Data or mathematics? Solutions to semantic problems in artificial intelligence

Abstract

Data support is already driving the development of artificial intelligence. But it cannot solve the semantic problem of artificial intelligence. This requires improving the semantic understanding ability of artificial intelligence. Therefore, a question answering system based on semantic problem processing is proposed in this study. The question answering system utilizes an improved unsupervised method to extract keywords. This technology integrates the semantic feature information of text into traditional word graph model algorithms. On this basis, semantic similarity information is used to calculate and allocate the initial values and edge weights of each node in the PageRank model. And corresponding restart probability matrices and transition probability matrices are constructed for iterative calculation and keyword extraction. Simultaneously, an improved semantic dependency tree was utilized for answer extraction. The improved keyword extraction method shows a decreasing trend in $P$ and $R$ values. The improved answer extraction method has a maximum $P$ -value of 0.876 in the training set and 0.852 in the test set. In a question answering system based on keyword and answer extraction, the improved method has lower loss function values and running time. The improved method has a larger area under ROC. The results of the validation analysis confirm that the improved method in this experiment has high accuracy and robustness when dealing with semantic problems.

Keywords

Data logic artificial intelligence semantic issues keyword answer extraction

1. Introduction

Data support is the foundation for the development of artificial intelligence (AI). When dealing with the natural language (NL), AI cannot correctly handle the semantics of NL due to the insufficient understanding [1]. This has led to a debate on the issue of AI semantics problem (SP). However, data support cannot completely solve the SP of AI [2]. Therefore, a thorough study of AI language is also needed. In the research of it, mathematical methods such as mathematical logic can be used to solve SP [3]. In essence, AI pays more attention to the study of mathematical logic [4]. Therefore, to solve the data logic problem of AI, it is also necessary to focus on the processing of SP. Therefore, in this study, the focus was on improving the methods of Artificial Intelligence Semantic Problem (AISP). Question and Answer System (QA) is one of the main application forms of AI. The NL processing capability in QA is a key factor affecting its application. However, among the existing QA language processing methods, it is not yet possible to fully understand and master the more complex human language [5]. Therefore, it is necessary to further improve the NL processing method in QA. The correct analysis of the semantics of natural language is the main direction for improving NL processing methods. In the experiment, a QAbased text processing method was introduced to improve the processing ability of AINL. Based on this, the semantic understanding ability of intelligent QA is enhanced, thereby improving the autonomous decision-making ability of AI. The article is mainly divided into five parts. It first provides a brief introduction to the research background of the article, which leads to the following content. The literature review analyzes and summarizes the relevant methods. The method section mainly provides a detailed explanation of the methods used. The next part is the validation of the performance of the method. The conclusion is a summary of the entire text, as well as a description of the shortcomings and future prospects. The system response performance test was conducted through actual cases, and the results showed that the unsupervised keyword extraction algorithm that integrates semantic features is beneficial for improving the efficiency of keyword extraction. Effectively integrating multiple model information simultaneously can improve the progress of answer extraction. These results demonstrate that the improved algorithm proposed in this article has played a certain positive role in the dialogue effect of the human-machine question and answer model.

2. Literature review

Based on the growth of data and technological updates, AI has achieved continuous development and improvement [6]. QA is an important product of the development of AI. It is an important channel for communication between humans and machines. QA has been well applied in various fields through continuous development and improvement. The processing of NL is the main factor affecting the effectiveness of QA display applications [7]. In the existing QA, simple human-machine dialogue can already be carried out. However, due to the complexity of language, QA is still unable to provide very accurate answers during human-machine communication. Therefore, researchers have made different improvements to the NL processing methods in QA. Researchers used a combined NL processing method to process text vectors in QA. Through example analysis, it has been proven that NL processing methods based on knowledge graphs and other methods have high accuracy in text processing [8]. This combined NL processing method can improve the runtime while ensuring the accuracy of the method. Lauriola et al. investigated the application of deep learning technology in NLP and found that deep learning could affect various tasks in NLP. They emphasized the main applications and limitations of deep learning in NLP [9]. NL processing affects the analysis results of data, which has a significant impact on the results of medical research. Researchers have improved the NL processing method for information extraction in pathological reports [10]. This further enhances the application value of NL processing in human society. Semantic similarity is a key and challenging aspect of NL processing. Besher et al. utilized a dialogue system for contextual semantic analysis. They used encoders to deeply analyze the semantics of the text. This improves its ability to infer contextual semantics. This SP processing method can provide users with more accurate answers [11]. Kumar and other scholars have combined deep learning methods with NL processing. The combined method can measure semantic similarity. Through this method, hidden semantic relationships can be captured. This improves user retrieval efficiency and speed of problem solving [12].

