1 Introduction
Informed decision-making on a controversial issue usually requires considering several pro and con arguments. To answer the question “Is organic food healthier?”, for example, people may query a search engine that retrieves arguments from diverse sources such as news editorials, debate portals, and social media discussions, which can then be compared and weighed. However, given the constant stream of digital information, this process may be time-intensive and overwhelming. Search engines and similar support systems may therefore benefit from employing argument summarization, that is, the generated summaries may aid the decision-making by helping users quickly choose relevant arguments with a specific stance towards the topic.
Argument summarization has been tackled both for single documents syed:2020 and multiple documents bhatia:2014; egan:2016. A specific multi-document scenario introduced by bar-haim:2020a is key point analysis where the goal is to map a collection of arguments to a set of salient key points (say, high-level arguments) to provide a quantitative summary of these arguments.
The Key Point Analysis (KPA) shared task by roni:2021111https://2021.argmining.org/shared_task_ibm, last accessed: 2021-08-08 invited systems for two complementary subtasks: matching arguments to key points and generating key points from a given set of arguments (Section 3
). As part of this shared task, we present an approach with two complementary components, one for each subtask. For key point matching, we propose a model that learns a semantic embedding space where instances that match are closer to each other while non-matching instances are further away from each other. We learn to embed instances by utilizing a contrastive loss function in a siamese neural network
bromley:1994. For the key point generation, we present a graph-based extractive summarization approach similar to the work of alshomary:2020a. It utilizes a PageRank variant to rank sentences in the input arguments by quality and predicts the top-ranked sentences to be key points. In an additional experiment, we also investigated an approach that performs aspect identification on arguments, followed by aspect clustering to ensure diversity. Finally, arguments with the best coverage of these diverse aspects are extracted as key points.Our approaches yielded the top performance among all submissions to the shared task in both quantitative and qualitative evaluation conducted by the organizers of the shared task (Section 5).222The code is available under https://github.com/webis-de/ArgMining-21.
2 Related Work
In summarization, arguments are relatively understudied compared to other document types such as news articles or scientific literature, but a few approaches have come up in the last years.
In an extractive manner, argument mining has been employed to identify the main claim as the summary of an argument petasis:2016; daxenberger:2017. wang:2016 used a sequence-to-sequence model for the abstractive summarization of arguments from online debate portals. A complementary task of generating conclusions as informative argument summaries was introduced by syed:2021. Similar to alshomary:2020b who inferred a conclusion’s target with a triplet neural network, we rely on contrastive learning here, using a siamese network though. Also, we build upon ideas of alshomary:2020a who proposed a graph-based model using PageRank page:1999 that extracts the argument’s conclusion and the main supporting reason as an extractive summary. All these works represent the single-document summarization paradigm where only one argument is summarized at a time, whereas the given shared task is a multi-document summarization setting.
The first approaches to multi-document argument summarization aimed to identify the main points of online discussions. Among these, egan:2016 grouped verb frames into pattern clusters that serve as input to a structured summarization pipeline, whereas Misra:2016 proposed a more condensed approach by directly extracting argumentative sentences, summarized by similarity clustering. bar-haim:2020a continued this line of research by introducing the notion of key points and contributing the ArgsKP corpus, a collection of arguments mapped to manually-created key points. These key points are concise and self-contained sentences that capture the gist of the arguments. Later, bar-haim:2020b proposed a quantitative argument summarization framework that automatically extracts key points from a set of arguments. Building upon this research, our approach aims to increase the quality of such generated key points, including a strong relation identifier between arguments and key points.
3 Task Description
In the context of computational argumentation, bar-haim:2020a introduced the notion of a key point as a high-level argument that resembles a natural language summary of a collection of more descriptive arguments. Specifically, the authors defined a good key point as being “general enough to match a significant portion of the arguments, yet informative enough to make a useful summary.” In this context, the KPA shared task consists of two subtasks as described below:
-
Key point matching. Given a set of arguments on a certain topic that are grouped by their stance and a set of key points, assign each argument to a key point.
-
Key point generation and matching. Given a set of arguments on a certain topic that are grouped by their stance, first generate five to ten key points summarizing the arguments. Then, match each argument in the set to the generated key points (as in the previous track).
Data
We start from the dataset provided by the organizers as described in roni:2021. The dataset contains 28 controversial topics, with 6515 arguments and a total of 243 key points. For each argument, its stance towards the topic as well as a quality score are given. Each topic is represented by at least three key points, with at least one key point per stance and at least three arguments matched to a key point. From the given arguments, 4.7% are unmatched, 67.5% belong to a single key point, and 5.0% belong to multiple key points. The remaining 22.8% of the arguments have ambiguous labels, meaning that the annotators could not agree on a correct matching to the key points. The final dataset contains 24,093 argument-key point pairs, of which 20.7% are labeled as matching. To develop our approach, we use the split as provided by the organizers with 24 topics for training, four topics for validation, and three topics for testing.
