Political Rhetoric Project

In order to answer the temporal question, we first need to understand the topics he shared with his electorate in his radio talks. We used a machine learning approach called topic modelling to identify the most relevant and frequent topics.

Topic modelling is an unsupervised machine learning technique that is capable of detecting word and phrase patterns, and automatically cluster word groups and expressions to characterise a corpus.  It provides a means to reinforce our intuition that words such as ‘play’ and ‘performance’ often coincide with words like ‘audience’ and ‘theatre’ in a document, rather than ‘invasion’ or ‘explosion’. That’s not to say that these terms never would – a theatre review of “War of the Worlds” would have such terms in the same document. However, given a document collection, we would likely find that these 6 words would naturally form 2 clusters – one cluster with terms pertaining to ‘the arts’ and one cluster around ‘war’ – which we would refer to as “topics”.

There are a number of ways we can determine these clusters of terms and we have used a method called LDA (Latent Dirichlet Allocation) because it helps us discover a couple of things about a document collection. It provides us a way to uncover which terms have a propensity to be collocated together – these are our “topics” – but additionally it models how these topics are distributed over the document collection. We can discover which topics are talked about together and in what proportion throughout the document collection. This is a valuable feature in describing a compilation of related texts, as we have done with Malcolm Fraser’s radio speeches.

In building a model one of the key factors to determine is the optimal number of topics to extract from the speeches in order to extract topics that are unambiguous, distinct and sensible. There are certain measures that help determine the right value for this hyperparameter, such as topic coherence. For this project we leveraged our domain knowledge to identify the right number of topics. We started by using a grid search approach and came up with 4 models that contained 10, 30, 60 and 90 topics. Subject matter experts concluded that the 60 topic model had the most topics that were both granular and concise enough to become categories of their own without overlapping too much. They have labelled these topics on the interactive graph below.

In the topic model we have created for Fraser’s speeches, each word in a speech is attributable to one of the 60 topics, and each speech contains a number of these identified topics. However each of these identified topics is weighted differently. So when we choose the highest weighted topic for a speech, we identify the dominant topic. When it comes to visualising these speeches, we use a specific approach of dimensionality reduction called t-distributed stochastic neighbor embedding on the topic distribution for the speeches. This allows us to visualise those speeches in a 2 dimensional space as shown in the interactive graph below.

This section focuses on analysing the frequency of topics over the decade 1960-1970. The graph below shows the frequency of the topics where they have been the dominant one for a speech.

This fairly simple visualisation shows us some useful points about Malcolm Fraser during this decade.  The Vietnam War is easily the most frequent dominant topic, indeed it is three times more frequent than any other topic. Malcolm Fraser was Minister for the Army between 1966 and 1968, which can explain this predominance to some extent (see below for more discussion around the temporal analysis). But he was also Minister for Education between 1968 and 1969, and that topic appears far less frequently. Both the Vietnam War and education reform were controversial and prominent topics during these years. So it would appear that it is not solely his ministerial responsibilities that determined which topics he spoke about to his constituents.

The next most dominant topic is Parliament, a label given to speeches that sought to explain procedures or events that take place within Parliament. Then there are a series of topics that could be described as local or rural affairs; Wannon was a country district with a high portion of farmers (including Fraser's family). Overall, it looks like Fraser's messages to his constituents over that decade were a mix of international affairs, ministerial duties and parliamentary issues, and rural/local issues. Given that Fraser was a backbencher for just over half the decade and had a rural constituency, this mix of topics might be considered surprising. It demonstrates that Fraser held a keen interest in international affairs and wished to discuss them with his constituents from early in his career.

The most obvious aspect of this graph is the spike in discussions of the Vietnam War, beginning in 1964, the year that the National Service Act was passed. Between 1966 and 1968 it is by far the most frequent topic, which coincides with his term as Minister for the Army. By 1967, 30% of the discussion is about the Vietnam War despite that year also being the year with the most speeches recorded.

If we remove Vietnam from the graph, no other single topic dominates over time. There is an interesting consolidation of topics (defence policy/equipment, parliament, education, and foreign policy) in 1969, but for the rest of the period it seems most of the topics were generally quite stable and almost equal in distribution. There is certainly some overlap between the timeline and the temporal topic distribution. We finally see some correlation here with his time as Minister for Education, with a spike in the education topic in 1969, the year that he announced substantial changes to federal funding for non-government schools.

The temporal analysis seems to confirm what the topic modelling had suggested; that Fraser spoke to his constituents regularly about a range of topics over time. That might sound like an obvious statement, but in some senses it is what the graph is not showing that is the interesting feature. This time period saw Fraser move from being a very young backbencher from a rural electorate to a minister with experience of several portfolios, a national profile and soon to challenge successfully for party leadership. Therefore, one might have expected to see a a clear topic trend towards issues that might be considered to be of more national or international significance. With the exception of the early discussions of the Portland wool sales and marketing, this was not the case.

With the second hierarchy, we broke the topics into politics, economics and administration (meaning topics that didn't seem overtly political, such as the development of the national capital city).

Clearly these topics are concomitant (eg industrial relations is both an economic and a political topic, national service is both a national and international topic), but for the sake of distinguishing trends over time, we sought to distribute them as best we could based on our understanding of each of the 60 topics identified by the model.

The theme-based hierarchy has a more noticeable decline of economics-themed topics, ending the decade with a big gap between topic themes that are political and those that are economic. This analysis is somewhat unexpected, given the topics placed into the economic category are ones that generally retained relevance over the decade.