The correct judgment of semantic similarity affects the accuracy of QA keyword and answer extraction. Semantic dependency tree is an important method for extracting semantic relationships in text. Researchers have improved the existing semantic dependency tree [13]. They introduced attention mechanisms into the basic model. And neural networks were combined in the experiment to extract relationships. In the validation results of the dataset, the improved method can effectively improve F1 scores. Shi et al. improved the traditional dependency tree relationship extraction strategy and introduced graph neural networks, which could adaptively obtain dependency paths related to entity relationships and reduce noise in the data. A large number of numerical experiments verified the effectiveness of the new method [14]. Wang et al. combined neural networks with semantic dependency trees. The new model can correctly handle the information of network text. This model can effectively extract structural information from text and remove excess structural information. The experimental results in the dataset confirm that the model has higher information processing ability when processing text information [15]. And the unsupervised learning shows good performance in data processing. Some scholars introduced the advantages of unsupervised learning in biological research. After summarizing and comparing the methods, they emphasized the importance of unsupervised learning in data processing. And they also pointed out some limitations of this method and the challenges it faced [16].

The above research results were summarized and analyzed in the experiment. The results show that language processing ability is the main direction to improve the effectiveness of QA applications. This requires proper handling of the semantics problem of NL. Research has shown that methods such as semantic dependency trees can solve the semantics problem of AI. Therefore, in this experiment, an improved unsupervised method was used for keyword extraction. Then, a combination of multiple models was adopted to improve the answer extraction method based on semantic dependency trees. It is hoped to further address the SP issue with artificial intelligence question answering system (AIQA).

3. Solution of AISP based on keyword extraction method

With the development of mathematical logic, AI has gradually derived various application forms. AIQA in machine translation is one of the application forms of AI. When applying AIQA, it can raise questions in the form of NL. Then answer the question using the same NL method through the computer system. The SP in AIQA is inevitable. The accuracy of semantics is related to the accuracy of computer answers [17]. AIQA needs to have a correct understanding and expression of NL [18]. If natural semantics cannot be understood correctly, the answers provided by QA can create obstacles in human-machine dialogue and communication. Therefore, further improvement of AIQA is needed. In the following research content, it is proposed to use keyword and answer extraction methods to solve SP in AIQA.

3.1 QA and related language processing technologies in AI

The emergence of AIQA is one of the more mature application forms of NL. The earliest people proposed using computers to establish a QA for patient psychotherapy. The continuous development of technology has made QA a hybrid multifunctional NL interaction model. It can provide users with relevant knowledge queries and problem-solving. The main steps of QA are shown in Fig. 1.

Figure 1.

Framework diagram of the question answering system.

In the main steps of QA mentioned above, the understanding ability of the problem is first included. This section mainly analyzes various issues of user input, and conducts type judgment, grammar analysis, and keyword extraction. Then there is the information retrieval capability of QA. This section mainly uses the keywords extracted from the problem understanding step to locate candidate answers, in order to improve the accuracy of answer acquisition in the answer extraction step. The third step is answer extraction, which is an important technique in the entire QA process. Its purpose is to provide users with corresponding answers based on problem understanding and information retrieval. Due to its late start, there is still a significant gap between domestic research on QA and international level. The reason may be due to the complexity of Chinese semantics and the lack of corresponding language processing capabilities. In the existing NL processing technologies, it mainly involves the syntax, semantics, etc. of NL [19]. The NL processing methods mainly used include deep learning, structure tree, text vectorization, etc. The keyword extraction methods can include the supervised keyword extraction, the unsupervised keyword extraction, and the semisupervised keyword extraction. Supervised keyword extraction methods include genetic algorithms, support vector machines, etc. Unsupervised keyword extraction methods have strong scalability and are commonly used in keyword extraction techniques. This method can extract corresponding keywords based on word sorting, and has high convenience. Figure 2a shows the general pattern of unsupervised methods.

Figure 2.

Unsupervised keyword extraction method and PageRank.

