Building Smarter Chatbots: Using AI to Generate Reflective and Personalized Responses

With the advent of artificial intelligence-based tools, chatbots have been integral for user interactions. However, in most cases, chatbots provide users with a generic response to their queries. This lack of personalization does not capture any behavioral signals from the user. This not only leads to a suboptimal user experience but also a lot of missed opportunities for the organizations to convert these users from being return customers. This is a critical bottleneck for a lot of organizations to solve, as this prevents free users from becoming paid customers.

The first interaction these users have with these chatbots is more often than not the only opportunity to capture the user’s attention. For example, for a user visiting a mental health support application and interacting with its chatbot, it's vital to capture the user’s emotional state and potential needs and provide an empathetic response to make the user feel heard. Missing this window of opportunity is not only impacting the organization’s customer acquisition but also, leading to users not being led to a care pathway that could help improve their lives significantly.

Solving such a problem would allow organizations to improve their user conversions by providing more intellectually enhanced personalized experiences improving user satisfaction. Aligning intervention cues with unique user contexts might aid in giving support that feels individualized rather than generic.

Every user query can be thought as an input data to the chatbot algorithm. On the basis of that input query we want to formulate a response that is not only reflective of the original input but also has additional description and empathy. This requires the machine learning model to be trained on extensive query-response datasets to enable forming adequate responses. A large set of such curated datasets can be found at this Github repository.    

Figure 1 is a high level description of the process to generate a more personalized response to a user query.

Figure 1: High Level Process Description

For example, a user asks: “How may I deal with feeling depressed?”—this will be our query. Based on this, our trained machine learning model may generate a response saying: “I would like to find a problem that might be bothering you and try to understand how that makes you feel.” Then we would move to step 3, and convert the query to a reflection: “So, you want to understand how you may deal with feeling depressed,” and then prepend it to the model response.

This leads us to our final response: “So, you want to understand how you may deal with feeling depressed. I would like to find a problem that might be bothering you and try to understand how that makes you feel.”

Typical conversational systems like the one we are discussing are a part of language modeling problems. The main goal of such a model is to make it learn the likelihood that a would occur depending on the preceding sequence of words. 

We can use an n-gram level language model and bi-directional long short term memory (LSTM) to predict a natural language response.

A typical training dataset would include a question and its response. Figure 2, describes the data pre-processing we would apply on the question and response.

Figure 2: Data Pre-Processing

Followed by the data pre-processing we would perform tokenization to produce a sequence of tokens. This process is typically known as n-gram sequenced token generation.

A machine learning model works with an uniform input shape, so we would perform sequence padding to make all the token sequences of an uniform length.

But, a model would need predictors and its corresponding label to learn. So, we generate an n-gram sequence which act as predictors and the next work of that sequence is considered as the label.

A simple model architecture to begin with can be as described in Figure 3. 

Figure 3: Simple bi-directional LSTM based architecture

The first layer can be a bi-directional LSTM with ReLU activation and 256 units which can be followed by a Dropout and finally a Dense layer with softmax activation. For loss, we can use a cross-entropy loss function with an Adam optimizer to train the model.

This is an introductory model architecture which can further be enhanced depending on data and requirements.

Developing such a system does come with significant challenges with lack of annotated data being the most prevalent one. Its typically painstaking to find datasets annotated with responses marked as being reflective. Also, proper punctuation and ensuring the generated response is grammatically accurate can be challenging.

Finally, adding empathy and emotion to the response when required needs additional model development and training which comes under sentiment-enhanced text generation and might be constrained by availability of annotated data

To further improve such a system we can refer to MISC coded data which will help us get annotated data comprising of simple and complex reflections known as RES and REC respectively. We can use this to annotate our data with simple and complex reflections. An intriguing work by Can D, Marín RA et. al. titled "It sounds like...: A natural language processing approach to detecting counselor reflections in motivational interviewing” can be a great work to implement given the availability of the right annotated data and can help us make significant advancement in generating reflective responses.