Our corpus consists of research papers published in the journal Global Environmental Change (GEC). The corpus includes all of the articles from the first volume (1990/1991) to Volume 20 (2010). In compiling the corpus we targeted only full-length articles and did not include non-research papers, such as book reviews. The corpus includes the main body text but excludes other sections of research papers such as the abstract or appendices. Tables and figures are not included, either. Mathematical symbols and equations have been replaced with the non-word EQSYM. The corpus includes 675 research papers and consists of 4.1 million words. Based on external criteria, it is a specialised corpus; it is large enough to preclude hands-on reading of all the texts in it as a way of surveying the corpus.

In taking a topic modelling approach, an initial decision we have to make is what to conceive as a text. In our study, we wished to take a more fine-grained approach than would be captured by considering each research paper as a single text. A research paper may contain several topics. A paper, for instance, might have a high frequency of theme-related words such as environment or pollution at the beginning, while the same paper may have a high frequency of method-related words such as analyze or experiment in the middle. To capture the within-paper shift of topic of this kind, each paper was divided into multiple blocks, each constituting a ‘text’ as defined in the topic model. More specifically, each block included a minimum of 300 words, and a block was not permitted to cut across a paragraph boundary, on the assumption that paragraphs themselves are topic-based units. For example, suppose that a paper consists of six paragraphs, and the number of words in each paragraph is as follows:

Topic transition within text-blocks is assumed to be smaller than that between text-blocks because neighbouring paragraphs tend to belong to the same section and are likely to be topically related. This, however, does not mean that topic modelling assumes topically uniform text-blocks. Rather, a strength of the technique is that each text-block includes a mixture of topics (Blei, 2012). Therefore, even if a text-block includes paragraphs with different topics, it does not significantly affect the identification of topics in the corpus.

Most of the previous literature using topic models excludes the words in the standard stop words list (e.g., Marshall, 2013) because function words and pronouns provide little information about the topic of texts. Those words, however, are potentially informative in characterising research papers linguistically; pronouns, for example, enact engagement between writer and reader (Hyland, 2005). We therefore excluded far fewer words as stop words: only prepositions, articles, and, it, as, that, be, have and do (and the inflected forms of the last three verbs). This ensured that we retain the potentially important insights brought by closed-class words. We also excluded one-letter words because they included much noise such as various abbreviations (e.g., p for ‘page’) and statistical values (e.g., t, F and r).

All the words were stemmed (e.g., require → requir, analysis → analysi) with the Porter stemming algorithm (Porter, 1980). Stemming was employed rather than lemmatisation because lemmatisation requires a dictionary, and specialised corpora tend to include words that are too infrequent to be included in a typical lemmatisation dictionary. For instance, acequias, biogenics and Carpathians are not lemmatised as acequia, biogenic and Carpathian, respectively, in Someya's list.⁵ Removing inflectional and derivational morphemes through stemming may collapse the words that should ideally be distinguished and lead to information loss (Sinclair, 1991). Without stemming, however, input data (i.e., document-term matrix) can be too sparse and we may not be able to target as many words as we can when stemming is applied (see the paragraph below). Here, we follow a common practice in topic modelling and opt for retaining as many root words as possible. The effect of stemming, however, has not been investigated in topic modelling literature, and there is no doubt that the issue should be addressed in future research.

To ensure the reliability of results, the topic model only targeted the 7,758 word types that occurred in at least 0.1 percent of all the text-blocks. Stemming and the removal of short or infrequent words are common pre-processing steps in topic models (e.g., DiMaggio et al., 2013; and Marshall, 2013). Each 300+ word text-block was assigned with information on where in the paper it appeared (e.g., 70 percent from the beginning of the paper).

Table 2 shows the numbers of papers and text-blocks across publication years. After excluding the stop words mentioned above, the corpus included 10,555 text-blocks with the average length of 242 words (SD=50).

Table 2: Numbers of papers and text-blocks across years in the GEC Corpus.