In the unsupervised keyword extraction method, it mainly includes the preprocessing stage of text, screening of candidate word sets, sorting of candidate words, evaluation of keywords, and final keyword acquisition. The main keyword extraction methods used in unsupervised methods include word graph models, such as the PageRank algorithm in Fig. 2b. This method can iteratively calculate word nodes, in which the random walk strategy is applied. And this method calculates the score of word nodes to represent the importance of words. The importance of words based on the calculated score ranking is determined. Due to its high scalability and convenience, the PageRank algorithm was chosen as the fundamental technology for keyword extraction in QA in this study.

3.2 Improved keyword extraction methods in AIQA

The keyword extraction method is related to the performance of the entire QA semantic understanding. In this section, the PageRank algorithm mentioned above was selected as the basic method for keyword extraction. However, this method is easy to ignore the potential semantic information contained between text and words in iterative calculation. This can lead to insufficient understanding of textual information, thereby reducing the semantic comprehension ability of QA. To solve this problem, this experiment uses the improved unsupervised method and text vectorization method to extract text information. The obtained text and word order information were simultaneously input into the word graph model. Keyword extraction was performed in this model. Figure 3 shows the structure of the improved keyword extraction method.

Figure 3.

System technical structure diagram.

Among the improved keyword extraction methods, the unsupervised keyword extraction method is the main core architecture of the improved method. The improved method is detailed in Fig. 4. In the improved unsupervised keyword extraction method, it mainly includes text preprocessing, construction of co-occurrence word graph, semantic representation of text, scoring of semantic features, and keyword extraction. The text preprocessing stage is to screen candidate words that meet the criteria. The semantic representation of text is to obtain the vector between semantics and text. The scoring method for semantic features is to score information on semantics and word order. Finally, the word with the highest score was selected as the keyword.

Figure 4.

Flow chart of an improved unsupervised keyword extraction technology.

In the representation of text semantics, it is necessary to use the cosine similarity in Eq. (1) to calculate semantic similarity.

$\displaystyle\textit{sim}(v_{j},t)=\frac{\sum\nolimits_{k=1}^{n}{(v_{jk}\times t% _{k})}}{\sqrt{\sum\nolimits_{k=1}^{n}{(v_{jk})^{2}\times\sum\nolimits_{k=1}^{n% }{(t_{k})}}^{2}}}$ (1)

In Eq. (1), $v_{j}$ is the vector of the word $j$ , and $t$ represents the vector of the document. $n=512$ , which is represented as a spatial dimension value. The closer the semantic similarity between documents and words is to 1, the higher the semantic similarity between the two, and the greater the likelihood of the word being selected as a keyword. The scoring method for semantic features is to score information on semantics and word order. To improve the accuracy of PageRank algorithm keyword extraction, the semantic and word order information of the text was integrated into the experiment. In the experiment, the iterative scoring method was used to modify the transfer and restart probabilities. Then, the nodes were rated according to Eq. (2). The scoring method is based on the random walk strategy.

$\displaystyle TR(u_{i})=(1-\beta)\cdot r(u_{i})+\beta\sum\limits_{j:u_{j}\to u% _{i}}{\frac{e(u_{j},u_{i})}{\textit{out}(u_{j})}}\cdot TR(u_{j})$ (2)

In Eq. (2), $TR(u_{i})$ means the score in the co-occurrence word graph of word $u_{i}$ . $TR(u_{j})$ is the score in the co-occurrence word graph of word $u_{j}$ . $\beta$ is the damping coefficient, and $\beta\in[0,1]$ . GG $r(u_{i})$ means the restarting probability. $\frac{e(u_{j},u_{i})}{\textit{out}(u_{j})}$ means the word transition probability matrix value. Equation (3) is the calculation of $\textit{out}(u_{j})$ .

$\displaystyle\textit{out}(u_{j})=_{uf:uj\to uf}e(u_{j},u_{f})$ (3)

$u_{f}$ in Eq. (3) is the word node connecting the word $u_{j}$ . Equation (4) is the calculation of $e(u_{j},u_{i})$ .

$\displaystyle e(u_{j},u_{i})=\textit{sim}(u_{j},t)+\textit{sim}(u_{i},t)$ (4)

In Eq. (4), $\textit{sim}(u_{j},t)$ is the semantic similarity between $t$ and $u_{j}$ . $\textit{sim}(u_{i},t)$ means the semantic similarity between $t$ and $u_{i}$ . Equation (5) is the calculation of restart probability.

$\displaystyle r(u_{i})=\frac{\textit{sim}(u_{j},t)}{\sum\nolimits_{k=1}^{m}{% \textit{sim}(u_{k},t)}}$ (5)

In Eq. (5), $\textit{sim}(u_{k},t)$ refers to the semantic similarity between $t$ and $u_{k}$ . Based on the above score calculation, the word with the highest score is ultimately selected as the keyword.

