It is imperative for Large language models (LLMs) to follow instructions with elaborate requirements (i.e. Complex Instructions Following). Yet, it remains under-explored how to enhance the ability of LLMs to follow complex instructions with multiple constraints. To bridge the gap, we initially study what training data is effective in enhancing complex constraints following abilities. We found that training LLMs with instructions containing multiple constraints enhances their understanding of complex instructions, especially those with lower complexity levels. Additionally, we further propose methods addressing how to obtain and utilize the effective training data. Finally, we conduct extensive experiments to prove the effectiveness of our methods in terms of overall performance and training efficiency. We also demonstrate that our methods improve models’ ability to follow instructions generally and generalize effectively across out-of-domain, in domain, and adversarial settings, while maintaining general capabilities.

Aspect Sentiment Quad Prediction (ASQP) aims to predict all quads (aspect term, aspect category, opinion term, sentiment polarity) for a given review, which is the most representative and challenging task in aspect-based sentiment analysis. A key challenge in the ASQP task is the scarcity of labeled data, which limits the performance of existing methods. To tackle this issue, we propose a self-training framework with a pseudo-label scorer, wherein a scorer assesses the match between reviews and their pseudo-labels, aiming to filter out mismatches and thereby enhance the effectiveness of self-training. We highlight two critical aspects to ensure the scorer’s effectiveness and reliability: the quality of the training dataset and its model architecture. To this end, we create a human-annotated comparison dataset and train a generative model on it using ranking-based objectives. Extensive experiments on public ASQP datasets reveal that using our scorer can greatly and consistently improve the effectiveness of self-training. Moreover, we explore the possibility of replacing humans with large language models for comparison dataset annotation, and experiments demonstrate its feasibility. We will release our code and data via GitHub.

To obtain a better understanding of user preferences in providing tailored services, dialogue systems have to generate semi-structured interviews that require flexible dialogue control while following a topic guide to accomplish the purpose of the interview. Toward this goal, this study proposes a semantics-aware GPT-3 fine-tuning model that generates interviews to acquire users’ food preferences. The model was trained using dialogue history and semantic representation constructed from the communicative function and semantic content of the utterance. Using two baseline models: zero-shot ChatGPT and fine-tuned GPT-3, we conducted a user study for subjective evaluations alongside automatic objective evaluations. In the user study, in impression rating, the outputs of the proposed model were superior to those of baseline models and comparable to real human interviews in terms of eliciting the interviewees’ food preferences.

Dialogue systems that aim to acquire user models through interactions with users need to have interviewing functionality. In this study, we propose a method to generate interview dialogues to build a dialogue system that acquires user preferences for food. First, we collected 118 text-based dialogues between the interviewer and customer and annotated the communicative function and semantic content of the utterances. Next, using the corpus as training data, we created a classification model for the communicative function of the interviewer’s next utterance and a generative model that predicts the semantic content of the utterance based on the dialogue history. By representing semantic content as a sequence of tokens, we evaluated the semantic content prediction model using BLEU. The results demonstrated that the semantic content produced by the proposed method was closer to the ground truth than the semantic content transformed from the output text generated by the retrieval model and GPT-2. Further, we present some examples of dialogue generation by applying model outputs to template-based sentence generation.