Modeling Event Background for If-Then Commonsense Reasoning Using Context-aware Variational Autoencoder

Understanding event and event-centered commonsense reasoning are crucial for natural language processing (NLP). Given an observed event, it is trivial for human to infer its intents and effects, while this type of If-Then reasoning still remains challenging for NLP systems. To facilitate this, a If-Then commonsense reasoning dataset Atomic is proposed, together with an RNN-based Seq2Seq model to conduct such reasoning. However, two fundamental problems still need to be addressed: first, the intents of an event may be multiple, while the generations of RNN-based Seq2Seq models are always semantically close; second, external knowledge of the event background may be necessary for understanding events and conducting the If-Then reasoning. To address these issues, we propose a novel context-aware variational autoencoder effectively learning event background information to guide the If-Then reasoning. Experimental results show that our approach improves the accuracy and diversity of inferences compared with state-of-the-art baseline methods.


Introduction
Recently, event-centered commonsense knowledge has attracted much attention (Chambers and Jurafsky, 2008;Segers et al., 2016;Wang et al., 2017;Li et al., 2018), because of understanding events is an important component of NLP.Given a daily-life event, human can easily understand it and reason about its causes, effects, and so on.However, it still remains a challenging task for NLP systems.This is partly due to most of them are trained for task-specific datasets or objectives, which results in models that are adapt at finding task-specific underlying correlation patterns but have limited capability in simple and explainable commonsense reasoning (Sap et al., 2018).* To facilitate this, Rashkin et al. ( 2018) build the Event2Mind dataset and Sap et al. (2018) present the Atomic dataset, mainly focus on nine If-Then reasoning types to describe causes, effects, intents and participant characteristic about events.Together with these datasets, a simple RNN-based encoder-decoder framework is proposed to conduct the If-Then reasoning.

Corresponding author
However, there still remains two challenging problems.First, as illustrated in Figure 1, given an event "PersonX finds a new job", the plausible feeling of PersonX about that event could be multiple (such as "needy/stressed out" and "relieved/joyful").Previous work showed that for the one-to-many problem, conventional RNN-based encoder-decoder models tend to generate generic responses, rather than meaningful and specific answers (Li et al., 2016;Serban et al., 2016).
Second, as a commonsense reasoning problem, rich background knowledge is necessary for generating reasonable inferences.For example, as shown in Figure 1, the feeling of PersonX upon the event "PersonX finds a new job" could be multiple.However, after given a context "PersonX was fired", the plausible inferences would be narrowed down to "needy" or "stressed out".
To better solve these problems, we propose a context-aware variational autoencoder (CWVAE) together with a two-stage training procedure.Vari-ational Autoencoder (VAE) based models have shown great potential in modeling the one-tomany problem and generate diversified inferences (Bowman et al., 2015;Zhao et al., 2017).
In addition to the traditional VAE structure, we introduces an extra context-aware latent variable in CWVAE to learn the event background knowledge.In the pretrain stage, CWVAE is trained on an auxiliary dataset (consists of three narrative story corpora and contains rich event background knowledge), to learn the event background information by using the context-aware latent variable.Subsequently, in the finetune stage, CWVAE is trained on the task-specific dataset to adapt the event background information to each specific aspect of If-Then inferential target (e.g., intents, reactions, etc.).
Experiments on the Event2Mind and Atomic dataset show that our proposed approach outperforms baseline methods in both the accuracy and diversity of inferences.The code is released at https://github.com/sjcfr/CWVAE.

