Читать книгу Domain-Sensitive Temporal Tagging - Jannik Strötgen - Страница 12

CHAPTER 1

Introduction

Temporal tagging is a specific task in natural language processing (NLP), in which temporal expressions are extracted from text documents and normalized to some standard format. Since temporal expressions are prevalent in many types of documents and because temporal information is an important dimension in any information space, applications of several domains can benefit from the output of temporal taggers.

This book covers the topic of temporal tagging and is structured as follows. In this chapter, we describe the task of temporal tagging, and then present some examples of NLP and NLP-related application scenarios in which temporal information can be exploited to provide more meaningful and useful results. In Chapter 2, we provide background knowledge and cover basic concepts related to temporal information. The foundations of temporal tagging are described in Chapter 3, and temporal tagging of different types of documents and thus domain-sensitive temporal tagging are explained in Chapter 4. An overview of existing techniques and tools for temporal tagging including our own system HeidelTime is provided in Chapter 5. Finally, future research directions are discussed in Chapter 6. However, to guarantee the correct understanding of two important terms frequently used in this book, we start with defining the concepts “temporal expression” and “value of a temporal expression”.

• A temporal expression is either an expression referring to a date or time of any granularity (e.g., “March 11, 2007”, “yesterday”, “June 2016”, “20th century”, “9 pm”), an expression referring to a duration (e.g., “three years”, “several months”), or an expression referring to the periodical aspect of an event (e.g., “every Monday”, “twice a week”).

• The value (of a temporal expression) covers the (most important) semantics of the temporal expression in a standard format, that is, the normalized information of the expression.

Examples of and more details about different types of temporal expressions and annotation standards for temporal expressions will be covered later in this book, but these definitions are crucial to understand the task of temporal tagging, which is defined and explained next.

1.1 THE TASK OF TEMPORAL TAGGING

Temporal tagging addresses the extraction, classification, and normalization of temporal expressions occurring in text documents. It is a prerequisite of the full task of temporal annotation (temporal information extraction), which concerns the detection and interpretation of temporal expressions, events, and temporal relations between events and between temporal expressions and events [Verhagen et al., 2009]. However, temporal tagging is not only valuable in the context of temporal information extraction, but also in many research areas and application scenarios as will be detailed in Section 1.2.

In general, temporal tagging can be considered as a specific type of named entity recognition and normalization. Although the three standard named entity types are person, organization, and location [Nadeau and Sekine, 2007], “the notion of named entity is commonly extended to include things that are not entities per se, but nevertheless have practical importance and do have characteristic signatures that signal their presence” [Jurafsky and Martin, 2008, p. 762]. Thus, further types of information are sometimes also covered under the named entity umbrella, for example, genes and proteins, numbers, and temporal expressions.

The classical tasks of named entity recognition (NER) tools are to identify the spans of named entities in texts and to classify the extracted named entities into pre-defined classes of entities. Thus, the normalization of entities to a unique identifier or some value in a standard format is only performed if the named entities’ normalization—depending on the type of entity also referred to as disambiguation, linking, or resolution—is addressed, too. In contrast, a temporal tagger identifies the spans of temporal expressions in texts and normalizes the expressions according to some standard format. Depending on the annotation specifications, expressions are also sometimes classified according to their type, e.g., whether an expression is a date (e.g., May 3, 2009) or a duration (e.g., three days). However, this classification of temporal expressions can be considered as a part of the normalization process and thus, one can specify the two subtasks of temporal tagging as follows.

• Extraction: given a text, determine the spans of all temporal expressions.

• Normalization: given a text and a set of extracted temporal expressions, assign the temporal semantics to each expression in the form of normalized values in a standard format that adheres to some annotation specification.

