Research On Intelligent Question Answering Method Based On Deep Learning

Intelligent question answering takes on various forms in real-life scenarios.Early research in intelligent question answering primarily relied on rule-based and statistical-based methods.However,these methods often depend on predefined rules and literal features for learning text feature representation,which cannot effectively capture the deep semantic knowledge contained in the text.With the rapid development of deep learning techniques,intelligent question answering methods based on deep learning have emerged.This thesis primarily focuses on the widely applicable matching-based question answering methods and conducts research in this area.Matching-based question answering methods mainly model the semantic matching relationship between two texts,aiming to judge the semantic matching between them.Currently,deep learning-based matching-based question answering methods can be roughly classified into three categories: methods based on feature representation,methods based on feature interaction,and methods based on pre-trained models.Existing work indicates that the performance of the latter two methods is superior to the first one.However,most existing methods often only perform shallow-level feature interactions between texts,adopt a single-level contrastive learning strategy,and use attention networks lacking directional awareness for text representation.This thesis focuses on addressing the aforementioned issues through research.The contributions of this thesis are mainly reflected in the following three aspects:⑴ This thesis proposes a question semantic matching method based on multiple alignment and feature augmentation.Addressing the issue of insufficient capturing of deep feature interactions within questions and between questions in existing approaches,this thesis introduces a question semantic matching model based on multiple alignment and feature augmentation.The model utilizes character and word granularities as multi-granular inputs and performs multiple rounds of attention alignment to capture deep semantic features.It also employs feature augmentation operations to generate the final question semantic matching representation.The thesis conducts a series of experiments on two Chinese datasets,and through analysis of the experimental results,the question semantic matching model based on multiple alignment and feature augmentation demonstrates improved performance compared to baseline methods.⑵ This thesis proposes a question semantic matching method based on a two-stage contrastive learning framework.Addressing the limitation of existing approaches that only perform contrastive learning at a single level(sentence-level or pair level)in text tasks,this thesis introduces a question semantic matching model based on a two-stage contrastive learning framework.The first stage of the model is used to obtain a well-distributed semantic feature representation at the sentence level,which is then passed to the second stage for pair level contrastive learning.Additionally,coarse-grained interaction features are utilized to enhance fine-grained interaction features,generating the final semantic matching representation.Through extensive experiments and analysis of the results,the question semantic matching model based on the two-stage contrastive learning framework demonstrates good performance compared to baseline methods.⑶ This thesis proposes a question-answer semantic matching method based on direction-awareness.Addressing the issue that existing approaches neglect the perception of contextual information in different directions when applying attention mechanisms to text tasks,this thesis introduces a direction-aware question-answer semantic matching model.The model represents the current text using semantic information from both preceding and succeeding directions,and captures the interaction features between question and answer pairs through attention alignment.Experimental results demonstrate that this approach effectively enhances the perception of different directional contexts in text,thereby improving the model’s prediction accuracy.