Background
Before specifically describing two dataset --Event2Mind and Atomic used in this paper as well as the If-Then reasoning task, for clarity, we define the following terminologies: Base event: the prerequisite event in If-Then reasoning, organized as a verb phrase with a predicate and its arguments, such as the event "PersonX finds a new job" shown in Figure 1.
Inference dimension: a particular If-Then reasoning type, e.g., intents, effects of the base event.Details are shown in Table 1 and Table 2. Target: the inferential results.For example, as shown in Figure 1, given a base event "Per-sonX finds a new job" and one inference dimension "xReact", the targets could be "relieved" or "needy".Notice that each inference dimension can have multiple targets.Event2Mind Dataset contains 25K base events and 300K targets, annotated through crowdsourcing.Event2Mind is organized in a hierarchical form: each base event has three types of inference dimensions, and given a base event, under one of inference dimensions, several targets may simultaneously exist.Atomic dataset shares the same hierarchical structure as Event2Mind, while scales up the size of dataset and expands the scope to nine types of inference dimensions.ditional variational autoencoder (CVAE) (Sohn et al., 2015) is an extension of VAE on the conditional generation problem.As shown in Figure 2 (a), CVAE characterizes the conditional oneto-many generation problem using three random variables: event x, target y and a latent variable z, which is used for modeling the latent distribution of semantic over targets given an event.
Hence, under a certain inference dimension, with regard to the latent semantic variable z, the conditional generation problem could be expressed as p(y|x) = p(y|x, z)p(z|x)dz.CVAE models p(y|x, z) and p(z|x) using deep neural networks (parameterized by θ) p θ (y|x, z) and p θ (z|x).Then as illustrated in Figure 2 (b), y could be generated from x and z.
CVAE is trained to maximize the conditional likelihood p(y|x), which involves an intractable marginalization over the latent variable z.Instead, following Kingma and Welling (2014), a practical way is to introduce another deep network (parameterized by φ) q φ (z|x, y) to approximate the true posterior distribution p(z|x, y) and maximize the evidence lower bound (ELBO) of the loglikelihood function: Therefore, CVAE is composed of three neural networks in general.We refer to p θ (z|x) as a prior network, q φ (z|x, y) as a recognition network, and p θ (y|x, z) as a neural decoder.

Context-aware Variational Autoencoder
Traditional CVAE can model the event-target relation.In other words, given an observed event, CVAE can generate its corresponding targets.While in this paper we model the If-Then reasoning as a [(background), event]-target process.It means that in addition to the observed event, we also want to involve the event background knowledge (which can be learned from event contexts) to generate the reasonable targets.
To this end, we propose a context-aware variational autoencoder (CWVAE), with two additional latent variables: a context-acquiring latent variable z c to directly acquire context information, and a context-aware latent variable z c to learn background knowledge from z c , as shown in Figure 3 (a).However, the event context information is absent in the Event2Mind and Atomic dataset.To learn from the external event context information, we design the following two-stage training procedure for CWVAE.
Pretrain: Learning Event Background Knowledge from Auxiliary Dataset In the pretrain stage, CWVAE is trained on three narrative story corpora with rich event context information.As shown in Figure 3 (a), context-acquiring latent variable z c is directly conditioned on the context c.Hence, z c could be employed for acquiring background knowledge from event contexts.Then, we minimize the distance between z c and the contextaware latent variable z c , by which the event background knowledge is transferred from z c to z c .