4 Approach
Our approach consists of two components, each corresponding to one subtask of the KPA shared task. The first subtask of matching arguments to key points is modeled as a contrastive learning task using a siamese neural network. The second subtask requires generating key points for a collection of arguments and then matching them to the arguments. We investigated two models for this subtask: One is a graph-based extractive summarization model utilizing PageRank page:1999 to extract sentences representing the key points; the other identifies aspects from the arguments and selects the most representative sentences that maximize the coverage of these aspects as the key points.
4.1 Key Point Matching
Conceptually, we consider pairs of arguments and key points that are close to each other in a semantic embedding space as possible candidates for matching. Furthermore, we seek to transform this space into a new embedding space where matching pairs are closer and the non-matching ones are more distant from each other (Figure LABEL:key-point-argument-mapping.ai). To do so, we utilize a siamese neural network with a contrastive loss function.
Specifically, in the training phase, the input is a topic along with a key point, an argument, and a label (matching or not). First, we use a pretrained language model to encode the tokens of the argument as well as those of the concatenation of the topic and the key point. Then, we pass their embeddings through a siamese neural network, which is a mean-pooling layer that aggregates the token embeddings of each input, resulting in two sentence-level embeddings. We compute the contrastive loss using these embeddings as follows:
where
is the cosine similarity of the embeddings, and
reflects whether a pair matches (1) or not (0).4.2 Key Point Generation
Our primary model for key point generation is a graph-based extractive summarization model. Additionally, we also investigate clustering the aspects of the given collection of arguments.
Graph-based Summarization
Following the work of alshomary:2020a, we first construct an undirected graph with the arguments’ sentences as nodes. As a filtering step, we compute argument quality scores for each sentence as toledo:2019 and exclude low-quality arguments from the graph. Next, we employ our key point matching model (Section 4.1) to compute the edge weight between two nodes as the pairwise matching score of the corresponding sentences. Only nodes with a score above a defined threshold are connected via an edge. An example graph is sketched in Figure LABEL:key-point-pagerank.ai. Finally, we use a variant of PageRank page:1999 to compute an importance score for each sentence as follows:
(1) | ||||
where is a damping factor that can be configured to bias the algorithm towards the argument quality score or the matching score . To ensure diversity, we iterate through the ranked list of sentences (in descending order), adding a sentence to the final set of key points if its maximum matching score with the already selected candidates is below a certain threshold.
Aspect Clustering
Extracting key points is conceptually similar to identifying aspects (bar-haim:2020a), which inspired our clustering approach that selects representative sentences from multiple aspect clusters as the final key points. We employ the tagger of schiller:2021 to extract the arguments’ aspects (on average, 2.1 aspects per argument). To tackle the lack of diversity, we follow Heinisch:2021 and create diverse aspect clusters by projecting the extracted aspect phrases to an embedding space. Next, we model the candidate selection of argument sentences as the set cover problem. Specifically, the final set of key points summarizing the arguments for a given topic and stance maximizes the coverage of the set of arguments’ aspects. To this end, we apply greedy approximation for selecting our candidates, where an argument sentence is chosen if it covers the maximum number of unique aspect clusters while having the smallest overlap with the clusters covered by the already selected candidates. Also, to avoid redundant key points, we compute its semantic similarity to the already selected candidates in each candidate selection step, and we add it to the final set if its score is below a certain threshold.
5 Experiments and Evaluation
In the following, we present implementation details of our two components, and we report on their quantitative and qualitative results.
5.1 Key Point Matching
We employed RoBERTa-large liu:2019 for encoding the tokens of the two inputs of key point matching to the siamese neural network, which acts as a mean-pooling layer and projects the encoder outputs (matrix of token embeddings) into a sentence embedding of size 768. We used Sentence-BERT reimers:2019b
to train our model for 10 epochs, with batch size 32, and maximum input length of 70, leaving all other parameters to their defaults.
For automatic evaluation, we computed both strict and relaxed mean Average Precision (mAP) following roni:2021. In cases where there is no majority label for matching, the relaxed mAP considers them to be a match while the strict mAP considers them as not matching. In the development phase, we trained our model on the training split and evaluated on the validation split provided by the organizers. The strict and relaxed mAP on the validation set were 0.84 and 0.96 respectively. For the final submission, we did a five-fold cross validation on the combined data (training and validation splits) creating an ensemble for the matching (as per the mean score).
