Abstract:
Conversational agents implemented for advising, question asking, in general for knowledge
generation purposes play relevant role in terms of accessibility, reliability, and gives cost efficient
auto services. Conversational agents are not simple to design, since different domains can need
different algorithms. Extracting information of certain law topic (such as law articles, criminal
stories, user selected law topics…) was achieved by the built model. This study aimed on
designing and implementing adaptive bot model which can facilitate a communication of
Ethiopian published laws by Federal Government of Ethiopia via solving the problem of taskoriented
models of generating relevant responses either topic responses or naturalness of
communication. The handcrafted and limited corpus of the model emphasizes usage of ensemble
architecture which is Amharic word to vector (Word2Vec) and Sequence to Sequence modeling
with rules integrated on top of it. This task oriented conversational bot identifies entities such as
questions for general knowledge, what-if situation is happened, and auto-generated law advices,
criminal stories for advising purpose. In order to design the artificial intelligence bot, syntactic and
semantic structure of data was under consideration, and to track user conversation memory
networks with sequence to sequence modeling also embedded. Four neural networks were
examined in order to select a fitting neural network for the intended model. The reason of using
word embedding’s is for the purpose of semantic extraction of words and sequence to sequence
model in generating responses. For word embedding’s, we built custom Amharic Word2Vec of
law documents and news obtained from websites, which produces word embedding then finally
we used it for extracting semantic similar words. To show the performance and results of the law
conversational bot, perplexity, accuracy measurement, and f-1 score metrics and user-acceptance
methods were used. The model throughput correctly classifies the intent with 73.12% accuracy.
The results show that the ensemble architecture based on sequence to sequence modeling having
rules on top of it with word vector implementation gives relevant information for user utterances.
Besides it, supportive multi-turn replies from the model is not considered by the model, and
including multi-turn responses, developing automatic question generation as well as creating
Amharic Word2Vec which incorporates massive text data can be supportive of the law bot model
for replying user intended information which is an implementation of future task.