PQF-Game: A Restricted-Domain QAS in video games community by leveraging keywords filtering and enhancing the Pointer-Generator model

Variability in answer quality, frequent question repetition, fragmented information, unstructured content, and the complexity of natural language are the primary issues in online video game forums. Unfortunately, traditional Q&A systems are inadequate for forum discussions due to their inability to process unstructured, contextual, and informal language. We propose the Pointer-Question-Filter-Game (PQF-Game), a novel language model designed for Restricted-Domain Question Answering Systems (RDQASs) in video game forums. In the answer recognition module of PQF-Game, we inserted a keyword filter pipeline using an attention-based Neural Matching Model (aNMM) to select relevant input keywords during similarity scoring. In the answer generation module, we integrated a question layer and a document layer within the Pointer-Generator architecture to improve semantic focus, reduce redundancy, and address out-of-vocabulary (OOV) issues. We evaluated our model using GameFAQs and Steam datasets. The PQF-Game achieved the highest ROUGE score, ROUGE-1 = 0.3372, ROUGE-2 = 0.1574, and ROUGE-L = 0.2658 when applying a 900-dimension skip-gram word embedding, compared to Pointer-Generator, Pointer-Gen + Cov, word-lvt5k-1sent, and state-of-the-art models, i.e., BART, T5, and PEGASUS. Although user input may contain typos, slang, or unseen words, the question layer successfully preserves the semantic aspect of the question and documents. Our proposed system demonstrates the ability to generate meaningful and accurate responses, as confirmed through human evaluation. Our key contribution lies in combining an aNMM-based keyword filtering pipeline with dual-attention mechanisms in the Pointer-Generator to enhance RDQASs performance.