AI’s Next Evolution: Teaching Common-Sense Reasoning to Language Models for Culturally-Informed Responses
AI language models like ChatGPT have made significant strides in their capabilities, ranging from passing exams to generating creative content. However, one crucial aspect where these models still lag behind humans is in the realm of reasoning. In a recent Q&A session, Dr. Vered Shwartz and Mehar Bhatia shed light on the importance of teaching AI models common-sense reasoning and the need for diverse datasets to train them.
Currently, large language models like ChatGPT learn by processing vast amounts of text from the internet, enabling them to generate information based on recognized patterns. However, their understanding is limited to what is explicitly mentioned in the documents they’ve analyzed. Humans, on the other hand, possess reasoning abilities that extend beyond explicit information. We can utilize logic and common sense to derive meaning and make inferences.
Reasoning abilities are acquired by humans from an early age. We learn from our surroundings, environment, and situations. For example, we instinctively know not to use a blender late at night to avoid disturbing others. These reasoning skills are invaluable when it comes to AI models, as we cannot manually instruct them on every common-sense rule for every specific context.
Integrating common-sense reasoning into existing models like ChatGPT would enhance their accuracy and make them more effective tools for human interactions. While current AI models do exhibit some level of common-sense reasoning, there is still much room for improvement. For example, the latest version of ChatGPT can correctly differentiate between a child’s mud pie as a dessert and an adult’s mud pie as a face covered in dirt. This showcases its contextual understanding to some extent.
However, common-sense reasoning in AI models is far from perfect. Training them solely on massive amounts of data can only take us so far. Human intervention and the provision of appropriate data are necessary to refine these models. Unfortunately, most English text data available on the web originates from North America, leading to a North American bias in English language models. This bias can result in the models lacking knowledge of concepts from other cultures or even perpetuating stereotypes.
To address this issue, Dr. Shwartz, Ms. Bhatia, and their team conducted a study on training common-sense reasoning models using diverse datasets from cultures such as India, Nigeria, and South Korea. The results were promising, with the culturally-informed models providing more accurate responses. For instance, when shown an image of a woman in Somalia getting a henna tattoo and asked for the reason, the model trained on diverse data correctly suggested marriage, whereas the earlier model had wrongly assumed a purchase intention.
Additionally, ChatGPT displayed a lack of cultural awareness when responding to a hypothetical situation involving tipping in Spain. The model suggested that a four percent tip indicated dissatisfaction with the service, disregarding the cultural difference where tipping is not common in Spain.
The implications of language models with biased cultural norms are significant. Inaccurate and discriminatory information could be provided to or about people from diverse backgrounds. Moreover, individuals from different cultures using products powered by English models might have to adapt their inputs to fit North American norms, leading to suboptimal performance and exclusion.
Ensuring inclusivity, diversity, and cultural sensitivity in AI technologies is crucial. The ongoing research conducted by the team aims to achieve this goal. Canada, being a culturally diverse country, emphasizes the necessity of AI tools that cater to all cultures and norms.
In conclusion, the evolution of AI language models lies in equipping them with common-sense reasoning abilities. By training these models on datasets that represent diverse cultures, we can improve their accuracy and foster inclusivity. The quest for culturally-informed AI models is essential for ensuring a fair and comprehensive representation of global perspectives.