Figure 1.1 illustrates the two tasks of a temporal tagger. Given a text document (left), determine the temporal expressions (middle), and assign a normalized value in a standard format to each identified temporal expression (right). In Chapter 3, we will give an overview of existing annotation standards for temporal expressions. These define what should be considered as a temporal expression and how temporal expressions are to be normalized. Before that, however, we will first outline some application scenarios in which temporal expressions can be exploited, and then have a closer look at the concept of time in Chapter 2.

1.2 APPLICATION EXAMPLES EXPLOITING TEMPORAL TAGGING

For well-known NLP tasks such as named entity recognition (NER), there are many motivating application scenarios described in the literature. In the following, to illustrate the utility of temporal tagging, we present some use cases, in which applications can easily exploit extracted and normalized temporal information and benefit from the output of temporal taggers and thus from the value of temporal information in general.

Figure 1.1: The two tasks of temporal tagging: extraction and normalization.

TEMPORAL TAGGING FOR INFORMATION EXTRACTION

In many text documents, events play an important role. Typically, events happen at some specific time and some specific place [Strötgen and Gertz, 2012a]. The importance of temporal information when organizing and summarizing extracted events is intuitive: given a text document with event mentions, the chronological ordering of the described events obviously benefits from normalized temporal expressions. Similar to temporal information, geographic information is also important in this context. However, the geographic aspect of events is out of the scope of this book.

As illustrated in Figure 1.2, many documents do not mention events in a chronological order. Typically, sections about specific topics are used and contain temporally overlapping content. Further examples are biographies that often contain temporally overlapping sections about, for instance, “private life” and “professional life”, and news articles that report on recent happenings before referring to events that have happened in the past. An example of such a news article is shown in Figure 1.3.

Similar to the task of summarizing and ordering events extracted from documents, temporal fact extraction also requires temporal tagging output. For instance, when collecting facts for a knowledge base, it should be taken into account that most facts are not static but either evolve with time or are valid only during a particular time period [Kuzey and Weikum, 2012]. For instance, “Bill Clinton” holdsPoliticalPosition “President of the United States” is a correct fact but only valid for a specific time period.

While extracting events and temporal relations from single documents has a rather long tradition and was, for instance, addressed in the TempEval competitions at SemEval 2007 [Verhagen et al., 2007], 2010 [Verhagen et al., 2010], and 2013 [UzZaman et al., 2013], research was more recently extended to perform cross-document temporal relation extraction, as in the Timeline task of SemEval 2015 [Minard et al., 2015].¹ A further indication of the importance of temporal tagging in the context of information extraction is the fact that at the 2015 SemEval competition, in addition to the Timeline task, three additional shared tasks were organized, in which extracted and normalized temporal expressions are a prerequisite to successfully address the tasks: QA TempEval² [Llorens et al., 2015], Clinical TempEval³ [Bethard et al., 2015], and Diachronic Text Evaluation⁴ [Popescu and Strapparava, 2015].

TEMPORAL TAGGING FOR TOPIC DETECTION AND TRACKING

The goal of topic detection and tracking (TDT) is to organize news documents in an event-based way by building clusters of topics [Allan, 2002]. In this context, a topic is typically defined as “a seminal event or activity, along with all directly related events and activities” [Fiscus and Doddington, 2002]. For instance, the very first news article about a plane crash opens a new topic, and following news articles such as reports about the number of fatalities belong to the same topic. In contrast, news articles reporting about another plane crash do not belong to the same cluster. To decide whether an upcoming news document belongs to an existing cluster or opens a new cluster, the similarity between documents is typically determined based on some information extracted from the documents. For instance, Makkonen et al. [2003] create event vectors consisting of (i) names, (ii) locations, (iii) temporals, and (iv) content words.

Figure 1.2: Excerpts of the Wikipedia page about “Heidelberg University” and a timeline to which occurring temporal expressions are mapped. The content is not reported in a chronological order due to different topical sections about Heidelberg University. Thus, temporal tagging is crucial to correctly extract and order event information in a chronological way.

Figure 1.3: Excerpts of the CNNMoney article of Figure 1.1. After reporting on a recent happening, it refers to an event from the past in its last paragraph. Again temporal tagging is crucial to correctly extract and order event information.

