INS: An Interactive Chinese News Synthesis System

TL;DR

This paper introduces INS, an interactive system that automatically or interactively synthesizes news articles into overview summaries, aiding users and editors in efficiently understanding and creating news content.

Contribution

The paper presents a novel interactive news synthesis system that combines automatic generation with user interaction, enhancing news overview creation and editor assistance.

Findings

INS performs well in topic representation and synthesis quality.

User study shows high satisfaction and usefulness of INS.

System demonstrates effective news synthesis capabilities.

Abstract

Nowadays, we are surrounded by more and more online news articles. Tens or hundreds of news articles need to be read if we wish to explore a hot news event or topic. So it is of vital importance to automatically synthesize a batch of news articles related to the event or topic into a new synthesis article (or overview article) for reader's convenience. It is so challenging to make news synthesis fully automatic that there is no successful solution by now. In this paper, we put forward a novel Interactive News Synthesis system (i.e. INS), which can help generate news overview articles automatically or by interacting with users. More importantly, INS can serve as a tool for editors to help them finish their jobs. In our experiments, INS performs well on both topic representation and synthesis article generation. A user study also demonstrates the usefulness and users' satisfaction with the INS tool. A demo video is available at \url{https://youtu.be/7ItteKW3GEk}.

Tables

Table 1. Table 1: Manual evaluation results

Method	Read.	Struc.	Topic.	Redun.	Overall
Lead-sen	4.214	3.500	4.107	3.964	3.964
Coverage-sen	3.250	2.714	4.179	3.179	3.107
Centroid-sen	4.179	3.714	3.964	4.250	4.036
TextRank-sen	3.929	3.464	3.929	3.929	3.786
Centroid-blk	5.036	4.857	4.179	4.571	4.929
TextRank-blk	4.893	4.536	4.393	4.643	4.679
INS	5.321	5.179	5.286	5.214	5.179

Full text

1 INTRODUCTION

In the last decade, news websites and apps become more and more popular, which can provide us an extremely large volume of news articles. Even for a single news event, there are usually tens or hundreds of related news articles published online. In order to have a complete image of a news event, we have to look through all of the related news articles, which is very time-consuming and inefficient. With a large amount of time spent, what we get is just fragmented information scattered in different news articles.

Ideally, if there exists an overview article about an event, we will fully and efficiently understand the event by reading it. However, such overview articles are not easy to write, even for professional editors, and the writing of them is very time-consuming. Though existing multi-document summarization systems can produce short summaries by selecting several representative sentences, they cannot produce long overview articles with good structure and high quality. So it is of vital importance to design a system to help editors or users to efficiently synthesize a batch of related news articles into a long news overview article, either in a fully automatic way or in a semi-automatic way.

To achieve the above goal, we put forward a novel Interactive News Synthesis system (i.e. INS), which can help generate Chinese news overview articles automatically or by interacting with users. Given a news event or topic, INS first crawls news articles from major Chinese news websites, detects different subtopics and represents them with easy-to-understand labels. Afterward, a span of text for each subtopic will be generated and the news synthesis article will be organized accordingly. It is noteworthy that INS can interact with users in different stages.

We automatically evaluate the key component of the INS system, i.e., subtopic detection and representation, and evaluation results demonstrate its efficacy. Human evaluation is employed and a user study is performed to demonstrate the usefulness and users’ satisfaction of the INS tool.

2 Related Work

One of the related fields is document summarization. The methods can be divided into extractive methods Gillick and Favre (2009); Lin and Bilmes (2010); Berg-Kirkpatrick et al. (2011); Sipos et al. (2012); Woodsend and Lapata (2012); Wan and Zhang (2014); Nallapati et al. (2017); Ren et al. (2017) and abstractive methods Rush et al. (2015); Nallapati et al. (2016); Tan et al. (2017).

There are several pilot studies on producing long articles from a batch of news articles or web pagesYao et al. (2011); Zhang and Wan (2017); Liu et al. (2018). However, the generated overview articles do not have good structures and there are no interaction functions.

There are some attempts of adding interaction functions into the traditional document summarization tasks Jones et al. (2002); Leuski et al. (2003). However, the above work focuses on producing short summaries and the generation of long news overview articles is more challenging. Moreover, in the above work, the keyphrases to represent salient information are extracted based on some heuristic rules or simple clues, and they are usually not good subtopic representations.

3 System Overview and User Interaction

Our INS system aims to produce a long overview article for a specific news event by synthesizing a number of existing Chinese news articles. It consists of the following components:

News Fetcher: According to the input news event, INS first crawls relevant news articles by using a popular Chinese news search engine (e.g. Baidu News), and extracts the title and body text for each news article. The number of news articles can be set by users and it is set to 100 by default.

