You talk what you read: Understanding News Comment Behavior
by Dispositional and Situational Attribution

In this subsection, our goal is to mine dispositional factor from users’ historical comments. The comment is mainly composed of aspect and opinon terms Pontiki et al. (2016), for example, in the sentence Great China!, the aspect term is “China”, and the opinion term is “Great”. Therefore, we model the dispositional factor as the user’s preferences of aspect and opinion.

In this subsection, our goal is to detect situational factor from the news. Since not all sentences contribute equally to motivate users to comment, we introduced an attention-based method using weighted aspect vector $v_{a}$ to measure importance of news sentences with respect to a specific user.

Considering the opinion topic preference, we design a comment decoder to dynamically integrate the opinion vector and the news context vector to generate comments. The formula of the decoder is defined by:

$\displaystyle state_{t}=LSTM_{dec}([c^{X}_{t};e(y_{t}-1);M_{t}],state_{t-1}),$

(21)

where $[\cdot;\cdot]$ denotes vector concatenation. The comment word is then sampled from output distribution based on the concatenation of decoder state as:

$$\tilde{y}_{t}\sim softmax({W_{y}}(state_{t})).$$

(22)

During training, cross-entropy loss $\mathcal{L}_{ce}$ is employed as the optimization objective. The decoder takes the embedding of the previously decoded word $e(y_{t}-1)$ , the context vector $c^{X}_{t}$ and the dynamic opinion state $M_{t}$ as input to update its state. The context vector $c^{X}_{t}$ is a weighted sum of encoder’s sentence representations calculated by:

	$\displaystyle c^{X}_{t}=$	$\displaystyle\sum\nolimits_{i}\alpha_{t,i}s_{i},$		(23)
	$\displaystyle\alpha_{t,i}=$	$\displaystyle\frac{g_{i}\odot e_{t,i}}{\sum_{j}g_{j}\odot e_{t,j}},$		(24)
	$\displaystyle e_{t,i}=$	$\displaystyle softmax(state_{t-1}{{W_{\alpha}}}s_{i}).$		(25)

The dynamic opinion state $M_{t}$ is initialized by the opinion vetor $v_{s}$ which is calculated similar with $v_{a}$(see Eqn. (19)), and decays by a certain amount at each time step. This process is described as:

	$\displaystyle M_{t}=$	$\displaystyle g^{u}_{t}\odot M_{t-1},$		(26)
	$\displaystyle g^{u}_{t}=$	$\displaystyle sigmoid({W_{o}}(state_{t})),$		(27)
	$\displaystyle M_{0}=$	$\displaystyle v_{s},$		(28)
	$\displaystyle\widehat{z}_{s}=$	$\displaystyle softmax({W_{s}}[pf_{s};\sum\nolimits_{i}g_{i}s_{i}]),$		(29)

where $\odot$ denotes element-wise multiplication. Once the decoding process is completed, the opinion state is expressed on context vector completely, and the comment is generated.

The above three modules are jointly trained with the following overall loss function

$\displaystyle\mathcal{L}=\mathcal{L}_{ce}+\lambda_{1}\mathcal{L}_{a}+\lambda_{2}\mathcal{L}_{s},$

(30)

where $\lambda_{1}$, $\lambda_{2}$ are hyperparameters to balance between different modules. Then we jointly train all components according to Eqn. (30). After dispositional and situational comment attribution, we can obtain $\widehat{z}_{a}$, $\widehat{z}_{s}$, $g_{i}$ and the decoder attention $e_{i}(\text{the mean of }e_{t,i})$ to support the following experiments and applications.

Quantitative evaluation results are shown in Table 4. The proposed DS-Attributor outperforms the baselines on all 5 evaluation metrics. This demonstrates the advantage of exploring the dispositional factors in modeling news comment behaviors. In Table 6 and Table 7, we illustrate some example aspect and opinion topics discovered from news interaction history. We can see that aspect topics describe different focuses and interests of users, and opinion topics help understand user sentimental preferences. Among the baseline methods, Hierarchical-Attention generally performs better than Seq2seq and Graph2seq. A possible reason is that Hierarchical-Attention captured and aggregated the key information in the news through hierarchical attention mechanism.

On all 5 evaluation metrics, DS-Attributor achieved superior performance than the two variants, showing the contribution of important sentence measurement and opinion integration. Key observations include: (1) The performance of DS-Attributor(w/o IM) decreased significantly on BLEU and METEOR, which indicates that leveraging weighted aspect vetor $v_{a}$ is beneficial to remove irrelevant information and detect users’ focused aspects of specific news. (2) When opinion vector is removed, DS-Attributor(w/o OV) performs poorly on ROUGE-L and CIDEr, which shows that mining users’ opinion preference does provide prior information to understand the sentimental tendency and thus help accurately predict comment reaction.