Vectorization is the process of associating words and vectors, by vectorizing text to obtain a low dimensional floating-point vector. Each dimension in the obtained feature vector represents the grammar or semantic feature information of the text or words. In addition, the semantic information and word order information between text and words can be reflected through feature vectors. By combining the obtained word order information, semantic information between text and words can be fully explored. There are currently many methods for obtaining vector representations, among which the main representative models include Word2vec, Doc2vec, and Send2vec. During the process of word vector training for the Word2vec model, there are two models: Continuous Bag of Words Model (CBOW) and Skip gram model. The CBOW model, as shown in Fig. 5, is divided into input layer, projection layer, and output layer. The principle of its idea is to take the intermediate word of a paragraph of text as the output and the context of the word as the input, and predict the probability of the occurrence of the intermediate word by using the context.

Figure 5.

CBOW model architecture.

The Skip-gram framework model, as shown in Fig. 6, is also divided into input layer, projection layer, and output layer. Its model philosophy is opposite to CBOW, which predicts the probability of context based on the current word information of the sentence that appears. The co-occurrence information of words was considered in the training process of both models mentioned above. The word vectors obtained after training preserve the semantic correlation between words, and both can obtain semantic information of words.

Figure 6.

Skip-gram model architecture.

In order to vectorize words and analyze word embedding techniques, this article selects the Word2vec public model for vector representation of words, and uses the Skip gram model during the word vector training process. The candidate set words obtained during the preprocessing stage will be mapped to the vector space through the Skip gram model in the experiment. Based on existing experimental experience, during the training, the spatial dimension of the word vector is set to 512.

3.3 Improved answer extraction methods in AIQA

The answer extraction method in AIQA is to obtain accurate answers to user questions. The main answer extraction methods include word bags, pattern matching, etc. These methods are easy to implement and understand, but there is a significant semantic difference between the obtained answer and the true answer. The answer extraction method based on semantic dependency tree has good application results in QA. This method can analyze the semantic dependency relationship between questions and answers, thereby obtaining the semantic relationship of words. Therefore, this study chose semantic dependency trees as the basis for the answer extraction method. In the experiment, it is assumed that $D_{1}$ is the dependency tree structure in the problem, which includes $l_{1n}$ subtrees. So all subtrees contained in the dependency tree structure are $d_{11},d_{12},\ldots,d_{1n}$ . The result in Eq. (6) is obtained after $D_{1}$ vectorization.

$\displaystyle h_{(}D_{1})=(h_{1}(D_{1}),h_{2}(D_{1}),\ldots,h_{n}(D_{1}))$ (6)

In Eq. (6), $h_{1}(D_{1})$ is the total occurrences of subtree $D_{1}$ in the original semantic dependency tree. In the experiment, it is assumed that $D_{2}$ is the dependency tree structure in the answer, which includes $l_{2m}$ subtrees. So all subtrees contained in the dependency tree structure are $d_{21},d_{22},\ldots,d_{2n}$ . The result in Eq. (7) is obtained after $D_{2}$ vectorization.

$\displaystyle h_{(}D_{2})=(h_{1}(D_{2}),h_{2}(D_{2}),\ldots,h_{m}(D_{2}))$ (7)

In Eq. (7), $h_{1}(D_{2})$ is the total occurrences of subtree $D_{2}$ in the original semantic dependency tree. According to Eqs (6) and (7), the similarity in semantic dependency tree can be obtained in Eq. (8).

$\displaystyle\textit{sim }D(D_{1},D_{2})=\frac{\sum\nolimits_{l=1}^{l=\max(n,m% )}{h_{l}(D_{l})\times h_{l}(D_{l})}}{\sqrt{\sum\nolimits_{l=1}^{n}{h_{l}(D_{l}% )^{2}\times\sum\nolimits_{l=1}^{m}{h_{l}(D_{l})}}^{2}}}$ (8)

Semantic dependency trees can improve the similarity calculation, thereby mining complete semantic structure information. However, a simple semantic dependency tree disrupts the similarity of some semantic relationships when examining semantic feature information. This leads to insufficient accuracy in obtaining answers. Therefore, this study has improved the method. Figure 7 shows the system structure diagram of the improved method.

Figure 7.

System structure diagram of improved answer extraction method.

