E-commerce Query-based Generation based on User Review

In order to obtain context vector of the model input, we use three separate encoders for item reviews, review attribute snippets and review ratings. For item reviews $\mathbf{X}$, we first obtain the embedding vector of the input and pass them through a vanilla bi-directional GRU to get the last layer hidden states as the context vector $\mathbf{H}\in\mathbb{R}^{l_{r},l_{s},l_{emb}}$:

$$\mathbf{X}_{emb}=embedding(\mathbf{X})$$

(2)

$$\mathbf{H}=\overrightarrow{GRU(\mathbf{X}_{emb})}+\overleftarrow{GRU(\mathbf{X}_{emb})}$$

(3)

Where $\mathbf{X}_{emb}\in\mathbb{R}^{l_{r},l_{s}}$ stands for the embedding vectors for reviews under the specified item, $l_{r}$ and $l_{s}$ stand for the number of reviews and the length of each review. Similarly for attribute snippet and review rating input, we obtain context vector $\mathbf{S}$ and $\mathbf{U}$. Since attribute snippet and rating contains no positional information, we only obtain their embeddings as their context vector.

Snippet-conditioned Generation. Based on the bi-directional attention method introduced in Seo et al.’s workSeo et al. (2018), we use a bi-directional attention mechanism between item review context vector and snippet context vector to obtain snippet-conditioned review context vector:

$$\mathbf{A}_{i,j}^{H}=\alpha_{1}(\mathbf{H}_{i},\mathbf{S}_{j})$$

(4)

$$\mathbf{a}_{i}^{H}=softmax(\mathbf{A}_{i,:});\mathbf{a}_{j}^{S}=softmax(\mathbf{A}_{:,j})$$

(5)

Where $\mathbf{H}_{i}$ and $\mathbf{S}_{j}$ represent the $i$ word in item review and $j$ word in attribute snippet and $\mathbf{A}_{i,j}^{H}$ is the attention weight matrix between review word $i$ and snippet word $j$ calculated using attention method $\alpha_{1}$. In this paper, we use $\alpha_{1}(x,y)=v_{\alpha_{1}}^{T}[x,y,x\odot y]$, where $v_{\alpha_{1}}$ is a trainable projection vector. By applying softmax normalization, we get attention scores between review word $i$ and each word from the snippets denoted as $\mathbf{a}_{i}^{H}$, and $\mathbf{a}_{j}^{S}$ vice versa. After this, we compute the weighted context vector by aggregating the dot product of the attention score and input context vector of both review and snippets:

$$\tilde{\mathbf{H}}_{i}=\sum_{j}{\mathbf{a}_{i,j}^{H}\mathbf{H}_{i}};\tilde{\mathbf{S}}_{j}=\sum_{i}{\mathbf{a}_{j,i}^{S}\mathbf{S}_{j}}$$

(6)

$$\tilde{\mathbf{H}}_{S}=\alpha_{2}(\mathbf{H},\tilde{\mathbf{H}},\tilde{\mathbf{S}})\odot\mathbf{H}$$

(7)

Here $\mathbf{a}_{i,j}^{H}$ denotes the $j$ element in $\mathbf{a}_{i}^{H}$ and $\mathbf{a}_{j,i}^{H}$ denotes the $i$ element in $\mathbf{a}_{j}^{H}$. $\tilde{\mathbf{H}}_{S}$ denotes the weighted context for item review text, and similarly to equation 1 $\alpha_{2}$ denotes the attention method. Here we use $\alpha_{2}(x,y,z)=v_{\alpha_{2}}^{T}[x;y;x\odot y;x\odot z]$ to emphasize on the original input context.

Rating-conditioned Generation. To enforce rating condition over generation, we perform attention weighting between query input rating embedding $\mathbf{r}$ and each input item review rating embedding $\mathbf{R}$:

$$\tilde{\mathbf{R}}_{t}=\beta_{1}(\mathbf{r},\mathbf{R}_{t})\odot\mathbf{R}_{t}$$

(8)

Where $\tilde{\mathbf{R}}_{t}$ denotes weighted rating context for item review $t$, and we use attention method $\beta_{1}(x,y)=v_{\beta_{1}}^{T}(x\odot y)$ to emphasize the rating similarity between reviews.

We use bi-directional GRU for decoding. To ensure the generated output is relevant to the input, we calcualte attention using the context vector we obtained above and the decoder hidden states $h$ and project the representation onto the vocabulary dimension:

$$\mathbf{a}_{t}^{H}=softmax(\mathbf{W}_{H}^{T}[\tilde{\mathbf{H}}_{t};h]+\mathbf{b}_{H})$$

(9)

$$\mathbf{a}_{t}^{R}=softmax(\mathbf{W}_{R}^{T}[\tilde{\mathbf{R}}_{t};h]+\mathbf{b}_{R})$$

(10)

$$P(w_{t})=softmax(\mathbf{W}^{T}[h;\mathbf{a}_{t}^{H}\odot\tilde{\mathbf{H}}_{t};\mathbf{a}_{t}^{R}\odot\tilde{\mathbf{R}}_{t}]+\mathbf{b})$$

(11)

Where the $P(w_{t})$ denotes the probability of word $w_{t}$ appearing at this decoding step, and $\mathbf{W}$ and $\mathbf{b}$ denote the representation of a dense layer.