In general, ambiguous expressions—such as “Tuesday”, “Friday”, and “March” in the news article shown in Figure 1.3—are quite frequent in news documents. To be able to exploit information about temporal expressions occurring in documents, temporal tagging is again a prerequisite because not just the detection but in particular the normalization of temporal expressions is crucial for successful topic detection and tracking.

TEMPORAL TAGGING FOR INFORMATION RETRIEVAL

During recent years, the value of temporal information has been increasingly exploited in the context of information retrieval research and applications [Alonso et al., 2007, 2011, Campos et al., 2014, Derczynski et al., 2015, Kanhabua et al., 2015]. Note, however, that there are different types of temporal information that can be used in information retrieval scenarios. The two main aspects are (i) time as a dimension of relevance and (ii) time as query topic.

On the one hand, when time is used as a dimension of relevance, temporal tagging is not needed. However, information about the document creation time is typically utilized to improve the ranking of documents. For example, for news-related queries, the freshness of search results may be important [see, e.g., Li and Croft, 2003]. In addition to improving search results, time as contextual information can be used to perform time-sensitive query auto-completion [Sengstock and Gertz, 2011, Shokouhi and Radinsky, 2012].

On the other hand, temporal tagging plays a crucial role when time is a query topic. No matter whether the temporal part of a query is provided explicitly or implicitly, temporal expressions occurring in potentially relevant documents have to be detected, normalized, and compared to the temporal aspect of the query. Berberich et al. [2010], for instance, integrate temporal expressions into a language modeling approach, and Strötgen and Gertz [2012a] present a query model to explicitly formulate temporal queries in a flexible way. Note that time as query topic must be handled by search engines, because temporal queries occur frequently as was shown by some query log analyses of web search engines: Nunes et al. [2008] found 1.5% queries with explicit temporal information, Metzler et al. [2009] determined 7% queries with implicit temporal intent, and Zhang et al. [2010] reported 13.8% for queries with explicit time and 17.1% with implicit time.

Note that sometimes the document creation time of a document might be a good indicator for detecting whether a document is relevant for a given query. However, using a temporal tagger to analyze the documents’ content is often crucial to successfully find relevant documents. For instance, both documents shown in Figure 1.4 can be considered as relevant for the example information need “Germanwings” with the time interval of interest being set to “1st of March 2015 to 30th of April 2015”. While the first document is a news document also published during the time interval of interest, the second document is a news article published in November 2015, that is, outside of the time interval of interest. However, both documents contain temporal expressions referring to the Germanwings plane crash in March 2015 (“Tuesday” and “March”, respectively), and they thus satisfy the information need.

Figure 1.4: Temporal information retrieval example. Given the query 〈“Germanwings”, “1st of March 2015 to 30th of April 2015”〉, both documents can be identified as relevant if a temporal tagger is used to extract and normalize the temporal expressions in the documents’ content.

A further interesting observation from Figure 1.4 is that the term “Tuesday” in the first document refers to a date within the time interval of interest (March 24, 2015) while the same term in the second document does not (here, it refers to November 10, 2015).

TEMPORAL TAGGING FOR QUESTION ANSWERING

A further area in which time is a crucial dimension is question answering. While this is one commonality with information retrieval, the two tasks share further aspects: In both areas, a user is faced with an information need, and the goal of both information retrieval and question answering is to satisfy this information need. In contrast, the main differences between them is that in information retrieval, the information need is typically formulated as a query consisting of keywords—possibly enriched with time intervals of interest in the area of temporal information retrieval—but in question answering, the information need is formulated as a natural language question. Analogously, the presentation of results is also different: in information retrieval, a ranked list of relevant documents is typically presented to the user while in question answering, the answer to the information need is directly provided.