Table 2: Numbers of papers and text-blocks across years in the GEC Corpus.

We used the topicmodels package (Grün and Hornik, 2011) in R (R Core Team, 2015) to build the topic models. There is no agreed way to automatically decide the number of topics (but see, for example, Ponweiser, 2012, for attempts). In other words, the decision on how many topics a corpus will be deemed to contain is a subjective one and the answer may be defended on the grounds of usefulness but not on the grounds of accuracy. As an exploratory step, therefore, we built models with 40, 50, 60, 70, 80, 90 and 100 topics, and we inspected them to decide the appropriate level of granularity with which to explore the data. We decided to use the model with sixty topics because the model with fifty topics lacked some of the potentially interesting topics that we observed in the sixty-topic model and the one with seventy topics included some apparently redundant topics. Appendix A shows the ten words (or word stems) with the highest probability of occurrence in each of the sixty topics.

This list of topics and other findings in this paper were deduced from the following pieces of information:

Figure 1 demonstrates topic distribution in some of the text-blocks and papers. In Figures 1A and 1B, the horizontal axis represents sixty topics and the vertical axis represents the corresponding probability in each text-block (1A) or paper (1B). In Figure 1A, the panel label is the identifier of the text-block. For instance, ‘1993\_3\_2\_Glantz\_0.91’ indicates it is a block of paragraphs taken from the paper whose first author is Glantz and which was published in 1993 in Volume 3 Issue 2 of GEC, and the block is located at the point 91 percent away from the beginning of the paper (i.e., towards the end). This within-paper location indicates where the middle word in the block falls in the paper and was calculated by dividing the sum of the number of words before the block and half of the number of words in the block divided by the number of words in the paper. Figure 1A shows that different topics are prominent in different text-blocks, and that while some text-blocks have one very prominent topic and the other topics are weak (e.g., Topic 37 in the final panel), others have multiple prominent topics (e.g., Topics 1, 22 and 25 in the second panel).

Figure 1: Topic distribution in text-blocks, papers and words.

Figure 1B shows topic distribution at the level of papers. For this purpose, we averaged topic probability across all the text-blocks taken from the paper. Here, we chose four papers that have Topic 10 as the most prominent topic. The top ten keywords of Topic 10 are as follows: water, river, basin, suppli, flow, irrig, resourc, avail, use and stress. The topic can justifiably be summarised as ‘water’, and indeed, the titles of the four papers signal that water is their main topic:

Figure 1C visually represents the topic probability of individual words in a selection of the topics. Each row represents a topic with its interpretative label as given on the left (see Appendix A for the complete list of topics). The shading in each cell indicates probability, with a darker shade corresponding to higher probability. We can tell that topic probability for any given word is highly skewed: a word has a few prominent topics at most and has negligible probability for most of the topics. The word polici (policy), for instance, is highly probable in Topic 36 (Environmental policy actors, makers) but is practically absent in the other topics. Although still skewed, some words have a decent level of probability in multiple topics. For the word area, a fair amount of probability mass is allocated to Topics 19 (Forestry management), 23 (Wetlands, coastal, flooding) and 39 (About ecosystems and biodiversity). This means that the word is relatively frequent in the three topics compared to the other topics. The figure also shows that individual topics are characterised by just a few keywords. Topic 3 (Emission regulations), for example, includes a high frequency of carbon, emiss (emission), greenhous (greenhouse) and level, but not other words. Thus, these are the distinctive keywords of the topic. In this manner, topic models link topics and their keywords.

Appendix A shows the labels and keywords of our sixty topics. Many topics are straightforwardly thematic topics, such as Topic 3 labelled as ‘Emissions regulation’ and with keywords like emiss, reduct, greenhous and co2. Not every topic is thematic, however. Topic 30, for instance, has been labelled ‘Hypothetical discussion’ and captures the co-occurrence of the words that are often used in expressing speculation, such as if, would, could, possibl and potenti. This topic does not correspond to a topic in its usual sense of the word but represents the manner in which people write. In this way, topic models go beyond indicating textual ‘aboutness’, and give additional information about register and style (see Rhody, 2012).

