An Evaluation of Chinese Human-Computer Dialogue Technology

With the development of artificial intelligence, human-computer dialogue technology has become increasingly popular and has attracted growing attention [1]. Human-computer dialogue systems are conversation agents, which are normally divided into two classes [2, 3]: task-oriented dialogue systems [4, 5, 6] and none-task-oriented systems [7, 8]. In this paper, we mainly focus on task-oriented dialogue systems.

There are two important tasks in a task-oriented dialogue system. One is concerned with classification of a user's intents, which is a text categorization task. Its purpose is to recognize the user's chat intentions, such as task-based interaction, a knowledge quiz or chit-chat. It is the foundation for building a large and complex human-machine dialogue system [9] and it is a clear but difficult task because of a limited number of corpora available for training the algorithms and difficulties in understanding semantic meanings. Recently, there have been some evaluations with user intent classification tasks. For example, Task 1 in the 17th China National Conference on Computational Linguistics (CCL2018)^①, which is based on Chinese corpora, is a user intent classification task in the customer service field. They provide some open data to allow participants to build systems and then test them on hidden data sets. However, the range of data sets they provide is limited to Q&A data from China Mobile Communications Group Co., Ltd.^②, including the query categories, data processing categories and business consulting categories.

The other is to accomplish tasks in a specific domain in a human-computer dialogue. A complete human-computer dialogue system should be capable of understanding the tasks that users want to accomplish and assist them in completing a specific domain task, such as inquiring for train information or booking a ticket. This is a fairly complex task, which can fully reflect the intelligence of a human-machine dialogue system. Another challenge is how to evaluate and compare these systems, and what influencing factors we need to pay attention to. A similar evaluation based on English corpora is the 6th Dialog System Technology Challenges (DSTC6) held in 2017 [10]. In DSTC6, participants need to build a system that responds to a user's utterances based on the context of the conversation, where they can use external data. Both objective and subjective indicators are used to evaluate the submitted systems [11]. However, the focus of the task for participants in DSTC6 is on text generation instead of the complete process of accomplishing the given task. As far as we know, the last manual evaluation of the end-to-end task-based dialogue system was the Spoken Dialog Challenge 2010 [12], which was held eight years ago.

In short, in order to promote the development of the evaluation technology for human-computer dialogue systems, and to attract more people to pay attention to the above two key issues in human-computer dialogue systems, the Second Evaluation of Chinese Human-Computer Dialogue Technology was held during the 7th China National Conference on Social Media Processing^③ (SMP2018-ECDT), which consists of two tasks:

1. User intent classification. There are 31 categories in total, which include one chit-chat category and 30 vertical categories of 30 specific tasks such as accessing apps and inquiring about the weather. The submitted systems need to determine which category the user's input belongs to among all of the 31 categories.

2. Online testing of task-oriented dialogues. The submitted systems should complete the corresponding tasks about tickets inquiring or reservation through online real-time dialogues with testers.

This Evaluation has not only automatic evaluation (for user intent classification tasks) but also online manual testing (for online testing of task-oriented dialogues). Compared with CCL2018 Task 1 and DSTC6, this Evaluation bears the following features:

• Compared to CCL2018 Task1, as organizers of the competition our data set contains a more comprehensive and more general set of tags, not just in one area. Specifically, we provide a data set which contains 31 user intents that appear frequently in general-purpose chatbots.

• Compared to DSTC6, we select several reviewers to evaluate the complete process of accomplishing a given task, and the reviewers will give their scores for a submitted system during each process.

• In order to avoid revealing the hidden test set and thereby reducing the possibility of manual intervention, we have modified the traditional evaluation method to allow the participating teams to set up services to respond to our requests so that the participants do not have to submit the code. At the same time, in order to avoid participants obtain the complete test set, we add a lot of noise to the test set.

In addition, compared to SMP2017-ECDT [13], this year we add new data sets for each of the two tasks. Our data sets provided by iFLYTEK Corporation^④ are all labeled manually. Different from Task 1 last year, we cancel the evaluation of the closed domain and only remain the open domain evaluation. The difference between the closed domain and the open domain is that users can not only use the provided training data but also collect data by themselves in the open domain. However, there is no guarantee that the participating teams will just use the evaluation data provided by us for training and developing their systems if we do not ask them to provide the code.