Architecture of CWVAE
As shown in Figure 4, CWVAE is mainly composed of four parts: a neural encoder that provides distributed representations of base events/targets, a recognition network for inferring q φ (z|x, y), q φ (z c |x, c) and q φ (z|z c , x), a prior network for modeling p θ (z c |x) and p θ (z|x, z c ), and a neural decoder that integrates the information from z and z c to generate targets.Neural Encoder We employ a bidirectional GRU as neural encoder, which encodes context c, event x and target y into distributed representations h c = {h c 1 , . . ., h c lc }, h x = {h x 1 , . . ., h x lx } and h y = {h y 1 , . . ., h y ly }, where l c , l x and l y is the length of c, x and y, respectively.Recognition Network The recognition network models q φ (z|x, y), q φ (z c |x, c), q φ (z|z c , x) based on h x , h y and h c .
Following traditional VAE, the abovementioned three distributions are assumed to be multivariate Gaussian distribution with a diagonal covariance structure: where µ denotes the mean of the distribution, σ denotes the standard deviation of the distribution, and I denotes the identity matrix.
Given h x , h y and h c , we propose a novel attention-based inferer (ABI) module to estimate the mean and standard deviation of q φ (z c |x, c), q φ (z c |x, y) and q φ (z|x, y): Briefly, through the attention mechanism, ABI can capture the semantic interaction between input sequences, and estimate the parameters of distributions based on it.We will introduce the specific structure of ABI in below.Prior Network Prior Network models p θ (z c |x) and p θ (z|x, z c ) based on h x .The distribution of p θ (z c |x) and p θ (z|x, z c ) are still assumed to be multivariate Gaussian, whereas the parameters are different: where µ denotes the mean of the distribution, σ denotes the standard deviation of the distribution and I denotes the identity matrix.
Then the attention-based inferer module is still employed to estimate parameters of distributions: Neural Decoder Given the base event x, the semantic latent variable z, and the context-aware latent variable z c , the neural decoder defines the generation probability of y as following: where p(y j |y < j, z, z c , x) = g(y j−1 , s j−1 , e j ), g(•) is an attention-based feed forward model, e j = i α ji h x i is the context vector and s j−1 is the hidden state of the decoder.We obtain g(•) and e j the same way as Bahdanau et al. (2014).Whereas our decoder differs from Bahdanau et al. (2014) in that our model integrates the context-aware latent variable z c and semantic latent variable z in the computation of , where E yj is the word embeddings of target words.
Note that through concatenating z and z c with E yj and s j−1 , s j could be affected by contextaware latent variable z c and semantic latent variable z.This allows model to directly access to the event background knowledge from z c .In addition, the randomness of z and z c would increase the diversity of model generation.Attention-based Inferer Attention mechanism has shown strong ability in capturing semantic interactions (Gong et al., 2017).Inspired by the coattention mechanism (Parikh et al., 2016), we propose an attention-based inferer (ABI) to estimate the mean and standard deviation of a distribution belongs to p θ (•) or q φ (•) by capturing semantic interactions of input sequences.
Specifically, given two input sequences (e.g., representations of contexts and events) a = {a 1 , . . ., a la } and b = {b 1 , . . ., b l b } with length l a and l b , we first obtain the attention scores from each side through: where W a ∈ R d×da and W b ∈ R d×d b are parameter weights.With these attention scores, the context vectors of both sequences are given by: Then we perform a mean pooling operation on context vectors of both sequences: To obtain the mean and standard deviation, the pooled context vectors ca and cb which carry semantic interaction between two sequences, are concatenated and projected into a latent semantic space through a nonlinear transformation: Finally the mean and standard deviation are generated through a nonlinear transformation over h z :

Optimizing
With the incorporation of z c , the original loglikelihood could be decomposed as: Then following traditional CVAE, the ELBO of CWVAE is defined as follows: which is the objective function at the finetune stage.
While in the pretrain stage, as we aim to learn background knowledge through minimizing the distance between z c and z c , in addition to L ELBO , a context-aware regulation term is introduced: where the context aware regularization term is the KL distance between z and z c .Through minimizing the context aware regularization term, we aim to pass event context knowledge from z c to the context aware latent variable z c .

Training Details
To test the performance of CWVAE, we split the Event2Mind and Atomic dataset into training, development and test sets (80%, 10%, 10%) in the same way as Rashkin et al. ( 2018) and Sap et al. (2018), respectively.
We initialize the embedding layer from 300d GloVe word embeddings.The neural encoder is chosen to be biGRU with 300 hidden units.For the ABI module, size of W a and W b is set to be 100 × d a and 100 × d b respectively.The dimension of z c , z c and z is all set as 40.The neural decoder is set to be GRU with 300d hidden state.Regulation coefficient λ of context-aware regulation term is set to be 0.1.Models are trained using an Adam optimizer (Kinga and Adam, 2015) with a learning rate of 0.001.

Auxiliary Dataset
The auxiliary dataset is built upon three humanwritten story corpora: ROCStories (Mostafazadeh

Event
Inference Target 1 jason had been really stressed out at work. 2 he decided he needed a different kind of job. 3 jason applied for a job in a different field.4 he got the job .
5 jason was much happier at his new job .Table 3: An example for the construction of auxiliary dataset.For a five-sentence-paragraph, the first three sentences are taken as event context, while the fourth and fifth sentence is taken as base event and target respectively.
et al., 2016), VIST (Huang et al., 2016) and Writ-ingPrompts (Fan et al., 2018).ROCStories and VIST are composed of short stories with five sentences.We filter out stories of more than 1,000 words in WritingPrompts, and cut the remaining stories into five-sentence-paragraphs.
For each five-sentence-paragraph, we define the first three sentences as contexts of the base event, the fourth sentence as the base event, and the fifth sentence as the inference target.For example, as shown in Table 3, the first three sentences describe a context that Jason was unsatisfied about his job and applied for a new job.Hence, after happening the event "he got the job", a plausible react about the event could be "jason was much happier at his new job".In total, the auxiliary dataset contains 192,316 (context, event, target) triples.

