforming traditional pruning methods. Moreover, when coupled with quantization, MKA delivers even greater compression. Specifically, on the MMLU dataset using the Llama3-8B model, MKA achieves a compression ratio of 43.75% with a minimal performance decrease of only 2.82\%. The proposed MKA method offers a resource-efficient and performance-preserving model compression technique for LLMs.

Vitou Phy, Yang Zhao, Akiko Aizawa

Many automatic evaluation metrics have been proposed to score the overall quality of a response in open-domain dialogue. Generally, the overall quality is comprised of various aspects, such as relevancy, specificity, and empathy, and the importance of each aspect differs according to the task. For instance, specificity is mandatory in a food-ordering dialogue task, whereas fluency is preferred in a language-teaching dialogue system. However, existing metrics are not designed to cope with such flexibility. For example, BLEU score fundamentally relies only on word overlapping, whereas BERTScore relies on semantic similarity between reference and candidate response. Thus, they are not guaranteed to capture the required aspects, i.e., specificity. To design a metric that is flexible to a task, we first propose making these qualities manageable by grouping them into three groups: understandability, sensibleness, and likability, where likability is a combination of qualities that are essential for a task. We also propose a simple method to composite metrics of each aspect to obtain a single metric called USL-H, which stands for Understandability, Sensibleness, and Likability in Hierarchy. We demonstrated that USL-H score achieves good correlations with human judgment and maintains its configurability towards different aspects and metrics.

Ryosuke Kohita, Akifumi Wachi, Yang Zhao et al.

Unsupervised methods are promising for abstractive text summarization in that the parallel corpora is not required. However, their performance is still far from being satisfied, therefore research on promising solutions is on-going. In this paper, we propose a new approach based on Q-learning with an edit-based summarization. The method combines two key modules to form an Editorial Agent and Language Model converter (EALM). The agent predicts edit actions (e.t., delete, keep, and replace), and then the LM converter deterministically generates a summary on the basis of the action signals. Q-learning is leveraged to train the agent to produce proper edit actions. Experimental results show that EALM delivered competitive performance compared with the previous encoder-decoder-based methods, even with truly zero paired data (i.e., no validation set). Defining the task as Q-learning enables us not only to develop a competitive method but also to make the latest techniques in reinforcement learning available for unsupervised summarization. We also conduct qualitative analysis, providing insights into future study on unsupervised summarizers.

Xiaomian Kang, Yang Zhao, Jiajun Zhang et al.

Document-level neural machine translation has yielded attractive improvements. However, majority of existing methods roughly use all context sentences in a fixed scope. They neglect the fact that different source sentences need different sizes of context. To address this problem, we propose an effective approach to select dynamic context so that the document-level translation model can utilize the more useful selected context sentences to produce better translations. Specifically, we introduce a selection module that is independent of the translation module to score each candidate context sentence. Then, we propose two strategies to explicitly select a variable number of context sentences and feed them into the translation module. We train the two modules end-to-end via reinforcement learning. A novel reward is proposed to encourage the selection and utilization of dynamic context sentences. Experiments demonstrate that our approach can select adaptive context sentences for different source sentences, and significantly improves the performance of document-level translation methods.

Yang Zhao, Ping Yu, Suchismit Mahapatra et al.

Variational autoencoders (VAEs) are essential tools in end-to-end representation learning. However, the sequential text generation common pitfall with VAEs is that the model tends to ignore latent variables with a strong auto-regressive decoder. In this paper, we propose a principled approach to alleviate this issue by applying a discretized bottleneck to enforce an implicit latent feature matching in a more compact latent space. We impose a shared discrete latent space where each input is learned to choose a combination of latent atoms as a regularized latent representation. Our model endows a promising capability to model underlying semantics of discrete sequences and thus provide more interpretative latent structures. Empirically, we demonstrate our model's efficiency and effectiveness on a broad range of tasks, including language modeling, unaligned text style transfer, dialog response generation, and neural machine translation.

Jiajun Zhang, Long Zhou, Yang Zhao et al.