The type of co-occurrence in topic models can be well understood in comparison to multidimensional analysis (MDA; Biber, 1988), another latent model that is often used in corpus linguistics. In MDA, analysts assume that there are latent (i.e., unobserved) dimensions that give rise to the co-occurrences of linguistic features. In topic models, we assume that latent topics invite word co-occurrences. In both cases, ‘co-occurrence’ takes the span of a few hundred to a few thousand words. In this sense, topic models differ from collocations, where the span is typically much shorter. Co-occurrences in topic models and in MDA often have situational reasons. In other words, in topic models, words co-occur typically because they are topically related and words under the same topic tend to co-occur, while in MDA, linguistic features co-occur because they are functionally related and those that serve the same function tend to co-occur (Biber, 1995).

As discussed above, the sixty ‘topics’ identified in the GEC corpus give different kinds of information about the corpus. Appendix A shows the ten words (or word stems) with the highest probability of occurrence in each topic. For convenience, each topic is also given a mnemonic label. Between them, the topics encapsulate and delineate what might be called the themes of the corpus. These include (in no particular order):

This is by no means a comprehensive listing. It confirms and expands on the information given on the journal website:⁶ this is a research journal about the natural world, human beings, and the interactions between nature and humans. The sixty topics encompass the scope of the journal, and between them give the observer a good intuitive ‘feel for’ the journal content.

As with any list of words, some more specific observations might be made. For example, a relatively large number of words refer to individuals or groups of people, but these tend to be at a high level of generality (e.g., people) or abstraction (e.g., actor, decision-maker, committee and stakeholder). More importantly, the topic lists serve to organise the words so that each word type is nuanced by the words it co-occurs with. For example, the natural entities of rivers, forests and oceans (Topics 10, 19 and 23) are transformed into entities used by or impacting on humankind: river co-occurs with irrigate/irrigation (10); forest co-occurs with conservation and timber (19); sea co-occurs with flood and impact (23). Most strikingly, perhaps, words to do with risk and its mitigation (problems and solutions) occur in no fewer than fifteen out of the sixty topics. One topic (20) connects general negative words such as risk, hazard and disaster with the human-related words health and people and with reduce/reduction – words associated with the mitigation of negative effects. As examples of other topics, Topic 3 connects carbon with mitigation, Topic 10 connects water and river with stress, Topic 15 connects environment with problem. Forest is connected with conservation (Topic 19). Sea and coastal are linked with both protect and loss (Topic 23). Vegetation is linked with conservation (Topic 39) and pollution is linked with control (Topic 45). Topic 54 connects ecology with resilience, while Topic 55 links climate change with response, adapt(ation) and mitigation. These co-occurrences provide detail of how natural and human entities are connected in the journal, and how entities are connected both with problems and the ways they may be addressed.

We now turn to the question of how the topics are distributed within papers. This gives information about the organisation of papers in the journal. Figure 2A shows the distribution of each of the sixty topics. The horizontal axis represents the within-paper position, where 0 is the beginning of the paper and 1 is the end of it. Each line indicates the predicted probability of each topic based on the generalised additive model (Wood, 2006) that models topic probability based on text-block position. The cubic regression spline was used as the smoothing basis. We can see that different topics behave differently. Some topics are prominent at the beginning of the paper, while others are prominent at the end. Yet others show a U-shaped pattern or randomly fluctuate.

Figure 2: Within-paper distribution of topic probability.

Figure 2B illustrates the distribution of the six topics whose relative probability decreases most radically from the beginning to the end of the paper (i.e., the topics with the lowest standardised slopes). The panels were ordered such that the topic with the most dramatic probability decrease (Topic 50: et, al, 2005, 2003, etc.) comes first, followed by the topic with the second most dramatic decrease (Topic 53: al, et, 1996, 1995, etc.), and so forth.

