Using random walks for question-focused sentence retrieval

Abstract

We consider the problem of questionfocused sentence retrieval from complex news articles describing multi-event stories published over time. Annotators generated a list of questions central to understanding each story in our corpus. Because of the dynamic nature of the stories, many questions are time-sensitive (e.g. "How many victims have been found?") Judges found sentences providing an answer to each question. To address the sentence retrieval problem, we apply a stochastic, graph-based method for comparing the relative importance of the textual units, which was previously used successfully for generic summarization. Currently, we present a topic-sensitive version of our method and hypothesize that it can outperform a competitive baseline, which compares the similarity of each sentence to the input question via IDF-weighted word overlap. In our experiments, the method achieves a TRDR score that is significantly higher than that of the baseline.

Networks and Natural Language Processing

“…For example, both sentences and words can be added as vertices in the same graph. Edges can represent cooccurrence (such as two words that appear in the same sentence or in the same dictionary definition), collocation (for example, two words that appear immediately next to each other or that may be separated by a conjunction), syntactic structure (for example, the parent and child in a syntactic dependency), and lexical similarity (for example, cosine between the vector representations of two sentences).In terms of graph-based algorithms, the main methods used so far can be classified into: (1) semisupervised classification (Zhu and Ghahramani 2002;Zhu and Lafferty 2005;Toutanova, Manning, and Ng 2004;Radev 2004;Otterbacher, Erkan, and Radev 2005), where random walks or relaxation are applied on mixed sets of labeled and unlabeled nodes; (2) network analysis (Masucci and Rodgers 2006;, Caldeira et al 2006), where network properties such as diameter, centrality, and so on, are calculated; (3) graph-based clustering methods (Pang and Lee 2004, Widdows andDorow 2002), such as min-cut methods; (4) minimum spanning-tree algorithms (McDonald et al 2005).In this article, we overview several I n a cohesive text, language unitswhether they are words, phrases, or entire sentences-are connected through a variety of relations, which contribute to the overall meaning of the text and maintain the cohesive structure of the text and the discourse unity. Since the early ages of artificial intelligence, associative or semantic networks have been proposed as representations that enable the storage of such language units and the relations that interconnect them and that allow for a variety of inference and reasoning processes, simulating some of the functionalities of the human mind.…”

“…In terms of graph-based algorithms, the main methods used so far can be classified into: (1) semisupervised classification (Zhu and Ghahramani 2002;Zhu and Lafferty 2005;Toutanova, Manning, and Ng 2004;Radev 2004;Otterbacher, Erkan, and Radev 2005), where random walks or relaxation are applied on mixed sets of labeled and unlabeled nodes; (2) network analysis (Masucci and Rodgers 2006;, Caldeira et al 2006), where network properties such as diameter, centrality, and so on, are calculated; (3) graph-based clustering methods (Pang and Lee 2004, Widdows andDorow 2002), such as min-cut methods; (4) minimum spanning-tree algorithms (McDonald et al 2005).…”

Automatic Summarization from Multiple Documents

“…Graphs have been used to determine salient parts of text [Mih04,ER04b,ER04c] or query related sentences [OER05]. Lexical relationships [MR06] or rhetorical structure [Mar00] and even non-apparent information [Lam05] have been represented with graphs.…”

Discourse oriented summarization

“…Otterbacher et al, 2005;Jagarlamudi et al, 2006). These authors argue that there is a prior probability that a passage in a document is salient, independent of the query, and that the salience value of a passage is a weighted average of the relevance based on the document alone and the query-relevance.…”

“…If each sentence in the source text is (manually) rated for importance, an extractive summary can be evaluated by calculating the sum of the ratings of its sentences Wolf and Gibson, 2004). An indirect method of rating sentences was used by (Otterbacher et al, 2005), who asked annotators to come up with a list of questions to information key to understanding the story. Then, each sentence in the source which answered those questions was marked as relevant.…”