ZhiweiJiang's Research Group at our institute haverecentlymadesomeprogress in the field of text quality assessment. Their achievements have been accepted as full papers at two top international conferences: The 61st Annual Meeting of the Association for Computational Linguistics (ACL2023, CCF-A conference) and The 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR2023, CCF-A conference).

The three published papers are as follows:

1. "Improving Domain Generalization for Prompt-Aware Essay Scoring via Disentangled Representation Learning". This research aims to enhance the generalization ability of Automated Essay Scoring (AES) models, enabling them to perform better on unseen essay prompts. To achieve this, a prompt-aware neural AES model is proposed to extract comprehensive representations for essay scoring, including prompt-invariant and prompt-specific features. Additionally, a disentangled representation learning framework is designed and introduced, which disentanglesprompt-invariant information from prompt-specific information through a contrastive norm-angular alignment strategyand a counterfactual self-training strategy. Extensive experimental results on the ASAP and TOEFL11 datasets demonstrate the effectiveness of this approach in domain generalization settings. The paper has been published at the ACL2023 conference. Colleagues interested in this research are welcome to contact:jzw@nju.edu.cn.

2. "Aggregating Multiple Heuristic Signals as Supervision for Unsupervised Automated Essay Scoring". This research introduces a novel unsupervised Automated Essay Scoring (AES) method called ULRA. The core idea of ULRA is to utilize multiple heuristic quality signals as pseudo-groundtruth labels and then train a neural AES model by learning from the aggregation of these signals. To aggregate these inconsistent quality signals into a unified supervision, ULRA formulates the AES task as a ranking problem and employs a specialized Deep Pairwise Rank Aggregation (DPRA) loss for training. In the DPRA loss, each signal has a learnable confidence weight to address conflicts among signals, and the neural AES model is trained in a pairwise way todisentangle the cascade effect among partial-orderpairs. Experimental results on the ASAPdataset demonstrate that ULRA achievesthe state-of-the-art performance compared withprevious unsupervised methods in terms ofboth transductive and inductive settings, and is comparable to many domain-adapted supervised models, showcasing its effectiveness. The paper has been published at the ACL2023 conference. Colleagues interested in this research are welcome to contact:jzw@nju.edu.cn.

3. "Unsupervised Readability Assessment via Learning from Weak Readability Signals". This research introduces a novel unsupervised framework for readability assessment called LWRS, which aims to evaluate the reading difficulty of text without the need for manually-labeled data for model training. The framework utilizes a set of heuristic signals to describe different aspects of text readability, guiding the model to output readability scores for ranking. Compared to methods relying on manually-labeled data, the LWRS multi-signal learning model effectively leverages multiple heuristic signals for model training and employs a pairwise ranking paradigm to reduce cascading coupling among partial order pairs.Furthermore, the study proposes a strategy based on signal consensus distribution to determine the most representative readability signals. Experimental results demonstrate that LWRS outperforms each heuristic signal and their combinations significantly, and can evenperform comparably with some supervised methods. Moreover, after being trained on one dataset, LWRS can be effectively applied to other datasets, including those in other languages, showcasing its strong generalization capability and broad potential for application.The paper has been published at the SIGIR2023 conference. Colleagues interested in this research are welcome to contact:jzw@nju.edu.cn.