The figure also demonstrates the 95 percent confidence interval of the probability. The first two topics (50 and 53) are related to in-text citations, as exemplified by such keywords as et and al, and numbers representing years, and, thus, it is natural that their probability is high at the beginning of papers, where a literature review is typically located. Topic 38 appears to cover the overview of the paper, with such keywords as studi, this, paper, approach and discuss. Topics 27, 15 and 49 are more directly related to the contextualisation of papers. The keywords of Topic 27 include temporal expressions such as year, period, recent, centuri, decad and past, which provide the historical context of the paper. Topics 15 and 49 are similar in that they are both on specific issues (global environmental security issues and global warming, respectively). All six topics help to situate the paper in a wider context, and, thus, are more prominent at the beginning of papers.

Figure 2C similarly shows the topics whose relative probability most radically increases towards the end of the paper. They are all prominent in the discussion and ‘future research’ sections of the paper. Topic 40 directly discusses findings with such keywords as more, than, less, rather, signific, high and differ. Topic 30 relates to hypothetical discussion, as mentioned earlier, and is used to offer implications and speculations of the paper. Topic 11 is similarly related to discussion of the future, while Topic 12 encompasses the overall implication of the paper with words like manag, plan, strategi, institut, learn and implement as keywords. Topic 42 is another non-thematic topic that includes words related to discussion and evaluation. Here, we succeeded in identifying the paper structure with the topic model.

Topic models can also inform us of the chronological change within a journal (Blei and Lafferty, 2006; and Priva and Austerweil, 2015). Figure 3 shows the chronological topic transition obtained in a similar manner to Figure 2. Instead of the smooth curve based on generalised additive models, however, Figure 3 draws the average probability of each topic in each year. Figure 3A illustrates the topic transition of all of the sixty topics. We can observe a variety of patterns: different topics tend to be prominent in different years.

Figure 3: Chronological transition of topic probability.

Figure 3B shows the transition of the six topics that are prominent in early years but decline in later years. These were identified in the same manner as in Figure 2B. The prominent topics in early years tend to describe particular problems. Topic 45 deals with pollution issues, while Topic 15 addresses environmental security. Topic 35 discusses problems in developing and developed countries, and Topics 49 and 5 are related to global warming and energy use, respectively. Topic 11 describes the predicted and potential impacts of the issues.

Figure 3C shows the transition of the topics that are prominent only in recent years. Topic 50 is prominent in the latter half only because it characterises in-text citations after 2000. The other prominent topics tend to address the people vulnerable to environmental change and the ways in which humans tackle environmental issues. Topic 9 is about how people can adapt to climate change and who are vulnerable to it. Topic 56 discusses the impact of environmental change on farmers, while Topic 12 is related to environmental management. Topic 24 deals with how environmental issues are discussed in the media, and Topic 18 pertains to how local communities adapt to environmental change with their local knowledge and traditions. The shift of the prominent topics above suggests that GEC set its research agenda in the first years by identifying environmental problems and in later years started to address those research agendas.

Topic models can also help to identify different types of papers. GEC is an interdisciplinary journal and includes a wide range of topics. We hypothesised that some papers focus on a single, perhaps specialised, topic, while others may address a variety of topics. To examine this possibility, we identified two papers with the highest and the lowest relative entropy (Gries, 2013), which in this case selects a paper whose topic distribution is heavily skewed and one where the distribution is relatively even.

Figure 4 illustrates the topic distribution of the two papers. The upper panels show the topic profiles of the whole papers, while the lower panels show the topic profiles at the level of text-blocks. In one paper, 2008_18_3_Hof, there is one very prominent topic (Topic 22: Explaining cost–benefit analyses in figures, especially damage), and all the others are nearly negligible. This tendency applies to individual text-blocks as well. In the other paper, 1992_2_2_Dahlberg, although a few topics tend to be more prominent than others, there is no single topic that is the strongest throughout the paper. The topic profiles, thus, suggest that Hof et al.'s paper focusses on a single topic throughout the paper, while Dahlberg's paper includes a number of topics. This is indeed what we find.

