The impressive performance of large language models (LLMs) has attracted considerable attention from the academic and industrial communities. Besides how to construct and train LLMs, how to effectively evaluate and compare the capacity of LLMs has also been well recognized as an important yet difficult problem. Existing paradigms rely on either human annotators or model-based evaluators to evaluate the performance of LLMs on different tasks. However, these paradigms often suffer from high cost, low generalizability, and inherited biases in practice, which make them incapable of supporting the sustainable development of LLMs in long term. In order to address these issues, inspired by the peer review systems widely used in academic publication process, we propose a novel framework that can automatically evaluate LLMs through a peer-review process. Specifically, for the evaluation of a specific task, we first construct a small qualification exam to select "reviewers" from a couple of powerful LLMs. Then, to actually evaluate the "submissions" written by different candidate LLMs, i.e., the evaluatees, we use the reviewer LLMs to rate or compare the submissions. The final ranking of evaluatee LLMs is generated based on the results provided by all reviewers. We conducted extensive experiments on text summarization tasks with eleven LLMs including GPT-4. The results demonstrate the existence of biasness when evaluating using a single LLM. Also, our PRE model outperforms all the baselines, illustrating the effectiveness of the peer review mechanism.
Unfortunately, the official English (sub)task results reported in the NTCIR-14 WWW-2, NTCIR-15 WWW-3, and NTCIR-16 WWW-4 overview papers are incorrect due to noise in the official qrels files; this paper reports results based on the corrected qrels files. The noise is due to a fatal bug in the backend of our relevance assessment interface. More specifically, at WWW-2, WWW-3, and WWW-4, two versions of pool files were created for each English topic: a PRI ("prioritised") file, which uses the NTCIRPOOL script to prioritise likely relevant documents, and a RND ("randomised") file, which randomises the pooled documents. This was done for the purpose of studying the effect of document ordering for relevance assessors. However, the programmer who wrote the interface backend assumed that a combination of a topic ID and a document rank in the pool file uniquely determines a document ID; this is obviously incorrect as we have two versions of pool files. The outcome is that all the PRI-based relevance labels for the WWW-2 test collection are incorrect (while all the RND-based relevance labels are correct), and all the RND-based relevance labels for the WWW-3 and WWW-4 test collections are incorrect (while all the PRI-based relevance labels are correct). This bug was finally discovered at the NTCIR-16 WWW-4 task when the first seven authors of this paper served as Gold assessors (i.e., topic creators who define what is relevant) and closely examined the disagreements with Bronze assessors (i.e., non-topic-creators; non-experts). We would like to apologise to the WWW participants and the NTCIR chairs for the inconvenience and confusion caused due to this bug.
Conversational Search has been paid much attention recently with the increasing popularity of intelligent user interfaces. However, compared with the endeavour in designing effective conversational search algorithms, relatively much fewer researchers have focused on the construction of benchmark datasets. For most existing datasets, the information needs are defined by researchers and search requests are not proposed by actual users. Meanwhile, these datasets usually focus on the conversations between users and agents (systems), while largely ignores the search behaviors of agents before they return response to users. To overcome these problems, we construct a Chinese Open-Domain Conversational Search Behavior Dataset (ConvSearch) based on Wizard-of-Oz paradigm in the field study scenario. We develop a novel conversational search platform to collect dialogue contents, annotate dialogue quality and candidate search results and record agent search behaviors. 25 search agents and 51 users are recruited for the field study that lasts about 45 days. The ConvSearch dataset contains 1,131 dialogues together with annotated search results and corresponding search behaviors. We also provide the intent labels of each search behavior iteration to support intent understanding related researches. The dataset is already open to public for academic usage.