Baselines
We compared our proposed model with the following four baseline methods: • RNN-based Seq2Seq proposed by Sap et al. (2018) for the If-Then reasoning on Atomic.Note that, for each baseline method, we train distinct models for each distinct inference dimension, respectively.

Evaluation Metrics
Automatic Evaluation We first compare the perplexity of CWVAE with baseline methods.Perplexity measures the probability of model to regenerate the exact targets, which is particular suit-  able for evaluating the model performance on oneto-many problem (Serban et al., 2017).Further, we employ BLEU score to evaluate the accuracy of generations (Papineni et al., 2002), and the number of distinct n-gram to evaluate the diversity of generations (Li et al., 2016).The distinct is normalized to [0, 1] by dividing the total number of generated tokens.
Human Evaluation Since automatic evaluation of generations is still a challenging task (Liu et al., 2016), we also conduct human evaluations on the model performance.Five human experts are employed to evaluate the coherence, diversity and fluency of generated targets.Experts are asked to vote for if a generation is fluent or coherent for each generated target, and give a 1-5 score for the diversity of generations.For both Event2Mind and Atomic datasets, 100 events are randomly selected from the test set.For each method, top 10 generated targets of each base event are used for eval- uation.Finally we report three overall averaged scores of coherence, diversity and fluency on both datasets, respectively.

Overall Results
We list the perplexity and BLEU score of CWVAE and baseline methods on Event2Mind and Atomic in Table 4 and Table 6, respectively, and show the distinct-1 and distinct-2 score on Event2Mind and Atomic in Table 5 and Table 7, respectively.We find that: (1) As shown in Table 5 and Table 7, comparison between RNN-based Seq2Seq and variationalbased methods, including Variational Seq2Seq, VRNMT, CWVAE-unpretrained and CWVAE shows that, variational-based methods could increase the diversity of generations.This confirms one of our motivations that variational-based methods could capture the latent semantic distribution within targets and increase the diversity of If-Then reasoning.
(2) Comparing CWVAE-unpretrained with other baseline methods shows that, in general CW-VAE improves the accuracy and diversity on both dataset.These results indicate the efficiency of CWVAE in capturing the latent semantic distribution of targets, and generate more reasonable inferential results.
(3) Comparison between CWVAE and CWVAE-unpretrained shows that the pre-   train stage could enhance the performance of CWVAE in both the accuracy and diversity.This is mainly because event knowledge could offer the guidance for If-Then reasoning.In the pretrain stage, CWVAE could capture the event background knowledge through context-aware latent variable, and such knowledge could be be adapted to our task through the fintune stage.
To further evaluate the effectiveness of our proposed approach, we also conduct human evaluations, the results of which are shown in Table 8 and Table 9 on coherence and fluency.The main reasons are twofold: first, the CWVAE has advantage in capturing the semantic distribution of targets; second, event background learned from the pretrain stage is helpful for the If-Then reasoning.

Case Study
Table 10 provides an example model generations given the base event "PersonX works tirelessly" and the inference dimension "xIntent".The generations under CWVAE mainly contain four kinds of semantics: (1) be productive, (2) finish his work soon, (3) accomplish goal, (4) earn more money.While the semantics of generations using baseline RNN-based Seq2Seq model is relatively limited.Furthermore, the first three kinds of semantic overlap the three ground truth targets, and the fourth kind of semantic is in accordance with daily-life commonsense.Compared to RNNbased Seq2Seq model, our approach can increase the diversity and rationality of generations, meanwhile keep the accuracy.
5 Related Work