The improved method is based on a semantic dependency tree, which integrates word position semantic feature information into the semantic dependency tree. In this fusion process, it is necessary to consider the word frequency information of semantic features. Semantic feature information and word frequency information can effectively measure the semantic similarity of keywords between sentences. In the improved answer extraction method, a combination of multiple models can be used to introduce these two pieces of information into the semantic dependency tree. This operation takes into account the fragmentation of partial semantic relationship similarity in the semantic dependency tree. This can to some extent compensate for the accuracy of similarity calculation in semantic dependency trees. This further improves the performance of answer extraction methods based on semantic dependency trees. Figure 8 shows the specific flowchart of the improvement method.

Figure 8.

Improved answer extraction algorithm flowchart.

In the improved answer extraction method, it is necessary to conduct preliminary work on the questions raised by users. After completing the preliminary processing work, it is necessary to return to the corresponding candidate answer set. For questions and candidate answers, $D_{1}$ and $D_{2}$ need to be established and vectorized. Then, it is necessary to calculate $\textit{sim }D(D_{1},D_{2})$ between the question and answer in $D_{1}$ and $D_{2}$ . Next, Eq. (9) is used to calculate the position and frequency information of answer keywords.

$\displaystyle\textit{pos}(u_{i})=\sum\nolimits_{k=1}^{tf(u_{i})}{\frac{1}{p_{k% }(u_{i})}}$ (9)

In Eq. (9), $tf(u_{i})$ means the total occurrences of $u_{i}$ in the sentence. $p_{k}(u_{i})$ means the position where $u_{i}$ appears for the $k$ -th time in a sentence. Simultaneously, Eq. (10) is used to calculate the positional similarity between question and answer.

$\displaystyle\textit{sim }P(q,w)=\sum\nolimits_{i=1}^{n}\textit{pos}_{i}tf_{i}$ (10)

To incorporate location information into similarity calculation, the logarithmic linear model in Eq. (11) was introduced.

$\displaystyle S=\prod_{i=1}^{n}h_{i}^{f_{i}}$ (11)

In Eq. (11), $h_{i}$ means the keyword feature of $i$ . $f_{i}$ means the weight value corresponding to $i$ . By combining the above calculation methods, the improved semantic similarity calculation in Eq. (12) can be obtained.

$\displaystyle\textit{sim }DP(q,w)=\left(\frac{\sum\nolimits_{l=1}^{l=\max(n,m)% }{h_{l}(D_{l})\times h_{l}(D_{l})}}{\sqrt{\sum\nolimits_{l=1}^{n}{h_{i}(D_{i})% ^{2}\times\sum\nolimits_{j=1}^{m}{h_{j}(D_{j})}}^{2}}}\right)^{f_{i}}\times% \left(\sum\nolimits_{i=1}^{n}{\textit{pos}_{i}tf_{i}}\right)^{2}$ (12)

The minimum error rate was selected for parameter adjustment. Equation (13) was assumed to be the answer sentence extracted from QA.

$\displaystyle RA(\lambda(f))=w_{1},w_{2},\ldots,w_{m}$ (13)

In Eq. (13), $\lambda(f)$ means the weight estimation function, and $w$ is the number of keywords. So Eq. (14) is the reference answer to the question.

$\displaystyle CA=F(w)$ (14)

The minimum error function in Eq. (15) can be obtained from Eqs (13) and (14).

$\displaystyle E=\frac{\textit{Factual questions with incorrect answers}}{% \textit{General factual issues}}$ (15)

From this, the minimum error rate function in Eq. (16) can be obtained.

$\displaystyle f=\arg\min_{f}E(RA(\lambda(f)),CA)$ (16)

Based on the above calculation process, the information of position and word frequency can be integrated into the calculation of semantic similarity in the experiment. This can improve answer extraction accuracy in QA.

4. Validation analysis of AISP solutions

In order to validate the performance of the studied method, all validation experiments were conducted in the same experimental environment. The specific experimental environment is shown as follows: the processor is Intel(R) Core(TM) i5-9300H CPU @ 2.40GHz, RAM is 8G, Backend development language is Python, the compilation platform is PyCharm, the learning framework platform is TensorFlow and TensorFlow is WeChat. The keyword dataset for the validation analysis experiment is sourced from NetEase News. This includes content from multiple fields such as politics, society, education, and the arts. During the validation experiment, the keywords annotated in the dataset were used as the control content for the validation experiment. The selected keyword is considered to be the correct extraction result only when and only when the selected keyword using the method is completely consistent with the annotated keyword.

