Reinforcement Learning from Human Feedback (RLHF) is a crucial approach to aligning language models with human values and intentions. A fundamental challenge in this method lies in ensuring that the reward model accurately understands and evaluates human preferences. Current methods rely on ranking losses to teach the reward model to assess preferences, but they are susceptible to noise and ambiguous data, often failing to deeply understand human intentions. To address this issue, we introduce contrastive learning into the reward modeling process. In addition to supervised ranking loss, we introduce an unsupervised contrastive loss to enable the reward model to fully capture the distinctions in contrastive data. Experimental results demonstrate that the proposed contrastive learning-based reward modeling method effectively enhances the generalization of the reward model, stabilizes the reinforcement learning training process, and improves the final alignment with human preferences.

Large language models (LLMs) have demonstrated proficiency across various natural language processing (NLP) tasks but often require additional training, such as continual pre-training and supervised fine-tuning. However, the costs associated with this, primarily due to their large parameter count, remain high. This paper proposes leveraging sparsity in pre-trained LLMs to expedite this training process. By observing sparsity in activated neurons during forward iterations, we identify the potential for computational speed-ups by excluding inactive neurons. We address associated challenges by extending existing neuron importance evaluation metrics and introducing a ladder omission rate scheduler. Our experiments on Llama-2 demonstrate that Sparsity-Accelerated Training (SAT) achieves comparable or superior performance to standard training while significantly accelerating the process. Specifically, SAT achieves a 45% throughput improvement in continual pre-training and saves 38% training time in supervised fine-tuning. It offers a simple, hardware-agnostic, and easily deployable framework for additional LLM training.

Retrieval-augmented generation (RAG) has emerged as a popular solution to mitigate the hallucination issues of large language models. However, existing studies on RAG seldom address the issue of predictive uncertainty, i.e., how likely it is that a RAG model’s prediction is incorrect, resulting in uncontrollable risks in real-world applications. In this work, we emphasize the importance of risk control, ensuring that RAG models proactively refuse to answer questions with low confidence. Our research identifies two critical latent factors affecting RAG’s confidence in its predictions: the quality of the retrieved results and the manner in which these results are utilized. To guide RAG models in assessing their own confidence based on these two latent factors, we develop a counterfactual prompting framework that induces the models to alter these factors and analyzes the effect on their answers. We also introduce a benchmarking procedure to collect answers with the option to abstain, facilitating a series of experiments. For evaluation, we introduce several risk-related metrics and the experimental results demonstrate the effectiveness of our approach. Our code and benchmark dataset are available at https://github.com/ict-bigdatalab/RC-RAG.

In Reinforcement Learning from Human Feedback (RLHF), the reward model plays a crucial role in aligning language model outputs with human values. The human preference data used to train the reward model consists of a prompt and a response pair, with humans annotating which response better aligns with human value preferences. Due to the complexity and subjectivity of the annotation task, multiple organizations including OpenAI and Anthropic report significant noise in the human preference datasets, leading to instability and deviation in reward model training from human values. We discover that the difference in scores assigned to response pairs by the reward model effectively indicates the quality of data, and data of varying qualities show significant distinctions in reward model training. We introduce a method that automatically adjusts reward modeling based on data quality, reducing the impact of noise and making full use of dataset. Experiments on multiple human preference datasets demonstrate that our method stabilizes reward model training and significantly enhances the alignment performance of RLHF.

Text summarization tasks commonly employ Pre-trained Language Models (PLMs) to fit diverse standard datasets. While these PLMs excel in automatic evaluations, they frequently underperform in human evaluations, indicating a deviation between their generated summaries and human summarization preferences. This discrepancy is likely due to the low quality of fine-tuning datasets and the limited availability of high-quality human-annotated data that reflect true human preference. To address this challenge, we introduce a novel human summarization preference alignment framework AlignSum. This framework consists of three parts: Firstly, we construct a Data Pymarid with extractive, abstractive, and human-annotated summary data. Secondly, we conduct the Gaussian Resampling to remove summaries with extreme lengths. Finally, we implement the two-stage hierarchical fine-tuning with Data Pymarid after Gaussian Resampling. We apply AlignSum to PLMs on the human-annotated CNN/DailyMail and BBC XSum datasets. Experiments show that with AlignSum, PLMs like BART-Large surpass 175B GPT-3 in both automatic and human evaluations. This demonstrates that AlignSum significantly enhances the alignment of language models with human summarization preferences.

