Traditional product recommender systems often model user preferences based on historical information such as user inquiries and interactions with products. However, traditional systems cannot handle problems like unclear user demands and time-varying user preferences. In the conversational recommender system (CRS), the system obtains the explicit demands and dynamic preferences of users through real-time multi-turn interactions, so as to provide users with more accurate recommendation services. In recent years, many researches about CRS have been carried out. However, to apply CRS to real-world e-commerce platforms, there are still many challenges to overcome:

At first, the construction of CRS relies on the mining of product attributes and sales strategies of experts. However, existing automatic strategy mining algorithms usually produce results with coarse granularity and low accuracy. Such results cannot be directly used in practical applications and still requires complicated annotations. In addition, existing methods are not domain adaptive. When facing new dialogue scenarios or products from new categories, the mining system has to be reconstructed.

Secondly, the researches about conversation strategy module focus more on the interactive strategies aimed at product recommendation. These strategy modules cannot meet the requirements of real-world e-commerce platforms which have a package of marketing applications such as complementary-product recommendation, coupon recommendation, and selling point notification. A good strategy module should not only consider the mutual influence of different marketing applications and select the correct trigger time, but also guide the conversation topics and make users feel comfortable in the interactions. In addition to the conversion rate, the strategy module should also consider whether the customer's problems have been solved and his rating to the customer service.

To sum up, there are still many valuable problems to be overcome in the automatic conversational recommender system for e-commerce platforms. We hope to build an efficient CRS without laborious annotation and high operating costs, which can provide customers with more convenient shopping experience and bring higher revenue to the sellers.

传统商品推荐往往是基于用户的行为轨迹、询单query等历史信息来建模用户偏好，但存在用户诉求不明确、偏好建模缺乏时效性等问题。在对话推荐系统中，系统通过实时的多轮交互，获取用户的显式化诉求与动态偏好，从而为用户提供更精准的推荐服务。目前学术界在通用领域已经展开了广泛的研究，然而，在特定领域如电商背景下，实际应用过程中仍然存在着诸多的问题与挑战：

一方面，对话推荐系统的建设依赖于商品属性信息与专家导购策略的挖掘，而现有的自动挖掘算法，产出的结果通常粒度较粗、精准度不足，距离真实业务场景的要求仍然有一定差距。为了保证实际应用的效果，算法挖掘的商品属性信息与导购策略仍然需要人工深度参与标注，维护成本高昂。此外，现有方法缺乏领域迁移能力，在面对新商品类目或对话场景时需要重复建设，这也进一步影响了功能的推广。

另一方面，学术界关于交互式导购的决策模块更多聚焦于以商品推荐为目标的交互策略，无法满足新零售场景下复杂的营销导购需求。除了商品推荐，实际电商应用中还包含商品搭配、优惠活动、营销文案等丰富的营销功能。交互式推荐系统的决策模块需要考虑不同功能的触发时机与相互影响，并能在多轮对话中引导话题的走向；而决策的目标除了转化率之外，还包括用户对接待服务的满意度、用户问题的解决率等指标。

综上所述，在面向电商领域的全自动对话推荐系统方面，仍然存在许多兼具学术和业务价值的问题有待克服，具有重要的研究意义。我们希望在已有对话推荐系统的研究和业界应用基础上，实现轻运营投入、零人工配置、融合业务场景特色的交互导购对话系统，为用户提供更便捷的购物体验，为商家带来营收的增长。

1.Automatic mining of product attributes from the dialogues in e-commerce platforms. Attribute and attribute value mining is a kind of aspect mining. Existing researches focus on mining from sentences or paragraphs. Mining attributes and values from dialogues in the scenario of e-commerce customer service not only needs to consider the dialogue context, but also faces challenges such as informal and fragmented queries. And considering the huge amount of products in Taobao, it is hoped that the semi-supervised mining method can be used to reduce the cost.

2.Multi-task conversational strategy module for conversion rate and satisfaction rate. By combining the cutting-edge reinforcement learning and multi-task learning methods, we seek to build a conversational strategy module which can not only improve the conversion rate, but also make customers feel comfortable in the interactions.

3.Attractive response generation. Traditional response generation methods focus on generating readable and accurate responses. We hope to generate more attractive answers by combining user preferences, product properties and dialogue context, so as to improve the conversion rate of inquiries while solving the user's questions.

4.Dynamic update for user representation in Conversational Recommender System. During the interactions with customers, the customer's preferences to product attributes are updated in real time. We seek to build a method which can update the customer's implicit representation dynamically, so as to provide more accurate product recommendation.

1.电商场景下面向对话交互的属性挖掘。属性和属性值挖掘任务属于Aspect Mining的一种，学术界的研究数据通常是基于句子或篇章维度。在电商客服场景下挖掘对话交互中的商品属性，不仅需要考虑上下文语境，同时面临着成分缺省、表述随意等挑战。此外由于淘系商品数量巨大，希望能够利用半监督挖掘方法降低挖掘成本。

2.面向询单转化率和满意度的多目标对话策略模型。希望结合强化学习和多目标学习领域的前沿进展，实现多目标的对话策略模型，在提升询单转化率的同时提升智能客服的满意度。

3.面向导购目标的答案生成。传统的答案生成方法以文本内容的准确性为目标，我们希望结合用户偏好、商品特点和对话情景生成更具有吸引力的答案，在解决问题的同时提升询单转化率。

4.对话推荐系统中用户模型的动态更新。在与用户交互的过程中，智能客服不断更新着用户对于商品属性的偏好信息。我们希望能够基于这些动态偏好信息实时地调整用户模型，提供更精准的商品推荐。

• 商品属性抽取 (Product Attribute Value Extraction)
• 细粒度（属性）情感分析(Aspect-based Sentiment Analysis)
• 对话结构挖掘 (Dialogue Structure Mining)
• 探索与利用问题（Trade-offs between Exploration and Exploitation）
• 对话决策的迁移技术（Transfer Learning for Dialogue Policy）
• 交互话术生成（Multi-turn Script Generation）
• 多目标对话策略学习（Multi-task Conversational Strategies Learning）
• 对话推荐系统的评估与用户模拟（Evaluation and User simulation for CRS）