In the validation analysis of this chapter, other different indicators were used to validate the performance of the methods used above. The comparison indicators selected include precision ( $P$ ), recall ( $R$ ), and the ratio $F$ of the two. And the comparison of indicators between the methods and those in the literature was also examined. Figure 9 shows the validation experimental results of the keyword extraction method. Figure 9a shows the relationship between each evaluation indicator and the number of keywords. When the number of keywords increases, $P$ , $R$ , and $F$ indicators show significant changes. $P$ and $R$ decrease while F increases. Figure 9b–d show the comparison results between reference methods and the keyword extraction method in this experiment. $P$ values of each method show a decreasing trend, while $R$ values show an increasing trend. However, the improved keyword extraction method in this experiment has higher $P$ and $R$ values than the methods in references. In Fig. 9d, $F$ -values of each method show a trend of increasing first and then decreasing. The method in this experiment has the highest $F$ -value of 0.29 when keywords number is 3. This is because word order information plays an important role in understanding document information. Using word order information can better grasp the primary and secondary relationships of document components, the construction of phrases, and some potential feature information. Therefore, using word order information in the experiment can be more conducive to accurately grasping the semantic feature information of sentences in the process of text semantic similarity calculation.

Due to the lack of suitable datasets for performance verification in the existing Chinese QA, 300 questions were selected as the test set in the network in this experiment. Table 1 shows the relevant issues in the self-made test set. It mainly includes issues such as numbers, place names, person names, and time. In the experiment, 240 questions were selected as training set, and the remaining 60 questions were selected as testing set.

Table 1
Specific contents of the dataset

Problem classification standards and types of interrogative words				Experimental dataset statistics
Problem types (major categories)	Problem type (sub category)	Types of interrogative words	Example sentence	Person name category	68
Name	Specific person name/ description, etc.	Who/who/which person/ who/those people, etc.	Who is the author of “Water Margin”?	Digital class	56
Number	Quantity/number/area, etc.	How much/how many	What is the global surface area in square kilometers?	Time class	55
Time	Year/month/day/time, etc.	When/when to wait	Which year was the Daqing Oilfield discovered?	Place names	50
Place name	Mainland/country/address, etc.	Where/where/where, etc.	What is the longest river in the world?	Define class	34
Definition	What	What	What was Lu Xun’s original name?	Others	37
Others	–	–	What is the nickname of Kunming?	–	–

Figure 9.

Different evaluation index.

In the validation experiment of answer extraction method, $P$ is used as the evaluation indicator. Figure 10 shows the $P$ results of different methods. Figure 10a shows the $P$ values of different methods in training and testing sets. The improved answer extraction method has a higher $P$ -value. Its maximum $P$ value in the training set is 0.876, and its maximum $P$ value in the test set is 0.852. Figure 10b shows the $P$ values of each method as keywords increase. As keywords increase, the $P$ -value of the improved answer extraction method gradually decreases. But the downward trend of this method is lower than that of other methods. This is because the algorithm in this article integrates the semantic information and word order information of the text into the keyword extraction scoring of the word graph model. This method can effectively fuse the semantic feature information of text and improve the performance of keyword extraction.

Figure 10.

Comparison of $P$ value.

The validation experiment also tested the overall performance of QA. Figure 11 compares the loss function value and running time of the improved method during training. This method has lower loss function value and running time. The final value of its loss function is 0.22, and the final running time is 2.54 ms. This means that the improved method can achieve better robustness after training. And this method effectively reduces the running time after training.

Figure 11.

Loss function and running time.

As one of the evaluation indicators for machine learning and other methods, the receiver operating characteristic curve (ROC) can reflect the accuracy of the model. The area under the curve is larger, the corresponding method accuracy is higher. The PR curve is actually a curve made with precision and recall as variables, where recall is the horizontal axis and precision is the vertical axis. Similar to the ROC curve, the larger the area under the PR curve, the better the performance of the method. The ROC and PR of different methods are compared in Fig. 12. From Fig. 12a, the improved method has a larger area under the curve. And this method has smoother curve features. This represents that the improved method has higher accuracy and robustness. From Fig. 12b, it can also be seen that the method used in this study has a larger area under the PR curve, which is consistent with the results of the ROC curve. Both confirm the excellence of the research method. This is because the semantic features of word position and word frequency are integrated into the semantic similarity calculation of semantic dependency trees. This approach allows for a more comprehensive consideration of semantic feature factors, thereby improving the accuracy of answer extraction in the system and laying the foundation for optimizing the performance of the entire system.

