We would like to thank all the reviewers for their valuable feedback and constructive suggestions.

This is a summary of main concerns and how we addressed them.

Reviewer 1

> 1) Some related work in resolving different types of ellipsis should be discussed.

Thanks for pointing out the literature. We will add these discussions and refer to them in the final version.

> 2) SRL data set and utterance completion overlapping

We carefully ensure that SRL test data set does not overlap with utterance completion data set. We randomly choose utterances from a wider range of Alexa data set to form our SRL training set so it might overlap with the completion set.

> 3) train and test sets split, evaluation set, hyperparameters

We do cross validation on the data sets and tune hyper parameters such as epoch of both our model and baseline models to give their respectively best performance.

Reviewer 2

> 1) comparison with state-of-the-art

Thanks for the comments!
We might fail to mention the state-of-the-art methods clearly. For dialog act prediction and SRL, the model considering only original utterances with ellipsis is the state-of-the-art. Namely, Yu and Yu (2019) for dialog act prediction and He et al. (2017) for SRL.
Besides, the main contribution of our paper is that we address the challenge of resolving prevalent ellipsis in social conversations that previous state-of-the-art models did not focus on.

> 2) Responses in case 4 of Table 7 are the same but the acts are different.

This example shows that Hybrid-EL-CMP overcomes the problem that automatic completion utterances might obscure some responses. In this case, whether the user responds with a simple "yes" or "yes I have had a pet", after automatic completion the response would be the same as "yes I have had a pet". But since our ground truth dialog act is given on the original utterances, their corresponding dialog acts would be different, i.e. positive answer and positive answer/statement respectively. If we just infer from the automatically completed utterances, we can not accurately tell the original dialog act. However, by combining original utterances, we can address this problem and correctly identify the dialog act as positive answer and statement.

Review 3

> 1) It is unclear how expert knowledge is encoded into and used by the selection model in dialog act prediction.

We pre-define some dialog acts that are not suitable to be completed eg. hold, apology. If original utterances predict such dialog acts, we directly use them as the final output. Otherwise, we combine the predictions from original utterances and automatic completion utterances.

> 2) Captions for Table 2 and 3 should be clearly differentiated.

We apologize for not differentiating the captions. We will modify them to be one for utterances with no predicates and the other for subordinate clauses separately and clearly.

> 3) deciding factor for selection model in SRL between rule-based and probability-based expert knowledge

Our rule-based and probability-based expert knowledge are two independent selection methods. We show their separate results in Table 8 as Hybrid-EL-CMP1 and Hybrid-EL-CMP2 and compare their respective benefits in the case study part of SRL experiment.

> 4) mechanisms to handle errors propagated from completion model

Combining original utterances and automatically completed utterances is a mechanism to handle trickle down completion errors. Besides paraphrases and minor changes do not affect the corresponding dialog acts much. We will consider adding more mechanisms to handle these errors in the future.

> 5) the reason why expert knowledge performs better only during testing

When incorporating expert knowledge during training, we apply tensor masks on two branches of logits according to our pre-defined dialog act set. These tenser masks potentially make back propagation harder.

> Comments on grammar, space, abstract and conclusion part

Thanks for pointing them out. We will revise the paper to fix these problems.