The rest of the paper is organized as follows. We introduce two tasks in detail in Section 2 and describe the data sets of two tasks in Section 3. Parts of the evaluation results are given in Section 4 and finally the conclusion is drawn in Section 5.

In this section, we give a brief introduction to evaluation tasks.

The specific descriptions of Task 1 are as follows: build a system that can classify a user's input into the most relevant category, including chit-chat or task subcategories, e.g.,

In Task 1, participating teams do not need to consider the overall intention of multiple rounds of a task-based dialogue, but to pay attention to a single round of dialogue. In addition, they are provided with a template of an example system^⑤ to facilitate the unification of the interface.

There are many text categorization tasks that use F1-measure as evaluation indicators, such as [14, 15, 16]. In order to avoid the imbalance of category distribution and meanwhile take into account each category, we also evaluate submitted systems based on the F1-measure obtained from precision and recall. Specifically, we first construct a confusion matrix for calculating the Precision P_i and Recall R_i value of each category, and then take the average precision as $P¯=1N∑i=1NPi$ and take the average recall as $R¯=1N∑i=1NRi$⁠, and F1-measure is calculated by Equation (1):

where N denotes the total number of categories.

Task 2 of the Evaluation is described as follows: For a complex task on booking a flight, a train ticket, or a hotel room, build a system to guide the user to complete the corresponding task based on the given relevant database. In this evaluation, we evaluate submitted systems online manually. Research in [17] suggests that the use of crowdsourcing technology is feasible and it can provide reliable results, and our reviewers are professional testers from iFLYTEK Corporation, which will be more likely to produce accurate results. A complete intent of a flight reservation task is described as:

“帮我订一张从北京到上海的飞机票，早上或者中午都行”

“Booking a flight from Beijing to Shanghai in the morning or at noon”.

The whole dialogue process of this flight reservation task is shown in Table 1, where U denotes the utterance of the user and R denotes the response of the agent.

Considering a variety of important factors on evaluation of a task-oriented dialogue system, we use the following indicators to evaluate the submitted systems in Task 2:

• Task completion ratio: The number of tasks completed during the test divided by the total number of tasks.

• Average number of dialogue turns: The number of utterances during the process of completing a task.

• Satisfaction score: The subjective score of the system marked by the tester, including 5 integers from −2 to 2.

• Fluency degree of response: Subjective scoring, including 3 integers from −1 to 1.

• Uncovered data guidance capability: Subjective scoring, including 0 and 1.

The core purpose of a task-oriented dialogue system is to help users complete a specific task. Then, the two most direct indicators for evaluating a task-oriented dialogue system is the task completion ratio and the average number of dialogue turns [18, 19]. The task completion ratio indicates the completion of the task and is the most important indicator that can reflect the system's capabilities. In Task 2, a complete intent may contain multiple subtasks, such as booking a flight first, and then booking a train ticket, and at last booking a hotel room. In order to demonstrate the ability of the system to complete composite tasks, when all the subtasks of a composite task are completed, we mark the completion of the task. For the average number of dialogue turns, it is counted by the evaluation system. When the task completion ratio is the same, the smaller the number of dialogue rounds, the better the system performs. In order to ensure that the number of dialogue turns of unfinished subtasks must be greater than or equal the number of rounds of completed subtasks, we take the number of dialogue turns of unfinished subtasks as the theoretical maximum number. If the maximum number of rounds is exceeded during the test, the current round of testing will be terminated. The remaining indicators are the subjective scores of the three reviewers, all of which are average scores. They reflect the performance of the dialogue system in the three aspects, respectively.

Actually, the test method of this evaluation is not only applicable to the Chinese Human-Computer Dialogue Technology Evaluation but also can be applied to the same evaluation tasks in other languages without too much modification except for the corpus.