Table 2

Experimental results of each model on the STC-SeFun and Weibo datasets

Dataset	Model	Automatic evaluation indicators			Manual evaluation indicators
		BLEU1	BLEU2	CWS	Fluency	Information amount	Relativity
STC-SeFun	Reference 17	0.1577	0.1192	2.4003	1.5645	1.0427	1.1718
	Reference 18	0.1676	0.1266	2.4943	1.5929	1.0395	1.0637
	Reference 19	0.1651	0.1248	2.3747	1.6380	1.0112	0.9555
	This paper	0.1790	0.1378	3.0263	1.7609	1.3409	1.1897
Weibo	Reference 17	0.0988	0.0788	2.9978	1.4385	0.8505	0.4694
	Reference 18	0.1578	0.1275	3.9060	1.3230	0.8085	0.4442
	Reference 19	0.1527	0.1227	3.7000	1.2884	0.5219	0.1365
	This paper	0.1740	0.1428	4.3640	1.7178	0.9797	0.5429

Figure 12.

ROC and PR of different methods.

In order to further validate the practical application performance of the question answering system proposed in the experiment, automated evaluation indicators BLEU1, BLEU2, and CWS were used in Table 2, as well as scores for fluency, information amount, and relevance of manual evaluation indicators. BLEU reflects the degree of overlap between the responses generated by the model and the actual responses. CWS reflects the amount of information in the conversation response by generating the number of content words in the response through the model. From the experimental results, it can be seen that the model proposed in this article is superior to other models in both automated evaluation indicators and manual indicators. The most important thing to note is that the proposed model has much higher information content in both CWS and manual evaluation indicators than other models, indicating that the model can effectively improve the information content of conversation responses. Combining the fluency index, it indicates that the model in this paper can balance the fluency and information content of reply generation. This is because the improved unsupervised keyword extraction technique proposed in the article can accurately obtain vector representation information of text and words. On this basis, the PageRank model was used in the experiment to improve the accuracy of keyword extraction. At the same time, the improved method was used in the experiment to improve the accuracy of answer extraction. These tasks can accurately locate the user’s questions in the human-machine question answering system, thus laying the foundation for smooth dialogue.

5. Conclusion

Natural language processing is a branch of artificial intelligence that involves the design and implementation of systems and algorithms that can interact through human language [20]. Due to the latest advancements in deep learning, the performance of NLP applications has been unprecedentedly improved [21, 22]. AISP is mainly embodied in mathematical logic [23]. How to correctly understand NL is an urgent issue that AIQA needs to solve [24]. To solve the SP problem, an improved keyword extraction method and answer extraction method were proposed in this experiment to improve the NL processing ability of AI. In the experiment, an improved unsupervised method was used to extract keywords, and an improved semantic dependency tree was used to extract answers. In the validation analysis comparison results, the $P$ and $R$ values of the improved keyword extraction method showed a decreasing trend. The overall $F$ -value of this method shows a downward trend. When the keywords are 3, the maximum $F$ -value of this method is 0.29. The improved answer extraction method has a higher $P$ -value. Its maximum $P$ -value in training set is 0.876, and its maximum $P$ -value in test set is 0.852. As keywords increase, the $P$ -value of the improved answer extraction method gradually decreases. But the downward trend of this method is lower than that of other methods. The improved method has lower loss function value and running time. The final value of its loss function is 0.22, and the final running time is 2.54 ms. The improved method has a larger area under ROC. And this method has smoother curve features. The above validation analysis demonstrates that the improved method has higher accuracy and robustness. This method has higher operational efficiency. This experiment has made some improvements in handling QASP. However, there are still many key and difficult points, such as corpus being essential training data for QA. However, there is currently a lack of suitable Chinese language materials. Therefore, in subsequent research, it is necessary to establish a corpus suitable for Chinese dialogue. And there is a lack of research on synonym recognition in keyword extraction methods. Therefore, it is necessary to further expand the synonyms of keywords in order to facilitate subsequent information retrieval and matching.

References

Luo

Chen

Guo

. Semantic communications: Overview, open issues, and future research directions. IEEE Wirel Commun. 2022; 29(1): 210-219.

Pokhrel

. Learning from data streams for automation and orchestration of 6G industrial IoT: Toward a semantic communication framework. Neural Comput Appl. 2022; 34(18): 15197-15206.

Samant

Bachute

Gite

Kotecha

. Framework for deep learning-based language models using multi-task learning in natural language understanding: A systematic literature review and future directions. IEEE Access. 2022; 10: 17078-17097.