Recently, Large Language Models (LLMs) have been demonstrated to possess impressive capabilities in a variety of domains and tasks. We investigate the issue of prompt design in the multi-turn text-to-SQL task and attempt to enhance the LLMs’ reasoning capacity when generating SQL queries. In the conversational context, the current SQL query can be modified from the preceding SQL query with only a few operations due to the context dependency. We introduce our method called CoE-SQL which can prompt LLMs to generate the SQL query based on the previously generated SQL query with an edition chain. We also conduct extensive ablation studies to determine the optimal configuration of our approach. Our approach outperforms different in-context learning baselines stably and achieves state-of-the-art performances on two benchmarks SParC and CoSQL using LLMs, which is also competitive to the SOTA fine-tuned models.

Large language models have demonstrated their capabilities in storyline creation and human-like character role-playing. Current language model agents mainly focus on reasonable behaviors from the level of individuals, and their behaviors might be hard to constraint on the level of the whole storyline. In this paper we introduce IBSEN, a director-actor coordinate agent framework that generates drama scripts and makes the plot played by agents more controllable. The director agent writes plot outlines that the user desires to see, instructs the actor agents to role-play their characters, and reschedules the plot when human players participate in the scenario to ensure the plot is progressing towards the objective. To evaluate the framework, we create a novel drama plot that involves several actor agents and check the interactions between them under the instruction of the director agent. Evaluation results show that our framework could generate complete, diverse drama scripts from only a rough outline of plot objectives, meanwhile maintaining the characteristics of characters in the drama. Our codes and prompts are available at https://github.com/OpenDFM/ibsen.

The use of large language models (LLM), especially ChatGPT, to help with research has come into practice. Researchers use it for timely advice and hope to obtain in-depth feedback. However, can LLM be a qualified and reliable reviewer? Although there already exist several review-related datasets, few works have carefully and thoroughly inspected model’s capability as a reviewer, especially the correctness of generated reviews. In this paper, we first evaluate GPT-3.5 and GPT-4 (the current top-performing LLM) on 2 types of tasks under different settings: the score prediction task and the review generation task. In addition, we propose a dataset containing 197 review-revision multiple-choice questions (RR-MCQ) with detailed labels from the review-rebuttal forum in ICLR-2023. By asking questions from technical details to the overall presentation and quality, our RR-MCQ data provides a more complete model ability assessment. The results show that LLM is generally helpful, but great caution is needed as it always makes mistakes. Although it can give passable decisions (> 60% accuracy) on single options, completely correct answers are still rare (about 20%); models are still weak on long paper processing, zero-shot scoring, and giving critical feedback like human reviewers.

Large Language Models (LLMs) have ushered in a new era in Natural Language Processing, but their massive size demands effective compression techniques for practicality. Although numerous model compression techniques have been investigated, they typically rely on a calibration set that overlooks the multilingual context and results in significant accuracy degradation for low-resource languages. This paper introduces Multilingual Brain Surgeon (MBS), a novel calibration data sampling method for multilingual LLMs compression. MBS overcomes the English-centric limitations of existing methods by sampling calibration data from various languages proportionally to the language distribution of the model training datasets. Our experiments, conducted on the BLOOM multilingual LLM, demonstrate that MBS improves the performance of existing English-centric compression methods, especially for low-resource languages. We also uncover the dynamics of language interaction during compression, revealing that the larger the proportion of a language in the training set and the more similar the language is to the calibration language, the better performance the language retains after compression. In conclusion, MBS presents an innovative approach to compressing multilingual LLMs, addressing the performance disparities and improving the language inclusivity of existing compression techniques. Keywords: Large Language Model, Multilingual Model Compression

Exploring the generalizability of a text-to-SQL parser is essential for a system to automatically adapt the real-world databases. Previous investigation works mostly focus on lexical diversity, including the influence of the synonym and perturbations in both natural language questions and databases. However, the structural variability of database schema (DS), as a widely seen real-world scenario, is yet underexplored. Specifically, confronted with the same input question, the target SQL may be represented in different ways when the DS comes to a different structure. In this work, we provide in-depth discussions about the schema generalizability challenge of text-to-SQL tasks. We observe that current datasets are too templated to study schema generalization. To collect suitable test data, we propose a framework to generate novel text-to-SQL data via automatic and synchronous (DS, SQL) pair altering. When evaluating state-of-the-art text-to-SQL models on the synthetic samples, performance is significantly degraded, which demonstrates the limitation of current research regarding schema generalization.