The evaluation data set^⑥ in Task 1 is provided by iFLYTEK Corporation, all of which are labeled manually. Some specific examples of this data set are shown in Table 2. There are 31 categories of intent data and Table 3 shows how the data set is divided.

The data set of Task 2 contains information on flights, train tickets and hotels. It mainly includes the origin and destination of the flight or the train, the departure time and arrival time, the price, the type of tickets of the flight or the train, and the price and location of the hotel. Participants need to build a task-based dialogue system based on this information. In addition, we provide testers with some test cases and corresponding starting sentences that contain individual intentions and mixed intentions for tasks on booking air tickets, train tickets and hotels.

In this section, we show partial results of Task 1 and Task 2. Meanwhile, we analyze the results and summarize some frequently occurring problems of the two tasks. The complete leaderboards are shown in  Appendix A.

For Task 1, we have received 21 submitted systems in total, and part of the evaluation results are shown in Table 4.

After evaluating and ranking the submitted systems, we find that the average F1 score (0.8079) of the top five entries in this year's competition is much lower than that of last year (0.9268). The main reason is perhaps that the test set of this year is completely new and it is created later than the training set and the development set, which makes the test set in the different distribution with the training set and the development set. Therefore, the model trained in the training set performs worse on this year's test set than on last year's test set. This also indicates that many of the current models for text classification tasks have considerable losses after migration.

Since Task 2 is much more difficult and complex than Task 1, the number of submitted systems is also relatively small. A total of 10 systems are submitted in Task 2 (Table 5). The main reference indicators are C (task completion rate) and T (the average number of dialogue turns: the smaller the score T, the better the system). In this task, 34.29 is the theoretical maximum number of T, and the maximum penalty is made when C is zero.

The results shown in Table 5 are ranked by C firstly, then ranked by T, Sa, F and G in order. Among these indicators, C, Sa, F and G are manually labeled and T is calculated by the evaluation system. There are three reviewers to score each test case for each participating system. The final score for each indicator is the average of its scores on all test cases, given by reviewers or the evaluation system.

According to the results, this evaluation has been completed smoothly. Each participating team has verified their system on the provided data set and has achieved results that are consistent with their expectations. Through this evaluation, some key problems in the human-computer dialogue have attracted more people's attention. In addition, this evaluation mainly focuses on the application of human-computer dialogue systems, so it provides some references for the industry to solve the problem of constructing a human-computer dialogue system. In the meanwhile we found an interesting phenomenon from the evaluation results that the top three teams in the two tasks are almost all from the industry, which demonstrates the importance of experience in natural language processing evaluation tasks.

We introduce the Second Evaluation of Chinese Human-Computer Dialogue Technology, which has made some adjustments and improvements to solve the problems of the first session of the competition in 2017. In this paper, we introduce Task 1 and Task 2 of this Evaluation, respectively, and explain the updated indicators of the two tasks and the calculation methods of them. In addition, we illustrate the data sets of the two tasks. Finally, we show the evaluation results and analyze the problems in the evaluation.

This work was a collaboration between all of the authors. W. Zhang ([email protected]) is the leader of SMP 2018-ECDT, who drew the whole picture of the evaluation. W. Che ([email protected]), Z. Chen ([email protected]) and Y. Zhang ([email protected]) supervised the evaluation process. They summarized the conclusion part of this paper. Z. Zhao ([email protected], corresponding author) summarized the data sets and results of SMP2018-ECDT and drafted the paper. All the authors have made meaningful and valuable contributions in revising and proofreading the resulting manuscript.

We would like to thank Social Media Processing Committee of Chinese Information Processing Society of China (CIPS-SMP) for its strong support for this evaluation. Then we would especially thank Wenxia Feng, Xinyi Chen, Shu Fang, etc. They are the very serious and responsible testers from iFLYTEK Corporation, and it is them who complete the online evaluation of Task 2 patiently and impartially. Thanks to Huawei Technologies Co. Ltd. for providing financial support which sums up to RMB 50,000 as a bonus for this evaluation. Thanks to Lingzhi Li, Caihai Zhu, Yiming Cui, Haoyu Song and Yuanxing Liu for their indispensable support during the evaluation.