Finding Answers to Definition Questions on the Web

Alejandro Figueroa

PhD-Thesis Universität des Saarlandes 7/2010.


Fundamentally, question answering systems are designed for automatically responding to queries posed by users in natural language. The first step in the answering process is query analysis, and its goal is to classify the query in concert with a set of pre-specified types. Traditionally, these classes include: factoid, definition, and list. Systems thereafter chose the answering method in congruence with the class recognised in this early phase. In short, this thesis focuses exclusively on strategies to tackle definition questions (e.g., ``Who is Ben Bernanke?"). This sort of question has become especially interesting in recent years, due to its significant number of submissions to search engines. Most advances in definition question answering have been made under the umbrella of the Text REtrieval Conference (TREC). This is, more precisely, a framework for testing systems operating on a collection of news articles. Thus, the objective of chapter one is to describe this framework along with presenting additional introductory aspects of definition question answering including: (a) how definition questions are prompted by individuals; (b) the different conceptions of definition, and thus of answers; and (c) the various metrics exploited for assessing systems. Since the inception of TREC, systems have put to the test manifold approaches to discover answers, throwing some light onto several key aspects of this problem. On this account, chapter four goes over a selection of some notable TREC systems. This selection is not aimed at completeness, but rather at highlighting the leading features of these systems. For the most part, systems benefit from knowledge bases (e.g., Wikipedia) for obtaining descriptions about the concept being defined (a.k.a. definiendum). These descriptions are thereafter projected onto the array of candidate answers as a means of discerning the correct answer. In other words, these knowledge bases play the role of annotated resources, and most systems attempt to find the answer candidates across the collection of news articles that are more similar to these descriptions. The cornerstone of this thesis is the assumption that it is plausible to devise competitive, and hopefully better, systems without the necessity of annotated resources. Although this descriptive knowledge is helpful, it is the belief of the author that they are built on two wrong premises: It is arguable that senses or contexts related to the definiendum across knowledge bases are the same senses or contexts for the instances across the array of answer candidates. This observation also extends to the fact that not all descriptions within the group of putative answers are necessarily covered by knowledge bases, even though they might refer to the same contexts or senses. Finding an efficient projection strategy does not necessarily entail a good procedure for discerning descriptive knowledge, because it shifts the goal of the task to a ``more like this set" instead of analysing whether or not each candidate bears the characteristics of a description. In other words, the coverage given by knowledge bases for a specific definiendum is not wide enough to learn all the characteristics that typify its descriptions, so that systems are capable of identifying all answers within the set of candidates. From another angle, a conventional projection methodology can be seen as a finder of lexical analogies. All in all, this thesis investigates into models that disregard this kind of annotated resource and projection strategy. In effect, it is the belief of the author that a robust technique of this sort can be integrated with traditional projection methodologies, and in this way bringing about an enhancement in performance. The major contributions of this thesis are presented in chapters five, six and seven. There are several ways of understanding this structure. For example, chapter five presents a general framework for answering definition questions in several languages. The primary goal of this study is to design a lightweight definition question answering system operating on web-snippets and two languages: English and Spanish. The idea is to utilise web-snippets as a source of descriptive information in several languages, and the high degree of language independency is achieved by making allowances for as little linguistic knowledge as possible. To put it more precisely, this system accounts for statistical methods and a list of stop-words, as well as a set of language-dependent definition patterns. In detail, chapter five branches into two more specific studies. The first study is essentially aimed at capitalising on redundancy for detecting answers (e.g., word frequency counts across answer candidates). Although this type of feature has been widely used by TREC systems, this study focuses on its impact on different languages, and its benefits when applied to web-snippets instead of a collection of news documents. An additional motivation behind targeting web-snippets is the hope of studying systems working on more heterogenous corpora, without incurring the need of downloading full-documents. For instance, on the Internet, the number of distinct senses for the definiendum considerably increases, ergo making it necessary to consider a sense discrimination technique. For this purpose, the system presented in this chapter takes advantage of an unsupervised approach premised on Latent Semantic Analysis. Although