Figure 4: Topic profile of two distinct types of papers.

Hof et al.'s paper is entitled ‘Analysing the costs and benefits of climate policy: value judgements and scientific uncertainties’. The paper, as the title suggests, addresses the costs and benefits of climate policy, and more specifically, computationally models the impacts of climate policy under various parameter settings. The paper heavily draws on an earlier modelling work, called the Stern Review, that also computationally modelled the economic impacts of climate policy, and regards its results as the benchmark. The paper is closely focussed on the reporting and discussion of their modelling work.

Dahlberg's paper is titled ‘Renewable resources systems and regimes: key missing links in global change studies’. The paper, as mentioned earlier, contains a variety of topics, which are well illustrated in the abstract:

The author argues that:

Notice that the individual points above are not necessarily on the same theme. Thus, the abstract already signals that the paper includes a number of topics. In the main body, too, we can observe a topic shift by looking at the first sentences of two successive text-blocks:

It is not surprising that the most prominent topic of the former text-block is Topic 33 labelled ‘Land use description’ and that of the latter is Topic 34 labelled ‘Fishing trade’. The example here thus illustrates that topic models can identify papers with radically different thematic structures.

A further strength of the topic model is that it often reveals how different senses of polysemous words behave (see DiMaggio et al., 2013). We will illustrate this below with the word level as an example. Figure 5A demonstrates the within-paper change in the frequency of the word level. The line represents the fitted value of the generalised additive model that predicts the relative frequency of level in each text-block based on the within-paper position. Each observation (or text-block) was weighted by the number of words in the text-block. While the frequency of level fluctuates somewhat, we cannot observe a systematic pattern of change. This, however, is merely the aggregated pattern. Figure 5B illustrates the change in the probability of the seven topics where level is one of the top twenty keywords. Their interpretive labels are given below:

Figure 5: Within-paper distribution of the word level and the distribution of the probability in the seven topics where level is a keyword.

Although level is a keyword in these seven topics, its sense varies across the topics. In Topics 3 and 49, the word is used to refer to the degree of concentration, as in the case of ‘[t]he present base level of atmospheric CO₂ concentration’ (1993_3_4_Schulze_0.29847182425979). These topics behave similarly in Figure 5B in that they are clearly more prominent at the beginning of the paper than at the end. This is probably because the topics are on greenhouse gas emissions and the CO₂ level is often discussed to contextualise the paper.

In Topics 22, 44 and 59, the word refers to a position on a scale, as in ‘societies tend to dematerialize above a certain level of wealth’ (2010_20_4_Schandl_0.478278251599147). These topics show similar patterns in Figure 5B as well: their probability tends to be highest at approximately 60–70 percent from the beginning of the paper. This is because it is associated with the results section of the paper. For instance, a variable can be correlated with the levels of income or education. As in the case above, we can observe that similar senses of the word behave similarly within papers.

In Topic 23, level refers to a height or distance, as in the case of ‘Wetlands are sensitive to sea-level rise as their location is intimately linked to sea level’ (2004_14_1_Nicholls_0.420021895146576). The probability of this topic is high at around 60 percent as well. Here, again, the word, particularly in connection to the rise of sea levels, occurs in the results section of papers.

Finally, in Topic 32, the word refers to a relative rank on a scale, as in ‘local governments may feel they are left little option but to use their powers at the local level to respond to regional level concerns’ (1995_5_4_Millette_0.489480090419058). The probability of the topic is highest towards the end of the paper. This is presumably because the roles that the multiple levels of actors (e.g., international, national and regional) play are discussed in the conclusion of the papers. The discussion here illustrates that the topic model reveals the systematic pattern of the individual senses of a word that cannot be observed when the senses are aggregated and that, more generally, topic models can discriminate different senses of a word without any semantic information (see DiMaggio et al., 2013).

