Three Forms of Human-Machine Semantic Interaction in Artificial Intelligence
-
With the maturation of technologies like speech recognition and NLP, multi-turn dialogue systems are becoming a crucial bridge in human-machine interaction.
Currently, human-machine dialogue interactions in the market are mainly divided into three types: task-oriented, Q&A-oriented, and chit-chat-oriented.
- Task-oriented interactions aim to collect necessary parameter information based on user queries to complete tasks.
- Q&A-oriented interactions rely on model parsing to match answers from knowledge bases and provide them to users.
- Chit-chat-oriented interactions focus on emotional engagement, adapting to the user's current context to foster a closer connection.
In human-machine dialogue interactions, machines employ different semantic expression forms to process natural language. The primary forms are distributed semantics, frame semantics, and rule-based semantics.
01 Distributed Semantics
This can be simply understood as similar-sentence model parsing. Distributed semantics requires extensive annotated training data, where a large number of knowledge points form a vector space. When a user query arrives, the machine first tokenizes the sentence and matches it against a vocabulary. Alternatively, semantics can be interpreted as vectors, with the machine determining the optimal solution based on spatial vector distribution.
During model optimization, adjustments are made from both data and algorithmic perspectives. The optimal F-score is identified based on the model's precision and recall rates, and a threshold is set for the model.
Currently, most machine learning applications use distributed semantics, relying on data to handle semantic relationships. However, distributed semantics is relatively shallow and struggles with deeper semantic understanding.
02 Frame Semantics
Many voice assistant products on the market extract key parameters from user queries and fill them into predefined protocols to complete tasks.
For example, in the query "Check flights from Shanghai to Los Angeles on July 1," four information slots must be identified:
- The user's intent (flight information, excluding other transportation modes).
- The date parameter (July 1).
- The origin (Shanghai) and destination (Los Angeles), which must not be reversed.
Frame semantics relies heavily on identifying these parameters. Missing any slot can prevent the system from fulfilling the user's needs. To address this, "parallel slots" are often used. If a slot is missing, the system prompts the user to provide the necessary information.
03 Rule-Based Semantics
Rule-based semantics matches user queries against predefined expressions. When the rules are satisfied, the system provides a response.
For rule-based matching to work, rule-based logic must take precedence over similar-sentence matching. Rules also require extracting entity information from queries to ensure a match.
For example, a rule might involve symbols like
() {} . & | : #and fields to form a complex-looking but logically simple expression. The intent might be an action like "open stock details," with slots for stock types (A-shares, Hong Kong stocks, U.S. stocks) and a specific column. If the user's query fits the rule, the corresponding answer is provided.Different semantic processing forms have unique logics, but the ultimate goal is to complete user tasks or operations. With advancements in tokenization, entity extraction, and NLU, human-machine interaction will become more harmonious and efficient.