On the border between both areas lies so-called entity-oriented search [Balog et al., 2012]. A typical information retrieval query is to ask for a specific entity or fact about an entity. Thus, the goal of entity-oriented search is—as in question answering—to directly provide an answer, in the ideal case together with a justification, e.g., in the form of small text nuggets rather than full-length documents [Pasca, 2008]. An example of such a query with a temporal dimension is the query “Golden Gate bridge built” with the answer “1937”.

A research competition dealing with temporal (and geographic) information needs is NTCIR GeoTime [Gey et al., 2010, 2011]. As in question answering, the information needs are formulated as natural language questions. Due to the temporal and geographic focus of the competition, the questions contain “where” and “when” aspects. However, unlike in standard question answering, systems are not evaluated based on whether they provide the correct answer, but on whether or not the documents ranked in a result list answer the question and are thus relevant. That is, the evaluation is performed in an information retrieval fashion.

In contrast to entity-oriented search and GeoTime, which both directly benefit from extracted and normalized temporal expressions, time-related question answering often deals with more complex temporal phenomena [Pustejovsky et al., 2005]. Then, temporal tagging on its own is not sufficient but temporal reasoning is often necessary, for example, to answer questions of the form “did event x happen before event y?”. To be able to automatically answer such questions, the full task of temporal information extraction is required—including the subtasks of temporal tagging, event extraction, and temporal relation extraction. In the recent QA TempEval challenge at SemEval 2015, in which temporal information extraction systems were to be developed, the systems were evaluated solely based on how well they perform in answering such time-related questions for which temporal reasoning is important [Llorens et al., 2015]. In Chapter 3, we will detail how temporal taggers can be evaluated in general.

TEMPORAL TAGGING FOR SUMMARIZATION

While the value of temporal tagging for the above examples is quite straightforward, there are further application scenarios, in which temporal tagging can provide more indirect benefits. An example of such an application scenario is the document summarization task.

In the text summarization community, it is well known that coreference resolution is valuable to create better text summaries [Azzam et al., 1999, Steinberger et al., 2007]. Similar to coreference relations between (proper) nouns and pronouns, the relations between temporal expressions could also be taken into account to improve summaries. Assume the document that is to be summarized contains the following two sentences consecutively:

• s₁ = 〈In 2010, something unimportant happened.〉

• s₂ = 〈One year later, something important happened.〉

Obviously, good document summarizations should contain important information, that is, in our example s₂ should be part of the summary but s₁ should not be contained in the summary. However, without proper context information, the semantics of s₂ is unclear due to the ambiguity of “One year later”. To fully understand s₂, the reader requires a reference time to resolve the relative temporal expression. Unfortunately, this reference time is part of s₁ (“2010”).

One solution to address this issue is to include both sentences in a summary. However, this results in a summary containing unimportant content, so that a better approach is to exploit the information provided by a temporal tagger (in s₂ that “One year later” refers to 2011). In this way, the unimportant sentence s₁ could be skipped, and s₂ could be part of the summary in a slightly modified way, for instance, starting with “One year later (2011), something important happened”. Note that even for the first solution, some information about occurring temporal expressions is necessary, namely that s₁ contains the reference time of s₂.

1.3 SUMMARY OF THE CHAPTER

In the context of temporal tagging, two tasks can be distinguished: extraction and normalization of temporal expressions. In several NLP-related research areas, and thus in many applications, temporal tagging output can be exploited to improve the approaches. Note that for almost all applications and research topics exploiting temporal information, the normalization subtask is highly crucial.

¹Timeline: Cross-Document Event Ordering, http://alt.qcri.org/semeval2015/task4/ [last accessed: Nov 9, 2015].

²Question Answering TempEval, http://alt.qcri.org/semeval2015/task5/ [last accessed: Nov 9, 2015].

³Clinical TempEval, http://alt.qcri.org/semeval2015/task6/ [last accessed: Nov 9, 2015].

⁴Diachronic Text Evaluation, http://alt.qcri.org/semeval2015/task7/ [last accessed: Nov 9, 2015].