This subsection compares the topic model to the UCREL Semantic Analysis System (USAS) accessed through Wmatrix (Rayson, 2008). A critical difference between supervised semantic tagging such as USAS and the unsupervised topic model such as the one introduced in this paper is that the former assigns pre-specified categories to words whereas the latter finds (typically) semantically related groups of words in a bottom-up way. The granularity of semantic categories needs to be pre-determined in semantic tagging. Semantic tagging, thus, requires a sophisticated tagset and a dictionary. In topic models, however, the specified number of topics is identified inductively, and thus the granularity depends on the topical heterogeneity of the corpus and the number of topics identified in it. They do not require tagsets or dictionaries. Indeed, topic models can be run on any language, provided that the text can be tokenised.

To compare empirically the results of semantic tagging and topic models, we annotated our GEC corpus with USAS through Wmatrix. USAS assigns multiple tags to a token, but only the first candidate was retained. When the first candidate included two tags (i.e., double membership), both were retained as separate tags. Markers of the position on semantic scales (+ and –), those of semantic templates indicating multi-word units, and other symbols following the main tag (e.g., f standing for ‘female’) were removed.

To investigate chronological change in GEC, we identified key semantic fields in the first decade (1991–2000) and those in the second decade (2001–2010). This was performed through the Keyword List function in AntConc (Anthony, 2014) with one sub-corpus (e.g., papers published between 1991 and 2000) as the target corpus and the other (e.g., papers published between 2001 and 2010) as the reference corpus. Log-likelihood was used as the keyword statistic. To capture USAS tags with AntConc, a token was defined as a sequence of English alphabet characters, digits and full-stops.

Tables 3a and 3b list the resulting top ten key semantic tags in each decade and the five most frequent words for each tag in the target decade. For instance, the tag O1.3, which corresponds to the semantic category labelled as ‘Substances and materials generally: Gas’, was nearly three times as frequent in 1991–2000 papers as in 2001–2010 papers (33.7 versus 12.1 per 10,000 words), and the most frequent words with the tag in 1991–2000 papers were CO₂, gas, gases, methane and ozone.

Table 3a Key semantic fields across time.

Table 3a Key semantic fields across time.

Table 3b Key semantic fields across time.

Table 3b Key semantic fields across time.

Some findings match with the findings based on the topic model (e.g., ‘substance and materials’ in semantic tagging as a key semantic field in the first decade versus Topic 45 labelled as ‘Toxic substance and pollution management’ as the key topic). The semantic category in USAS, however, is sometimes too coarse for our corpus. W5, labelled as ‘Green issues’ and including environmental terms, was the third key category in 1991–2000 papers and the five most frequent words were environmental, environment, conservation, nature and pollution. When we look at which topics those five words are the keywords of, we notice that, as expected, environment(al) and pollution are included in Topics 15 and 45 – the two topics that showed the most dramatic decline. The word conservation, however, is included in Topic 19 (‘Forestry management’) and Topic 39 (‘About ecosystems and biodiversity’) as one of the top ten keywords, and the probability of these topics remains relatively unchanged. This suggests that while some green issues such as air pollution are on a declining trend, the trend does not apply to other issues such as forest conservation.

A similar observation can also be made regarding key semantic fields in the latter decade. A key semantic domain in 2001–2010 papers is ‘weather’ and includes such words as climate, rainfall, flood and climatic. When we look at the topics whose keywords include these words, we notice that while the occurrence of the word climate in Topic 55 (‘Mitigation, adaptation’) increases over time, the reverse is true for Topic 49 (‘Greenhouse gases, climate changes’).

Therefore, there are sub-patterns within the single semantic category that the topic model can distinguish but the USAS model cannot. The topic model can thus provide a more fine-grained view of the thematic structure of the corpus.

Another potentially comparable technique is keywords analysis. Keywords in keywords analysis are a list of words that are more frequent in a corpus than in the reference corpus, and have been often associated with textual ‘aboutness’ (Bondi, 2010; and Scott, 2010). Aboutness here, however, is defined with reference to the reference corpus, and represents how the target corpus is different from the reference corpus. The need to pre-specify the reference corpus is potentially a drawback of the technique.

