Challenges in Conversational AI and What We Can Do To Prevent Them
If we were to stare at a timeline chronicling the evolution of search, we could look into the past and see how it began with simple keyword recognition. Closer to the center, in the present day, we see how natural language processing (NLP) and machine learning help us recognize and predict user intent, finely tuning results to provide more accurate answers. But as we turn our heads toward the future, it is conversational search that will likely become the norm.
With 55% of U.S. households expected to own a smart speaker by 2022 , conversational search represents an obvious and exciting advancement in technology. However, it also poses several challenges and the same threats of bias we encounter with its text-based predecessor.
Overcoming Inherent Bias in Search
No matter how fair, open-minded or pro-equality people claim to be, inherent bias lives within them and comes to fruition through their actions. Whether it’s a bias toward the New York Yankees over the Boston Red Sox, action movies over romantic comedies or liberal media news outlets over conservative, bias is the byproduct of choice.
On the darker side of the spectrum, bias may reveal predilections toward a specific gender, ethnicity or socioeconomic status. Like it or not, bias plays a factor in how we search and interact with the Web and other data sources. Devices learn from user behavior, producing potentially tainted or one-sided results that lead to actions skewed in a particular direction and get dispersed out to the web of connected users.
Within text-based search, machine learning and natural language processing capabilities have made great strides toward understanding intent, but mind reading has yet to become an exact science. When it comes to conversational search, a whole new range of challenges and potential biases must be considered. Let us explore them.
- Language diversity: According to Ethnologue , about 23 languages are spoken by approximately half the world’s population, with more than 7,000 total languages in existence. While everyone across the world may not yet have access to a voice assistant or smart speaker, their differences must be accounted for in order for machines to properly learn and improve. Factors such as accents, dialects, slang, pronunciation, words/phrases with more than one meaning and emotional inflections from humor, sarcasm and irony must be discernible. As it stands, only of the world population has internet connectivity, and we are still far from reaching a large percentage of them.
- Multiple conversations: 63% of smart speaker owners placing the device in their living room , chances are it will eventually be exposed to multiple people engaged in overlapping conversations. In this type of scenario, voice assistants and smart speakers need to be able to differentiate between who is speaking and detect repetition to avoid recording duplicate data. Data redundancy and ghost identities will pollute the data in both singular and multiple languages.
- Competing assistants: The percentage of smart-speaker owners who said they owned three or more devices rose to , but the big-name companies have yet to find a way for these assistants to communicate with each other or distinguish between artificial and human voices. This means queries could be counted on numerous devices, resulting in more bias.
How Do These Challenges Affect Consumers?
With text-based search, consumers receive a list of relevant results to choose from, giving them the flexibility to choose what best suits their needs. However, with conversational search, users often expect only one result — the result .
When laced with bias, there’s no way to guarantee the accuracy of the results that voice-based search needs to deliver and popularity bias increases. While data bias will always exist to some extent as a product of user biases, businesses and developers can take a proactive approach to combat it on their end.
How Can Businesses and Developers Handle or Mitigate These Challenges?
Awareness is the first step in fighting bias. The second step is giving more control to the user by either mitigating the bias directly (e.g., with specific settings) or by using implicit/explicit feedback loops that will inform the search system of issues related to bias.
Language diversity is naturally achieved by increasing the languages handled by the systems and, today, that is driven by potential revenue rather than by the number of native speakers. If the main actors decide to invest today in the largest spoken markets, language diversity will be achieved sooner and potentially larger markets may become a future reality.
Handling multiple people and conversations well requires better identification techniques and contextual conversation tracking. In the case of doubt or when new unregistered people appear, the system could address them directly to resolve any ambiguity. These techniques will also need to recognize the voice of other smart speakers and identify the associated data as non-human.
Pro-activity Leads to Prosperity
Conversational search is still fairly fresh, but steadily and swiftly moving forward. It holds an unprecedented amount of possibility for businesses to understand consumers’ needs and for consumers to access and obtain what they want efficiently. Choice gives birth to bias and bias is the inevitable demise of choice because it limits knowledge and opportunity.
To ensure the playing field stays fair and accurate, businesses will have to incorporate a proactive approach to overcome the challenges that prevent long-term growth.
Originally published at https://www.forbes.com on July 30, 2019.