In this insightful interview with TechBullion, Siarhei delves into the transformative role of AI in user review analysis. He shares his expertise on how AI-driven tools are revolutionising the way businesses understand customer feedback, detect trends, and respond proactively to user needs. With the vast amount of data generated from online reviews, Siarhei discusses how AI-powered solutions are not only enhancing data accuracy but also enabling companies to make faster, more informed decisions to improve customer experience and drive growth.
Interviewer: We are glad to have you, Siarhei. It’s good to know you are with us. So, AI is all over the place, however, it is easy to envisage the misuse of AI for marketing purposes instead of value addition. What is the acceptable manner to build AI into a development team, while working on a user review processing feature, and not just throw in AI for the sake of it?
Siarhei: Thank you, it’s good to be here. It’s easy to agree with you. Yes indeed, the temptation of AI being at the forefront of attention is an evil that sometimes we all fight against. The basic question that teams should need to answer is, “What core problem are we trying to solve for the users ?” From a product development perspective the target should always be either addressing a problem or shifting the focus in accordance to the users and the patterns that they follow or possibilities mentioned in the reviews.
One concrete example of AI in action is the ability to provide recommendations based on the users’ reviews and behavioural data as done by AI at Netflix. Instead of simply tagging feedback as ‘good’ or ‘bad’, their natural language processing models take things a step forward by examining user satisfaction or lack thereof, at a more granular level such as satisfaction with how a certain character was developed or how the plot was paced. This allows Netflix, through its AI, to recommend more relevant content and to further learn what the users appreciate, by framing the feedback this way.
Another example is sentiment analysis and topic modelling on user reviews to identify recurrent issues such as questions regarding the cleanliness or responsiveness of the hosts, which is the case for Airbnb. Instead of relying on an aggregate star rating, Airbnb’s models do not overlook the points most related to any ready-for-use listing and take the measures necessary through the host platform. This AI-based approach has not only streamlined how they control quality but also how much more precise and contextualised recommendations have been made for the new users of Airbnb’s models.
In both cases, it is AI which sifts through massive collections of user feedback, finds patterns that might have an effect on the user, and lets this information go back into the development history. The intention here is not simply to put in AI as one feature, but rather to build feedback such that user sentiment can be acted upon and thus inform decisions which will convert insights to product improvement.
Interviewer: With GPT technology advancing rapidly, what advice can you give developers who want to pick tools that actually work rather than be deluded by the generative buzz? It was interesting to see how you mentioned NLP to be a key aspect for this type of analysis.
Siarhei: ChatGPT and other generative models such as LLMs have rather stormy hype, but they’re only an angle of AI and barely the most relevant. I mean, if we are about to analyse users’ reviews or a similar activity, then developers will need to use models that do the opposite – that’s interpretive or even analytical, not generative. A rough tip in this case would be a recommendation to turn to more specific ones first, for tasks but more often than not high-level NLP – sentiment analysis, topic modelling, and entity recognition – is required.
A comprehensive strategy uses several models to account for the multitude of pages of users’ reviews. So for instance a sentiment-analysis model will highlight the emotional background to convince, while topic modelling will allow for a clear outline of the set of reviews for “usability“, ‘performance‘, and ‘support‘ and others. Moreover, aspect-based sentiment analysis can decompose users’ sentiments expressed in topics to highlight which feature does the user like and which he dislikes more clearly. Such a multi-layer approach can get a more detailed target view; hence, development teams can address the priorities and adjust the details.
Moreover, developers need to expand their horizons beyond NLP and consider the possibility of clustering or regression based models, more in the case trends analysis over time. These techniques can assist in providing insights that would describe how particular issues may worsen or improve, thus making the analysis of feedback not only descriptive in nature but also predictive.
Interviewer: Those are powerful insights. For iOS developers specifically, integrating AI sometimes requires a level of access that is both technical and practical. Would you suggest any tools and resources to help in this integration for a start?
Siarhei: The ecosystem of Apple makes it easier. For example, CoreML provides a great benefit for iOS application developers because it enables them to include ML models within the applications. It is a mobile-centric approach that, in addition to efficacy, also promotes safety. In turn, CreateML helps in this regard because it allows the programs without deep knowledge in ML such as the majority of the teams without data scientists.
Another resource worth consideration is Apple’s Vision framework, which is particularly useful in image based applications. By utilising Vision together with CoreML, the AI is too reliant on the cloud and should instead be relied on too much, such as for real time video interpretation, face detection and recognition, and the classification of various content uploaded by users. It is the most effective way since AI computations remain on the device itself which is crucial for applications that are sensitive when it comes to personal data.
In addition to the world of Apple, Hugging Face provides access to pre-trained NLP models that can be trained exclusively for specific tasks such as reviewing a user. BERT and RoBERTa that can be found on Hugging Face are very effective in context modelling, which is important when dealing with user feedback that includes industry jargon or niche language references. These models can be used by iOS developers employing CoreML Converters to integrate more smoothly with their applications.
Interviewer: Once a developer has completed the exploration phase of a project and is ready to move on to the implementation stage, how can these models be integrated in such a way that they improve the experience for end users?
Siarhei: We begin with experimentation to recognize what works but it’s important to progress quickly onwards to what can be measured. In order to reach that point, I would advise to go with an iterative and validation driven approach. Use a phased manner: to begin with, evaluate your selected equity model using a limited set of user feedback data that contains actual real-world opinions. Determine how the model performs and then fine-tune the output for relevance and accuracy according to the feedback you are analysing.
In this case, metrics are very important, and set parameters for success prior to the launch. For example, if sentiment analysis tools are employed, determine how good the model is in distinguishing between positive, negative and neutral reviews. A/B testing should be done to establish whether these insights actually assist developers in focusing on the features that users are more satisfied with. Model retraining should be done regularly as well; user language and expectations change, hence modifying the model based on fresh input will keep it current.
When operating, deploy the logging and error tracking feature. This allows you to manage the performance of the models through flagging for changes where retraining may be necessary. In addition, it would be useful to also implement such a feature that allows for non-technical team members, such as product management, to view an aggregate of AI results as well. This enables all teams working together and the AI-related results actually determine the product roadmap.
Interviewer: In your opinion, how can AI transform insights gained from the app user reviews into app enhancements? Please share a more advanced example.
Siarhei: Certainly. For example, let’s take a fitness app that requires users to take pictures for workout sharing or to track the workout. Users might be providing feedback that they have issues related to picture quality or issues related to image uploading or formats. A standard design answer for such feedback could be ‗ implementation of images upload features‘, however AI does not stop there; it expands to capturing the various sub issues posed.
For instance, CoreML computer vision models might study an image’s properties at the time of transferring an image it might successfully detect a failed attempt of high-error images transferring and advise the user about ideal conditions for the photo. NLP also goes through some of the reviews, and searches comments on issues of usability concerning the image upload, while other comments differentiate whether there was frustration from the image upload or just a slight annoyance.
Thus this makes it possible to address all complaints as intelligently as possible: for instance, the developers before adding change of supporting multiple formats integrated a photo preview feature, or the balance of the file afterward is improved. AI allows the transformation of what was broadly framed as an image upload issue in the past into well defined and specific enhancement implementation, which will unavoidably improve user experience.
Interviewer: This has been an eye-opening and enlightening discussion, Siarhei. It’s evident that it is the prudent use of AI that yields radical results. Thank you for your knowledgeable input.
Siarhei: My pleasure. The true potential of AI can be realised if it is developed to meet specific user requirements and needs.
