How to realize the theme analysis of short texts? Python3 implementation

I am very interested in this topic. Say a few words about Mark's study.

When reading, people don't necessarily presuppose one or several themes, but gradually enter the author's mind according to the scenes or categories involved in the text, and perhaps the last sentence reveals that it is just a joke.

I think there are two main shortcomings of LDA, one is the number of theme buckets, and the other is the statistics of word order.

One way to improve thinking is to replace the topic bucket with a dictionary, that is, to extract the possible scenes or categories of each word for analysis and divergent thinking, instead of choosing from the topic bucket. For example: Apple, Price, Jobs. The possible categories of apples are fruit, agriculture, economy and mobile phones. Overlapping with the following parts of speech to form a topic neural network, and the analysis effect will be continuously improved through the improvement of AI dictionary in the future.

One is to increase the consideration of word order, capture the plot content, and analyze the plot through the changes of theme and scene before and after.

Just an idea, for reference only.