Research On Discourse Representation Based On Event Semantic And Its Applications

Recently, the analysis of word and sentence level has been paid more attentions in Natural Language Processing. Various technologies, such as lexical analysis, word semantic analysis and syntactic analysis, have been gradually mature. However, as the most important research content in Natural Language Processing, discourse learning and understanding is still not optimistic. The main reason may be that some ambiguity and polysmy phenomena exist in the combination process of different granularity semantic units of discourse text, namely, words, phrases and sentences. To resolve this ambiguity, more knowledge reasoning need be conducted based on these semantic processing units. Therefore, how to mine more abundant knowledge from the unstructured text and apply them in discourse learning and reasoning is still a threshold of natural language understanding.Previous studies on discourse learning tasks, such as automatic summarization, topic analysis and information retrieval, etc., regard words, phrases or sentences as basic pro^ssing unit, and have achieved greater success in both academy and industry. Compared with phrases or sentences, the event, as a special form of knowledge representation, can play a more important role in discourse learning and understanding. From the semantic level, words or phrases are fine-grained. The disambiguation is required when this kind of units are applied into specific semantic analysis tasks. Similar to sentence, each event has a certain semantic in most of cases. From the granularity of basic processing unit, sentences are coarse-grained. The sparseness problem appears in the process of semantic analysis, and brings some difficulties to information statistic and reasoning. So, in this paper, we regard the events as the trade-off between words and sentences. We focus on exploiting events for discourse representation, and conduct some studies on some discourse tasks, such as headline generation, multi-document summarization and topic representation.Different from word or sentence, which has a natural form, the structured event needs be extracted from discourse in particular using the reasonable technology. Events should be identified from the discourse as more as possible. It can bring great convenience to the knowledge reasoning, and improve the quality of discourse learning and understanding. However, the discourse representation and learning based on events still faces some challenges. On one hand, we need a unified form of event structure, and can extract the structured event from the text exactly. On the other hand, some studies should be conducted on how to overcome the disadvantages of traditional methods based on words or sentences, and how to take advantages of the events as basic processing units for knowledge reasoning in discourse understanding. Therefore, this paper first focuses on event extraction in open domain, and then conducts some studies on how to exploit event semantics for discourse learning and understanding in several discourse tasks, respectively.Our works are described as follows.1. Open Domain Event Extraction via Double PropagationMost previous studies on event extraction focus on news articles in special domain. In open domain, event extraction is simply regarded as a preprocessing step in the studies based on events. These works directly exploit the syntactic rules like dependency relations or entity relations based on open relation extraction tools. They neglect the characteristic of language, and depend on the performance of dependency parsers or open relation extraction tools in excess. To some extent, the incorrect results or low recall may give some limitation to the effect of discourse learning and understanding. This paper proposes to exploit double propagation to combine event extraction and event pattern generation. This unsupervised method does not require seed events or seed event patterns. In iterative propagation process, the candidate events are used to reinforce the generation of standard event patterns, while the standard patterns are used to guide the modification of candidate events and the extraction of new events. Experimental results on two different scale corpora show the effectiveness of the proposed method.2. Event-Driven Headline GenerationHeadline generation, as a special text summarization task, is challenging in not only informativeness and readability, but also the length reduction. Extractive models focus on tailoring important human-written sentences and give less informative results. In contrast, abstractive models use sentence synthesis technology based on some salient phrases, but it is more difficult to ensure the grammaticality of the results. This paper proposes an event-driven model to alleviate the disadvantages of both extractive and abstractive models. This model exploits a novel multi-sentence compression algorithm to fuse a set of salient event to generate the headline for the document. Firstly, a bipartite graph between lexical chains and events is constructed, and salience information of both sentences and phrases is exploited to learn the salient events. Then, based on these events, a directed acyclic word graph is constructed and a beam search algorithm is designed to find the title based on path scoring. Experimental results on standard dataset show learning salient event is benefit to choose important candidate sentences and the constraints of structured event in path searching are also useful to generate the final title. Compared with previous state-of-the-art systems, the proposed model achieves the best results.3. Multi-Document Abstractive Summarization Using Event GuidanceThis paper proposes an abstractive multi-document summarization method via submodular maximization using event guidance. Different from extractive approaches, abstractive approaches require deeper text analysis and can generate new sentences to convey the important content from documents. Previous methods exploited more fine-grained syntactic units than sentences, namely, noun/verb phrases, to construct new sentences, but the results are still very inaccurate because the phrases have little grammatical information. This paper proposes to exploit the event information to guide the splitting of subtopics and the generation of candidate summary sentences. On one hand, the clusters are generated through event clustering based on event semantic similarity. This method is much less noise than the traditional sentence clustering. On the other hand, the structured event information is naturally introduced into multi-sentence compression to constraint the choice of vertices in path searching. Finally, three submodular functions are combined to select some preferable paths to construct the final summary. Experimental results on standard datasets demonstrate the event information can facilitate not only the sentence clustering, but also the compression candidate generation. The quality of our summary outperforms the state-of-art abstractive system.4. Topic Analysis Integrated with Event KnowledgeThis paper explores to exploit structured events for topic analysis. Most previous models are based on words or phrases. This form of topic representation has a poor interpretability, due to the lack of deep semantic information. To address the problem, two topic models based on Biterm Topic Model are proposed in this paper. Event semantic knowledge is incorporated into these models using two different ways. The first model exploits Generalized Polya Urn model to increase the probability of assigning same topic to similar events. Differently, the second model introduces an indicator variable for each biterm, and exploits event semantic information to solve the topic assignment of the events in one biterm more reasonably. This paper not only directly evaluates the topic models based on two metrics, namely topic coherence and KL-divergence, but also conducts the external evaluation by carrying out text classification task based on the results of topic representation. The experimental results demonstrate two topic models effectively diminish the sparseness from two perspectives: event co-occurrence and semantic relatedness. Compared to the topic analysis based on words, the semantic information of event effectively promotes the topic quality and improves the interpretability and topic discrimination of topic representation.