The cross-domain text-to-SQL task aims to build a system that can parse user questions into SQL on complete unseen databases, and the single-domain text-to-SQL task evaluates the performance on identical databases. Both of these setups confront unavoidable difficulties in real-world applications. To this end, we introduce the cross-schema text-to-SQL task, where the databases of evaluation data are different from that in the training data but come from the same domain. Furthermore, we present CSS, a large-scale CrosS-Schema Chinese text-to-SQL dataset, to carry on corresponding studies. CSS originally consisted of 4,340 question/SQL pairs across 2 databases. In order to generalize models to different medical systems, we extend CSS and create 19 new databases along with 29,280 corresponding dataset examples. Moreover, CSS is also a large corpus for single-domain Chinese text-to-SQL studies. We present the data collection approach and a series of analyses of the data statistics. To show the potential and usefulness of CSS, benchmarking baselines have been conducted and reported. Our dataset is publicly available at https://huggingface.co/datasets/zhanghanchong/css.

Recently Large Language Models (LLMs) have been proven to have strong abilities in various domains and tasks. We study the problem of prompt designing in the text-to-SQL task and attempt to improve the LLMs’ reasoning ability when generating SQL queries. Besides the trivial few-shot in-context learning setting, we design our chain-of-thought (CoT) prompt with a similar method to schema linking. We provide a method named ACT-SQL to automatically generate auto-CoT exemplars and thus the whole process doesn’t need manual labeling. Our approach is cost-saving since we only use the LLMs’ API call once when generating one SQL query. Furthermore, we extend our in-context learning method to the multi-turn text-to-SQL task. The experiment results show that the LLMs’ performance can benefit from our ACT-SQL approach. Our approach achieves SOTA performance on the Spider dev set among existing in-context learning approaches.

This paper presents an ontology-aware pretrained language model (OPAL) for end-to-end task-oriented dialogue (TOD). Unlike chit-chat dialogue models, task-oriented dialogue models fulfill at least two task-specific modules: Dialogue state tracker (DST) and response generator (RG). The dialogue state consists of the domain-slot-value triples, which are regarded as the user’s constraints to search the domain-related databases. The large-scale task-oriented dialogue data with the annotated structured dialogue state usually are inaccessible. It prevents the development of the pretrained language model for the task-oriented dialogue. We propose a simple yet effective pretraining method to alleviate this problem, which consists of two pretraining phases. The first phase is to pretrain on large-scale contextual text data, where the structured information of the text is extracted by the information extracting tool. To bridge the gap between the pretraining method and downstream tasks, we design two pretraining tasks: ontology-like triple recovery and next-text generation, which simulates the DST and RG, respectively. The second phase is to fine-tune the pretrained model on the TOD data. The experimental results show that our proposed method achieves an exciting boost and obtains competitive performance even without any TOD data on CamRest676 and MultiWOZ benchmarks.

Recently, the structural reading comprehension (SRC) task on web pages has attracted increasing research interests. Although previous SRC work has leveraged extra information such as HTML tags or XPaths, the informative topology of web pages is not effectively exploited. In this work, we propose a Topological Information Enhanced model (TIE), which transforms the token-level task into a tag-level task by introducing a two-stage process (i.e. node locating and answer refining). Based on that, TIE integrates Graph Attention Network (GAT) and Pre-trained Language Model (PLM) to leverage the topological information of both logical structures and spatial structures. Experimental results demonstrate that our model outperforms strong baselines and achieves state-of-the-art performances on the web-based SRC benchmark WebSRC at the time of writing. The code of TIE will be publicly available at https://github.com/X-LANCE/TIE.

With the development of pre-trained language models, remarkable success has been witnessed in dialogue understanding (DU). However, current DU approaches usually employ independent models for each distinct DU task, without considering shared knowledge across different DU tasks. In this paper, we propose a unified generative dialogue understanding framework, named UniDU, to achieve effective information exchange across diverse DU tasks. Here, we reformulate all DU tasks into a unified prompt-based generative model paradigm. More importantly, a novel model-agnostic multi-task training strategy (MATS) is introduced to dynamically adapt the weights of diverse tasks for best knowlege sharing during training, based on the nature and available data of each task. Experiments on ten DU datasets covering five fundamental DU tasks show that the proposed UniDU framework largely outperforms task-specific well-designed methods on all tasks. MATS also reveals the knowledge sharing structure of these tasks. Finally, UniDU obtains promising performance on unseen dialogue domain, showing great potential of generalization.

