Known NLU problems


Problems we know and are already working on solving

... but it's worth knowing they exist now.

  1. The prefix problem - come / become, send / resend, underpay / overpay - if the rest of the sentence is identical, the algorithm is not able to distinguish these two phrases at this stage by only the prefix difference. Hence, if possible, it is worth grouping such verbs together for the same intent.

  2. Difference in tense or some other grammatical category - I would like to apply/ I’m applying/ I have applied — at the moment it is one and the same for the bot, because the model does not store and learn information about grammatical categories (morpho-syntactic or syntactic). Hence, it is better to avoid phrases constituting the context of the statement, e.g. I have submitted an application and I am waiting for your decision, because they may be confused with I would like to submit an application and replace them better, e.g. I would like to know about the decision regarding the submitted application.

  3. For the same reason, it is also better to build simple phrases with one verb, and then add compound sentences after, because there is a high probability that although people speak this way when talking to a human, they will avoid flowery phrases when talking to a bot, because they will be afraid of not being understood; We recommend adding complex sentences only after analyzing the logs with actual conversations with the bot and determining when and how people use this type of structure most often.

  4. Polysemy - verb ambiguity depending on the object, e.g. report the damage vs report a complaint or change the personal data (update) vs change the policy (to a new one). Another example: Failure report [application/website/application] vs failure report [at home] (NOTIFICATION OF DAMAGE vs TECHNICAL PROBLEMS). At the moment, we either group these types of phrases together and redirect them to the appropriate intent by entity, or we agree to the resulting conflicts and wait for the result of testing a new system solution.

  5. Perspective as the only difference in meaning to have something/an offer for someone (for you), e.g. Do you have x in your offer? vs I have an offer X for you. Unfortunately, at present the model cannot cope with this type of problem, because time differences, verb conjugations (personal endings) do not matter for it (for the same reason it confuses double intents with contexts cf. Points 2 and 3) . These kinds of things can nowadays be solved as a rule with MTT entities. We are currently researching other solutions to this problem.