In order to better understand how dispositional and situational attribution contribute to the comment behavior, as shown in Table 5, we visualize the generated comments for one specific news regarding the requisition of a university in Wuhan as a medical isolation place. Compared to the baselines only leveraging modeling the situational factors, DS-Attributor can generate appropriate comments according to different users considering the dispositional factors mined from their interaction history. It is shown that the generated comments for three example users from DS-Attributor contain diverse and more clear focuses. Regarding comment-3 which expresses complaint about the decision of dormintory requisition, we examine the corresponding situational and dispositional factors. In particular, for situational factor, we highlight the news sentence with blue color that users focus most on with the highest attention value $g_{i}$. From situational attribution we can see that the user is concerned about dormitories and personal belongings in this news. For dispositional factor, from the aspect and opinion distributions $\widehat{z}_{a}$, $\widehat{z}_{s}$, we find that this user has the highest preference for Aspect#38 and Opinion#4 . As shown in Table 6 and Table 7, Aspect#38 talks about school, and Opinion#4 indicates negative sentiment. This gives rise to the dissatisfaction in the comment and helps detect the user’s actual focus in the news.

In this subsection, we introduce a useful application aggregating the predicted comments to forecast the audience’s focus and opinion for future news. We will use an example of news to illustrate this application which describes Vietnam exposed a large-scale protest in the streets of the people. Specifically, for the given news, 200 users are selected as test subjects to predict their focused aspect distribution $\widehat{z}_{a}$ and corresponding opinion distribution $\widehat{z}_{s}$. For simplicity, we obtain above two topics distribution of 200 users, and analyse the topics with the highest weights respectively.

We observe that most people concentrated on Aspect#39 which talks about social topics(e.g., politics, pretest, etc.). The keywords of generated comments on Aspect#39 are shown in the Figure 2(left), which are closely related to the news content. As for people’s opinions on Aspect#39, we visualize the opinion distribution in the Figure 2(right). Specifically, most people express positive emotions, such as “protect”, “alive” in Opinion#2 and “hope”, “solve” in Opinion#3, and the others express opposition to this matter(e.g., no, disbelief, etc.). Therefore, DS-Attributor provides the possibility to predict people’s reaction before or at the early stage of news release. By examining the users from certain community, By examining the users from a certain community, we can also support fine-grained aspect-opinion forecast. This will enable timely and effective public opinion management.

DS-Attributor derives users’ aspect preferences as by-product, which helps understand the subjective focus on news. Therefore, instead of objective news summarization as most current studies conduct by only analyzing the correlation between news sentences, we can exploit the derived user aspect preference to support a novel subjective news summarization. Specifically, we introduce subjective user factors into the traditional objective news summarization solution, by fusing the score $g_{i}$ (see section 3.2) and the decoder attention $e_{i}$ (see section 3.3) to update the similarity matrix $w(i,j)$ of standard TextRank Mihalcea and Tarau (2004) as

$$w(i,j)=\alpha_{1}w_{s}(i,j)+\alpha_{2}w_{g}(i,j)+\alpha_{3}w_{e}(i,j),$$

(31)

where $w_{s}(i,j)$ is cosine similarity of two sentence vectors, $w_{g}(i,j)$ is defined as

$$w_{g}(i,j)=\frac{g_{j}w_{s}(i,j)}{\sum_{k}g_{k}w_{s}(i,k)},$$

(32)

$w_{e}(i,j)$ is defined similar to $w_{e}(i,j)$, and $\alpha_{1}$, $\alpha_{2}$, $\alpha_{3}$ are coefficients. The final sentence importance score is estimated after performing TextRank. With ROUGE-L as the evaluation metric, we compared three summarization strategies on 100 news: (1) Standard TextRank: to extract top-k sentences as summary without reader factors. (2) Single-user: to randomly select one from 20 users’ top-k results $m$ times, and average the ROUGE-L scores. (3) Multi-user: to randomly select $n$ users’ summary results and choose sentences by voting each time, repeat $m$ times and average the ROUGE-L scores.

The evaluation results of different methods are shown in Figure 3.

From the results, we have the following conclusions: (1) reader-aware summarization strategies show better performance than standard TextRank, because subjective reader factor is useful to extract the news article highlight. (2) the multi-reader strategy achieves superior performance when $k$ is small, which shows that the common interest of multiple readers is beneficial to mining the main purpose of the news. Users’ interests disperse as $k$ increases, where multi-user strategy obtains close performance to single-user strategy but still clearly outperforms the standard TextRank-based solution. Note that this evaluation is conducted with news title as the ground-truth. In practical scenarios, by exploiting the dispositional preference for a specific individual or group of users, we can develop applications like customized and even personalized news summarization.