Justifying Citizens' Observations and Concerns with Digital Twins for Smart Cities

Metropolises are complex systems where human factors substantially prevail over automatic and IT components. Complex systems imply complex interactions, where the effect of specific building or economic policies' is difficult to predict. Urban redevelopment works may not have the same impact on different districts of the same city, either because of the social composition, the neighbourhood's history, or the logistical and architectural impossibilities.

It is therefore of the utmost importance to ensure all citizens' well-being in the most equitable way.

One way to ensure this is to leverage online discussions (e.g. X formerly Twitter) about events within a city and compare them to golden data collected from a smart city (e.g. traffic information, weather sensors) or trusted sources of information (e.g. police reports, newspapers). For this, we envision City Digital Twins (CDTs) being continuously informed in a real city via real-time data connections.

However, previous literature does not consider textual data into the CDT, which we will in this project. We envision the usage of deterministic and explainable logic-driven AI for determining and motivating the outcome of a decision. We could struggle with identifying entities if there is not enough context provided.

For example, "Newcastle" could refer to one of three places: "Newcastle upon Tyne", "Newcastle under Lyme", and "Newcastle, Australia". So, we must ensure that full-texts have enough context for our analysis. This will then require the usage of CommonSense networks for motivating a specific reconstruction upon the definition of a specific context of interest, while determining the interest referred by the user.

Golden data will further narrow the set of possible interpretations of interest to the ones related to Newcastle upon Tyne specifically. For this, we need to determine the similarity between two or multiple texts, where this cannot directly exploit existing vector semantic tools. There exist limitations when involving a learning-based approach, for example, struggles with negation, meaning embedding must be provided for a given neural network (NN) to determine if two sentences are similar or not. Provided below are examples of different negations with the same entity as the subject:

1. "There is no traffic in Newcastle city centre" 2. "There is no traffic in Newcastle" 3. "There is traffic in Newcastle but not in the city centre"

Looking at the list above; 1. says there is not any traffic specifically in the city centre but does not clarify if there is traffic anywhere else, 2. says there is not traffic anywhere in Newcastle, and 3. says there is traffic everywhere but the city centre. We can easily decipher where the negation is occurring in these examples but AI struggles with this problem.

By using dependency graphs, we can tackle this problem by precisely determining what the negation is referring to, then apply rewriting rules for generating a declarative and machine-readable interpretation of the text. On top of this, noise complaints from online discussions could be compared to traffic data (our golden data), represented as a time series, to see if the time of the online post matches the time of the traffic congestion from our golden data, thus reflecting discrepancy between different means of observation.

The assessment of these discrepancies cannot be learned, as topics might vary with time. Information Retrieval-based Query Answering (QA) relies on the vast amount of text on the web or in collections of scientific papers, which neural reading comprehension algorithms read to draw an answer directly from a given text. Knowledge-based QA instead builds a semantic representation of the query, such as mapping a given question to the logical representation, which generalised graph grammars can help us achieve.

Overall, we envision exploiting such technology for bridging the gap between the user's understanding of the tool and the outcome of the data