In the research paper \Towards Conversation Entailment : An Empirical Investigation",
the authors Chen Zhang and Joyce Y. Chai investigate the problem of conversation entailment [which
is the process of inferring hypotheses from conversational scripts]. For this they examine two levels
of semantic representations - basic representation based on the syntactic parsing from conversation
utterances and augmented representation which takes into consideration conversation structures.
The authors further explored two techniques to capture long distance relations between language
constituents - implicit modeling based on the length of distance and explicit modeling based on
actual patterns of relations. Some of the key observations of these experiments are :
1) The use of explicit model results in better entailment prediction accuracy when compared to the
use of implicit model. This suggests that the explicit modeling captures the semantic relationship
better than implicit ...view middle of the document...
This suggests that hypothesis type dependent models are
4) The modeling of long distance relations between language constituents appears only effective
when conversation structure is incorporated in the representation.
Critique 1 To simplify the model, the paper makes an assumption that a noun term from the conversation and a
verb term from the hypothesis (or vice versa) wouldn't be aligned to one another during the alignment
model. While modeling argument consistency using implicit and explicit modeling, the lack of
noun-verb alignments can have a negative impact on the overall semantic information available.
Since deeper and better semantics have a positive impact on entailment prediction accuracy, the
effectiveness of predicting whether a hypothesis entails a conversation or not is compromised to a
certain extent by this assumption.
Critique 2 In this digital age where conversations are abundant, the paper takes into account only a small
amount of conversation data. Although pre-processing the data to organize it in the conversation-
hypothesis form is a huge task, large data can help immensely towards accurate prediction.
Suggestion 1 It was observed in the paper that hypotheses type dependent models could prove beneficial. To
make use of this observation, instead of building a model which is common for all the hypotheses
we can do that following -
1) The hypotheses can be subjected to a classification algorithm - which associates the hypotheses
into one of the class among the pre-decided genres like fact, intent, desire etc.
2) Based on the type of hypothesis, we can select the corresponding entailment model to predict if
a hypothesis can be entailed from a particular conversation.
To get started with this approach, large amount of conversation data is required. But in this digital
age, that shouldn't be much of an issue. Initially, we can start out with only a few types of hypothesis
and keep expanding the model by including new genres of hypotheses as we go.
Conclusion The area of conversation entailment is an area of study in natural language processing with a
great scope for improvement. Also, the techniques used for conversation entailment can have great
practical application in the related studies of information extraction, question answering and summarization.