Event-Centered Commonsense Reasoning
Understanding events and constructing eventcentered commonsense knowledge are crucial to many NLP applications, such as intention recognition (Goldwasser and Zhang, 2016) and dialog generation (Wen et al., 2017).Recently a growing number of studies focus on event-centered commonsense reasoning, which mainly concentrates on two areas, script event prediction and story ending generation/choosing. Script event prediction concerns with the temporal relationships between script events (Granroth-Wilding and Clark, 2016), which requires models to choose a correct subsequent triple-organized event among the candidates (Wang et al., 2017).Prior work mainly focused on modeling event pairs (Granroth-Wilding and Clark, 2016), event chains (Wang et al., 2017) and event graph (Li et al., 2018) to predict the subse-quent event.Story ending generation focuses on generating plausible story endings (Mostafazadeh et al., 2016), which requires models to understand the story context, and keep generated endings logically consistent with it (Peng et al., 2017;Guan et al., 2019).The above tasks mainly investigate the logical orders of events, whereas the If-Then reasoning task focuses on inferring the mental state of event participants.

Variational AutoEncoder-Decoder Based
Natural Language Generation VAE (Kingma and Welling, 2014) has been widely applied in various of text generation tasks, such as dialogue and machine translation.In dialogue generation, Zhao et al. ( 2017) adapts VAE with encoder-decoder framework to model the latent semantic distribution of answers, which can increase the diversity of generations.For the task of machine translation, Su et al. (2018) and Zhang et al. (2016) employ a latent variable to capture the semantic interaction between the source and target sentence, and regard the latent variable as a supplementation of attention mechanism.While Wang et al. (2019) use the latent variable to model topic distributions in text generation.In this paper, we introduce an additional context-aware latent variable to effectively learn background knowledge and conduct If-Then reasoning on the guidance of it.

Conclusion
In this paper, we propose a novel context-aware VAE (CWVAE) framework with two training stages for If-Then commonsense reasoning.By introducing an additional context-aware latent variable, CWVAE is able to learn external background knowledge, and conduct If-Then reasoning under its guidance.In the pretrain stage, CWVAE learns event background knowledge, then in the finetune stage CWVAE adapts such knowledge to each inference dimension.Experimental results demonstrate that CWVAE outperforms baseline methods in both the accuracy and diversity of generations.

Figure 1 :
Figure 1: A illustration of two challenging problems in If-Then reasoning.(a) Given an observed event, the feelings about this event could be multiple.(b) Background knowledge is need for generating reasonable inferences, which is absent in the dataset (marked by dashed lines).

Figure 2 :
Figure 2: Illustration of inference and generation process of CVAE in a directed graph.Dashed lines represent the inference of z.Solid lines represent the generation process.

Figure 3 :
Figure 3: Illustration of pretrain, finetune and generation process of CWVAE in a directed graph.Dashed lines represent the inference of z, zc and z c .Solid lines represent the generation process.Red circle denotes the context-aware latent variable.

Finetune:
Adapt Event Background Knowledge to Each Inference Dimension In the finetune stage, as shown in Figure 3 (b), CWVAE is trained on the Event2Mind and Atomic dataset without the event context information.Pretrained CWVAE is finetuned to learn the specific inferential knowledge of each inference dimension.After the training procedure, as shown in Figure 3 (c), samples of z is generated based on x and samples of z c , where z c contains rich event background knowledge helpful for If-Then reasoning.

Figure 4 :
Figure 4: Architecture of CWVAE.We mark Neural encoder in green, prior network in blue, recognition network in brown and neural decoder in orange, respectively.

Table 2 :
Hierarchical structure of Atomic dataset.For spe-

Table 5 :
Distinct-1 and distinct-2 scores for the top 10 generations under each inference dimension of Event2Mind.The the best result for each dimension is emboldened.

Table 6 :
Average perplexity and BLEU scores (reported in percentages) for the top 10 generations under each inference dimension of Atomic.The the best result for each dimension is emboldened.

Table 7 :
Distinct-1 and distinct-2 scores for the top 10 generations under each inference dimension of Atomic.The the best result for each dimension is emboldened.

Table 8 :
Human evaluation results on Event2Mind.

Table 9 :
Human evaluation results on Atomic.

Table 10 :
. On both datasets, CWVAEbased methods achieve consistent better coherence, diversity and fluency performances.While comparing with CWVAE-Unpretrained, the pretrain procedure could improves the performance An example of inferences made by CWVAE and RNN-based Seq2Seq model under inference dimension "xIntent".