the outcome of this study shows that sense discrimination is hard to achieve when operating solely on web snippets, it also reveals that they are a fruitful source of descriptive knowledge, and that their extraction poses exciting challenges. The second branch extends this first study by exploiting multilingual knowledge bases (i.e., Wikipedia) for ranking putative answers. Generally speaking, it makes use of word association norms deduced from sentences that match definitions patterns across Wikipedia. In order to adhere to the premise of not profiting from articles related to a specific definiendum, these sentences are anonymised by replacing the concept with a placeholder, and the word norms are learnt from all training sentences, instead of only from the Wikipedia page about the particular definiendum. The results of this study signify that this use of these resources can also be beneficial; in particular, they reveal that word association norms are a cost-efficient solution. However, the size of the corpus markedly decreases for languages different from English, thus indicating their insufficiency to design models for other languages. Later, chapter six gets more specific and deals only with the ranking of answer candidates in English. The reason for abandoning the idea of Spanish is the sparseness observed across both the redundancy from the Internet and the training material mined from Wikipedia. This sparseness is considerably greater than in the case of English, and it makes learning powerful statistical models more difficult. This chapter presents a novel way of modeling definitions grounded on n-gram language models inferred from the lexicalised dependency tree representation of the training material acquired in the study of chapter five. These models are contextual in the sense that they are built in relation to the semantic of the sentence. Generally, these semantics can be perceived as the distinct types of definienda (e.g., footballer, language, artist, disease, and tree). This study, in addition, investigates the effect of some features on these context models (i.e., named entities, and part-of-speech tags). Overall, the results obtained by this approach are encouraging, in particular in terms of increasing the accuracy of the pattern matching. However, in all likelihood, it was experimentally observed that a training corpus comprising only positive examples (descriptions) is not enough to achieve perfect accuracy, because these models cannot deduce the characteristics that typify non-descriptive content. More essential, as future work, context models give the chance to study how different contexts can be amalgamated (smoothed) in agreement with their semantic similarities in order to ameliorate the performance. Subsequently, chapter seven gets even more specific and it searches for the set of properties that can aid in discriminating descriptions from other kinds of texts. Note that this study regards all kinds of descriptions, including those mismatching definition patters. In so doing, Maximum Entropy models are constructed on top of an automatically acquired large-scale training corpus, which encompasses descriptions from Wikipedia and non-descriptions from the Internet. Roughly speaking, different models are constructed as a means of studying the impact of assorted properties: surface, named entities, part-of-speech tags, chunks, and more interestingly, attributes derived from the lexicalised dependency graphs. In general, results corroborate the efficiency of features taken from dependency graphs, especially the root node and n-gram paths. Experiments conducted on testing sets of various characteristics suggest that it is also plausible to find attributes that can port to other corpora. The second and the third are extra chapters. The former examines different strategies to trawl the Web for descriptive knowledge. In essence, this chapter touches on several strategies geared towards boosting the recall of descriptive sentences across web snippets, especially sentences that match widespread definition patterns. This is a side, but instrumental study to the core of this thesis, as it is necessary for systems targeted at the Internet to develop effective crawling techniques. On the contrary, chapter three has two goals: (a) presenting some components used by the strategies outlined in the last three chapters, this way helping to focus on key aspects of the ranking methodologies, and hence to clearly present the relevant aspects of approaches laid out in these three chapters; and (b) fleshing out some characteristics that make separating the genuine from the misleading answer candidates difficult; particularly, across sentences matching definition patterns. Chapter three is helpful for understanding part of the linguistic phenomena that the posterior chapters deal with. On a final note about the organisation of this thesis, since there is a myriad of techniques, chapter six and seven start dissecting the related work closer to each strategy. The main contribution of each chapter begins at section 6.5 and 7.6, respectively. These two sections start with a discussion and comparison between the proposed methods and the related work presented in their corresponding preceding sections. This organisation is directed at facilitating the contextualisation of the proposed approaches as there are different question answering systems with manifold characteristics.


dissertation.pdf (pdf, 3 MB)

German Research Center for Artificial Intelligence
Deutsches Forschungszentrum für Künstliche Intelligenz