A Case Study

A case study with a concept-set {hand, sink, wash, soap} for qualitative analysis of machine generations.

Transfer Learining Study

We study how fine-tuned CommonGen (CG) models can benefit commonsense-centric downstream tasks such as Commonsense Question Answering (CSQA) with their generative commonsense reasoning ability. We use several trained CG models to generate choice-specific context and see their influences. Check Sec. 5.3. for more.


We use the AMT platform for collecting such sentences for covered the top-ranked 2,500 concept-sets in the sampled results from large visual caption copora. Each of them is assigned to at least three different workers. Furthermore, we use the remaining concept-sets as the training examples, for which we use the associated captions as the target outputs. There are on average 4 sentences for each example in dev and test sets, which provide a more diverse test-bed for further automatic and manual evaluation. We highlight the ratio of novel concept compositions (i.e., concept, concept-pair, and concept-triple) in dev/test, which never (co-)occur in training examples. This makes CommonGen challenging in terms of compositional generalization ability.


Connectivity Distribution of Concept-Sets

We here introduce deeper analysis of the dataset by utilizing the largest commonsense knowledge graph (KG), ConceptNet as an tool to study connectivity and relation types. Obviously, if the concepts inside a concept-set is more densely connected with each other on the KG, then it is easier to write a scene about them. In each 5-concept-set (i.e. a concept-set consists of five concepts), there are 10 unique pairs of concepts, the connections of which we are interested in. If we look at the one-hop links on the KG, about 60% of the 5-concept-sets have less than one link among all concept-pairs. On the other hand, if we consider two-hop links, then nearly 50% of them are almost fully connected (i.e. each pair of concepts has connections). These two observations together suggest that the CommonGen has a reasonable difficulty: the concepts are not too distant or too close, and reasoning about the associated scenes is thus neither too difficult nor too trivial.

Relation Distribution of Concept-Sets

Furthermore, the relation types of such connections can also tell us what kinds of commonsense knowledge are potentially useful for relational reasoning towards generation. We report the frequency of different relation types of the one/two-hop connections among concept-pairs in the dev and test examples.

In both cases, we find most frequent relation types are about 1) spatial knowledge (e.g. AtLocation, LocatedNear), 2) object properties (e.g. UsedFor, CapableOf, PartOf, ReceiveAction), 3) human behavior and social conventions (e.g. CauseDesire, Motivated), 4) temporal knowledge (e.g. First/Last-SubEvent, Perquisite), and 5) other general commonsense (e.g. RelatedTo, HasContext, IsA).
One/two-hop relation frequency in the \textsc{CommonGen} dev.\&test sets on ConceptNet.



    title = "{C}ommon{G}en: A Constrained Text Generation Challenge for Generative Commonsense Reasoning",
    author = "Lin, Bill Yuchen  and
      Zhou, Wangchunshu  and
      Shen, Ming  and
      Zhou, Pei  and
      Bhagavatula, Chandra  and
      Choi, Yejin  and
      Ren, Xiang",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2020",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.findings-emnlp.165",
    pages = "1823--1840",