In a depression-diagnosis-directed clinical session, doctors initiate a conversation with ample emotional support that guides the patients to expose their symptoms based on clinical diagnosis criteria. Such a dialogue system is distinguished from existing single-purpose human-machine dialog systems, as it combines task-oriented and chit-chats with uniqueness in dialogue topics and procedures. However, due to the social stigma associated with mental illness, the dialogue data related to depression consultation and diagnosis are rarely disclosed. Based on clinical depression diagnostic criteria ICD-11 and DSM-5, we designed a 3-phase procedure to construct D⁴: a Chinese Dialogue Dataset for Depression-Diagnosis-Oriented Chat, which simulates the dialogue between doctors and patients during the diagnosis of depression, including diagnosis results and symptom summary given by professional psychiatrists for each conversation. Upon the newly-constructed dataset, four tasks mirroring the depression diagnosis process are established: response generation, topic prediction, dialog summary, and severity classification of depressive episode and suicide risk. Multi-scale evaluation results demonstrate that a more empathy-driven and diagnostic-accurate consultation dialogue system trained on our dataset can be achieved compared to rule-based bots.

Task-oriented dialogue (TOD) systems have been widely used by mobile phone intelligent assistants to accomplish tasks such as calendar scheduling or hotel reservation. Current TOD systems usually focus on multi-turn text/speech interaction, then they would call back-end APIs designed for TODs to perform the task. However, this API-based architecture greatly limits the information-searching capability of intelligent assistants and may even lead to task failure if TOD-specific APIs are not available or the task is too complicated to be executed by the provided APIs. In this paper, we propose a new TOD architecture: GUI-based task-oriented dialogue system (GUI-TOD). A GUI-TOD system can directly perform GUI operations on real APPs and execute tasks without invoking TOD-specific backend APIs. Furthermore, we release META-GUI, a dataset for training a Multi-modal convErsaTional Agent on mobile GUI. We also propose a multi-model action prediction and response model, which show promising results on META-GUI. The dataset, codes and leaderboard are publicly available.

Thanks to the development of pre-trained language models, multitask learning (MTL) methods achieve a great success in natural language understanding area.However, current MTL methods pay more attention to task selection or model design to fuse as much knowledge as possible, while intrinsic task correlation is often neglected. It is important to learn sharing strategy among multiple tasks rather than sharing everything.%The MTL model is directly shared among all the tasks. %For example, in traditional MTL methods, the last classification layers or the decoder layers are manually separated. More deeply, In this paper, we propose AdapterShare, an adapter differentiation method to explicitly model the task correlation among multiple tasks. AdapterShare is automatically learned based on the gradients on tiny held-out validation data. Compared to single-task learning and fully shared MTL methods, our proposed method obtains obvious performance improvement. Compared to the existing MTL method AdapterFusion, AdapterShare achieves absolute 1.90 average points improvement on five dialogue understanding tasks and 2.33 points gain on NLU tasks.

This work aims to tackle the challenging heterogeneous graph encoding problem in the text-to-SQL task. Previous methods are typically node-centric and merely utilize different weight matrices to parameterize edge types, which 1) ignore the rich semantics embedded in the topological structure of edges, and 2) fail to distinguish local and non-local relations for each node. To this end, we propose a Line Graph Enhanced Text-to-SQL (LGESQL) model to mine the underlying relational features without constructing meta-paths. By virtue of the line graph, messages propagate more efficiently through not only connections between nodes, but also the topology of directed edges. Furthermore, both local and non-local relations are integrated distinctively during the graph iteration. We also design an auxiliary task called graph pruning to improve the discriminative capability of the encoder. Our framework achieves state-of-the-art results (62.8% with Glove, 72.0% with Electra) on the cross-domain text-to-SQL benchmark Spider at the time of writing.

Given a database schema, Text-to-SQL aims to translate a natural language question into the corresponding SQL query. Under the setup of cross-domain, traditional semantic parsing models struggle to adapt to unseen database schemas. To improve the model generalization capability for rare and unseen schemas, we propose a new architecture, ShadowGNN, which processes schemas at abstract and semantic levels. By ignoring names of semantic items in databases, abstract schemas are exploited in a well-designed graph projection neural network to obtain delexicalized representation of question and schema. Based on the domain-independent representations, a relation-aware transformer is utilized to further extract logical linking between question and schema. Finally, a SQL decoder with context-free grammar is applied. On the challenging Text-to-SQL benchmark Spider, empirical results show that ShadowGNN outperforms state-of-the-art models. When the annotated data is extremely limited (only 10% training set), ShadowGNN gets over absolute 5% performance gain, which shows its powerful generalization ability. Our implementation will be open-sourced at https://github.com/WowCZ/shadowgnn

