In TripMantu we have added command line based interface in the backend and have intergated it with whatsApp.. See the link
As predicted most of the queries are usually on the general-purpose text that the user types. These are difficult to crack because the system doesn’t know the intent of the text and since auto-complete isn’t available, the system has to predict from a vast set of tokens for each word.
Considering we are from the travel domain, we can have a pool of nouns for the destinations and dip for each word in the sentence but that is very costly as travel destinations are close to 100 thousand from across the globe.
One way is to do a confined search within the tile based on the first identified destination and search the rest of the words in the tile that the first destination belongs to. This can work if users are always trying to search within a confined location and can be a possible candidate for local groups. If you see the tile is similar to the index that can be used for the optimal search.
The second option is to get rid of all verbs, auxiliary verbs, pronouns, and adjectives and apply the noun search in the destination map. Out of the best matches that we receive, see which is the closest match in terms of
- number of characters.
- percentage of matching characters.
- Difference in the length by variance of 2-3 characters.
- A string comparison is a bad option considering that the user may type one wrong letter and the complete prediction goes for a toss.
The third option would be to go the NLP way, where the analysis is done word by word and based on the grammar, the system identifies the possible destinations that the user is trying to convey.
The fourth option is to rely on the premium NLP APIs from Google or chatGPT. For now, lets keep this out and continue with the first three options, till we get funding 🙂
Google API to get the places is more based on autocomplete and less on text analysis and more ever they expect the training set to be provided. So we tried with OpenAI. With OpenAI we could retrieve the correct places and feed to our algorithm. Though still many corner cases are there but the integration works for more than 90% of cases. Failed cases act as training ground and for each of the failed cases we go back and tune the parameters and enhance our regression cases.
-Written by Parasa Kiran
Parasa Kiran has a diverse skill set and a successful track record in driving innovation and delivering impactful solutions across various domains, particularly in rule engines, AI, telecom, Ux, and web technologies.