Zhuang

Wang

Delgrande

. A generalisation of AGM contraction and revision to fragments of first-order logic. J Artif Intell Res 2019; 64: 147-179.

Matsui

Suzuki

Ando

Kitai

Haga

Masuhara

Kawakubo

. A natural language processing model for supporting sustainable development goals: Translating semantics, visualizing nexus, and connecting stakeholders. Sustain Sci. 2022; 17(3): 969-985.

Guo

Ren

Zhang

Zou

Liang

. Artificial intelligence for histological subtype classification of breast cancer: Combining multiscale feature maps and the recurrent attention model. Histopathology. 2022; 80(5): 836-846.

Zhang

Zhao

Wei

Zhang

Flanagan

Xiong

. Context-based semantic communication via dynamic programming. IEEE T Cogn Commun. 2022; 8(3): 1453-1467.

Phan

THV

. Developing a BERT based triple classification model using knowledge graph embedding for question answering system. Appl Intell. 2022; 52(1): 636-651.

Lauriola

Lavelli

Aiolli

. An introduction to deep learning in natural language processing: Models, techniques, and tools. Neurocomputing. 2022; 470: 443-456.

10.

Khurana

Koli

Khatter

Singh

. Natural language processing: State of the art, current trends and challenges. Multimed Tools Appl. 2023; 82(3): 3713-3744.

11.

Al-Besher

Kumar

Sangeetha

Butsa

. Bert for conversational question answering systems using semantic similarity estimation. Cmc-Comput Mater Con. 2022; 70(3): 4763-4780.

12.

Kumar

Steyvers

Balota

. A critical review of network-based and distributional approaches to semantic memory structure and processes. Top Cogn Sci. 2022; 14(1): 54-77.

13.

Lei

Song

Kou

. Neural attentional relation extraction with dual dependency trees. J Comput Sci Technol. 2022; 37(6): 1369-1381.

14.

Shi

Xiao

Quan

Lei

Niu

. Distant supervision relation extraction via adaptive dependency-path and additional knowledge graph supervision. Neural Net. 2021; 134: 42-53.

15.

Wang

Liu

Wei

. Weighted graph convolution over dependency trees for nontaxonomic relation extraction on public opinion information. Appl Intell. 2022; 52(3): 3403-3417.

16.

Yan

Wang

. Unsupervised and semi-supervised learning: The next frontier in machine learning for plant systems biology. Plant J. 2022; 111(6): 1527-1538.

17.

Sujatha

Subramaniam

Rene Robin

. A new design of multimedia big data retrieval enabled by deep feature learning and Adaptive Semantic Similarity Function. Multimedia Syst. 2022; 28(3): 1039-1058.

18.

Nawaz

Shi

Gan

Hirpa

Dong

Zheng

. Temporal moment localization via natural language by utilizing video question answers as a special variant and bypassing Nlp for corpora. IEEE T Circ Syst Vid. 2022; 32(9): 6174-6185.

19.

Huang

Gao

Tao

. Toward semantic communications: Deep learning-based image semantic coding. IEEE J Sel Area Comm. 2023; 41(1): 55-71.

20.

Perboli

Gajetti

Fedorov

Giudice

. Natural Language Processing for the identification of Human factors in aviation accidents causes: An application to the SHEL methodology. Expert Syst Appl. 2021; 186: 115694.

21.

Hur

Hardefeldt

Verspoor

Baldwin

Gilkerson

. Overcoming challenges in extracting prescribing habits from veterinary clinics using big data and deep learning. Aust Vet J. 2022; 100(5): 220-222.

22.

Wang

Zhuang

Sun

Zhang

Lin

Xia

. Adaptively sharing multi-levels of distributed representations in multi-task learning. Inform Sci. 2022; 591: 226-234.

23.

Luo

Chen

Guo

. Semantic communications: Overview, open issues, and future research directions. IEEE Wirel Commun. 2022; 29(1): 210-219.

24.

Zhou

Xie

Huang

. DFM: A parameter-shared deep fused model for knowledge base question answering. Information Sciences. 2021; 547: 103-118.

Data or mathematics? Solutions to semantic problems in artificial intelligence

Abstract

Keywords

1. Introduction

2. Literature review

3. Solution of AISP based on keyword extraction method

3.1 QA and related language processing technologies in AI

Table 1 Specific contents of the dataset

References

Table 1
Specific contents of the dataset