Sememes are defined as the atomic units to describe the semantic meaning of concepts. Due to the difficulty of manually annotating sememes and the inconsistency of annotations between experts, the lexical sememe prediction task has been proposed. However, previous methods heavily rely on word or character embeddings, and ignore the fine-grained information. In this paper, we propose a novel pre-training method which is designed to better incorporate the internal information of Chinese character. The Glyph enhanced Chinese Character representation (GCC) is used to assist sememe prediction. We experiment and evaluate our model on HowNet, which is a famous sememe knowledge base. The experimental results show that our method outperforms existing non-external information models.

Web search is an essential way for humans to obtain information, but it’s still a great challenge for machines to understand the contents of web pages. In this paper, we introduce the task of web-based structural reading comprehension. Given a web page and a question about it, the task is to find an answer from the web page. This task requires a system not only to understand the semantics of texts but also the structure of the web page. Moreover, we proposed WebSRC, a novel Web-based Structural Reading Comprehension dataset. WebSRC consists of 400K question-answer pairs, which are collected from 6.4K web pages with corresponding HTML source code, screenshots, and metadata. Each question in WebSRC requires a certain structural understanding of a web page to answer, and the answer is either a text span on the web page or yes/no. We evaluate various strong baselines on our dataset to show the difficulty of our task. We also investigate the usefulness of structural information and visual features. Our dataset and baselines have been publicly available.

Efficient structure encoding for graphs with labeled edges is an important yet challenging point in many graph-based models. This work focuses on AMR-to-text generation – A graph-to-sequence task aiming to recover natural language from Abstract Meaning Representations (AMR). Existing graph-to-sequence approaches generally utilize graph neural networks as their encoders, which have two limitations: 1) The message propagation process in AMR graphs is only guided by the first-order adjacency information. 2) The relationships between labeled edges are not fully considered. In this work, we propose a novel graph encoding framework which can effectively explore the edge relations. We also adopt graph attention networks with higher-order neighborhood information to encode the rich structure in AMR graphs. Experiment results show that our approach obtains new state-of-the-art performance on English AMR benchmark datasets. The ablation analyses also demonstrate that both edge relations and higher-order information are beneficial to graph-to-sequence modeling.

Chinese short text matching usually employs word sequences rather than character sequences to get better performance. However, Chinese word segmentation can be erroneous, ambiguous or inconsistent, which consequently hurts the final matching performance. To address this problem, we propose neural graph matching networks, a novel sentence matching framework capable of dealing with multi-granular input information. Instead of a character sequence or a single word sequence, paired word lattices formed from multiple word segmentation hypotheses are used as input and the model learns a graph representation according to an attentive graph matching mechanism. Experiments on two Chinese datasets show that our models outperform the state-of-the-art short text matching models.

One daunting problem for semantic parsing is the scarcity of annotation. Aiming to reduce nontrivial human labor, we propose a two-stage semantic parsing framework, where the first stage utilizes an unsupervised paraphrase model to convert an unlabeled natural language utterance into the canonical utterance. The downstream naive semantic parser accepts the intermediate output and returns the target logical form. Furthermore, the entire training process is split into two phases: pre-training and cycle learning. Three tailored self-supervised tasks are introduced throughout training to activate the unsupervised paraphrase model. Experimental results on benchmarks Overnight and GeoGranno demonstrate that our framework is effective and compatible with supervised training.

Dialogue state tracking (DST) aims at estimating the current dialogue state given all the preceding conversation. For multi-domain DST, the data sparsity problem is a major obstacle due to increased numbers of state candidates and dialogue lengths. To encode the dialogue context efficiently, we utilize the previous dialogue state (predicted) and the current dialogue utterance as the input for DST. To consider relations among different domain-slots, the schema graph involving prior knowledge is exploited. In this paper, a novel context and schema fusion network is proposed to encode the dialogue context and schema graph by using internal and external attention mechanisms. Experiment results show that our approach can outperform strong baselines, and the previous state-of-the-art method (SOM-DST) can also be improved by our proposed schema graph.

Pattern-based labeling methods have achieved promising results in alleviating the inevitable labeling noises of distantly supervised neural relation extraction. However, these methods require significant expert labor to write relation-specific patterns, which makes them too sophisticated to generalize quickly. To ease the labor-intensive workload of pattern writing and enable the quick generalization to new relation types, we propose a neural pattern diagnosis framework, DIAG-NRE, that can automatically summarize and refine high-quality relational patterns from noise data with human experts in the loop. To demonstrate the effectiveness of DIAG-NRE, we apply it to two real-world datasets and present both significant and interpretable improvements over state-of-the-art methods.

