The difference between PGCN and TextGCN is that PGCN includes sequence
position information while TextGCN does not. In the 20NG dataset, PGCN
is increased by 1.11% (an improvement of 0.74% on the R8 dataset;
1.46% on the R52 dataset; Ohsumed and MR increased by 3.72% and
10.88%). At the same time, position information is added into GAT for
experiment (i.e . the PBGAT model in Table 3). As can be seen from
Table 3, the classification accuracy of PBGAT is higher than that of
PGCN on the five datasets; adding position information into the network
can significantly improve the classification accuracy, especially in the
sentiment classification dataset MR. According to the analysis presented
herein, because the task of emotion classification is closely related to
word order, the effect on MR is significantly improved: for example, “I
like this actor but I don’t like this movie” without the position
information, the model cannot tell where the actor is in relation to the
movie. Once the positions are reversed, the emotion of the whole
sentence is reversed. Therefore, the improvement of sentiment analysis
dataset is the most obvious, which shows the necessity of adding
position information.
To verify the effectiveness of edge features in improving network
performance, a PBGCN and a PEGCN are compared here. The PEGCN model
processes the adjacency matrix based on PBGCN and makes full use of the
multi-dimensional features of the edge. The only difference between the
two is the processing of the edge features. As seen from Table 3, the
classification accuracy of PEGCN on five benchmark datasets has been
improved, and the optimal classification effect has been achieved on all
datasets. In particular, for the 20NG and MR datasets, the accuracy is
improved by 2.62% and 1.89%, respectively. The analysis of this study:
Compared with the edge features represented by the adjacency matrix in
the past, the tensor after dimension enhancement has richer semantics.
Discretized points can only indicate whether there is a connection
between nodes, that is, connectivity features; while the edge matrix
after dimension enhancement has P additional dimensional features, which
can be used to represent more information between nodes, such as
connection types, connection strength, etc., making full use of edge
features helps the network to train better node representations, thereby
improving the classification task. It can be seen from Table 3 that the
full extraction of edge features has a certain effect on GCN to improve
the accuracy of text classification.
To make a comparison with similar models, BERTGCN26 is
used as the baseline model: BERTGCN is the model combining BERT and GCN.
The model proposed herein is compared with the model also combined with
BERT and GCN (Table 4).
Table 4. Comparison of Classification Accuracy of Similar Models.
metric: accuracy (%)