AI: New Dataset, CIDER for Common sense Inferencing in Two Person Conversations Dialogues

In this article we discuss an research paper by (Ghosal et al, 2021) [1] named “CIDER: Commonsense Inference for Dialogue Explanation and Reasoning”.   

Aim:

The authors have developed a database for conversation-based text understanding for two person conversations and a deep learning-based models have also been evaluated for performance  for NLP tasks including inference, span detection and multiple-choice span detection.

My Analysis About Article:

The researchers had logically created a database of two-person conversation dialogues in which several key issues have been kept in mind. Firstly, the conversations were annotated by research students who were told to annotate the dialogues between entities in form of triplet of form: cause, action and impact. The triplets can be present in a dialogue between two people in following two ways:

1. Explicit Triplets- Which can be parsed by parser or basic NLP tools of language understanding.
2. Implicit Triplets-These triplets have to be extracted using common-sense knowledge apart from dialogue understanding and sometimes may require several steps of intermediate triplet detection.

The annotations by the three research students were verified by each other and the best or majority votes were performed in case of any clashes in annotating a dialogue in form of triplets. There are various kinds of relation and causes that are studied.

Further, the datasets were tested on already developed models for the following tasks.

1. Natural Language Inferencing or Textual Entailment task performed on DNLI dataset. If finds out if the conclusion is true or false given the dialogue.
2. Span Generation: Finds out the conclusion part given the relation and premise.
3. Multiple Step Span Generation: Same as span generation but here it is like a multiple choice question answering and one correct inference is to be selected.

My Comments:

Other such datasets viz. Concept, Glucose, Atomic have been recently been developed but neither this problem was covered fully in these datasets nor such annotated combinations for two-person dialogues especially for common-sense reasoning were there.

This is a great work in conversational analysis especially for implicit relations and can be further expanded.

References

Ghosal, D., Hong, P., Shen, S., Majumder, N., Mihalcea, R., & Poria, S. (2021). CIDER: Commonsense Inference for Dialogue Explanation and Reasoning. arXiv preprint arXiv:2106.00510.