In sequence to sequence generation tasks (e.g. machine translation and abstractive summarization), inference is generally performed in a left-to-right manner to produce the result token by token. The neural approaches, such as LSTM and self-attention networks, are now able to make full use of all the predicted history hypotheses from left side during inference, but cannot meanwhile access any future (right side) information and usually generate unbalanced outputs in which left parts are much more accurate than right ones. In this work, we propose a synchronous bidirectional inference model to generate outputs using both left-to-right and right-to-left decoding simultaneously and interactively. First, we introduce a novel beam search algorithm that facilitates synchronous bidirectional decoding. Then, we present the core approach which enables left-to-right and right-to-left decoding to interact with each other, so as to utilize both the history and future predictions simultaneously during inference. We apply the proposed model to both LSTM and self-attention networks. In addition, we propose two strategies for parameter optimization. The extensive experiments on machine translation and abstractive summarization demonstrate that our synchronous bidirectional inference model can achieve remarkable improvements over the strong baselines.

Yang Zhao, Yining Wang, Jiajun Zhang et al.

Neural Machine Translation (NMT) has drawn much attention due to its promising translation performance recently. However, several studies indicate that NMT often generates fluent but unfaithful translations. In this paper, we propose a method to alleviate this problem by using a phrase table as recommendation memory. The main idea is to add bonus to words worthy of recommendation, so that NMT can make correct predictions. Specifically, we first derive a prefix tree to accommodate all the candidate target phrases by searching the phrase translation table according to the source sentence. Then, we construct a recommendation word set by matching between candidate target phrases and previously translated target words by NMT. After that, we determine the specific bonus value for each recommendable word by using the attention vector and phrase translation probability. Finally, we integrate this bonus value into NMT to improve the translation results. The extensive experiments demonstrate that the proposed methods obtain remarkable improvements over the strong attentionbased NMT.

Zhong Meng, Jinyu Li, Zhuo Chen et al.

We propose a novel adversarial multi-task learning scheme, aiming at actively curtailing the inter-talker feature variability while maximizing its senone discriminability so as to enhance the performance of a deep neural network (DNN) based ASR system. We call the scheme speaker-invariant training (SIT). In SIT, a DNN acoustic model and a speaker classifier network are jointly optimized to minimize the senone (tied triphone state) classification loss, and simultaneously mini-maximize the speaker classification loss. A speaker-invariant and senone-discriminative deep feature is learned through this adversarial multi-task learning. With SIT, a canonical DNN acoustic model with significantly reduced variance in its output probabilities is learned with no explicit speaker-independent (SI) transformations or speaker-specific representations used in training or testing. Evaluated on the CHiME-3 dataset, the SIT achieves 4.99% relative word error rate (WER) improvement over the conventional SI acoustic model. With additional unsupervised speaker adaptation, the speaker-adapted (SA) SIT model achieves 4.86% relative WER gain over the SA SI acoustic model.

Yining Wang, Yang Zhao, Jiajun Zhang et al.

While neural machine translation (NMT) has become the new paradigm, the parameter optimization requires large-scale parallel data which is scarce in many domains and language pairs. In this paper, we address a new translation scenario in which there only exists monolingual corpora and phrase pairs. We propose a new method towards translation with partially aligned sentence pairs which are derived from the phrase pairs and monolingual corpora. To make full use of the partially aligned corpora, we adapt the conventional NMT training method in two aspects. On one hand, different generation strategies are designed for aligned and unaligned target words. On the other hand, a different objective function is designed to model the partially aligned parts. The experiments demonstrate that our method can achieve a relatively good result in such a translation scenario, and tiny bitexts can boost translation quality to a large extent.

Xiaoyu Shen, Hui Su, Yanran Li et al.

Deep latent variable models have been shown to facilitate the response generation for open-domain dialog systems. However, these latent variables are highly randomized, leading to uncontrollable generated responses. In this paper, we propose a framework allowing conditional response generation based on specific attributes. These attributes can be either manually assigned or automatically detected. Moreover, the dialog states for both speakers are modeled separately in order to reflect personal features. We validate this framework on two different scenarios, where the attribute refers to genericness and sentiment states respectively. The experiment result testified the potential of our model, where meaningful responses can be generated in accordance with the specified attributes.