Recently, deep reinforcement learning (DRL) has been used for dialogue policy optimization. However, many DRL-based policies are not sample-efficient. Most recent advances focus on improving DRL optimization algorithms to address this issue. Here, we take an alternative route of designing neural network structure that is better suited for DRL-based dialogue management. The proposed structured deep reinforcement learning is based on graph neural networks (GNN), which consists of some sub-networks, each one for a node on a directed graph. The graph is defined according to the domain ontology and each node can be considered as a sub-agent. During decision making, these sub-agents have internal message exchange between neighbors on the graph. We also propose an approach to jointly optimize the graph structure as well as the parameters of GNN. Experiments show that structured DRL significantly outperforms previous state-of-the-art approaches in almost all of the 18 tasks of the PyDial benchmark.

Dialogue state tracking (DST), when formulated as a supervised learning problem, relies on labelled data. Since dialogue state annotation usually requires labelling all turns of a single dialogue and utilizing context information, it is very expensive to annotate all available unlabelled data. In this paper, a novel cost-sensitive active learning framework is proposed based on a set of new dialogue-level query strategies. This is the first attempt to apply active learning for dialogue state tracking. Experiments on DSTC2 show that active learning with mixed data query strategies can effectively achieve the same DST performance with significantly less data annotation compared to traditional training approaches.

Dialogue state tracker is the core part of a spoken dialogue system. It estimates the beliefs of possible user’s goals at every dialogue turn. However, for most current approaches, it’s difficult to scale to large dialogue domains. They have one or more of following limitations: (a) Some models don’t work in the situation where slot values in ontology changes dynamically; (b) The number of model parameters is proportional to the number of slots; (c) Some models extract features based on hand-crafted lexicons. To tackle these challenges, we propose StateNet, a universal dialogue state tracker. It is independent of the number of values, shares parameters across all slots, and uses pre-trained word vectors instead of explicit semantic dictionaries. Our experiments on two datasets show that our approach not only overcomes the limitations, but also significantly outperforms the performance of state-of-the-art approaches.

The key to building an evolvable dialogue system in real-world scenarios is to ensure an affordable on-line dialogue policy learning, which requires the on-line learning process to be safe, efficient and economical. But in reality, due to the scarcity of real interaction data, the dialogue system usually grows slowly. Besides, the poor initial dialogue policy easily leads to bad user experience and incurs a failure of attracting users to contribute training data, so that the learning process is unsustainable. To accurately depict this, two quantitative metrics are proposed to assess safety and efficiency issues. For solving the unsustainable learning problem, we proposed a complete companion teaching framework incorporating the guidance from the human teacher. Since the human teaching is expensive, we compared various teaching schemes answering the question how and when to teach, to economically utilize teaching budget, so that make the online learning process affordable.

Hand-crafted rules and reinforcement learning (RL) are two popular choices to obtain dialogue policy. The rule-based policy is often reliable within predefined scope but not self-adaptable, whereas RL is evolvable with data but often suffers from a bad initial performance. We employ a companion learning framework to integrate the two approaches for on-line dialogue policy learning, in which a pre-defined rule-based policy acts as a “teacher” and guides a data-driven RL system by giving example actions as well as additional rewards. A novel agent-aware dropout Deep Q-Network (AAD-DQN) is proposed to address the problem of when to consult the teacher and how to learn from the teacher’s experiences. AAD-DQN, as a data-driven student policy, provides (1) two separate experience memories for student and teacher, (2) an uncertainty estimated by dropout to control the timing of consultation and learning. Simulation experiments showed that the proposed approach can significantly improve both safetyand efficiency of on-line policy optimization compared to other companion learning approaches as well as supervised pre-training using static dialogue corpus.

On-line dialogue policy learning is the key for building evolvable conversational agent in real world scenarios. Poor initial policy can easily lead to bad user experience and consequently fail to attract sufficient users for policy training. A novel framework, companion teaching, is proposed to include a human teacher in the dialogue policy training loop to address the cold start problem. Here, dialogue policy is trained using not only user’s reward, but also teacher’s example action as well as estimated immediate reward at turn level. Simulation experiments showed that, with small number of human teaching dialogues, the proposed approach can effectively improve user experience at the beginning and smoothly lead to good performance with more user interaction data.