Subtopic Finder: This component aims to discover the subtopics in the news articles and represent them with some informative labels. We extract n-grams that meet some demands as candidate labels and leverage a regression model to predict a score for each candidate label. Top 20 labels will be chosen and merged, after which each label represents a specific subtopic.

Article Synthesizer: This component aims to produce a span of text with moderate length for each subtopic, and then assemble the texts of selected subtopics to form the final news overview article. We first split the original news articles into coherent text blocks. The text blocks are then matched with the subtopic label and ranked by using the topic-sensitive TextRank algorithm and the MMR redundancy removal method. We select one or several top-ranked text blocks to describe each subtopic. Finally, all the selected text blocks are assembled to form the final overview article, with the subtopic labels used as the subtitles of the blocks.

The graphic user interface of INS is shown in Figure 1 with the input topic of ‘俄罗斯世界杯/2018 FIFA World Cup Russia’.

INS can interact with users in different stages: Users can choose and re-order some of the labels according to their preferences; Users can choose and edit the text blocks for each chosen subtopic; Users can edit on the final news synthesis result to get a more excellent article. Note that the interactions are optional. For ‘lazy’ users, INS can generate the final synthesis article with just one click.

4 Subtopic Finder

There are some existing unsupervised methods for detecting subtopics from documents, for example, text clustering and topic models. They have some drawbacks in common: (1) The number of subtopics (clusters) needs to be manually set beforehand and inappropriate subtopic number will affect the final results significantly. (2) It is hard to represent each subtopic, which is very important for users’ understanding and selection of the subtopics.

Different from the above methods which first detect the subtopics or clusters and then extract labels for each subtopic, we decide to directly find subtopic labels by using supervised learning, and the labels are informative and easy to understand. We first extract candidate labels from the news articles and then use a regression model to assign a rating score to each candidate. After that, we merge the top labels to obtain the final subtopic labels.

Candidate Label Extraction: We extract n-grams from the original news articles as our candidate labels, where n ranges from 1 to 3. We employ pyltp for Chinese word segmentation and POS tagging. An n-gram will be a candidate if it meets the following requirements: (1) Its term frequency is higher than a min-count threshold. We set min-count to 25 for unigrams and 10 for bigrams and trigrams. (2) It is not a substring of the input topic name, which is too general to be a subtopic label. (3) It does not include time words and adverbs. (4) A candidate label of unigram is limited to only nouns and verbs. We use these rules to filter out the n-grams that are not suitable for representing subtopics.

Label Score Prediction: We adopt regression models to predict a score for each candidate label. We choose 12 features to describe each label and each feature is expected to indicate whether a candidate is good or not from some aspect. These features include Term Frequency-Inverse Document Frequency, Document Frequency, Number of words in the label, Number of Chinese characters in the label, Intra-Cluster Similarity, Cluster Entropy, Independence Entropy, Frequency of the label in news titles, Syntactic continuity, Number of nouns in the label, and topic model information.

In the training data, each candidate label is manually assigned with a score of 0~3, and a higher score indicates a better label. We then train regression models for label score prediction. We explore and compare different regression models, and finally choose support vector regression (SVR).

Label Merging: After label score prediction, the candidate labels are sorted by their predicted scores and the top 20 labels are reserved. We will merge similar labels according to the following rules: (1) If two n-grams share a common part which is more than one word, we will merge them into one label. (2) If one label is a substring of another, we only reserve the label with a higher score. (3) If two labels’ cosine similarity is greater than a threshold (0.65 in this study), we only reserve the label with a higher score. After merged, labels are shown to users and each label stands for a subtopic.

5 Article Synthesizer

In this step, we need to produce a span of text with moderate length to describe each subtopic and then combine all the texts into the final synthesis article. We first segment the news articles into coherent text blocks and then rank the relevant text blocks for each subtopic. One or more salient text blocks can be chosen to describe the corresponding subtopic.

Text Segmentation: Most of the methods for document summarization use sentences as the basic unit, which is not appropriate for long article generation. Synthesis articles using the sentence as the basic unit tend to be fragmented and hard to read. As a result, we use the text block as our basic unit. A text block is a set of several continuous sentences and it can cover a relative complete idea. We use our proposed SenTiling algorithm Zhang and Wan (2017) to segment texts, which is a variant of Hearst’s TextTiling algorithm Hearst (1997). After text segmentation, news articles are divided into text blocks, which include 2.3 sentences on average.

Text Block Ranking and Selection: For each subtopic, INS system first selects candidate text blocks using exact match. A text block that contains a subtopic label is assigned to the subtopic. Then INS uses a topic sensitive TextRank algorithm to rank candidate text blocks for each subtopic where TextRank(Mihalcea and Tarau (2004)) is a typical graph-based ranking algorithm applied in document summarization. We build a graph with the candidate text blocks as vertexes and the similarity between text blocks as the weight of the edge. We can determine the importance $WS(V_{i})$ of every vertex through a random walk with a restart. In our case, the restart probability of every vertex is set to the normalized similarity between each text block and the subtopic label so that the top text blocks are both important and relevant to the subtopic.

