Although Computer Vision has made significant progress through Deep Learning technology, the adoption of Deep Learning in the field of Natural Language Processing is still in its early stages of development.
Regarding Chatbots, since the Turing Test was passed by the chatbot Eugene in 2014, Canadian scholars improved upon the test’s shortcomings by proposing the Winograd Schema Challenge, which is currently the most authoritative AI competition.
The first round of this competition involves Pronoun disambiguation problems. For example, when humans analyze sentences, they use experience to understand the referents:
- The city council refused the demonstrators because they feared violence.
- The city council refused the demonstrators because they advocated violence.
This multiple-choice question has only two answers: does the pronoun “they” refer to “the city council” or “the demonstrators”? The AI should identify that in the first sentence it refers to the city council, and in the second sentence it refers to the demonstrators. From the question, it can be seen that the system cannot derive the answer through contextual understanding of the passage. In traditional implementations, this requires reasoning via a Knowledge Graph or using deep neural network models. To win the $25,000 prize in the competition, Accuracy must reach over 90%, but the current best score is only 58%, far lower than humans.
In addition to the fundamental issues affecting Chatbot Q&A quality mentioned above, several challenges remain unbroken:
- Universal Model Architecture: To integrate modules such as speech recognition, lexical analysis, syntactic analysis, semantic analysis, deep learning, answer retrieval, dialogue management, natural language generation, and speech synthesis, ensuring their compatibility, current Chatbot architectures and models are quite complex and difficult to manage. Developing universal architectures and models is the development goal for all peers in the industry.
- Affective Computing: From Sentiment Analysis of text to Affective Computing that identifies human emotions, such as happiness, anger, and sadness; this can make Chatbot interactions with humans warmer, and is currently a popular research direction in academia and industry.
- Open Domain: Current Chatbots can only perform well in specific domains. Constructing open-domain knowledge, or even allowing machines to self-learn without manually constructed knowledge, is also a direction that academia and industry are working on.
- End to end: Instead of chaining traditional modules, use Deep Learning to build concise end-to-end models; achieving direct output results from raw input data. However, at the same time, it must support multi-turn dialogue management, contextual situations, and Knowledge Graph reasoning, avoid safe answers, maintain consistency in Chatbot personality, and correctly perform coreference resolution. These challenges are recent goals for academia and industry.
- Generative Model: Current natural language generation technologies can be divided into retrieval-based, template-based, and generation-based methods. All three can incorporate Deep Learning technology. Currently, retrieval-based and template-based methods are the industry mainstream; while generation-based methods are still in the early stages of development, with high spatial and temporal complexity, but its Deep LearningSeq2seq modelis very suitable for text generation. It is believed that as computing power increases, it will gradually become the mainstream approach.