To compare empirically the topic model to keywords analysis, we identified keywords of the GEC papers published in the first decade and those in the second decade. As in the identification of key semantic fields earlier, we had one sub-corpus (e.g., 1991–2000) as our target corpus and identified the keywords using the other sub-corpus (e.g., 2001–2010) as the reference corpus. Log-likelihood was employed as the keywords statistic, and the analysis was undertaken using AntConc. All the words were stemmed. Digits were included in the token definition because numbers were occasionally present in the keywords of our topic model and among the words that are frequent in the key semantic tags identified earlier.

Table 4a and Table 4b show the top twenty keywords of each decade. The word figur is the top keyword in 1991–2000 papers only because figures were referred to as, for instance, Figure 1, until the 1998 volume but as Fig. 1 afterwards. Numbers representing years after 1999 occupy eleven out of the twenty keywords in the second decade because they represent in-text citations after 2000. The keywords suggest that over time GEC came to deal less with the emission of greenhouse gases, such as methane, CFC and CO₂, and their impact on the global environment and more with vulnerability, adaptation and resilience related to environmental change. These are in line with our observation based on the topic model.

Table 4a Keywords across time, 1991–2000.

Table 4a Keywords across time, 1991–2000.

Table 4b Keywords across time, 2001–2010.

Table 4b Keywords across time, 2001–2010.

The topic model, however, often brings us more easily interpretable findings. From keyword analysis, we can observe that one of the keywords in the latter decade is household. The word, however, is difficult to interpret because there is no other keyword in the list that seems to be related to it in the first instance. When we look at Topic 56 (‘Households, village level’), which includes household as one of the keywords and is a topic that is prominent in later years, we can see that household co-occurs with such words as farmer, farm, village and livelihood. These words suggest that households in this context refer to those of farmers in villages. Combined with the increasing probability of Topic 9 (‘Vulnerability, adaptive capacity’) in later years, we can hypothesise that the households of farmers in villages are vulnerable to environmental change and need to adapt, and that this topic is on an increasing trend.

Part of the difficulty in interpreting the word household is due to the small number of keywords considered. The twenty-first keyword is livelihood and the thirty-second is farmer, both of which facilitate the interpretation of household in the same manner as above. However, it is only the topic model that automatically groups related words. In keywords analysis, researchers still need to reason that household is perhaps not related to some keywords like fig or water, but vulner and adapt are closely relevant. The topic model automates this process.

In collocation networks (Brezina et al., 2015; and Williams, 1998, 2002), analysts specify a node word as the starting point and investigate the network of words where the edges represent collocation. Collocation networks are similar to topic models in that they both identify word co-occurrences.

There are, however, notable differences between the two techniques. First of all, collocation typically looks at co-occurrence patterns within an immediate environment around the node word (e.g., five words to the left and the right of the node word), whereas topic models target co-occurrences within more extensive texts. As a result, each method captures different aspects of meaning: collocation studies reflect the distributed or prosodic meaning associated with phraseology whereas topic models identify thematic meaning. Related to this, in collocation networks it is necessary to specify a node word, and the potentially subjective choice of the node word influences the aspect of the corpus that the technique can reveal. On the other hand, topic models target the entire corpus, reducing the arbitrariness of the analysis.

Yet another way to capture word co-occurrence patterns is through concgrams (Cheng et al., 2006; Cheng et al., 2009; and Warren, 2010). Concgrams identify the co-occurrence of words (e.g., environmental problems) that may be intervened by other words (e.g., environmental health problems) or may vary in position (e.g., problems in environmental policy). Comparison between concgrams and topic models is more or less similar to the comparison between collocation networks and topic models. Since the typical span used in concgrams is much smaller than the span used in topic models (i.e., text), the topic model can identify what is similar to themes in a corpus while the congram discloses more local meaning. Also, concgrams require analysts to choose a target word to analyse, which potentially introduces arbitrariness.