After ranking, the top text blocks may be similar and this leads to redundancy. We try to solve this problem by using the MMR criterion Carbonell and Goldstein (1998).

Now we have chosen the text blocks to describe each subtopic. Then we rearrange the text blocks according to the following strategies: (1) If two text blocks are extracted from one article, we sort them by the original order. (2) If two text blocks are extracted from different articles, we put the text blocks written earlier in the front. Because news event usually changes over time, this strategy can partly reconstruct the event process.

Synthesis Article Construction: Now we have labels for different subtopics and the corresponding text for each subtopic. INS then constructs the final news synthesis article. If users choose and rank the labels, INS will use their preferred order. If users choose and edit the text blocks, INS will use them instead of the default ones. Otherwise, INS chooses several labels with the highest scores and use the highly ranked text blocks of the subtopic labels for news synthesis. In this way, INS produces the final news synthesis article. Users can fix it to get their own article.

6 EVALUATION

Data Set: We chose 20 news topics and crawled about 100 Chinese news articles for each news topic, 1969 articles in total. The news topics cover a wide range of fields, including politics(e.g., 萨德系统/THAAD Missile System), technology(e.g., 百度无人驾驶汽车/Baidu Self-driving Car), society(e.g., 江歌刘鑫案/The Case of Jiangge and Liuxin), entertainment(e.g., 绝地求生/PLAYERUNKNOWN’S BATTLEGROUND), and sports(e.g., 俄罗斯世界杯/2018 FIFA World Cup Russia).

In order to train and test the regression models for detecting subtopic labels, we extracted all the n-grams that meet the previously-mentioned requirements and got about 220 candidate labels for each topic on average. Then we tagged them manually, with each label assigned with a score from 0 to 3. A higher score means a better label. The majority of n-grams are assigned with 0 and the labels with nonzero scores account for 26.3% in total. The labels with nonzero scores are considered acceptable subtopic labels.

Evaluation on Subtopic Labels: We leverage all 12 features to train the SVR model, and use 20-fold cross-validation, with 19 news topics for training and the rest one for validation in turn. The values of P@5, P@10 and P@20 for SVR are 0.722, 0.693 and 0.619, respectively. The results indicate the majority of top 20 labels can represent subtopics. Other inappropriate labels can be filtered out by interaction with users.

Evaluation on News Synthesis Articles: There are no gold reference synthesis articles for each news topic, and it is also hard to manually write a few reference articles. Thus we choose to conduct a manual evaluation of the final news synthesis articles. We use the multi-document summarization methods implemented in PKUSUMSUM Zhang et al. (2016) as baselines.

The summarization methods we choose include Lead, Coverage, Centroid, and TextRank. Centroid and TextRank can work on either the sentence unit (i.e., Centroid-sen, TextRank-sen) or the text block unit (i.e., Centroid-blk, TextRank-blk). Thus we have six baselines. The baselines are compared with INS that does not involve user interaction. Each baseline and INS generate a synthesis article with 1000 words for each news event and 10 news topics will be manually evaluated. We employ 16 Chinese college students as judges. Each student evaluates one or two news topics. We make sure each student judge all the 7 articles for each news topic and each news topic is evaluated by 3 students. The judges are asked to give a rating score between 1 and 6 from the following aspects: readability, structure, topic diversity, redundancy removal, and overall responsiveness. The average results are shown in Table 1. From the table, we can see that INS achieves the best performance in every aspect, which proves the effectiveness of our INS system. We also perform pairwise t-tests when comparing our system with baselines, and find that INS significantly outperforms sentence-based methods in every aspect and block-based methods in almost all aspects (p-value $<$ 0.05).

Note that in the above comparison, INS does not involve any user interaction. We believe after the interaction with users at different stages, the quality of the synthesis articles will be much improved.

User Study on INS: We further performed a user study on INS by employing 10 users to experience it and give their judgments. The rating scores given by users on various aspects (e.g., usefulness, GUI, satisfaction, assistance) are generally high. This indicates that the INS system is useful and users are satisfied with the system.

In addition to scoring, we asked them to give their comments on INS. The majority consider it as an excellent helper. They said subtopics are helpful to know about the given topic clearly, and INS can filter redundant information out. Besides, they think INS provide them with enough and useful interactions to produce synthesis articles. One user pointed out that the GUI of INS can be further improved.

In the future, we will concentrate on improving the performance and beautifying the user interface. We also plan to deploy the system in several Chinese news media.

Acknowledgment

This work was supported by National Natural Science Foundation of China (61772036) and Key Laboratory of Science, Technology and Standard in Press Industry (Key Laboratory of Intelligent Press Media Technology). We appreciate the anonymous reviewers for their helpful comments. Xiaojun Wan is the corresponding author.