We use negative log likelihood as the generation loss:

$$L_{gen}(\theta_{g})=-\sum_{i}^{N}w_{t}log\hat{w_{t}}$$

(12)

Where $L_{gen}(\theta_{g})$ is the loss for model parameters $\theta_{g}$ at each decoding step.

To further ensure that the generated answers is consistent to the input rating, we employ an auxiliary classifier that takes in the generated answer and computes the cross entropy loss between the predicted and the groundtruth rating. To ensure that the classifier is strong enough, we pre-train it on all training sentences and formulate the target as a sentiment analysis task.

And to handle cases when user simply wants a overall review by not specifying the input rating, we add an additional ⟨PAD⟩ token during training that signals the generator to generate an overall review. In these circumstances, the classification cross entropy loss will not be computed and optimized (since the classifier only has ratings from one to five but not the ⟨PAD⟩ token).

The classification loss is

$$L_{cls}(\theta_{c})=-\sum_{c=1}^{5}{y_{c}\log{p_{c}}}$$

(13)

where $y_{c}$ is the groundtruth probability (either 0 or 1) for rating $c$, $p_{c}$ is the prediction probability for rating $c$, and $\theta_{c}$ is the classifier parameters at each timestamp.

We combine the classification loss with the generation loss as the final loss function of our model:

$$L(\theta_{g},\theta_{c})=-\lambda\sum_{i}^{N}w_{t}log\hat{w_{t}}-(1-\lambda)\sum_{c=1}^{5}{y_{c}\log{p_{c}}}$$

(14)

Where $\lambda\in[0,1]$ is a hyper-parameter.

We use the Amazon Review Data¹¹1https://nijianmo.github.io/amazon/index.html to train and validate our model performance in review generation. We use data under Amazon Fashion subcategory and filter out items with less than 20 reviews to ensure robustness of data. For review attribute snippet extraction, we use the noun phrase extraction method introduced by Pan et al. Pan et al. (2017).

Our model is implemented using Pytorch²²2https://www.pytorch.org. Data tokenization is conducted using NLTK³³3https://www.nltk.org. We set the hidden size of both the encoder and decoder to 512 and layer number to 1. For optimizer we use Adam with a learning rate of 0.0002. The embedding size is set to 512. We also apply a gradient clipping of $[-5,5]$ to prevent exploding gradient. The maximum input review number is set to 20 and the maximum length of review and attribute snippet is also set to 20. We truncate the review and words exceeding the maximum limit. We use beam search with a beam size of 5 during decoding, and the decoding maximum length is set to 15.

We compare with several baselines and previous work of similar target.

Random. It randomly selects one review from all reviews of the target product.

NN-rating. A Nearest Neighbor based method that first selects all reviews of the query product. Then, if user provides a target rating, it randomly selects one of review that has the same rating. Otherwise, it randomly selects one amongst all of them.

Seq2seq. A sequence-to-sequence model with attention mechanism Luong et al. (2015a). This model, as ours, takes input all reviews but encodes them without snippet and rating information. The decoder attends to the rating, the snippets and the encoder outputs to produce context vector during decode.

Dong. Model by Dong et al. Dong et al. (2017). The model takes input the productID and rating (not the userID as in the original paper due to a user might not be in the training set). The model is then trained to generate review as similar as possible to the original review.

Ni. Reference-based model with aspect-planning proposed by Ni et al. Ni et al. (2019) on recommendation justification that predicts what review a specific user would give to a specific item. It is based on historic reviews of the user, reviews of the item and fine-grained aspects of the item. We replace the fine-grained aspects input by the extracted snippets. Again, since we target at QA system and has no access to historic user review, we remove the historic user review input module. Hence, the final input would be the snippets along with the item reviews.

For evaluation, we use BLEUPapineni et al. (2002), METEORBanerjee and Lavie (2005) and ROUGE-LLin (2004) scores as our evaluation metrics.

As shown in table 1, our model achieves the best performance on BLEU-4, METEOR and ROUGE-L scores. This shows an improvement in both diversity and precision of generation using our proposed method. We also observe a drop of performance by our model in lower dimension phrase matching based criterion including BLEU-1 to 3. This might be explained as a compromise in establishing a language model with higher diversity, which can be further testified with the higher METEOR and ROUGE-L score.