This book concentrates on the important and difficult task of finding temporal expressions in text, and interpreting them. Interpreting these expressions is tough: There is almost always information missing, such as a year in a date, and they tend only to be as precise as needed in the situation where they’re embedded – having a lot of dependence on context. We are often quite happy to talk about, say, meeting one another at six, without reference to what six is (a date? an hour?), and which hour it might be (a.m.? p.m.?) or even day. This missing context balances out any of the advantages one could imagine associated with having so many closed-class terms around, which times often contain, like months of the year or days of the week.
Strötgen and Gertz book addresses this domain of problems in a satisfying and well-written way, taking the problem as cast in the authoritative and standard ISO-TimeML standard. Being leaders in the field, the depth of the content will readily satisfy experts, though the authors have taken care to use simple introductory language and introduce basic, fundamental concepts along the way without labouring points. This makes the book easily accessible to newcomers, and with the absence of other works addressing this well-developed area, this book is certainly the authoritative work in the domain.
Intensive focus is given to two special areas. The individual parts of the temporal expression recognition process are explained in detail, as well as the interaction between them. This is often crudely treated as a sentence chunking problem or, even worse, a ‘novel’ new system for tagging is invented each year that fails to capture any of the nuances of the underlying problem. This work sets out clearly the specific interactions and parts of the recognition process, which while possible to cast as a chunking problem, is much better understood and easily leveraged. For example, the idea of the final annotation and the interpretation can become mingled; do we annotate ‘24h’ as a duration of 24 with unit hour, or of 1 with unit day? There are many niche cases and a spectrum of specificities between the text and the calendar, and this book explains how, through annotation schema, to best interpret and annotate temporal expressions in these cases.
The second area receiving special focus is the anatomy of temporal expressions and their interaction with their dataset. This is again examined and explained in detail; readers can only benefit from learning about why these expressions interact at both clause and sentence level, as well as paragraph and document level. Being able to interpret reference time through a document, which provides the temporal context, is tough, though critical to getting the right interpretation of an ambiguous expression. For example, ‘last year’ is only really useful if you know what the current year was at that point in the discourse. Similarly, understanding the dominant tense and aspect are crucial parts to understanding times.
Strötgen and Gertz address state-of-the-art concepts and challenges, like social media and clinical temporal expression extraction. Successfully extracting temporal information in many of these situations can benefit from casting the task in a slightly different way. The authors investigate many such framings, highlighting the pros and cons, and also the various resources available to tackle each of them. For example, clinical text can be very short and terse, and comprise a different set of temporal expressions; reference time may be expressed in a different way, as we see with the Clinical TempEval approach to temporal ordering; and events can be things like changing the kind of medication, which can be expressed with just two words and no verbs. Similarly, social media presents different challenges, often having an explicit document creation time (DCT – a critical concept to this kind of grounding) in its metadata, though very little further information, and of course the usual heightened chance of unorthodox orthography or birth of new terms.
This gives a strong and far-reaching catalogue of temporal expression resources and tools, more than I’ve ever seen gathered in one place before, with an objective and subjective description of each. This makes it better than a reference resource, guiding the reader through where the author thinks resources will be useful. Some of the opinions aren’t ones that would come to me instinctively, but even in these cases, there is always a well-explained point in the book to support the information here, and everything is quite principled.
Multilingual temporal expressions are addressed, though not exhaustively, and that’s OK. The critical point here is the shift over the past half-decade to separating out the language-independent behaviours and interactions in temporal language, which the book does well. For example, the closed class of 12 months can safely be represented as a global structure across languages; but the idea of weekdays might be a bit off in some places (like for those places that have eight). We can’t always assume the first day of the week is Monday, and of course there’s the unusual MM/DD/YY ordering which again needs regional knowledge to get best results. Certainly, interpreting ‘half seven’ to be 19:30 is a brave leap of hope, best accomplished again by some local effort. Strötgen and Gertz provide clear guidelines on how to customise and what parts of normalisation to draw out of a language-general implementation, illustrated with examples from many languages – French, English, Vietnamese and the rest.
Nearing the final part of the book we find descriptions of tools, with plenty of concrete advice for practitioners, giving pointers on how to pick tools or approaches, for any given situation, and why. The book rounds off with big important challenges in temporal tagging, which we all need solutions to!
All in all, this makes for a comprehensive reference work, well-organised and accessible to masters students or even final-year bachelors students, with useful concepts explained from the ground up. An excellent book in the area. Now I need an event extraction one and temporal ordering one to accompany it on the office shelf.
