Natural language processing methods for eliciting implicit user needs from online reviews
Yi Han · 2023
This dissertation contributes a series of innovative natural language processing (NLP) methods designed to uncover implicit and latent needs from customer reviews, with the goal of facilitating concept generation in the early stages of product design. Recognizing customer needs as crucial elements in the initial phases of the design process, current methodologies primarily rely on focus groups, interviews, or primary user feedback, inherently constraining opinions within a narrow demographic spectrum and set of products. The motivation for this research stems from (1) the lack of systematic computational methods to transform large-scale customer review data into novel design knowledge and insights; (2) the absence of specialized text summarization models aimed at extracting and succinctly encapsulating domain-specific information from customer reviews; and (3) the lack of a formal mechanism to integrate design knowledge into large language models for extracting aspect-specific implicit and latent opinions. This dissertation aims to address these knowledge gaps by building and validating several NLP models to solve an array of problems related to user need elicitation from online reviews and other similar textual user-generated data. The first study collects customer-generated reviews from online platforms and conducts unsupervised aspect-based sentiment expression analysis using two novel algorithms: part of speech (POS) tagging with window set" (PWS) andlanguage parser with syntax rules" (PSR). These algorithms not only determine the sentiment polarity but also extract the customer's sentiment expression. The study then delves into exploring clustering information based on opinion text and sentiment intensity. In the second study, an effective and scalable methodology is proposed for the automated and comprehensive extraction of user needs at the attribute level. Leveraging the state-of-the-art pretrained deep language model, BERT (Bidirectional Encoder Representations from Transformers), the methodology integrates two additional convolutional layers for post-training and named-entity recognition (NER) layers to extract attributes, descriptions, and sentiment words from extensive online user review corpora. The machine translation algorithm BLEU (BiLingual Evaluation Understudy) is employed to extract need expressions in predetermined part-of-speech combinations, such as adjective-noun and verb-noun. The third study formulates an efficient computational framework for abstractive opinion summarization, guided by specific product attributes and sentiment polarities. This involves the automatic generation of a synthetic training dataset that encapsulated various degrees of granularity and polarity. An innovative hierarchical multi-instance attribute-sentiment deep-learning-based inference model (MAS) is developed to construct a high-quality synthetic dataset, refining a pre-trained language model T5 (Text-to-Text Transfer Transformer) for abstractive summary generation. The fourth study introduces a novel task: generating a newly annotated dataset with five distinct labels encompassing aspect, category, opinion, sentiment, and implicit indicator (ACOSI). The study contributes a unified model capable of simultaneously extracting all five labels in a generative manner. Additionally, a novel position encoding method incorporating domain knowledge (DKG) into the transformer model is designed. A benchmark, integrating Rouge scores with domain knowledge, is introduced to assess performance within the design domain. The dissertation introduces several innovative models that augment the effectiveness of modern state-of-the-art NLP models in identifying implicit needs. The ultimate outcome of this research is the establishment of intelligent data-driven methodologies that extract information related to implicit needs on a large scale, based on myriad online reviews, comments, forum discussions, and similar datasets. These methods prove to be beneficial for designers in the early stages of the design process.--Author's abstract