Approach | R-1 | R-2 | R-L |
---|---|---|---|
Graph-based Summarization | 19.8 | 3.5 | 18.0 |
Aspect Clustering | 18.9 | 4.7 | 17.1 |
5.2 Key Point Generation
For the graph-based summarization model, we employed Spacy honnibal:2020 to split the arguments into sentences. Similar to bar-haim:2020b, only sentences with a minimum of 5 and a maximum of 20 tokens, and not starting with a pronoun, were used for building the graph. Argument quality scores for each sentence were obtained from Project Debater’s API toledo:2019333Available under: https://early-access-program.debater.res.ibm.com/. We selected the thresholds for the parameters d, qual and match in Equation 1 as 0.2, 0.8 and 0.4 respectively, optimizing for ROUGE lin:2004. In particular, we computed ROUGE-L between the ground-truth key points and the top 10 ranked sentences as our predictions, averaged over all the topic and stance combinations in the training split. We excluded sentences with a matching score higher than 0.8 with the selected candidates to minimize redundancy.
Topic | Stance | Graph-based Summarization | Aspect Clustering |
---|---|---|---|
Routine child vaccinations should be mandatory | Pro | (1) Child vaccinations should be mandatory to provide decent health care to all. (2) Vaccines help children grow up healthy and avoid dangerous diseases. (3) Child vaccinations should be mandatory so our children will be safe and protected. | (1) Child vaccination is needed for children, they get sick too. (2) Routine child vaccinations should be mandatory to prevent the disease. (3) Yes as they protect children from life threatening and highly infectious diseases. |
Routine child vaccinations should be mandatory | Con | (1) Vaccination should exclude children to avoid the side effects that can appear on them. (2) Parents should have the freedom to decide what they consider best for their children. (3) The child population has a low degree of vulnerability, so vaccination is not urgent yet. | (1) Child vaccination shouldn’t be mandatory because the virus isn’t effective in children. (2) Child vaccinations should not be mandatory because vaccines are expensive. (3) It has not been 100% proven if the vaccine is effective. |
For aspect clustering, we created 15 clusters per topic and stance combination. After greedy approximation of the candidate sentences, we removed redundant ones using a threshold of 0.65 for the normalized BERTScore Zhang:2020 with the previously selected candidates.
Comparison of both approaches
To select our primary approach for key point generation, we first performed an automatic evaluation of the aforementioned models on the test set using ROUGE (Table 1). Additionally, we performed a manual evaluation via pairwise comparison of the extracted key points for both models for a given topic and stance.
Examples of key points from both the models are shown in Table 2. The key points from graph-based summarization model are relatively longer. This also improves their informativeness, matching findings of syed:2021. For the aspect clustering, we observe that the key points are more focused on specific aspects such as “disease” (for Pro) and “effectiveness” (for Con). In a real-world application, this may provide the flexibility to choose key points by aspects of interest to the end-user, especially with further improvement of aspect tagger by avoiding non-essential extracted phrases as “mandatory”. Hence, given the task of generating a quantitative summary of a collection of arguments, we believe that the graph-based summary provides a more comprehensive overview and chose this as our preferred approach for key point generation.
5.3 Shared Task’s Evaluation Results
KP Matching | KP Generation | |||
---|---|---|---|---|
Approach | mAP/Rank | Rel. | Rep. | Pol. |
bar_h | 0.885/1 | 2 | 1 | 1 |
mspl (ours) | 0.818/2 | 2 | 1 | 2 |
sohanpat | 0.491/3 | 4 | 4 | 2 |
peratham | 0.443/4 | 1 | 3 | 4 |
In key point matching, our approach obtained a strict mAP of 0.789 and a relaxed mAP of 0.927 on the test set, the best result among all participating approaches. For the second track, in addition to evaluating the key point matching task, the shared task organizers manually evaluated the generated key points through a crowdsourcing study in which submitted approaches were ranked according to the quality of their generated key points. Table 3 presents the evaluation results of the top three submitted approaches, along with the reference approach of bar-haim:2020b. Among the submitted approaches, our approach was ranked the best in both the key point generation task as well as the key point matching task. For complete details on the evaluation, we refer to the task organizers’ report roni:2021.
6 Conclusion
This paper has presented a framework to tackle the key point analysis of arguments. For matching arguments to key points, we achieved the best performance in the KPA shared task via contrastive learning. For key point generation, we developed a graph-based extractive summarization model that output informative key points of high quality for a collection of arguments. We see abstractive key point generation as part of our future work.