Word vector language model based on emergencies

Abstract

The invention provides a word vector language model based on emergencies. According to the language model, a traditional Word2Vec model is used for training a context; wherein the training comprises the steps of calculating input hidden layer information in an input layer of the model, adding vector representation of an emergency into the input layer, and obtaining final hidden layer representation through weighted summation to enable the vector representation of the emergency and the final hidden layer representation to jointly influence the final hidden layer representation; the generated hidden layer representation is not only related to the context, but also related to the emergency. The invention provides a new word vector model related to emergencies. The word vector model is used for modeling text stream data accommodating emergencies. According to the method, the word vector model with emergency characteristics can be learned to identify semantic changes, and emergency vector representation is added to improve semantic correlation.

Description

technical field [0001] The invention relates to the technical field of dynamic word vector generation models, in particular to a word vector language model based on emergencies. Background technique [0002] The semantics of words change over time, and many factors affect the semantics of words, including cultural changes, the emergence of new technologies, etc. For example, the word "Amazon" originally referred to the tropical rainforest. When the word "Amazon" is used, people tend to think it refers to e-commerce. [0003] Emergent events refer to events that are suddenly mentioned many times in a short period of time, such as "Notre Dame de Paris fire" and "Taiwan earthquake". In the emergency event detection model, the emergency event is generally expressed as <target word, start time, end time>. In previous work, people only focus on the representation of semantics in bursts, we propose to combine bursts and word vectors to capture and understand the correlation...

AI Technical Summary

Problems solved by technology

This approach is well motivated, but for some words can lead to over-smoothing of differences between word senses that have changed over time

[0011] 3) Although the non-random initialization method solves the alignment problem, it makes the words have rich semantic information of the previous corpus, and the ability to express semantic information of current emergencies is weakened

Images